filename
stringlengths 23
23
| markdown
stringlengths 18k
147k
|
|---|---|
10.1126_science.abc4787
|
RES EARCH
CORRELATED ELECTRONS
Observation of a critical charge mode in a
strange metal
Hisao Kobayashi1,2*, Yui Sakaguchi1, Hayato Kitagawa1,2, Momoko Oura1,2, Shugo Ikeda1,2,
Kentaro Kuga3, Shintaro Suzuki3, Satoru Nakatsuji3,4,5,6*, Ryo Masuda2,7,8, Yasuhiro Kobayashi2,7,
Makoto Seto2,7, Yoshitaka Yoda9, Kenji Tamasaku2, Yashar Komijani10,11,
Premala Chandra11, Piers Coleman11,12*
Understanding the strange metallic behavior that develops at the brink of localization in quantum
materials requires probing the underlying electronic charge dynamics. Using synchrotron radiation–
based Mössbauer spectroscopy, we studied the charge fluctuations of the strange metal phase
of b-YbAlB4 as a function of temperature and pressure. We found that the usual single absorption peak
in the Fermi-liquid regime splits into two peaks upon entering the critical regime. We interpret this
spectrum as a single nuclear transition, modulated by nearby electronic valence fluctuations
whose long time scales are further enhanced by the formation of charged polarons. These critical
charge fluctuations may prove to be a distinct signature of strange metals.
T he strange metal (SM) is a ubiquitous
state of matter found to develop in quan-
tum materials with strong correlations,
often appearing as a fan-shaped region
of the phase diagram centered around
an unstable quantum critical (QC) point. The
characteristics of SMs include a logarithmic
temperature (T) dependence of specific heat
C/T ~ –logT, a linear-in-T resistivity r(T) ~
T (1), and a strong violation of Kohlers law
in the magnetotransport (2–4). These prop-
erties and their universality defy the stan-
dard concept of quasiparticle excitations,
which is central to the Fermi liquid (FL) theory
of metals. This enigma has prompted a wide
range of proposals for the origin of SM be-
havior, including Fermi surface instabilities
(1, 5–7), valence quantum criticality (8), charge
stripes (9), and nematicity (10–12); it has also
motivated approaches such as holographic
duality (13–15) and simulation by use of cold
atoms (16).
Although the spin dynamics at quantum
criticality has been extensively studied, little
1Graduate School of Material Science, University of Hyogo,
3-2-1 Koto, Hyogo 678-1297, Japan. 2RIKEN SPring-8 Center,
Hyogo 679-5148, Japan. 3Institute for Solid State Physics,
University of Tokyo, Kashiwa 277-8581, Japan. 4Department
of Physics, University of Tokyo, Hongo, Bunkyo-ku, Tokyo
113-0033, Japan. 5Trans-scale Quantum Science Institute,
University of Tokyo, Bunkyo-ku, Tokyo 113-0033, Japan.
6Institute for Quantum Matter and Department of Physics
and Astronomy, Johns Hopkins University, Baltimore, MD
21218, USA. 7Institute for Integrated Radiation and Nuclear
Science, Kyoto University, Osaka 590-0494, Japan.
8Graduate School of Science and Technology, Hirosaki
University, Aomori 036-8561 Japan. 9Japan Synchrotron
Radiation Research Institute, Hyogo 679-5198, Japan.
10Department of Physics, University of Cincinnati, Cincinnati,
OH 45221-0011, USA. 11Department of Physics and
Astronomy, Rutgers University, Piscataway, NJ 08854, USA.
12Hubbard Theory Consortium, Department of Physics, Royal
Holloway, University of London, Egham, Surrey
TW20 0EX, UK.
*Corresponding author. Email: kobayash@sci.u-hyogo.ac.jp
(H.K.); satoru@phys.s.u-tokyo.ac.jp (S.N.); coleman@physics.
rutgers.edu (P.C.)
is known experimentally about the charge
dynamics because appropriate laboratory
probes are scarce. Conventionally, charge
dynamics are studied with optical spec-
troscopy (17), but these methods probe only
the low-momenta, divergence-free trans-
verse components of the current density that,
by the continuity equation, do not couple to
fluctuations in the charge density. Longitu-
dinal current fluctuations can be probed by
means of electron energy loss spectroscopy
(EELS) but to date are limited to energies
above the Debye energy because of difficul-
ties in subtracting the phonon background
in the signal (18–20). A classic method to
detect low-frequency longitudinal charge
dynamics is Mössbauer spectroscopy, suc-
cessfully used in the past to detect the slowing
of the charge dynamics at charge-ordering
transitions of europium- and iron-based com-
pounds (21, 22).
However, the widespread adoption of
Mössbauer methods has long been hindered
by the lack of suitable radioisotope sources.
To overcome these difficulties, a new genera-
tion of Mössbauer spectroscopy has recently
been developed by using synchrotron radia-
tion (SR) (23). SR-based Mössbauer spectros-
copy (Fig. 1A) can be used for a wide range of
Mössbauer isotopes, providing improved en-
ergy resolution for the isotopes with shorter
lifetimes; it offers an unprecedented capability
to select a particular nuclear transition, taking
advantage of the perfectly polarized SR. This
approach presents an ideal probe to resolve
longitudinal charge dynamics in materials for
which conventional Mössbauer techniques are
inapplicable.
We report a direct observation of critical
charge dynamics in a SM regime by using
SR-based 174Yb Mössbauer spectroscopy. The
heavy fermion metal b-YbAlB4 provides an
ideal platform to study the SM regime at am-
bient pressure in a stoichiometric crystal
(3, 24). In b-YbAlB4, core level x-ray studies
have established the presence of an interme-
diate valence state caused by valence fluc-
tuations between two ionic configurations:
Yb2þ⇌Yb3þ þ e(cid:2) (25). Usually, in heavy fer-
mion compounds, such valence fluctuations
are too fast to be observed with Mössbauer
spectroscopy (26–29), but we show that this is
not the case in the SM regime.
Using synchrotron radiation–based Mössbauer
spectroscopy to study charge fluctuations
Mössbauer spectroscopy measures the shift
in a nuclear absorption line caused by changes
in the local (q-integrated) charge density. The
characteristic time scale of the measurement
is the lifetime of the nuclear excited state, t0 ~
2.5 ns in 174Yb. Charge fluctuations that are
much shorter in time than t0 produce a single
motionally narrowed absorption line, whereas
charge fluctuations that are much longer in
time than t0 produce a double peak absorp-
tion line, corresponding to the two different
valence states of the Yb ion (Fig. 1C). By fitting
the Mössbauer absorption line shape, one can
detect charge fluctuations with time scales in
the range of ~0.1t0 to ~10t0 (30).
b-YbAlB4 exhibits quantum criticality with-
out tuning in an intermediate valence state
(25), and the application of an infinitesi-
mal magnetic field B tunes the SM into a
FL with kBTFL ~ mBB, where kB, TFL, and mB
are the Boltzmann constant, FL temperature,
and the Bohr magneton, respectively. The
slope of the linear-in-T resistivity r(T) ~ T
over T between 0.5 and 25 K at ambient
pressure corresponds to a nearly quantum-
saturated scattering rate t(cid:2)1
¼ 0:4 (cid:3) kBT =ℏ
tr
(30), thus establishing b-YbAlB4 as a system
with Planckian dissipation (31). This anom-
alous r(T) and its extension over a broad
pressure (p) range from ambient pressure to
p* ~ 0.5 GPa (Fig. 1B) (3, 24, 32) provides an
excellent setting for high-precision measure-
ments of the critical charge fluctuations, likely
of relevance to the broader family of SMs.
Measuring charge dynamics in b-YbAlB4
We investigated how the QC behavior in the
SM regime affects the charge dynamics, follow-
ing their evolution as the SM regime at am-
bient pressure transforms into a FL regime
under pressure. At 20 K and ambient pressure
(Fig. 2A), the Mössbauer spectra exhibit a
single line feature. However, below T* ~ 10 K,
as one enters the QC region, this peak broad-
ens into a two-peak structure, with 5s signifi-
cance (30). This two-peak structure observed
for p < 0.7 GPa at 2 K coalesces into a single
peak at around p ~1.2 GPa, ultimately sharpen-
ing into an almost resolution-limited peak at
p = 2.3 GPa, which is characteristic of a FL
(Fig. 2B) (30).
Kobayashi et al., Science 379, 908–912 (2023)
3 March 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
The local symmetry at the Yb site of
b-YbAlB4 with the orthorhombic structure
allows us to rule out a nuclear origin of the
double-peak structure. For ckk0
(the pro-
pagation vector of the incident x-ray), the
symmetry selects two degenerate nuclear
transitions Ig ¼ 0→I z
¼ T1 from the five E2
e
nuclear transitions (DIz = 0, ±1, and ±2) of
the 174Yb Mössbauer resonance (Fig. 1, A
and C) (33). The absence of magnetic order
in b-YbAlB4 (24, 32) also eliminates as ex-
planations magnetic and nonaxially sym-
metric quadrupolar hyperfine interactions
(30). This leaves a combination of the electric
monopole and axially symmetric quadrupolar
interactions—linking the hyperfine energy to
the valence state of the rare-earth ion—as the
only candidate for the observed splitting. The
presence of a Mössbauer line splitting then
implies a distribution of Yb valences within
the crystal. We argue that these result from
slow dynamic charge fluctuations.
All Yb sites are crystallographically equiva-
lent in b-YbAlB4, and SR x-ray diffraction mea-
surements (34) show that the lattice structure
does not change up to 3.5 GPa at 7 K; further-
more, the absence of any low-temperature
phase transitions rules out the possibility
of a charge density wave (30). Moreover, the
residual resistivity ratio (RRR) exceeds 100,
indicating the low levels of quenched disorder
in this material. Given that disorder broadens
the Mössbauer absorption peak, our ability to
resolve the double-peak structure is consistent
with this conclusion. An attempt to fit the
Mössbauer spectrum with two nuclear tran-
sitions (a static hyperfine interaction), using
a width corresponding to the experimental
energy resolution, fails to reconstruct the
feature at 2 K and ~0 mm/s (Fig. 2A, blue
dashed line). Thus, the two-peak structure
and line broadening observed for T < 5 K and
p < 0.7 GPa must derive from a single nuclear
transition that is dynamically modulated by
fluctuations between two different Yb charge
states (a time-dependent hyperfine interac-
tion) (Fig. 1C) (30).
We analyzed our Mössbauer spectra at
ambient pressure using a stochastic theory
(35–37) with a single nuclear transition
modulated by two different charge states
(30). The predicted spectra (Fig. 2A, red
lines) well reproduce the two-peak structure
in the spectra at low T and its subsequent
collapse into a single line with increasing T.
At ambient pressure, the extracted fluctua-
tion time tf between two different Yb charge
states is unusually long compared with the
electronic time scales, exhibiting a slow power-
law growth T–h (h ~ 0.2) on cooling below T*
(Fig. 2C). The energy difference between two
selected nuclear transitions is almost inde-
pendent of T up to 20 K (30), so that the de-
velopment of the two-peak structure in the
Fig. 1. Experimental setup and concept. (A) Schematic of our experimental setup for the synchrotron
radiation–based 174Yb Mössbauer spectroscopy (47). The 174Yb nuclear resonance (Eg = 76.471 keV) was
obtained from synchrotron radiation by use of a monochromator. The c axis of the single-crystalline b-YbAlB4
samples was aligned along the propagation vector k0 of the incident x-ray under both ambient and external
pressure. The single-crystalline YbB12 samples were cooled at 26 K. A Si avalanche photodiode (APD)
detector was used to measure delayed incoherent emission from 174Yb nuclei in the YbB12. (B) (Top)
Schematic phase diagram of b-YbAlB4 as a function of pressure at low temperatures. (Bottom) Illustration of
the crystal structure of b-YbAlB4 with a snapshot of the Yb valences, Yb2+ (large purple spheres) and Yb3+
(small red spheres, with arrows indicating magnetic moment). (C) (Top) Energy level diagram of a 174Yb
nucleus in Yb2+ and Yb3+ ions. The lowest excited states of a 174Yb nucleus lie in a Ie = 2 multiplet with a
lifetime t0 = 2.58 ns. The excited-state energies are perturbed by the electron charge distributions around
the nucleus; a spherically symmetric charge distribution (Yb2+) preserves the multiplet degeneracy, whereas
one with axial symmetry (Yb3+) splits the multiplet into two doublets and a singlet. The allowed Mössbauer
transitions are indicated with arrows, where the black arrows represent the two selected transitions for
photons travelling along the c axis. In b-YbAlB4, the valence of the Yb ions fluctuates between 2+ and 3+
states on a time scale tf. (Bottom) Schematic showing two different scenarios for Mössbauer absorption lines
depending on the relative time scales t0 and tf. If tf ≫ t0 (slow valence fluctuations), two distinct x-ray
frequencies are detected, leading to two distinct Mössbauer absorption lines (left), whereas if tf ≪ t0
(fast valence fluctuations), a single motionally narrowed x-ray frequency will be detected in the
Mössbauer absorption.
observed spectra must derive from the marked
low-T growth in tf. On the other hand the
gradual collapse of the two-peak structure in
the observed 174Yb Mössbauer spectra at 2K
with increasing p indicates that fluctuation
time scale tf becomes shorter as p increases
(Fig. 2B). The spectra at p < 1.2 GPa can be
analyzed and reconstructed with the same
stochastic model used at the ambient pres-
sure, whereas the spectrum observed at 2.3 GPa
was simply fit by using the static model. The
linewidth of this single absorption compo-
nent was found to be G = 1.11 mm/s, which is
slightly broader than the resolution limit G0 =
ℏ/t0 = 1.00 mm/s (3 mK), for 174Yb Mössbauer
spectroscopy (t0 = 2.58 ns, and ℏ is Planck’s
constant h divided by 2p).
tf gradually decreases with increasing p,
exhibiting a kink across ~p* in between 0.5
and 1 GPa, approaching the resolution limit
at 2.3 GPa (Fig. 2D). This is roughly consistent
with previous r(T) measurements in b-YbAlB4
(32); at T < 0.5 K and under p, r(T) displays
r ~ Ta with a = 3/2 below p*, and further
application of pressure increases the expo-
nent to a = 2, stabilizing a FL state at about
1 GPa (32). However, the FL temperature
TFL depends on p, and only for p ~ 2.3 GPa is
the system in the FL regime at T = 2 K (32).
Slow valence fluctuations
This consistency leads us to interpret the split
line shape observed in the Mössbauer spectra of
the SM as unusually slow valence fluctuations
Kobayashi et al., Science 379, 908–912 (2023)
3 March 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
Mössbauer (recoil-free) factor fLM in b-YbAlB4,
which is the equivalent of the Debye-Waller
factor in a usual scattering experiment. Gener-
ally, (cid:2)ln fLM ¼ k2
i, where Dz is an atomic
displacement from a regular position in a
crystal along the direction of k0 (40). The
expression for the variance in atomic posi-
tion is
h
0 Dz2
(cid:2)
Dz2
(cid:3)
º∫∞
0
dw
F wð Þ
w
1
Þ
ð
2coth bw=2
(cid:5)︷
(cid:4)
1
1
ew=T (cid:2) 1
2
þ
ð1Þ
ð
h
(cid:7)
Þ2T 2
iº 3=2 þ p=QD
where F(w) is the (partial) phonon density
of states. In a Debye model, F(w) º w2,
(cid:6)
which leads to Dz2
at
T ≪QD, where QD is the Debye temperature
(30). This Deybe relation holds above T*
at ambient pressure, where tf (~1.15ns) is
independent of T (Fig. 3A); from this, we
estimated QD = 95 K, corresponding to the
lattice response time tL ~ h/kBQD ~ 0.5 ps,
so that tf ≫tL . The estimated QD (= 95 K)
value is smaller than that (195 K) of a con-
ventional valence fluctuation metal YbAl2
(41). This indicates that the lattice vibra-
tions are softer in b-YbAlB4, which suggests
an enhanced effective coupling between
slow charge fluctuation modes and lattice
vibrations.
h
h
Þ2 Dz2
h
þ d Dz2
h
4 kBQD ∼ 1
ð
2 mYb kBQD=ℏ
The saturation of Dz2
Additionally, in the QC regime below T*,
where tf develops temperature-dependence,
i departs from this Debye behavior (Fig.
h
Dz2
3A), indicating an enhancement in the quan-
i ¼ Dz2
i
i,
h
tum fluctuations, Dz2
p
ffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Debye
∼ 0:014 Å root
i
h
d Dz2
of the Yb ions. The
mean square fluctuation observed here is
comparable with the quantum fluctuations of
the phonon mode, which is around 0.05 Å es-
i.
timated from 1
i is approximately constant at 2 K for p <
h
Dz2
p* and then drops when p > p* (Fig. 3B),
indicating that the anomalous vibrations of
i, disappear in the FL regime
the lattice, d Dz2
at low temperatures.
h
i for T < T* and p < p*
implies that the phonon spectrum F(w) has
changed its form to compensate the coth(bw/2)
term in the integral (1). This then suggests that
at energies and temperatures below T*, F(w)
acquires a temperature-dependence F(w, T) =
ϕ(w)tanh(w/2T) that cancels the coth(w/2T)
term in integral (1). The function tanh(w/2T) ~
w/2T for w ≪ T, and tanh(w/2T) ~ 1 for w ≫ T,
and thus has the marginal Fermi liquid (MFL)
form. This enhancement in phonon density
of states should be observable in inelastic
neutron scattering measurements. Because
the phonons are linearly coupled to the charge
density of the electrons, the appearance of
a MFL component in the phonon spectrum is
an indication of MFL behavior in the charge
fluctuations. The enhancement of tf through
polaron formation is crucial for slowing the
Fig. 2. Temperature and pressure dependence of synchrotron radiation–based 174Yb Mössbauer
spectra of b-YbAlB4. (A and B) Selected spectra as a function of (A) temperature (T) at ambient
pressure and (B) under external pressure (p) at 2 K. The c axis of the single-crystalline b-YbAlB4 samples
was aligned along the propagation vector k0 of the incident x-ray. The solid circles with error bars and
the red solid lines indicate the observed and the analytical spectra calculated as described in (30),
respectively. In (A), the dashed blue line in the spectrum at 2 K indicates the spectrum with two static
nuclear transitions expected with our experimental energy resolution, whereas the solid blue line shows
a fit to the wings of the line shape, discarding the double-peak structure in the center. The deviation
at the center corresponds to 5s statistical significance (30). (C) Temperature T and (D) pressure p
dependences of the refined fluctuation time tf between two different Yb charge states in b-YbAlB4.
(Inset) Log-log plots of tf versus T in b-YbAlB4. The dashed line indicates tf ~ T–0.2.
between the Yb2+ and Yb3+ ionic-like states in
b-YbAlB4, on a time scale tf > 1 ns that follows
an approximate power-law growth tf ~ T–0.2
with decreasing temperature below T*. The
Yb3+ ground state is a Jz = ±5/2 moment as
deduced by varying the incident angle of the
x-ray (30). The slow charge fluctuations extend
up to p*, beyond which a conventional valence
fluctuation state with rapid charge fluctuation
takes over in the pressured regime correspond-
ing to the FL regime.
The unusual aspect of the observed charge
dynamics is that not only are they slower than
the Planckian time tf ≫ ttr e 10(cid:2)2 ns at 2 K,
they are also slower than the characteristic
time scale of the lattice vibrations. There-
fore, the lattice is expected to adiabatically
respond to the associated charge redistri-
bution. Each valence fluctuation of Yb atoms
is then dressed by Np phonons, leading to
the formation of a polaron (38, 39) and re-
normalizing the matrix element for the charge
fluctuations; this provides a mechanism for
enhancing their time scale (tf →tf eNp ) (30).
Analysis of the Mössbauer spectra allowed
us to directly check this scenario. We used
the T-dependence of the absorption compo-
nents in the spectra to determine the Lamb-
Kobayashi et al., Science 379, 908–912 (2023)
3 March 2023
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Lamb-Mössbauer factor fLM for b-YbAlB4. (A and B) Shown is (lnfLM=lnf0
LM) as a function of (A) T2
at ambient pressure and (B) under external pressure p at 2K. The Lamb Mössbauer factor fLM is determined
from the intensity of the Mössbauer absorption line. Its logarithm (lnfLM=lnf0
squared fluctuations Dz2
in the z axis position of the nucleii in units of the zero-point fluctuations (T = 0) of
LM) measures the root-mean-
(cid:2)
(cid:3)
(cid:9)
(cid:10)
ER
kBQD
º (cid:2) 3
2
was obtained by extrapolating the high-temperature
the Debye model. In (A) and (B), lnf0
LM
behavior to zero temperature at ambient pressure, where ER ¼ E2
g
2mYbc2
line indicates a linear relation between lnfLM and T2. For 174Yb Mössbauer resonance of k0 = 38.75Å–1, Dz2
the Yb ions was evaluated in b-YbAlB4 from the T and p dependences of lnfLM=lnf0
LM by using QD = 95 K. In
(A) and (B), the Dz2
decreases to 1.7 ×
10−3 Å2 in the pressured regime corresponding to the FL regime, which is comparable with that for YbAl2 (41).
values (right axis) are ~2.6 × 10−3 Å2 in the SM regime. In (B), Dz2
is recoil energy. In (A), the dashed
for
(cid:10)
(cid:9)
(cid:2)
(cid:2)
(cid:3)
(cid:3)
(cid:2)
(cid:3)
charge fluctuations down to time scales ac-
cessible to Mössbauer spectroscopy.
Discussion and outlook
A possible interpretation of our results is
the QC tuning of a critical end point of a
classical valence transition (42) between
the Yb2+ and Yb3+ ionic states. Such first-
order valence transition lines, with second-
order end points, are well established in
rare earth compounds. It has been suggested
(42) that the tuning of such an end point to
zero temperature may provide an explana-
tion of the observed Mössbauer spectra.
An alternative interpretation is that the
observed valence fluctuation modes are an
intrinsic property of the SM regime con-
nected with a spin charge separation that
develops with the collapse of the f-electron
Fermi surface (43–46). This scenario sug-
gests that similar slow charge fluctuations
will be manifested in the Mössbauer spectra of
any partial Mott localization critical point,
such as in other heavy fermions and iron-
based superconductors.
We provide direct evidence for unusually
slow charge fluctuations in the SM regime
of b-YbAlB4 by using SR-based Mössbauer
spectroscopy. Because their time scales are
longer than that of the lattice response, we
have inferred polaronic formation in the
mixed valence regime (38, 39). Both the slow
charge fluctuation modes and the anomalous
vibrations of the lattice disappear in the
pressure-induced FL regime. It is natural
to expect that this observed slow charge
mode is connected to the linear resistivity
often observed in SMs. Various theoretical
approaches (13, 14) have suggested that the
previously unknown transport properties
of SMs are linked to the universal quantum
hydrodynamics of a Planckian metal. Be-
cause the local equilibrium is established at
the scale of Planckian time, it is natural to
regard the slow charge fluctuations detected
here as a possible signature of a distinct hydro-
dynamic mode. This would suggest that nano-
second charge fluctuations and anomalous
vibrations are not specific to b-YbAlB4 but
rather are universal properties of SM regimes
in quantum materials.
RE FERENCES AND NOTES
1. G. R. Stewart, Rev. Mod. Phys. 73, 797–855 (2001).
2. T. R. Chien, Z. Z. Wang, N. P. Ong, Phys. Rev. Lett. 67,
2088–2091 (1991).
3. S. Nakatsuji et al., Nat. Phys. 4, 603–607 (2008).
J. G. Analytis et al., Nat. Phys. 10, 194–197 (2014).
4.
5. S. Paschen et al., Nature 432, 881–885 (2004).
6. H. Shishido, R. Settai, H. Harima, Y. Ōnuki, J. Phys. Soc. Jpn.
74, 1103–1106 (2005).
7. P. Gegenwart, Q. Si, F. Steglich, Nat. Phys. 4, 186–197
(2008).
8. K. Kuga et al., Sci. Adv. 4, eaao3547 (2018).
9. F. Laliberté et al., Nat. Commun. 2, 432 (2011).
10. E. Fradkin, S. A. Kivelson, M. J. Lawler, J. P. Eisenstein,
A. P. Mackenzie, Annu. Rev. Condens. Matter Phys. 1, 153–178
(2010).
11. J.-H. Chu et al., Science 329, 824–826 (2010).
12. M. J. Lawler et al., Nature 466, 347–351 (2010).
13. J. Zaanen, Y.-W. Sun, Y. Liu, K. Schalm, Holographic Duality in
Condensed Matter Physics (Cambridge Univ. Press, 2016).
14. S. A. Hartnoll, A. Lucas, S. Sachdev, Holographic Quantum
Matter (MIT Press, 2018).
15. S. A. Hartnoll, A. P. Mackenzie, Rev. Mod. Phys. 94, 041002
(2022).
16. P. T. Brown et al., Science 363, 379–382 (2019).
17. L. Prochaska et al., Science 367, 285–288 (2020).
18. S. Vig et al., SciPost Phys. 3, 026 (2017).
19. M. Mitrano et al., Proc. Natl. Acad. Sci. U.S.A. 115, 5392–5396
(2018).
20. A. A. Husain et al., Phys. Rev. X 9, 041062 (2019).
21. O. Berkooz, M. Malamud, S. Shtrikman, Solid State Commun. 6,
185–188 (1968).
22. M. Takano, N. Nakanishi, Y. Takeda, S. Naka, J. Phys. Colloq.
40, C2-313 (1979).
23. M. Seto et al., Phys. Rev. Lett. 102, 217602 (2009).
24. Y. Matsumoto et al., Science 331, 316–319 (2011).
25. M. Okawa et al., Phys. Rev. Lett. 104, 247201 (2010).
26. C. M. Varma, Rev. Mod. Phys. 48, 219–238 (1976).
27. E. V. Sampathkumaran, Hyperfine Interact. 27, 183–192
(1986).
28. R. L. Cohen, M. Eibschütz, K. W. West, Phys. Rev. Lett. 24,
383–386 (1970).
29. I. Nowik, Hyperfine Interact. 13, 89–118 (1983).
30. Materials and methods are available as supplementary
materials.
31. A. Legros et al., Nat. Phys. 15, 142–147 (2019).
32. T. Tomita, K. Kuga, Y. Uwatoko, P. Coleman, S. Nakatsuji,
Science 349, 506–509 (2015).
33. J. P. Hannon, G. T. Trammell, M. Blume, D. Gibbs, Phys. Rev.
Lett. 61, 1245–1248 (1988).
34. Y. Sakaguchi et al., J. Phys. Soc. Jpn. 85, 023602
(2016).
35. P. W. Anderson, J. Phys. Soc. Jpn. 9, 316–339 (1954).
36. R. Kubo, J. Phys. Soc. Jpn. 9, 935–944 (1954).
37. M. Blume, Phys. Rev. 174, 351–358 (1968).
38. D. Sherrington, P. Riseborough, J. Phys. Colloq. 37, C4-255
(1976).
39. A. C. Hewson, D. M. Newns, J. Phys. C Solid State Phys. 12,
1665–1683 (1979).
40. G. T. Trammell, Phys. Rev. 126, 1045–1054 (1962).
41. D. Weschenfelder et al., Hyperfine Interact. 16, 743–745
(1983).
42. S. Watanabe, K. Miyake, Phys. Rev. Lett. 105, 186403
(2010).
43. M. Oshikawa, Phys. Rev. Lett. 84, 3370–3373 (2000).
44. T. Senthil, S. Sachdev, M. Vojta, Phys. Rev. Lett. 90, 216403
(2003).
45. J. H. Pixley, S. Kirchner, K. Ingersent, Q. Si, Phys. Rev. Lett.
109, 086403 (2012).
46. Y. Komijani, P. Coleman, Phys. Rev. Lett. 122, 217001 (2019).
47. R. Masuda et al., Appl. Phys. Lett. 104, 082411 (2014).
48. H. Kobayashi et al., Observation of a critical charge mode in a
strange metal. Zenodo (2023); doi:10.5281/zenodo.7542767.
AC KNOWLED GME NTS
We thank M. Takigawa for very useful discussions and F. Iga for
preparation of single-crystalline YbB12. Funding: The SR-based
174Yb Mössbauer experiments were performed at BL09XU and
BL19LXU on SPring-8 with the approval of the Japan Synchrotron
Radiation Research Institute (JASRI) (proposals 2011A1450,
2012B1521, 2013B1393, 2015A1458, 2016A1363, 2019B1597,
and 2020A1553) and RIKEN (proposals 2016110, 20170019,
20180019, and 20190025). This work is partially supported by
Grants-in-Aids for Scientific Research on Innovative Areas
(15H05882 and 15H05883) from the Ministry of Education,
Culture, Sports, Science, and Technology of Japan; by CREST
(JPMJCR18T3); Japan Science and Technology Agency; and by
Grants-in-Aid for Scientific Research (15K05182, 16H02209,
16H06345, 19H00650, and 23102723) from the Japanese Society
for the Promotion of Science (JSPS); the Canadian Institute
for Advanced Research; the National Science Foundation grant
DMR-1830707 (P.Co. and Y.Kom.) and by the US Department
of Energy (DOE), Office of Science, Basic Energy Sciences under
award DE-SC0020353 (P.Ch.). We work at the Institute for
Quantum Matter, an Energy Frontier Research Center funded by
DOE, Office of Science, Basic Energy Sciences under award
DE-SC0019331. P.Ch. and P.C. thank S. Nakatsuji and the Institute
for Solid State Physics (Tokyo) for hospitality when early stages of
this work were underway. P.Ch., P.Co., and Y.Kom. acknowledge
the Aspen Center for Physics and NSF grant PHY-1607611 where
this work was discussed and further developed. Author
contributions: H.Ko. designed the Synchrotron Mössbauer
experiment and performed it with Y.S., H.Ki., M.O., S.I., R.M.,
Kobayashi et al., Science 379, 908–912 (2023)
3 March 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
Y.Kob., M.S., Y.Y., and K.T. Sample synthesis and characterization
were performed by K.K., S.S., and S.N. Mössbauer analysis was
carried out by H.Ko. Theoretical interpretation was provided by
P.Ch., P.Co., and Y.Kom.; H.Ko., S.N., P.Ch., P.Co., and Y.Kom.
contributed to writing the manuscript. Competing interests: The
authors declare no competing interests, financial or otherwise.
Data and materials availability: All data and simulation codes
presented in this paper are deposited in Zenodo (48). License
information: Copyright © 2023 the authors, some rights reserved;
exclusive licensee American Association for the Advancement of
Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse
Materials and Methods
Supplementary Text
Figs. S1 to S12
References (49–75)
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abc4787
Submitted 28 April 2020; resubmitted 16 June 2021
Accepted 1 February 2023
10.1126/science.abc4787
Kobayashi et al., Science 379, 908–912 (2023)
3 March 2023
5 of 5
|
10.1126_science.abc8372
|
RES EARCH
SUPERCONDUCTIVITY
Restored strange metal phase through suppression of
charge density waves in underdoped YBa2Cu3O7–d
Eric Wahlberg1, Riccardo Arpaia1,2*, Götz Seibold3, Matteo Rossi2†, Roberto Fumagalli2,
Edoardo Trabaldo1, Nicholas B. Brookes4, Lucio Braicovich2,4, Sergio Caprara5,6, Ulf Gran7,
Giacomo Ghiringhelli2,8, Thilo Bauch1, Floriana Lombardi1*
The normal state of optimally doped cuprates is dominated by the “strange metal” phase
that shows a linear temperature (T) dependence of the resistivity persisting down to the
lowest T. For underdoped cuprates, this behavior is lost below the pseudogap temperature
T*, where charge density waves (CDWs), together with other intertwined local orders,
characterize the ground state. We found that the T-linear resistivity of highly strained,
ultrathin, underdoped YBa2Cu3O7–d films is restored when the CDW amplitude, detected by
resonant inelastic x-ray scattering, is suppressed. This observation suggests an intimate
connection between the onset of CDWs and the departure from T-linear resistivity in underdoped
cuprates. Our results illustrate the potential of using strain control to manipulate the ground
state of quantum materials.
scattering process (4, 7). Local interactions
among quasiparticles suggest that ordinary
metals cannot thermalize on such a short time
scale; reaching the Planckian limit would re-
quire every particle to be entangled with every
other particle in the system.
In cuprates, the T-linear resistivity is lost for
doping above and below the optimal doping.
In the overdoped regime, the recovery of an
almost T 2 dependence of the resistivity, typical
of a Fermi liquid, is a consequence of the
increased screening of the electron-electron
interactions caused by a higher charge carrier
density. In the underdoped regime, the devi-
ation from T-linear behavior happens at tem-
peratures close to T *, known as the pseudogap
temperature, where states are missing at the
Fermi energy (8). In the pseudogap region, the
high-temperature superconductor phase dia-
gram also hosts a plethora of intertwined elec-
tronic local ordering phenomena that break
rotational/translational symmetry (1, 9–13);
charge density wave (CDW) order (11) is the
most prominent one. The association between
the departure from the T-linear resistivity and
the occurrence of the pseudogap phenomenon
has long been speculated. However, there is
no consensus on the physics at play, nor on the
causality hierarchy among pseudogap, local
orders, and strange metal phenomenology
(14, 15). This is because the strange metal
exhibits its most salient features in transport
and its connection to spectral signatures is
elusive (16). More specifically, the challenge
is to disentangle the various possible mecha-
nisms leading to the breakdown of the T-linear
resistivity, such as the occurrence of the pseu-
dogap and the appearance of local orders such
as CDW. One way to address this challenge is
to tune the local properties of underdoped
high-temperature superconductors. In partic-
ular, the CDW can be strongly modified under
pressure (17, 18), strong magnetic fields (19),
and strain in crystals (20) and thin films (21).
To tune the ground state in thin films of
the cuprate material YBa2Cu3O7–d (YBCO), we
use the geometric modification of its unit cell
under the strong strain induced by the sub-
strate. We show that the T-linear resistivity
dependence is completely recovered (down to
the superconducting critical temperature Tc)
along the b axis when the CDW, detected by
resonant inelastic x-ray scattering (RIXS), is
suppressed along the a axis.
C uprate high-temperature superconduc-
tors belong to a class of materials where
strong electron-electron correlations play
a fundamental role, and whose uncon-
ventional properties might require aban-
doning traditional concepts of solid-state physics
for a proper description (1). The “strange metal”
phase of these superconductors is one of the
most striking manifestations of the strong cor-
relations. At optimal doping, this phase man-
ifests as a linear temperature dependence of
the resistivity that persists to the lowest T
when superconductivity is suppressed (1–4).
This behavior is fundamentally different from
that observed in more conventional metals (3),
where a T-linear dependence of the resistivity
is found only at high temperatures where
phonon scattering dominates the transport.
T-linear resistivity is also found in other strong-
ly correlated systems, including pnictides (5)
and magic-angle bilayer graphene (6).
Recent developments have attempted to
model this behavior by considering that the
scattering time approaches the fundamental
Planckian limit defined by t = ћ/kBT (where ћ
and kB are the reduced Planck and Boltzmann
constants) irrespective of the nature of the
1Quantum Device Physics Laboratory, Department of
Microtechnology and Nanoscience, Chalmers University of
Technology, SE-41296 Göteborg, Sweden. 2Dipartimento di
Fisica, Politecnico di Milano, I-20133 Milano, Italy. 3Institut
für Physik, BTU Cottbus-Senftenberg, D-03013 Cottbus,
Germany. 4ESRF, European Synchrotron, F-38043 Grenoble,
France. 5Dipartimento di Fisica, Università di Roma “La
Sapienza,” I-00185 Roma, Italy. 6CNR-ISC, I-00185 Roma,
Italy. 7Division of Subatomic, High-Energy and Plasma
Physics, Chalmers University of Technology, SE-41296
Göteborg, Sweden. 8CNR-SPIN, Dipartimento di Fisica,
Politecnico di Milano, I-20133 Milano, Italy.
*Corresponding author. Email: floriana.lombardi@chalmers.se
(F.L.); riccardo.arpaia@chalmers.se (R.A.)
†Present address: Stanford Institute for Materials and Energy
Sciences, SLAC National Accelerator Laboratory, Menlo Park, CA
94025, USA.
Fig. 1. Thickness dependence of the YBCO lattice parameters for films grown on MgO
substrates. (A) Lattice parameters (a, b, and c axes are indicated by squares, triangles, and
circles, respectively) of YBCO, measured by x-ray diffraction at 300 K, as function of the thickness
of films with hole doping p ≈ 0.12 grown on MgO. Thick lines are guides to the eye. (B) False-color
scanning electron microscope image of a typical device used to measure the resistivity r,
with the current I flowing at an angle f with respect to the YBCO [100] direction (a axis). Inset:
Overview of the patterned samples.
Wahlberg et al., Science 373, 1506–1510 (2021)
24 September 2021
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
Thin-film devices
The films we use span a wide range of hole-
doping p, going from the strongly underdoped
(p ≈ 0.10) up to the optimally doped (p ≈ 0.17)
regime (22, 23). The strain was modified both
by changing the substrate and by varying the
film thickness t in a range from 50 nm down
to 10 nm. We used (110)-oriented MgO and 1°
vicinal angle (001) SrTiO3 (STO) substrates to
grow untwinned YBCO films (24). When the
YBCO thickness is reduced to a few unit cells
(t = 10 nm), films grown on MgO are char-
acterized by a considerable elongation of
the b axis and contraction of the c axis, with
the total volume of the unit cell unchanged
with respect to relaxed systems (Fig. 1A). For
films grown on vicinal cut STO, the YBCO unit
cell is instead almost thickness-independent
(fig. S1).
Figure 1B shows a typical device used to
measure the resistivity as a function of tem-
perature. The devices are patterned at an
angle f with respect to the YBCO [100] direc-
tion by using a carbon mask in combination
with electron beam lithography and Ar+ ion
milling (22, 23, 25).
Angular dependence of in-plane resistivity
Figure 2A shows the temperature dependence
of the resistivity r, measured in two devices
oriented along the YBCO a and b axes (f = 0°
and 90°, respectively) realized in an under-
doped (p = 0.11) 50-nm-thick film grown on
an MgO substrate. The resistivity anisotropy
ratio at T = 290 K, defined by ra(290 K)/rb
(290 K), is ~1.2, a value fully compatible with
the level of hole doping (26). The temperature
TL, estimated as the temperature below which
the resistivity normalized to 290 K deviates by
1% from the linear fit, is approximately the
same for both devices, and it is very close to
the pseudogap temperature T * of single crys-
tals at comparable doping measured using
spectroscopic techniques (8, 27). For this film
thickness, we therefore have T a
L
≈T b
L
≈T (cid:1).
Figure 2B shows analogous data for two de-
vices patterned on a 10-nm-thick film on MgO
(grown with the same deposition conditions
as the 50-nm-thick film) with doping p com-
parable to Fig. 2A. The superconducting crit-
ical temperature Tc at this reduced thickness
remains almost unchanged (28) relative to 50-
nm-thick films. However, the overall r(T ) be-
havior is very different. We observe three
striking features: (i) The in-plane anisotropy
of the resistivity ra(290 K)/rb(290 K) is much
larger relative to the 50-nm-thick film, (ii) the
slopes ga,b of the high-temperature linear re-
sistivity along the a and b axes are very dif-
ferent, and (iii) the resistivity along the b axis
rb(T) has a much broader temperature range
of linearity, as is clearly highlighted by sub-
tracting from rb(T) the linear fit rL(T) = r0 +
gbT of the high-temperature resistive behavior
Fig. 2. Angular dependence of the YBCO in-plane resistivity as a function of the thickness of
films on MgO under strain. (A) r(T) of two devices, oriented along the a- and b-axis directions,
patterned on a 50-nm-thick underdoped (p = 0.11) film. At T = 290 K, ra/rb = 1.2. The dashed
lines are the linear fits of the curves for T > 260 K. The inset shows the determination of TL,
which is the temperature where the resistivity normalized to 290 K deviates by 1% from the linear fit
(r0 and g are the coefficients of the fit). (B) r(T) of two devices, oriented along the a- and b-axis
directions, patterned on a 10-nm-thick underdoped (p = 0.12) film. At T = 290 K, ra/rb = 2.1.
(C) Determination of TL in two slightly underdoped devices, oriented along the a- and b-axis
directions, patterned on a 10-nm-thick slightly underdoped (p = 0.147) film. r(T) along the
b axis is linear down to the onset of superconducting fluctuations around T = 100 K. (D) Simplified
model of the typical Fermi surface of cuprates (here we disregard the influence of the CuO chains).
For this kind of Fermi surface, the resistivity should be isotropic in the copper oxide planes.
(E) Hypothetical anisotropic model Fermi surface that is compatible with the anisotropic transport
of the 10-nm-thick devices.
(Fig. 2B, inset). Here, r0 and gb are respectively
the intercept at T = 0 and the slope of the high-
temperature linear dependence. At p = 0.14
and p = 0.147 (Fig. 2C and fig. S4C), for 10-nm-
thick films we find that the T-linear behavior
is completely recovered down to the super-
conducting transition. This finding is crucial
because it indicates that the “strange metal”
behavior is restored in ultrathin underdoped
YBCO. What are the conditions for this to
happen?
The most prominent structural effect we en-
counter by reducing the thickness of the films
is that the YBCO b axis expands, whereas the a
axis is only slightly modified; at the same time,
the total volume of the cell remains unaltered
as a consequence of a c-axis contraction. One
of the effects of the strain is therefore to in-
crease the orthorhombicity of few-nm-thick
films. For 10-nm-thick underdoped films (p ≈
0.12), the values of the lattice parameters a and
b are similar to those of YBCO with a much
higher doping (29), close to the optimal dop-
ing. However, this effect cannot explain the
anomalously high anisotropy ratio ra(290 K)/
rb(290 K) that we observe (Fig. 2B), a value
that in an optimally doped YBCO crystal is
mainly related to the presence of CuO chains
(26). Indeed, within a simple tight-binding de-
scription, an increase of ~0.02 Å in the b-axis
lattice parameter—as we observe upon re-
ducing the film thickness from 50 nm to 10 nm
(Fig. 1A)—would reduce the corresponding
hopping parameter between neighbor sites in
the b direction by only ~1% (30). Moreover, the
resulting (weak) modification of the electronic
structure would induce the opposite anisot-
ropy in the a- and b-axis resistivities, namely
rb > ra (because of the larger b-axis value),
relative to what we experimentally observe.
This rules out the increased orthorhombicity
as a possible direct explanation of the anisot-
ropy we observe.
The very different slopes of ra(T) and rb(T)
give a hint of the physics at play in the 10-nm-
thick films. Following Boltzmann transport
Wahlberg et al., Science 373, 1506–1510 (2021)
24 September 2021
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Angular dependence of the in-plane resistivity of the 10-nm YBCO on MgO as a function
of oxygen doping. (A) r(T) measured along the a-axis direction of 10-nm-thick films, with different
doping levels, grown on MgO substrates. TL is extracted as described in Fig. 1. The hatched
regions represent the temperature range above TL where at each p the T-linear resistivity regime
occurs. (B) Same as (A), but with r(T) measured along the b axis. (C) r(T), normalized to its value
at T = 290 K, for a 10-nm-thick film (p = 0.120) on MgO is shown as a function of f (i.e., with
respect to the a-axis direction). For each angle, the temperature TL has been extracted (black
spheres and dots). TL is approximately constant as a function of f, except around f = 90°, where
it exhibits a suppression. (D) TL values as a function of f for 10-nm-thick films with different
doping levels (p ≈ 0.099, 0.103, 0.108, 0.117, 0.118, 0.120, 0.123, 0.134, 0.140, or 0.147).
theory, the conductivity is given by
sa;b Tð Þ ¼ 2e2
X
k
v2
F;a;b
G kð Þ
(cid:3)n′F
f
g
ð1Þ
where vF,a,b is the Fermi velocity, G(k) is the
k-dependent scattering rate, and n′F is the
derivative of the Fermi distribution. Conse-
quently, the ratio G kð Þ=v2
F;a;b determines the
slope ga,b of the temperature-dependent film
resistivity. From the experimental data, we
have ga >> gb (at p = 0.11, ga = 4.9 mW·cm/K
and gb = 2.1 mW·cm/K), which indicates that
vF,b >> vF,a and therefore that the Fermi sur-
face is already strongly anisotropic at room
temperature. This is illustrated in Fig. 2, D
and E, which show respectively the typical
isotropic Fermi surface for the cuprates (where
the contribution of the chains is neglected)
and an anisotropic distorted Fermi surface,
compatible with our experimental resistivity
anisotropy. From Eq. 1, one may argue that
anisotropic elastic scattering processes [e.g.,
due to small-angle scattering from impurities
between CuO2 planes (31)] can also account
for the observed resistivity anisotropy. How-
ever, in this case G(k) would be determined by
the local density of states (31); that is, G would
still be proportional to 1/vF. Therefore, only an
anisotropy of the Fermi velocity, which breaks
the C4 symmetry of the crystal, can account for
the observed resistivity anisotropy.
A Fermi surface of the type shown in Fig. 2E
can be the consequence of a strong electronic
nematicity in the system. Such a state has been
extensively investigated from a theoretical point
of view (32–36) and it has been experimentally
found, by in-plane resistivity anisotropy mea-
surements, in tetragonal La2–xSrxCuO4 films
(37). We speculate that in the 10-nm thin films,
the strain-induced distortion of the cell plays
a fundamental role in stabilizing a nematic
ground state already at room temperature.
The wider T-linear behavior of rb(T) in ultra-
thin films is not observed in YBCO on vicinal
angle STO substrates. Here TL does not change;
it is the same along the a and b axes and going
from 50-nm-thick to 10-nm-thick films (see fig.
S3). This in agreement with the fact that the
slopes ga,b of ra(T) and rb(T) are comparable in
10- and 50-nm-thick films, so on STO substrates
the Fermi surface is not substantially modified
by strain at reduced thicknesses.
We have therefore arrived at the main result
of our paper: A specific strain in underdoped
10-nm-thick YBCO films on a MgO substrate
induces a nematic state that modifies the
Fermi surface already at room temperature.
But why would a distorted Fermi surface re-
cover the T-linear resistivity behavior along the
b axis? The answer to this question comes from
the study of the r(T) dependence as a func-
tion of doping. Figure 3 shows ra(T) and rb(T)
respectively as a function of the doping p (Fig.
3, A and B) and of the angle f (Fig. 3, C and D)
for 10-nm-thick YBCO films on MgO substrates
(as a function of the thickness in fig. S4).
The ra(T) curves are rather conventional
and in agreement with previous results (22, 26).
However, rb(T) (Fig. 3B) looks very different.
As anticipated earlier, for p ≈ 0.14 we have
completely recovered the strange metal be-
havior: The T-linear dependence extends through
the entire temperature range until super-
conductivity sets in. However, the situation
changes at lower doping: For p ≈ 0.10, the
extracted TL along b is close to the value along
the a axis (Fig. 2, C and D), and rb(T) shows a
pronounced upturn at low temperature before
the superconducting transition. The upturn of
the rb(T) in our 10-nm-thick YBCO films is
observed at higher doping relative to the ra(T)
(Fig. 3, A and B); indeed, it already appears at
p = 0.13. This upturn in the resistivity has
been attributed to the loss of high-mobility
electron pockets because of the CDW order
ending at p ≈ 0.08 (38) and/or to the proximity
to the antiferromagnetic Mott insulator through
spin density waves (39–41). The doping de-
pendence of rb(T) therefore points toward a
strong involvement of the CDW order in the
phenomenology we observe.
Resonant inelastic x-ray measurements of CDW
To characterize the CDW order in the YBCO
films, we used RIXS at the Cu L3 edge (~930 eV)
(23). We investigated films with thickness of
10 and 50 nm grown on both MgO and STO
at the doping p ≈ 0.125 where the intensity of
the CDW is strongest (11, 42). To isolate the
contribution of the CDW, we measured RIXS
spectra at T = 70 K (i.e., close to Tc), where
the CDW signal is maximized, and at T = 200 K,
a temperature where the CDW contribution is
negligible. The CDW peak has been explored
along a and b with two orthogonal cuts along
the H and K directions of reciprocal space,
centered around the wave vector of the CDW
qCDW
Þ. Shown in Fig. 4, A and
c
B, is the quasi-elastic component of the RIXS
spectra along the a and b axes for a 50-nm-
thick film on MgO, as a function of H, at T = 70 K
and T = 200 K. Along both directions, at T =
200 K only a broad peak is present (Fig. 4, A
¼ HCDW; KCDW
ð
Wahlberg et al., Science 373, 1506–1510 (2021)
24 September 2021
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
and B, red regions); at T = 70 K, the signal is
given by the sum of a broad peak (similar to
that measured at high temperature) and of
a narrow peak. This narrow, temperature-
dependent peak is a signature of CDW; the
broad, almost temperature-independent peak
is instead a signature of short-range charge
density fluctuations (CDFs) (43) (i.e., charge
modulations, precursors of CDWs), which are
present in the phase diagram at any tem-
perature and in a very broad doping range,
including the overdoped region. The same
measurements as in Fig. 4, A and B, are re-
ported in Fig. 4, C and D, for a 10-nm-thick
film on MgO. Here, along the b axis (Fig. 4D),
both the CDW and the CDF peaks are very
similar to those measured in the thick film.
Along the a axis (Fig. 4C), the situation is
markedly different. The broad-in-q// CDF peak
is unchanged with respect to the 50-nm-thick
sample; in contrast, the narrow CDW peak,
emerging at lower temperature, is almost neg-
ligible. This occurrence has been verified on
the same sample, measuring along the K di-
rection (fig. S5), and on other ultrathin films
on MgO with different doping (fig. S6).
Note that our films are not perfectly un-
twinned (the untwinning degree is ~85% for
films grown on MgO, ~90% for films grown
on STO). The small temperature-dependent
CDW signal measured along the a axis in the
10-nm films on MgO can be attributed to
twinned domains. Indeed, once the twinning
is taken into account, the actual CDW signal
becomes effectively negligible in the a-axis
direction. We conclude that in 10-nm-thick
films on MgO, the CDW is unidirectional
and directed along the b axis, whereas the
CDW in YBCO grown on STO is thickness-
independent (fig. S7). Strain-induced mod-
ifications of CDW have also been recently
observed in YBCO single crystals, where the
in-plane uniaxial compression along either
a or b gives rise to an enhancement of the
CDW in the orthogonal direction (44).
Discussion
Our results are in line with the theoretical
predictions of unidirectional CDWs in the pres-
ence of nematicity (34, 35), as a consequence
of the modified Fermi surface. Our experiment
additionally shows a clear correlation between
the T-linear dependence of r(T) and the charge
order: The recovery of the T-linear resistivity
along the b axis is associated with the sup-
pression of the CDW along the YBCO a axis. In
this scenario, only CDFs survive in the system
at any temperature, which might be relevant
both for Planckian metal (45) and marginal
Fermi liquid theories (46) of the strange metal.
The strong correlation between the CDW
and the breakdown of the T-linear behavior
is further supported by the coincidence of
the onset temperatures for the two pheno-
mena. Figure S8 shows the temperature de-
pendence of the CDW order for our 50- and
10-nm-thick films on MgO and STO substrates.
Within experimental error, TCDW (for p =
0.125) is the same as the temperature at
which the resistivity departs from the linearity
T a
L and the same as the pseudogap temper-
ature T* [taken from literature (8, 27)]. This is
illustrated in Fig. 5, which shows a revised
phase diagram for the 10-nm-thick films. Our
RIXS measurements therefore require the re-
vision of the common belief that the CDW is a
low-temperature phenomenon happening at
temperatures well below TL and T*, as indi-
cated in early studies (11, 42). The coincidence
among T *, TL, and TCDW is a common feature
of doping above p = 1/8 where the CDW order
is strong, as also supported by recent RIXS
data (43), which are included for completeness
in Fig. 5.
The quality of our samples allows us to ex-
clude the possibility that structural changes of
the CuO chains along the b axis in 10-nm films
might have a role in the phenomenology we
have observed. For overdoped films, where the
conductivity of the chains causes an upward
deviation from linearity in r(T), we observe the
same rb(T) dependence for 50-nm and 10-nm
films and very similar values of TL (fig. S2),
which excludes CuO chain modification effects
in our very thin films.
However, there is still an issue to be ad-
dressed: Why, in our films, does the disap-
pearance of the CDW along the YBCO a axis
lead to a recovery of the T-linear dependence
of the resistivity along the b axis (Figs. 2B and
Fig. 4. Thickness dependence of the CDW in YBCO thin films. (A and B) Quasi-elastic scans
measured at T = 70 K and T = 200 K on the 50-nm thin film on MgO along the (H, 0) direction
(i.e., a axis) (A) and the (H, KCDW) direction (i.e., b axis) (B). (C and D) Same as (A) and (B), but
on a 10-nm-thick sample. The measurements were performed along the (H, 0) direction (i.e., a axis)
(C) and the (H, KCDW) direction (i.e., b axis) (D). In the 10-nm-thick sample, the CDW intensity
along the a axis is almost negligible. If we take into account the percentage of twin domains
(~15%) present in our films, we conclude that in films a few unit cells thick on MgO, the CDW is
unidirectional along the b axis. The green line in the inset of each panel shows the direction of
the scan relative to the CDW peak in reciprocal space.
Wahlberg et al., Science 373, 1506–1510 (2021)
24 September 2021
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. Intertwined phases in the phase diagram
of YBCO ultrathin films. The antiferromagnetic
(AFM) and superconducting (SC) states are present
below TN and Tc, respectively [the Tc dome is
derived from (22)]. TL serves as a crossover line
between the quantum-critical region of the phase
diagram exhibiting strange metal properties (to the
right of TL) and the noncritical region, where the
linearity of the resistivity versus temperature is lost
(to the left of TL). Along the a axis, TL ¼ Ta
squares) is close to the pseudogap temperature
T* as determined by other techniques, whereas
along the b axis (blue triangles), TL is suppressed to
lower temperatures. The error bar on Tb
L highlights
the uncertainty of determining the correct value
owing to proximity to Tc. The lines are guides to the
eye. TCDW (stars) is the onset temperature of the
CDW, as determined by resonant inelastic x-ray
scattering. The open star is related to the doping
level investigated by RIXS in this manuscript (for
10-nm-thick films, revealed only along the b axis);
the solid stars are from (43).
L (pink
4C)? It is not uncommon in other compounds
that the occurrence of a unidirectional CDW
phenomenon affects the transport properties
in a direction orthogonal to the CDW q-vector
(47). In our experiment, the reason can be
linked to the nematic Fermi surface for very
thin films. The generic Fermi surface (Fig. 2D)
supports two scattering processes along the
b axis associated with the quasi-nesting prop-
erties of the CDW q-vector: one within the
Brillouin zone and the other connecting two
adjacent Brillouin zones at (p, 0) (fig. S9A).
For a conventional Fermi surface, the scatter-
ing rates of these two processes are compa-
rable (23). However, for the distorted Fermi
surface of Fig. 2E, the region around the (p, 0)
point has a much higher density of states rel-
ative to the unperturbed one (fig. S9). As a
consequence, the CDW scattering rate in-
volving two adjacent Brillouin zones becomes
larger than the scattering rate within the same
Brillouin zone by a factor of >4 (23). Follow-
ing the Boltzmann equation formalism, this
strongly affects the resistivity along the a axis
while leaving the resistivity along the b axis
unaltered.
≈T b
Finally, we consider the role of the pseudo-
gap in the departure of the T-linear resistivity
in underdoped cuprates. From the slopes of
r(T) along the a and b axes, ga,b, we infer that
the Fermi surface is anisotropic at any doping
(the ratio ga/gb is almost doping-independent).
For p ≤ 0.1, we observe that TL, which at
that doping level is much higher than TCDW
and comparable to T * [taken from litera-
ture (8, 27)], is isotropic (T a
L). This hints at
L
a pseudogap that is isotropic on a distorted
Fermi surface. Such an isotropic pseudogap
cannot explain the otherwise anisotropic TL
we observed at higher doping (p > 0.1). This
allows us to conclude that the pseudogap is
not the main effect responsible for the de-
parture of the linearity when the CDW order is
strong (p > 0.11), which is the range of doping
in our experiment. However, one cannot ex-
clude a major influence of the pseudogap on
the transport at lower doping when T * be-
comes much higher (close to room tempera-
ture and beyond) than TCDW.
RE FERENCES AND NOTES
1. B. Keimer, S. A. Kivelson, M. R. Norman, S. Uchida, J. Zaanen,
Nature 518, 179–186 (2015).
2. R. Daou et al., Nat. Phys. 5, 31–34 (2009).
3.
J. A. N. Bruin, H. Sakai, R. S. Perry, A. P. Mackenzie, Science
339, 804–807 (2013).
4. A. Legros et al., Nat. Phys. 15, 142–147 (2019).
5. S. Kasahara et al., Phys. Rev. B 81, 184519 (2010).
6. Y. Cao et al., Nature 556, 43–50 (2018).
7.
J. Zaanen, SciPost Phys. 6, 061 (2019).
8. T. Timusk, B. Statt, Rep. Prog. Phys. 62, 61–122
(1999).
9. C. M. Varma, Phys. Rev. B 55, 14554–14580 (1997).
10. Y. Sato et al., Nat. Phys. 13, 1074–1078 (2017).
11. G. Ghiringhelli et al., Science 337, 821–825 (2012).
12. J. M. Tranquada, B. J. Sternlieb, J. D. Axe, Y. Nakamura,
B. S. Uchida, Nature 375, 561–563 (1995).
13. B. Fauqué et al., Phys. Rev. Lett. 96, 197001 (2006).
14. E. Fradkin, S. A. Kivelson, J. M. Tranquada, Rev. Mod. Phys. 87,
457–482 (2015).
15. T. Wu et al., Nat. Commun. 6, 6438 (2015).
16. A. Alexandradinata et al., arXiv 2010.00584
[cond-mat.str-el] (1 October 2020).
17. O. Cyr-Choinière et al., Phys. Rev. B 98, 064513
(2018).
18. S. M. Souliou et al., Phys. Rev. B 97, 020503 (2018).
19. J. Chang et al., Nat. Phys. 8, 871–876 (2012).
20. H. H. Kim et al., Science 362, 1040–1044 (2018).
21. M. Bluschke et al., Nat. Commun. 9, 2978 (2018).
22. R. Arpaia, E. Andersson, E. Trabaldo, T. Bauch, F. Lombardi,
Phys. Rev. Mater. 2, 024804 (2018).
23. See supplementary materials.
24. R. Arpaia et al., Phys. Rev. Mater. 3, 114804 (2019).
25. S. Nawaz, R. Arpaia, F. Lombardi, T. Bauch, Phys. Rev. Lett.
110, 167004 (2013).
32. S. A. Kivelson et al., Rev. Mod. Phys. 75, 1201–1241
(2003).
33. A. V. Maharaj, I. Esterlis, Y. Zhang, B. J. Ramshaw,
S. A. Kivelson, Phys. Rev. B 96, 045132 (2017).
34. Y. Gannot, B. J. Ramshaw, S. A. Kivelson, Phys. Rev. B 100,
045128 (2019).
35. H. Yao, D. H. Lee, S. Kivelson, Phys. Rev. B 84, 012507
(2011).
36. Y. Zhang et al., Proc. Natl. Acad. Sci. U.S.A. 117, 68–72
(2020).
37. J. Wu, A. T. Bollinger, X. He, I. Božović, Nature 547, 432–435
(2017).
38. D. Leboeuf et al., Phys. Rev. B 83, 054506 (2011).
39. P. Bourgeois-Hope et al., arXiv 1910.08126
[cond-mat.supr-con] (17 October 2019).
40. S. Komiya, Y. Ando, Phys. Rev. B 70, 060503 (2004).
41. S. Ono et al., Phys. Rev. Lett. 85, 638–641 (2000).
42. S. Blanco-Canosa et al., Phys. Rev. B 90, 054513
(2014).
43. R. Arpaia et al., Science 365, 906–910 (2019).
44. H.-H. Kim et al., Phys. Rev. Lett. 126, 037002 (2021).
45. L. Delacrétaz, B. Goutéraux, S. Hartnoll, A. Karlsson,
SciPost Phys. 3, 025 (2017).
46. G. Seibold et al., Commun. Phys. 4, 7 (2021).
47. A. A. Sinchenko, P. D. Grigoriev, P. Lejay, P. Monceau,
Phys. Rev. Lett. 112, 036601 (2014).
48. E. Wahlberg et al., Restoring the strange metal phase via
suppression of charge density waves in an underdoped cuprate
superconductor; https://doi.org/10.5281/zenodo.4724109
(2021).
AC KNOWLED GME NTS
We thank M. Grilli, C. Di Castro, and M. Moretti for enlightening
discussions, D. Betto for initial tests at ID32 on ultrathin
YBCO films, and L. Martinelli for helping with complementary
RIXS measurements. This work was performed in part at
Myfab Chalmers. The RIXS experimental data were collected at
beamline ID32 of the European Synchrotron (ESRF) in
Grenoble, France, using the ERIXS spectrometer designed
jointly by the ESRF and Politecnico di Milano. Funding: Supported
by the European Union’s projects NANOCOHYBRI (Cost Action
CA16218) and OXiNEMS (Horizon 2020, grant agreement
828784), the Knut and Alice Wallenberg foundation (project
2014.0102), and the Swedish Research Council (VR) under
projects 2015-04368 (U.G.), 2018-04658 (F.L.), 2020-04945
(R.A.) and 2020-05184 (T.B.); Deutsche Forschungsgemeinschaft
grant SE 806/19-1 (G.S.); Italian Ministry of University
and Research PRIN project 2017Z8TS5B (G.G. and S.C.);
and the University of Rome Sapienza through projects
Ateneo 2018 (grant RM11816431DBA5AF), Ateneo 2019
(grant RM11916B56802AFE), and Ateneo 2020 (grant
RM120172A8CC7CC7) (S.C.). Author contributions: F.L.,
E.W., R.A., and T.B. conceived and designed the experiments;
E.W., R.A., and E.T. grew the samples; E.W. performed
the transport measurements; R.A., E.W., F.L., T.B., M.R.,
R.F., G.G., and N.B.B. performed the RIXS measurements;
R.A., G.G., and L.B. analyzed and interpreted the RIXS
experimental data; F.L., E.W., R.A., and T.B. interpreted the
transport measurements with the theoretical insights of U.G.;
G.S. and S.C. performed the modeling of the experimental
data and the theoretical calculations; and F.L., R.A., E.W.,
U.G., and T.B. wrote the manuscript with contributions from
all authors. Competing interests: The authors declare
no competing financial interests. G.G. is member of the
Science Advisory Council of the ESRF. Data and materials
availability: All experimental data shown in the main text
and in the supplementary materials are accessible at the
Zenodo repository (48).
26. Y. Ando, K. Segawa, S. Komiya, A. N. Lavrov, Phys. Rev. Lett.
SUPPLEMENTARY MATERIALS
88, 137005 (2002).
27. I. M. Vishik et al., Proc. Natl. Acad. Sci. U.S.A. 109,
18332–18337 (2012).
28. R. Arpaia et al., Phys. Rev. B 96, 064525 (2017).
29. J. D. Jorgensen et al., Phys. Rev. B 41, 1863–1877
(1990).
30. F. Liu, M. R. Press, S. N. Khanna, P. Jena, Phys. Rev. B 39,
6914–6924 (1989).
31. E. Abrahams, C. M. Varma, Proc. Natl. Acad. Sci. U.S.A. 97,
5714–5716 (2000).
https://science.org/doi/10.1126/science.abc8372
Materials and Methods
Supplementary Text
Figs. S1 to S9
References (49–64)
21 May 2020; resubmitted 8 February 2021
Accepted 22 July 2021
10.1126/science.abc8372
Wahlberg et al., Science 373, 1506–1510 (2021)
24 September 2021
5 of 5
|
10.1126_science.abc5667
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
STRUCTURAL BIOLOGY
Structural basis of antagonizing the vitamin K
catalytic cycle for anticoagulation
Shixuan Liu, Shuang Li, Guomin Shen, Narayanasami Sukumar, Andrzej M. Krezel, Weikai Li*
INTRODUCTION: Vitamin K antagonists (VKAs),
such as warfarin, are oral anticoagulants com-
monly used to treat and prevent thrombo-
embolic diseases, including stroke and heart
attack. Vitamin K supplementation, on the
other hand, has saved the lives of numerous
newborns with deficient hemostasis. Central
to these therapeutic processes is vitamin K ep-
oxide reductase (VKOR), an endoplasmic mem-
brane enzyme that generates the active form
of vitamin K to support blood coagulation.
VKAs inhibit VKOR catalysis, which is carried
out by two cysteine pairs, one directly reduc-
ing substrates and another mediating electron
transfers. VKOR homologs and paralogs (VKOR-
like) constitute a large family of integral mem-
brane thiol oxidoreductases, but understanding
their catalytic and inhibitory mechanisms
has been challenging in the absence of high-
resolution structures.
RATIONALE: Overdose of VKAs often causes
major or fatal bleeding, accounting for one-
third of hospitalizations for all adverse drug
S–S
S–S
Y139
N80
Warfarin
SH
S–S
Y139
N80
Cysteine
ct
adduct
Electron
transfer
nsfe
Vitamin K
epoxide or quinone
Partially
oxidized
g
cin
u
d
e
R
ule
c
ole
m
S–S
S S
Open
Catalytic
cycle
Closed
S–S
–S
Y139
N80
Fully
oxidized
Vitamin K
quinone or hydroquinone
S–S
S–S
Y139
N80
S–S
SH
N80
Y139
Cysteine
adduct
Electron
transfer
reactions in older adults. Despite decades of
clinical difficulties, mechanistic insights are
lacking for the antagonism of VKAs and for
their target enzyme VKOR, which manages to
transfer electrons across the water-membrane
interface to support its distinct activity of
epoxide reduction. Here, we present 11 crystal
structures of human VKOR and a VKOR-like
paralog with various substrates and antago-
nists in different functional states, revealing
nearly the entire catalytic cycle of VKOR en-
zymes and the action of VKAs.
RESULTS: Structures with four representative
VKAs show that their hydrogen bonding with
Asn80 and Tyr139 provides the recognition spe-
cificity in a largely hydrophobic pocket. This
VKA-binding pocket, also serving as the active
site, is surrounded by a four-transmembrane-
helix bundle and covered by a cap domain.
Mutations that destabilize the cap domain or
directly disrupt the warfarin-binding interac-
tions result in warfarin resistance. Metabolic
inactivation of warfarin is through a hydroxyl
modification that is energetically unfavorable
in the hydrophobic pocket. High-potency in-
hibition by “superwarfarins” is afforded by
their large side groups that bind to a tunnel
designated for the isoprenyl chain of vitamin
K. Structural comparison with a ligand-free
state suggests that local binding interactions
of warfarin lead to a global change from open
to closed protein conformation.
Substrate-bound structures reveal that a re-
duced active-site cysteine is required to form a
charge-transfer complex or covalent complex.
These substrate adducts form hydrogen bonds
with Asn80 and Tyr139 to facilitate the catalytic
chemistry, and their stable binding inter-
actions induce the closed protein confor-
mation. This conformational change brings
together the two pairs of cysteines, trigger-
ing electron transfer that enables the reduc-
tion of substrates.
CONCLUSION: The structures reveal an acti-
vation mechanism in which stably bound
substrate adducts trigger restructuring
of VKOR to promote electron transfer.
The high potency of VKAs results from
mimicking the conformational change at
the global level and the substrate hydrogen-
bonding interactions at the local level.
These mechanistic insights provide the ba-
sis to design new therapeutic strategies of
anticoagulation.▪
The catalytic cycle and inhibition of VKOR. Partially oxidized VKOR forms a cysteine adduct with
substrates, vitamin K epoxide, or quinone, whose binding induces a closed conformation, juxtaposing all
cysteines (S–S or SH) for unimpeded electron transfer. VKOR becomes fully oxidized with an open
conformation that releases reaction products, vitamin K quinone, or hydroquinone. Warfarin locks VKOR
in both redox states into the closed conformation. The luminal and transmembrane domains of VKOR are
shown as a pink hemisphere and gray cylinder, respectively. Y139, Tyr139; N80, Asn80.
The list of affiliations is available in the full article online.
*Corresponding author. Email: weikai@wustl.edu
Cite this article as S. Liu et al., Science 371, eabc5667
(2021). DOI: 10.1126/science.abc5667
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.abc5667
Liu et al., Science 371, 43 (2021)
1 January 2021
1 of 1
RES EARCH
R E S E A R C H A R T I C L E
◥
STRUCTURAL BIOLOGY
Structural basis of antagonizing the vitamin K
catalytic cycle for anticoagulation
Shixuan Liu1, Shuang Li1, Guomin Shen1*, Narayanasami Sukumar2, Andrzej M. Krezel1, Weikai Li1†
Vitamin K antagonists are widely used anticoagulants that target vitamin K epoxide reductases (VKOR), a
family of integral membrane enzymes. To elucidate their catalytic cycle and inhibitory mechanism, we report
11 x-ray crystal structures of human VKOR and pufferfish VKOR-like, with substrates and antagonists in different
redox states. Substrates entering the active site in a partially oxidized state form cysteine adducts that induce
an open-to-closed conformational change, triggering reduction. Binding and catalysis are facilitated by
hydrogen-bonding interactions in a hydrophobic pocket. The antagonists bind specifically to the same hydrogen-
bonding residues and induce a similar closed conformation. Thus, vitamin K antagonists act through mimicking
the key interactions and conformational changes required for the VKOR catalytic cycle.
V itamin K antagonists (VKAs) are oral
anticoagulants used to treat and prevent
thromboembolic diseases, including myo-
cardial infarction and stroke, the two
leading causes of human death and dis-
ability (1). Warfarin, a well-known VKA, is
among the most commonly used drugs world-
wide. However, warfarin overdose often causes
major and fatal bleeding (2). In older adults,
one-third of hospitalizations for adverse drug
reactions are due to warfarin use (3). Improving
the safety of anticoagulation therapy requires
the understanding of how VKAs inhibit their
target, vitamin K epoxide reductase (VKOR).
VKOR is an endoplasmic membrane enzyme
that sustains blood coagulation through the
vitamin K cycle (4). This cycle begins with the
epoxidation of vitamin K hydroquinone (KH2)
that drives the g-carboxylation of several co-
agulation factors, a posttranslational modifica-
tion required for their activity. To regenerate
this g-carboxylase cofactor, VKOR reduces the
vitamin K epoxide (KO) first to the quinone
(K) and then to the hydroquinone (KH2). Each
reduction step is coupled to the oxidation of
two active-site cysteines in VKOR (fig. S1A). To
return to their reactive state, these cysteines
are reduced in a process mediated by another
pair of conserved cysteines in VKOR (5–8). This
reductase activity is similar in a VKOR-like pa-
ralog, which plays a role in supporting bone
mineralization and inhibiting vascular calci-
fication (9, 10). In addition, VKOR homologs
catalyze disulfide-bond formation in many spe-
cies, constituting a major family of integral
1Department of Biochemistry and Molecular Biophysics,
Washington University School of Medicine, St. Louis, MO
63110, USA. 2NE-CAT, Cornell University, Argonne National
Laboratory, Argonne, IL 60439, USA.
*Present address: Institute of Hemostasis and Thrombosis, School
of Basic Medical Science, Henan University of Science and
Technology, Henan 471003, P. R. China.
†Corresponding author. Email: weikai@wustl.edu
membrane thiol oxidoreductases found from
bacteria to humans (11).
VKAs inhibit VKOR and VKOR-like en-
zymes at both their epoxide and quinone re-
duction steps. The naphthoquinone ring of
vitamin K resembles that of VKAs, which gen-
erally carry a pharmacophore of either 4-
hydroxycoumarin or 1,3-indandione (fig. S1, B
and C). Coumarin-based VKAs—including war-
farin, phenprocoumon, and acenocoumarol—
are the primary anticoagulants used in North
American and many European countries. An
indandione-based VKA, fluindione, is used
mostly in France (12). Coumarin and indan-
dione derivatives with large side groups, such
as brodifacoum and chlorophacinone (fig. S1D),
are known as superwarfarins that have a
long-lasting effect of inducing severe bleeding.
Owing to this high potency, brodifacoum is
one of the most used rat poisons. After decades
of VKA use, numerous VKOR mutations con-
ferring resistance have been identified in
rodents and humans. Patients carrying such
mutations require a high warfarin dosage that
varies by the type of mutations, adding another
complication to the anticoagulation therapy.
To reveal the molecular basis of antico-
agulation by VKAs, we determined crystal
structures of human VKOR (HsVKOR) with
representative VKAs and with the KO sub-
strate captured in different functional states.
We have also determined ligand-free, substrates-
bound, and warfarin-bound structures of a
VKOR-like homolog from the pufferfish Takifugu
rubripes. A total of 11 structures (table S1 and
fig. S2) illustrate almost the entire catalytic
cycle of these integral membrane oxidoreduc-
tases and reveal the mechanism of vitamin K
antagonism.
Overall structure of human VKOR with bound VKAs
HsVKOR contains a large and flexible endo-
plasmic reticulum (ER)–luminal region and is
a difficult target for structural characteriza-
tion because of its in vitro instability (13). We
found that prebound VKAs can stabilize the
HsVKOR protein in a detergent solution (Mate-
rials and methods and fig. S3, A and B). For
additional stabilization in vitro, we fused the
flexible N and C termini of HsVKOR to a split
superfolder green fluorescent protein (sfGFP)
(14), which also provides a scaffold for crystal-
lization and facilitates structure determination
(fig. S3, A, C, and D). This termini-restrained
construct is catalytically active and inhibitable
by warfarin, both in cells (Fig. 1A) and after
being purified with detergent (fig. S4, A and
C). The increased warfarin sensitivity in cells
and retained activity in detergent suggest that
the protein fold of HsVKOR is stabilized after
termini restraining. A combination of these
strategies allowed the structure determination
of HsVKOR at near-atomic resolution with
warfarin, phenindione (a close analog of flu-
indione), brodifacoum, and chlorophacinone
in three different crystallographic space groups
(table S2).
HsVKOR with these four different VKAs
bound adopts essentially the same overall struc-
ture (Fig. 1B and fig. S3E) under the different
crystal packing conditions. The transmembrane
helices (TMs) form a four-helix bundle that
creates a central pocket occupied by VKAs
(Fig. 1C). This same pocket is designated as
the active site of HsVKOR, where the catalytic
cysteines, Cys132 and Cys135, are located. The
coumarin or indandione ring of VKAs faces
the ER-luminal surface, and their side group
is buried in the transmembrane region of the
protein. TM1 and TM2 are connected by the
large ER-luminal region, which constitutes
almost one-third of the protein sequence. This
region starts from a helical extension of TM1
(named TM1e) and continues with a loop (loop 1)
and a b-hairpin. The two ends of the b-hairpin
are stapled together by a disulfide bond formed
between Cys43 and Cys51 (Fig. 1D), which are
the cysteine pair that mediates transfer of re-
ducing equivalents during catalysis (5–8). Cys51
connects to a cap domain that covers the top of
the central pocket. The cap domain consists of
a short helix (cap helix) and a loop (cap loop),
followed by an amphipathic anchor domain
that is buried partially in the membrane. This
anchor attaches the cap domain to the mem-
brane and thereby stabilizes its covering of the
central pocket. Residues whose mutation con-
fers strong warfarin resistance are either
located at this VKA-binding pocket or distrib-
uted in the other regions of the large luminal
domain (fig. S5).
Binding interactions of VKAs
The structures show that all VKAs are hydro-
gen bonded to Asn80 on TM2 and Tyr139 on
TM4 (Fig. 2, A and B). The Tyr139 hydroxyl
forms a hydrogen bond with the 4-hydroxyl
Liu et al., Science 371, eabc5667 (2021)
1 January 2021
1 of 10
RES EARCH | R E S E A R C H A R T I C L E
l
e
v
e
l
l
n
o
i
t
a
y
x
o
b
r
a
c
e
v
i
t
a
e
R
l
100%
80%
60%
40%
20%
0%
0.1
A
D
1
10
War (nM)
100
1000
Loop 1
TM1e
β-hairpin
C43-C51
TM1
Cap
TM2
C132-C135
Loop 3-4
TM4
War
TM3
B
HsVKOR
HsVKOR-sfGFP
TrVKORL
TrVKORL-sfGFP
β-hairpin
Loop 1
Cap helix
Cap
TM1e
Cap loop
Loop 3-4
C43-C51
C132-C135
War
TM1
C
ER lumen
Anchor
C132-C135
Cap
Central
pocket
War
o
90
TM4
TM2
TM3
ER
membrane
TM4
TM3
HsVKOR
Membrane
boundary
Tunnel
TM2
Anchor
Cytosol
Fig. 1. Overall structure of HsVKOR with bound warfarin. (A) Warfarin
inhibition of epoxide reductase activity in cultured cells. The inhibition curves
of untagged and sfGFP-fused HsVKOR and TrVKORL are compared. The activity
assay has been repeated three times. (B) Structure of HsVKOR (side view)
with warfarin (War). Secondary structure elements in the large ER-luminal
domain are named and presented in different colors. The four transmembrane
helices (TM1 to TM4) are shown in gray. The 4-hydroxycoumarin and side groups
of warfarin are colored in orange and yellow-green, respectively. The overall
structure of HsVKOR with warfarin and with other vitamin K antagonists is very
similar (fig. S3E). (C) Surface representation of the structure (exposed view).
Warfarin is bound at the central pocket that contains the active-site cysteines.
The cap helix forms the top part of the central pocket. The tunnel below is
expected to bind the isoprenyl chain of K or KO substrate. (D) Top view of
the structure.
group of warfarin and brodifacoum. The im-
portance of the 4-hydroxyl has been known
since the discovery of vitamin K antagoniza-
tion in 1939 because this coumarin modifi-
cation in spoiled sweet clover caused fatal
bleeding in cattle (15). Replacing the 4-hydroxyl
of warfarin with chemical groups that disable
hydrogen-bond formation results in almost no
HsVKOR inhibition (16). Another key hydro-
gen bond is formed between the 2-ketone group
of warfarin and the amide group of Asn80,
whose side chain is well positioned through
interacting with the protein backbone of the
cap loop (Fig. 2A). Mutations disrupting either
of these hydrogen bonds, such as Asn80Ala or
Tyr139Phe, result in strong warfarin resistance
(fig. S5A) (7). For phenindione and chloropha-
cinone, Asn80 and Tyr139 form hydrogen bonds
with the 1,3-diketones of their indandione ring
(Fig. 2B), which closely resemble the 2-ketone-
4-hydoxyl of warfarin or the 2,4-diketones of
its keto–enol tautomer (fig. S1B) (17). Thus, the
hydrogen bonding with the metapositioned
(1,3- or 2,4-) diketones or ketone-hydroxyl,
which provides the specificity of VKA recog-
nition, is a shared key feature underlying the
VKA inhibition of HsVKOR.
Except for these two hydrogen bonds, the
VKAs are bound in a largely hydrophobic pock-
et that excludes water molecules (Fig. 2, A and
C). The coumarin or indandione rings of VKAs
are stacked between Val54/Phe55 at the top and
Leu120 from the bottom and surrounded by
Trp59, Phe63, Leu124, and Leu128 (Fig. 2A). All of
these hydrophobic residues, when mutated,
render HsVKOR strongly resistant to warfarin
(fig. S5A) (7).
The large side groups of brodifacoum and
chlorophacinone afford their high-potency in-
hibition of HsVKOR. The side group of bro-
difacoum occupies the entire central pocket,
including a tunnel formed between TM2 and
TM3 (Fig. 2C), where the isoprenyl chain of
substrates should be bound (5). Hydrophobic
residues aligning along this tunnel—such as
Phe83, Phe87, and Tyr88—provide additional
binding interactions to the side group of bro-
difacoum; these residues do not interact with
warfarin, which only inhabits the inner end of
the tunnel. The indandiones show a similar
trend: The side group of chlorophacinone
occupies half of this tunnel, whereas phenin-
dione, with a nearly planar structure, cannot
reach to this tunnel. These differences indi-
cate that the large side groups of the super-
warfarins interact with the isoprenyl-chain
tunnel, increasing the binding area and re-
sulting in the strong inhibition (Fig. 2D).
On the basis of these structures, binding
interactions of warfarin metabolites can be
modeled. Warfarin is inactivated by cytochrome
P450 2C9 (CYP2C9), whose genotype is a ma-
jor predictor of warfarin dose (18). The CYP2C9
metabolism generates 7-hydroxywarfarin (71%)
and 6-hydroxywarfarin (22%). If these metab-
olites bind in the same way as warfarin does, a
polar 7-hydroxyl group would be energetically
unfavorable in the hydrophobic pocket, whereas
6-hydroxyl substitution would be within the
hydrogen bonding distance to the protein back-
bone, saturating the polarity (Fig. 2E). Con-
sistently, 6-hydroxywarfarin inhibits HsVKOR
to a similar level as warfarin, whereas 7-
hydroxywarfarin loses the inhibition efficacy
(Fig. 2D) (16, 19). Thus, the different binding
interactions of warfarin metabolites explain
their relative activity.
Stabilization of the cap domain
The cap domain conformation and VKA bind-
ing are mutually stabilized. The association
between the cap and TM domains requires
VKA-mediated interactions (Fig. 2A and movie
S1), which explains the stabilization effect of
bound VKAs during protein purification and
crystallization. Consistently, mass spectrometry–
based footprinting of HsVKOR in cells showed
Liu et al., Science 371, eabc5667 (2021)
1 January 2021
2 of 10
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Critical molecular interactions
between HsVKOR and vitamin K
antagonists. (A) The warfarin-binding
pocket. The Fo-Fc omit map of warfarin
is contoured at 3s (red mesh) and
–3s (blue mesh). Hydrogen bonding
interactions are indicated with dashed
green lines. (B) Similar hydrogen-bonding
patterns are observed for different VKAs
(Brod, brodifacoum; Phen, phenindione;
Chlo, chlorophacinone). Standard atomic
numbers for the key positions of VKAs
are indicated. (C) Side groups (yellow-
green) of VKAs bind differently to
the isoprenyl-chain tunnel (surface
view). (D) Inhibition curves of VKAs
and hydroxywarfarins against HsVKOR,
using the cell-based assay. (E) Models
of the 6- and 7-hydoxywarfarin in the
warfarin-binding pocket. The dashed
green line indicates a putative hydrogen
bond to the 6-hydroxyl group, and the
red curve represents a lack of favorable
interaction between the polar 7-hydroxyl
group and the hydrophobic protein
side chains.
A
HsVKOR
C43-C51
L128
Cap
C132-C135
War
V54
W59
F55
L124
TM4
N80
L120
B
C
Central pocket
1
War
4
3
2
N80
Tunnel
Y139
Y139
TM3
Central pocket
TM2
F63
Brod
4
1
23
N80
Tunnel
6-OH War
War Phen
7-OH War
Brod
Chlo
Y139
Phen
3
2
1
N80
Tunnel
0.1
1
10
102
103
104
War (nM)
Y139
HsVKOR
100%
80%
60%
40%
20%
0%
D
l
e
v
e
l
l
n
o
i
t
a
y
x
o
b
r
a
c
e
v
i
t
a
e
R
l
E
W59
L128
OH
7
OH
6
L124
War
L120
Y139
N80
Y139
Chlo
3
1
2
N80
Tunnel
that its structural flexibility is reduced with
bound warfarin (7, 20). Neighboring protein
regions also stabilize the cap domain conforma-
tion (Fig. 1D). At the horizontal direction, the cap
domain is saddled between TM1/TM1e and a
loop connecting TM3 and TM4 (loop 3–4).
TM1/TM1e interacts with loop 1 primarily
through Tyr25 and His28 (fig. S6A). Loop 1 and
b-hairpin in turn stabilize the cap helix, and
Asp44 is a central residue mediating such in-
teractions (fig. S6B). The helical conformation
of the cap helix is maintained by a hydrogen-
bonding network involving Asp44 and several
Ser residues (fig. S6C). In the vertical direction,
the anchor domain positions the cap domain
above the membrane surface, with Arg58 and
Glu67 forming a salt bridge (fig. S6D). To
anchor the cap loop in position, Asn80 from
TM2 provides several hydrogen bond inter-
actions (fig. S6E). Consistent with these struc-
tural observations, mutations such as Tyr25Ala,
His28Ala, and Asp44Ala show strong warfarin
resistance (fig. S5A) (7); their mechanism had
previously been enigmatic because these re-
sidues did not appear to interact with warfarin
(fig. S5B) (21, 22). The structures now reveal
that such mutations would disrupt critical in-
teractions that stabilize the cap domain. Taken
together, warfarin resistance is caused by two
distinct mechanisms: The resistant mutations
can either directly affect the warfarin binding
or destabilize the cap domain. This mechanis-
tic understanding may improve the ability to
predict warfarin dosage as a personalized me-
dicine (23); the U.S. Food and Drug Admin-
istration suggests pharmacogenomic tests for
prescribing warfarin, but the clinical benefits
of current dosing algorithms remain contro-
versial (24).
Warfarin inhibits both fully oxidized and
partially oxidized HsVKOR, which are the two
major redox states of HsVKOR in a cellular
environment (7). To mimic the partially oxi-
dized state in vitro, we determined the warfarin-
bound structure of a Cys43Ser mutant (fig. S7),
which generates a free Cys135 (fig. S2). In this
state, the Cys135 sulfhydryl is within hydrogen
bonding distance to the 4-hydroxyl group of
Liu et al., Science 371, eabc5667 (2021)
1 January 2021
3 of 10
RES EARCH | R E S E A R C H A R T I C L E
warfarin, and the cap helix is covalently at-
tached to TM4 through the Cys51−Cys132 disul-
fide, stabilizing the binding of warfarin in the
same orientation.
Warfarin-induced conformational changes
To understand the effect of warfarin binding,
we determined the structures of a VKOR-like
protein from T. rubripes (TrVKORL) in both
warfarin-bound and ligand-free states. TrVKORL
is catalytically active and inhibitable by war-
farin, both in cells and after being purified
(Fig. 1A and fig. S4, B and C). This protein
shares 73% sequence identity with human
VKOR-like (HsVKORL) and 52% with HsVKOR
(fig. S8); for simplicity of explanation, hereafter
we number conserved residues in these pro-
teins according to the HsVKOR sequence.
TrVKORL was chosen for crystallization be-
cause it has higher protein expression level
and better in vitro stability than those of
HsVKORL (25). Compared with HsVKOR, the
TrVKORL protein is stable in detergent mi-
celles even without a bound ligand, enabling the
capture of both the ligand-free and warfarin-
bound states.
The conformations of warfarin-bound
TrVKORL and HsVKOR in the fully oxidized
state are remarkably similar (root mean square
deviation 1.1 Å for all atoms in 149 super-
imposed residues) (Fig. 3, A and B). Both pro-
teins adopt a closed conformation, with loop
1 and b-hairpin folded near the active site and
the large luminal domain forming extensive in-
teractions with loop 3–4 (Fig. 3C). At the central
pocket, warfarin interacts with surrounding
residues in essentially the same way, and the
hydrogen bonding of warfarin to Asn80 and
Tyr139 is nearly identical (Fig. 3B). Of the
residues critical to warfarin inhibition or ca-
talysis, 90% are conserved between TrVKORL,
HsVKORL, and HsVKOR (fig. S8). The high se-
quence identity and structural similarity sug-
gest that these VKOR enzymes share essentially
the same catalytic and inhibitory mechanisms.
In the ligand-free state, the fully oxidized
TrVKORL shows an open conformation, form-
ing a luminal helix that is distant from the
active site (Fig. 3D). This helix restructures the
residues of the loop 1 and b-hairpin found in
the ligand-bound state. Moreover, the cap helix
becomes a loop in the absence of warfarin stabi-
lization, and TM1e loses a helical turn. In the
absence of these structural elements, few inter-
actions are observed between loop 3–4 and the
large luminal domain, a characteristic of the
open conformation.
Modeling the structural transition between
the ligand-free (open) and warfarin-bound
(closed) states suggests that the induced fit
prompted by warfarin brings Phe55 closer and
reorients Val54 in the cap domain, leading to
the formation of the cap helix (movie S2). This
A
β-hairpin
Loop 1
Cap
Loop
3-4
War
TM4
TM2
TM1
TM3
C
Closed
Loop 1
β-hairpin
TM1e
Cap
1-turn
D
Luminal helix
Open
Cap
TM1e
Anchor
Loop 3-4
War
HsVKOR
TrVKORL
TM4
TM1
TM2
TM3
Anchor
Loop 3-4
No ligand
TrVKORL
TrVKORL
TM1
TM3
TM2
TM4
Loop
TM1e
Anchor
B
E
L128
Cap
TrVKORL
C43-C51
V54
W59
L124
F55
C132-C135
War
N80
L120
HsVKOR
TM1e
C43-C51
o
90
90o
Loop 1
D44
β-hairpin
Closed
Cap
V54
Loop 3-4
TM3
C132-C135
F55
F
Luminal helix
D44
Cap
C43-C51
Open
Loop 3-4
V54
TM3
TM1e
F55
C132-C135
Y139
F63
War
TM1
TM2
TM4
No ligand
TM1
TM4
TM2
Fig. 3. TrVKORL structures showing warfarin-induced conformational
changes. (A) Overall structure of TrVKORL with warfarin (colored by structure
elements) is nearly the same as that of HsVKOR with warfarin (blue).
(B) Similarity of their warfarin binding interactions. TrVKORL residues are colored
gray (carbon atoms), and HsVKOR residues are blue. The Fo-Fc omit map of warfarin
is contoured at 3s (red mesh) and –3s (blue mesh). (C) TrVKORL with bound
warfarin adopts a closed conformation, in which the cap helix and b-hairpin are
formed and interact with loop 3–4. (D) The ligand-free TrVKORL adopts an open
conformation. In the luminal helix, residues converted from the loop 1 and b-hairpin
[in (C)] are colored in blue and purple, respectively. The TMs are colored in
dark gray to indicate the open conformation. (E to F) Detailed view of the
(E) closed and (F) open states. Warfarin binding induces the movements of
Cys43-Cys51, Val54, and Phe55 and promotes the stabilizing interactions of
Asp44 (dashed lines). A modeled structural transition is provided in movie S2.
Liu et al., Science 371, eabc5667 (2021)
1 January 2021
4 of 10
RES EARCH | R E S E A R C H A R T I C L E
helical conformation brings the Cys43−Cys51
disulfide close to the membrane surface to
interact with Phe55 (Fig. 3E). The reposition-
ing of Cys43−Cys51 induces the formation of
b-hairpin and breaks up the luminal helix
found in the open conformation. This change
also reorients Asp44, which becomes a central
residue holding the cap helix, b-hairpin, and
loop 1 together (fig. S6B). By contrast, in the
open conformation, Asp44 is located on the
luminal helix and does not participate in sta-
bilizing interactions (Fig. 3F). Taken together,
the local interactions with warfarin lead to
global changes of protein conformation: a
“domino effect” that culminates in the closed
conformation.
Substrate entrance
To understand the catalytic mechanism of
VKORs, we first determined the structures
of fully oxidized TrVKORL with KO and K.
The protein conformations in these states
are nearly identical to that in the ligand-free
state. The Fo-Fc maps show weak electron
densities (fig. S9) that suggest low-affinity
binding of the substrates, probably because
the oxidized enzyme is not at the state of
reacting with the substrates. The ring-shaped
part of the densities probably corresponds
to the substrate naphthoquinone rings, which
do not occupy the top of the central pocket but
are located near the isoprenyl-chain tunnel
(fig. S9), suggesting that the tunnel provides
the path of substrate entrance and/or product
release. For the substrates to be bound with
high affinity and positioned properly for ca-
talysis, a reactive Cys135 is required.
Substrate-induced reduction
To generate the reactive Cys135, we used a
Cys132Ser mutant of TrVKORL (fig. S2) and
determined a K-bound structure. In this cat-
alytic state, Cys135 is close to the naphtho-
quinone ring of K, whose isoprenyl chain
occupies the designated tunnel (Fig. 4A).
The sulfhydryl group of Cys135 is ~3 Å to the
C2 atom, which is consistent with the dis-
tance in a charge-transfer complex in which
the negative charge of Cys135 thiolate is par-
tially transferred to the quinone ring; similar
complexes have been found in other thiol oxi-
doreductases that use aromatic cofactors such
as quinone and flavin (26, 27). Formation of
this charge-transfer complex helps stabilize K
within Cys132Ser TrVKORL. The binding inter-
actions in turn induce the protein transition
from the open conformation in the ligand-free
state (Fig. 3D) to the closed conformation in
the catalytic state (Fig. 4, B and C). This con-
formational change brings Cys43−Cys51 close
to Cys132Ser and Cys135−K (Fig. 4B), a scenario
occurring as these four cysteines transfer
reducing equivalents in the wild-type pro-
tein (fig. S2). These residues are sequentially
A
Central
pocket
Tunnel
K
C132S
3 Å
C135
C2
TrVKORL C132S
o
180
C
β-hairpin
Loop 3-4
War
K
Cap
Anchor
TM1
TM2
TM3
TM4
B
Closed
β-hairpin
D
Cap
Loop 3-4
K
War
C43-C51
C132S
Electron
transfer path
C135
K
K
TM3
TM4
TM1
TM2
Y139
N80
Fig. 4. Structure of vitamin K–bound TrVKORL in catalytic state. (A) TrVKORL Cys132Ser mutant with K
(surface view of active site). The Fo-Fc omit map of K is contoured at 3s (red mesh) and –3s (light blue
mesh). The reduced Cys135 and K form a charge-transfer complex (dashed arrow). (B) Closed conformation
with formation of the charge-transfer complex. The unimpeded path between the four cysteines and K is
indicated with orange shading. (C) Superimposed structures of TrVKORL in the K-bound (colored by
structural elements) and warfarin-bound (as in Fig. 3C, blue) states. (D) Similar hydrogen-bonding
interactions with K (orange and yellow-green) and with warfarin (blue).
aligned within a short distance and thus are
physically posed for thiol-disulfide exchange.
These unimpeded exchanges facilitate the
passing of reducing equivalent to resolve the
Cys135−K adduct, resulting in the full reduc-
tion of K.
The closed conformation induced by the
proposed Cys135−K charge-transfer adduct is
very similar to that induced by warfarin, with
the cap domain, loop 1, and b-hairpin adopt-
ing essentially the same conformation in the
two structures (Fig. 4C). Similar to warfarin
binding interactions, Tyr139 and Asn80 form
hydrogen bonds with the para-positioned 1,4-
diketones of K (Fig. 4D). In addition, the iso-
prenyl chain of vitamin K interacts with Phe83,
Phe87, and Tyr88 along the tunnel, which are
the same residues binding the side group of
brodifacoum. These resemblances show that
warfarin (and superwarfarin) exploits features
that are tailored for the binding of substrates
and/or transition states of the reaction to
stably occupy the active site of VKORs.
We used a similar strategy to determine the
structure of HsVKOR with KO (Fig. 5A), this
time using a Cys43Ser mutation to generate
the reduced Cys135 (fig. S2). This mutant mi-
mics the state before attack of the Cys43 on
the Cys51-Cys132 disulfide while Cys135 is capable
of attacking the substrate. The distance be-
tween the sulfur atom of Cys135 and the C2
atom of the naphthoquinone ring in the elec-
tron density map is ~2 Å (Fig. 5B), which is
suggestive of a covalently bonded C–S adduct.
To investigate this possibility, we incubated
13C-labeled KO with the Cys135Ser and Cys43Ser
mutants of TrVKORL (chosen for their apo-
protein stability). Cys135Ser lacks the reactive
thiolate, whereas Cys43Ser is expected to carry
out an initial covalent step of the reaction
but not to proceed further without an exter-
nally supplied reducing agent (fig. S2). After
Liu et al., Science 371, eabc5667 (2021)
1 January 2021
5 of 10
RES EARCH | R E S E A R C H A R T I C L E
A
HsVKOR
C43S
Closed
Cap
TM1e
Loop 3-4
Anchor
C51-C132
C135
3-OH K
TM4
TM3
TM2
TM1
C
D
2.20
1.77
3-OH K
13C
KO
13C
R
R
1.50
K
13C
R
Mercapto 3-OH K
1.63-1.67
S
13C
R
1.8
H1
2.2
2.0
Mercapto 3-OH K
1.67
1.63
C13
10.6
10
14.7
15.5
24.0
15
20
25
[
p
p
m
]
1.6
1.4
[ppm]
DTT + KO
C43S + KO
C135S + KO
C13
22
23
24.0
24
Cap
F55
C51-C132
V54
C-S
3-OH K
C135
Y139
L120
N80
F63
3-OH K
10.6
WT
Y139F
N80A
B
E
%
y
t
i
v
i
t
c
a
e
v
i
t
a
e
R
l
100
75
50
25
0
WT
N80A Y139F
KO
K
2.20
2.2
1
H
2.6
2.4
14.7
15.5
1.77
1.50
2.0
1.8
1.6
[ppm]
H1
1.7
1.6
1.5
1.4
[ppm]
25
[
p
p
m
]
F
Partially oxidized (50%)
SH
43
51-132
S
S
SH
135
I
SH
43
Closed
S
51-132
S
S
135
KOH
or KH
II
KO
or K
Catalytic
cycle
or KH2
K
W
Closed
Reducing
molecule
S–S
43-51
W
S–S
132-135
Open
IVV
Fully oxidized (40%)
C13
10
12
14
16
[
p
p
m
]
Electron
transfer
43-51
S–S
132
SH
S
135
KOH
or KH
III
Fig. 5. Structure of HsVKOR bound to a KO adduct. (A) Overall structure of
HsVKOR Cys43Ser mutant cocrystallized with KO. The dashed line indicates the disordered
region between TM1e and cap domain. (B) Detailed view of the active site. The Fo-Fc
omit map of C135-KOH is contoured at 3s (red mesh) and –3s (blue mesh); 3-OH K is
modeled into the density on the basis of the geometric restraint and the catalytic chemistry
(28, 29). The short distance between the C2 atom of 3-OH K and the sulfhydryl group
of Cys135 suggests that they are connected by a C–S bond. (C) 2D 1H-13C HSQC spectrum
(full spectra are available in fig. S13) showing signals of 13C-labeled 2-methyl groups in
four products of KO reduction with dithiothreitol (DTT) (55). Chemical structures of
the products and the assigned chemical shifts are indicated. The chemical shifts of the
13C 2-methyl group are strongly influenced by the ring current field of the naphthoquinone
group (the 2-methyl is in plane with the naphthoquinone ring for K, in tetrahedral angle
for 3-OH K and in other angles for KO and mercapto 3-OH K). (D) Superimposed regions
of three 2D 1H-13C HSQC spectra (full spectra in figs. S13 to S15) showing mercapto
reduction products of 2-methyl-13C-labeled KO. Signals of products recovered from
TrVKORL Cys43Ser catalyzed reaction are shown in blue and those from Cys135Ser are
in red. The nonenzymatic (DTT-reduced) products signals are shown in black.
(E) Superimposed spectra (full spectra in figs. S16 to S18) of reaction products obtained
from wild-type HsVKOR (WT; green), Asn80Ala (blue), and Tyr139Phe mutants (orange)
with KO and DTT. An area of 2D HSQC spectrum containing 1H-13C peaks of 2-methyl-13C-
labeled K [2.20 to 15.5 parts per million (ppm)], KO (1.77 to 14.7 ppm), and 3-OH K (1.50
to 10.6 ppm) is shown. (Inset) Relative KO reduction activity of HsVKOR constructs in
microsomes (Materials and methods). (F) The catalytic cycle and inhibition of HsVKOR are
accompanied with redox-state and conformation changes. The large luminal domain is
shown as a hemisphere (pink), and the transmembrane domain is shown as a cylinder
(light gray, closed conformation; dark gray, open conformation). (State I) The partially
oxidized state with free Cys43 and free Cys135. (State II) Cys135 forms a stable adduct with
3-OH K (KOH) or K (KH), whose binding induces the closed conformation and the
juxtaposition of Cys43 for transfer of reducing equivalents (dashed arrow). (State III) The
reduced Cys132 attacks Cys135−K-OH (or KH) to generate K (or KH2). (State IV) The
fully oxidized state is in an open conformation to release K (or KH2). (Left) Warfarin (W)
competes with the substrates for the partially oxidized enzyme. Unlike the substrates,
warfarin binds also to the fully oxidized enzyme and removes it from the enzyme pool. The
bound warfarin locks HsVKOR in both of the redox states into a closed conformation.
Liu et al., Science 371, eabc5667 (2021)
1 January 2021
6 of 10
RES EARCH | R E S E A R C H A R T I C L E
proteolysis of the mutant proteins, different
vitamin K derivatives with the 13C label show
readily distinguishable positions in two-
dimensional (2D) 1H–13C correlated spectra
(Fig. 5, C and D, and fig. S10). Only the Cys43Ser
mutant enzyme produces derivatives with sig-
nals (Fig. 5D) that are consistent with covalent
adducts of 3-hydroxyl vitamin K (3-OH K) from
nucleophilic attack of Cys135 on the epoxide
(fig. S2). The crystal structure is consistent
with this covalent adduct, which has been
predicted from chemical modeling and quan-
tum chemistry simulation (28, 29).
Similar to the K-bound state, the 1,4-diketones
of 3-OH K retain hydrogen bonding to Tyr139
and Asn80 (Fig. 5B). Maintaining these hydro-
gen bonds is important for catalysis. In the
in vitro assay using dithiothreitol (DTT) as the
reductant, Asn80Ala shows nearly no enzyma-
tic activity, whereas Tyr139Phe remains largely
active but generates 3-OH K as a side product
(Fig. 5E), which is consistent with previous
reports (30, 31). In the cellular assay, Tyr139Phe
and Asn80Ala also lower the HsVKOR activity
(fig. S4, D and E). Taken together, Tyr139 and
Asn80, which are the same residues recognized
by VKAs (Fig. 2B), are also important for
catalysis.
In the HsVKOR structures, the naphtho-
quinone ring of 3-OH K rotates to an angle
different from that of warfarin (fig. S11A).
The cap helix also rotates and is barely main-
tained in a helical conformation (fig. S11B).
The cap loop and anchor domain interact dif-
ferently, with Arg58 interacting with Glu67 in
the 3-OH K state, whereas Arg58 inserts be-
tween the side chains of Glu67 and His68 when
warfarin is bound. Warfarin was previously
proposed to be a transition-state analog of
3-OH K because their chemical structures are
similar (16, 30). This proposal, however, is not
supported by the distinct structures in the
warfarin and 3-OH K bound states. The polar
3-hydroxyl probably has prevented Phe55 from
interacting with the naphthoquinone ring
of 3-OH K (fig. S11B); Phe55, however, forms
perpendicular p-p stacking interaction with
the coumarin ring of warfarin (Fig. 2A). More-
over, this 3-hydroxyl does not participate
in the hydrogen-bonding interaction as the
4-hydroxyl of warfarin does. However, the pro-
tein and ligand conformations in the warfarin-
bound state are quite similar to those in the
K-bound state (Fig. 4, C and D), suggesting
that warfarin mimics a later step of VKOR
catalysis.
Discussion
Although extensive clinical experience with
VKAs has been amassed over almost seven
decades (32), there are still questions as to
how these anticoagulants inhibit VKOR as well
as the mechanism of this family of enzymes
with natural substrates. Our structures reveal
steps that may correspond to the catalytic cycle
of VKOR (Fig. 5F and fig. S2) accompanied by
protein conformational changes (movie S3). In
the cellular environment, ~50% of HsVKOR
(and HsVKORL) is in the partially oxidized
state (7, 21). At this state, we propose that the
catalytic cycle begins with a reduced Cys135
(Fig. 5F, state I) that reacts with the substrates
(K or KO), forming a charge-transfer or cova-
lent complex that enables stable binding of the
substrate (Fig. 5F, state II). Interactions of
these adducts with the cap domain induce
the closed conformation; the energy required
for this conformational change probably orig-
inates from the binding and formation of the
substrate adducts. In this closed conforma-
tion, Cys43 is juxtaposed with Cys51−Cys132 in
an ideal position to undergo thiol-disulfide
exchange (Fig. 4B). The resulting reduced
Cys132 could in turn attack the Cys135−substrate
adduct to generate the fully reduced reaction
product (Fig. 5F, state III). Consequently, the
active site becomes fully oxidized, and the pro-
tein returns to its open conformation, result-
ing in a noncatalytic state that allows release
of the reduced product (Fig. 5F, state IV). K is
both a product and a substrate of VKOR ca-
talysis (fig. S1A). The temporary binding of K at
the low-affinity site (fig. S9) may allow K to
quickly access Cys135 once it is reduced again,
expediting the reduction of KO to KH2 through
the K intermediate. To return to the partially
oxidized state, VKOR is likely reduced by part-
ner proteins (33) or redox mediators such as
glutathione in the ER. Overall, the redox cycle
of VKOR catalysis is associated with open and
closed conformational changes. The closed con-
formation is induced by substrate binding and
is conducive to later catalytic steps, and this
mechanism thus differs from that known for
bacterial VKOR homolog or other membrane
and soluble thiol oxidoreductases, such as DsbB
and Ero1 (5, 34, 35).
The closed conformation induced by sub-
strate binding closely resembles that induced
by VKA binding (Fig. 4C). Comparison with
the ligand-free structure shows that the large
luminal regions of these proteins adopt inter-
converting structures with remarkable plasticity
and dynamics. The open and closed conforma-
tions are stabilized by electrostatic and hydro-
gen bonding interactions (fig. S6), which are
relatively weak and may be readily biased by
interaction with a ligand. Consequently, tran-
sitions between these conformations are used
by substrates to trigger the catalytic cycle and
by VKAs to achieve inhibition.
The binding of the VKAs and substrate ad-
ducts both require hydrogen bonding to Asn80
and Tyr139. Tyr139 mutations result in the accu-
mulation of a side-reaction product, 3-OH K,
indicating that Tyr139 not only participates in
substrate binding but also is essential to ca-
talysis. Another major role of Tyr139 and Asn80
is that their hydrogen bonding should increase
the redox potential of the naphthoquinone
(36), facilitating the substrate reduction. These
hydrogen bonds may also provide particularly
strong interactions to reaction transition states
or intermediates and to VKAs owing to the
hydrophobic environment at the active site
in the closed conformation induced by bound
substrates or VKAs. Overall, mimicking the
hydrogen-bonding interactions required in
the catalytic steps explains the very high af-
finity of VKA binding to VKOR, which is es-
sentially irreversible (37).
VKAs simultaneously block two major steps
of the catalytic cycle, the partially and fully
oxidized states (Fig. 5F). These states together
constitute ~90% of the HsVKOR cellular frac-
tion (7), and both states are now captured in
structures with warfarin bound (Fig. 2 and fig.
S7). By contrast, only the partially oxidized en-
zyme is catalytically reactive to substrates,
whereas the fully oxidized enzyme promotes
product release (Fig. 4 and fig. S9). VKAs and
substrates occupy the same binding pocket
and therefore should compete for the partially
oxidized enzyme. On the other hand, the sub-
strate reduction continuously generates the
fully oxidized enzyme, which is removed by
VKAs from further participating in the cat-
alytic cycle (Fig. 5F). Binding at both major
states of the catalytic cycle may explain the
unusual inhibition kinetics and the potency
of VKAs.
Overall, the crystal structures reported here
reveal the conformational control and reaction
chemistry of a family of integral membrane
oxidoreductases and elucidate the multifac-
eted actions of their antagonists. The struc-
tures also explain the mechanism of warfarin
resistance and the activity of warfarin metab-
olites, both of which contribute to warfarin
dose variation (23); incorporating resistance
profiles into warfarin dosage prediction may
render better safety and accuracy. New ther-
apeutic strategies of anticoagulation can be
designed on the basis of the different redox-
state requirements between the VKA action
and the substrate catalysis and on the basis
of the accompanied open and closed confor-
mational changes.
Materials and methods
Constructs and cloning
Codon optimized HsVKOR (accession code
Q9BQB6; 3-155) and TrVKORL (NP_001027940;
6-175) were tagged at their two ends by the
N- and C-halves of sfGFP (38). The DNAs
encoding these constructs were cloned into
a modified pPICZa expression vector (Invi-
trogen) by using a ligation-free polymerase
chain reaction (PCR)–based method (39). The
Cys43Ser mutant of HsVKOR and Cys138Ser mu-
tant of TrVKORL (corresponding to Cys132Ser
in HsVKOR) were generated by site-directed
Liu et al., Science 371, eabc5667 (2021)
1 January 2021
7 of 10
RES EARCH | R E S E A R C H A R T I C L E
mutagenesis (40). All nucleotide sequences
were verified by DNA sequencing.
Protein expression and purification
The plasmids were linearized and transformed
into Pichia pastoris by electroporation. Trans-
formants were selected by Zeocin resistance
on yeast extract peptone dextrose medium
with sorbitol (YPDS) agar plates. The expres-
sion levels of resistant clones were compared
through fluorescence-detection size-exclusion
chromatography (FSEC) analysis (41). Clones
with the highest expression levels were stored
at -80°C.
For large-scale protein expression, 1 L cul-
ture was grown in the buffered minimal glyc-
erol (BMG) media (1.2% glycerol, 0.34% yeast
nitrogen base, 1% ammonium sulfate, 0.4 mg/ml
biotin, and 100 mM potassium phosphate
pH 6.0) at 30°C for 20 hours. The growth
media was then changed to 1 L buffered mini-
mal methanol (BMM) media (0.34% yeast ni-
trogen base, 1% ammonium sulfate, 0.4 mg/ml
biotin, and 200 mM potassium phosphate pH
6.0), and the protein expression were induced
with 0.7% methanol. After 2 days at 25°C, the
cells were harvested by centrifugation and flash
frozen in liquid nitrogen.
Protein purification of wild-type TrVKORL
and Cys138Ser mutant used the following pro-
cedure. Frozen cells (20 g) were broken by a
ball mill (Retsch). The cell powder was resus-
pended in 40 ml lysis buffer (225 mM NaCl,
75 mM Tris-HCl pH 8.0, and 10 mg/ml DNase I).
Subsequently, 1.2 g n-dodecyl-b-D-maltoside
(DDM; 2% final concentration) was added to
solubilize the membranes by stirring for 3 hours
at 4°C. The suspension was centrifuged at
130,000 g for 20 min. The supernatant was
incubated with 3 ml TALON metal-affinity
resin (Clontech) for 3 hours at 4°C. The resin
was subsequently collected on a gravity-flow
column and washed with 60 ml wash buffer
(10 mM imidazole, 150 mM NaCl, 0.2% DDM,
and 20 mM Tris pH 8.0). The protein was
eluted with 10 ml elution buffer (250 mM imi-
dazole, 150 mM NaCl, 0.2% DDM, and 20 mM
Tris pH 8.0). The eluted protein was concen-
trated and applied to Superdex 200 for size-
exclusion chromatography (SEC) in a SEC buffer
containing 0.05% lauryl maltose neopentyl
glycol (LMNG, Anatrace), 150 mM NaCl, and
20 mM Tris pH 8.0. The peak fractions were
collected and the TrVKORL protein was con-
centrated to 40 mg/ml and immediately used
for crystallization.
This purification procedure was modified
for the TrVKORL-warfarin complex in the fol-
lowing ways. The complex was generated by
incubating the suspension of broken cells with
1 mM sodium salt of warfarin (Sigma) at 4°C
for 20 min before the DDM solubilization.
Warfarin at the 1 mM concentration was main-
tained throughout the membrane solubilization
and immobilized metal-affinity chromatogra-
phy (IMAC). However, the SEC buffer did not
contain warfarin because excess amount of
free warfarin interferes with crystallization.
To purify wild-type HsVKOR with VKAs, the
suspension of broken cells was incubated with
warfarin, phenindione, broadifacoum or chlo-
rophacinone at 4°C for 20 min before the
membrane solubilization. Warfarin was main-
tained at 1 mM concentration throughout the
purification process, including the IMAC and
SEC steps. The concentration of phenindione,
broadifacoum, and chlorophacinone was 100 mM
for IMAC and 500 mM or 1 mM for SEC. Puri-
fication of HsVKOR also required a lipid mixture,
constituted of 0.1 mg/ml (final concentration)
of 1-palmitoyl-2-oleoyl-glycero-3-phosphocho-
line (POPC), 1-palmitoyl-2-oleoyl-sn-glycero-3-
phosphoethanolamine (POPE) and 1-palmitoyl-
2-oleoyl-sn-glycero-3-phospho-(1'-rac-glycerol)
(POPG) (Avanti) at a 3:1:1 molar ratio, in the
IMAC and SEC buffers. The SEC buffer has the
lipids and VKAs dissolved in 0.05% LMNG,
150 mM NaCl, and 20 mM Tris pH 8.0.
Purification of the Cys43Ser mutant of
HsVKOR with warfarin used a similar pro-
tocol as that used for the wild-type protein
with warfarin. To facilitate crystallization, the
SEC buffer was changed to 200 mM potassium
formate, 0.05% LMNG, 0.1 mg/ml lipids, and
200 mM warfarin.
Purification of the Cys43Ser mutant of
HsVKOR with KO required the following mod-
ifications. The cell powder from 20 g cells was
resuspended with 100 ml lysis buffer and soni-
cated to break the cell membranes. The mem-
branes were collected by ultracentrifugation at
130,000 g for 45 min with a Beckman Ti45
rotor. The membrane pellet was homogenized
into 40 ml lysis buffer and solubilized in 1%
LMNG with 30 mM KO. KO at the same con-
centration was maintained in all the subse-
quent purification steps.
Crystallization and data collection
The wild-type and mutant HsVKOR and
TrVKORL with different ligands were crystal-
lized using the lipidic cubic phase (LCP) meth-
od (42). The protein and crystallization buffer
conditions used for different constructs and
different ligands are summarized in table S3.
For LCP crystallization, 40 mg/ml of protein
was mixed with monoolein at a 2:3 ratio (v/v).
For HsVKOR, all the ligands were added dur-
ing purification, because this protein is un-
stable without a bound ligand. For TrVKORL,
the ligands (except warfarin) were mixed with
monoolein for co-crystallization with TrVKORL.
The protein-monoolein mixture (0.1 ml) was
covered by crystallization buffer (0.8 ml), and
the LCP plates were incubated at 22°C. The
crystals usually appeared within a week and
grew to an optimal size after several weeks.
The crystals were harvested from the meso-
phase bolus and flash-frozen by direct transfer
to liquid nitrogen. Crystals embedded in the
lipid bolus generally showed good diffraction
under x-ray beams. The data were collected at
the NECAT 24ID-C and 24ID-E beam lines
at Advanced Photon Source (APS), which are
equipped with microdiffractometer-MD2 and
Dectris’s PILASTUS (24ID-C) and EIGER (24ID-E)
detectors.
Structure determination
The crystallographic data were processed, scaled
and reduced using the program HKL2000 (43)
or the program XDS (44) followed by POINT-
LESS and AIMLESS (45). For the datasets of
HsVKOR Cys43Ser-warfarin and Cys43Ser-KO,
it was necessary to remove the interference
from diffused diffraction of lipid bolus at-
tached to the LCP crystals by increasing the
outlier rejection probability in HKL2000 and
XDS. Molecular replacement with Phaser (46)
used sfGFP (PDB code 2B3P) (14) as the search
model. This initial sfGFP model was rigid-
body refined with REFMAC (47). The partial
phases were improved by solvent flattening, and
in some cases by 2-fold non-crystallographic
averaging, using the programs PARROT (48)
and DM (49). The density modified maps were
sufficient for automatic model building of the
HsVKOR or TrVKORL region (~80 to 90%
completeness) by the BUCCANEER software
(48). The models were manually built to com-
pletion and refined against the data by using
refinement programs installed in the PHENIX
suites (50).
Activity assays
Cell-based activity assays were performed as
previously described (51) using a human em-
bryonic kidney (HEK) 293 cell line that con-
tains a chimeric FIXgla-Protein C gene and has
endogenous VKOR and VKORL genes knocked
out. Dual-expression plasmids containing
HsVKOR, HsVKORL or TrVKORL constructs,
along with a luciferase gene, were transfected
into this double-knockout cell line. The car-
boxylation level of secreted FIXgla-PC in the
cell-culture medium was measured by a sand-
wich enzyme-linked immunosorbent assay
(ELISA), with luciferase activity serving as
the control for transfection efficiency. The
ELISA assay was conducted following a pro-
tocol previously reported (52). To obtain inhib-
ition curves of VKAs and warfarin metabolites,
the transfected cells were treated with 11 dif-
ferent concentrations of the compounds, with
the concentration range optimized according
to the dose response of each compound and
each construct. The median inhibitory con-
centrations (IC50) were analyzed by using
GraphPad Prism.
The activity assay using purified proteins
followed a previously described protocol (8)
with the following modifications. Purified
Liu et al., Science 371, eabc5667 (2021)
1 January 2021
8 of 10
RES EARCH | R E S E A R C H A R T I C L E
HsVKOR, HsVKORL, and TrVKORL proteins
(50 nM) were mixed with 50 mM K or 20 mM
KO in a buffer containing 20 mM Tris-HCl pH
7.5, 0.1 M NaCl, and 0.05% LMNG. The reac-
tion was initiated by adding 5 mM DTT. The
fluorescence of KH2 (excitation 250 nm and
emission 430 nm) was detected using a plate
reader (Molecular Devices).
Comparison of the cell-free activities of
HsVKOR constructs (Fig. 5E, inset) used mi-
crosomes to provide a native-like membrane
environment, thereby avoiding the use of de-
tergent that may affect the enzymatic activity.
Purification of microsomes followed a pre-
vious protocol (53) with the following mod-
ifications. Briefly, frozen Pichia cells were
applied to a ball mill (Retsch) to break the
cell walls. The cells were resuspended in
150 mM KCl and 50 mM HEPES pH 7.5, and
sonicated to disrupt the cell membrane. After
removal of cell debris, crude microsomes were
collected by ultracentrifugation at 138,000 g
for 1 hour. The resuspended microsomes were
purified through a step gradient of 35% and
60% sucrose (w/v), and the ER membranes
were enriched after ultracentrifugation at
138,000 g for 12 hours. Concentration of mi-
crosomal HsVKOR proteins was adjusted to
the same level by their peak heights on FSEC,
which used the fused GFP signal from sol-
ubilized microsomes. Catalysis of microsomal
HsVKOR was initiated in a buffer contain-
ing 5 mM DTT, 5 mM KO, 150 mM KCl, and
200 mM HEPES pH 7.5. The reactions were
carried out at 30°C for 2 hours and subse-
quently analyzed by means of high-performance
liquid chromatography (HPLC) for the rela-
tive levels of KO and K (54).
NMR experiments
The selectively labeled (13C 2-methyl) vitamin
K, purchased from Buchem BV, was used to
synthesize (13C 2-methyl) KO and mercapto
adducts (55). The 13C labeled 2-methyl KO was
used as a substrate for all enzymatic and non-
enzymatic reference reactions to facilitate the
identification of different chemical species.
Purified TrVKORL Cys49Ser and Cys141Ser
mutant proteins (corresponding to Cys43Ser
and Cys135Ser in HsVKOR) at 30 mM concen-
tration were incubated with 13C 2-methyl KO
(150 mM) in a buffer containing 20 mM HEPES
pH 7.5, 100 mM NaCl and 1 mM DDM at 4°C
overnight. Subsequently, 1 mM NEM (final con-
centration) was added to block free cysteines.
For complete proteolysis of the mutant pro-
teins, 5% (w/w) protease K was added, and the
digestion was carried out at room temperature
overnight. The vitamin K derivatives were ex-
tracted with isopropanol:hexane (3:2 v/v), and
the upper hexane phase was collected and
dried in argon. The residual was dissolved with
deuterated chloroform and analyzed by means
of nuclear magnetic resonance (NMR).
NMR experiments were carried out on a
Bruker Avance III 600 MHz (14.09 T) spec-
trometer equipped with the cryoprobe. Deu-
terated chloroform (D-100%, CIL) was used as
solvent for NMR experiments with all forms of
vitamin K. The assignments of resonances that
are available in previous literature (30, 55)
were confirmed and extended to additional
naphthoquinone protons and carbons using two-
dimensional 1H-13C HSQC (56) and ADEQUATE
experiments (57). The very limited quantities
of derivatized forms of vitamin K were not se-
parated into pure compounds. The mercapto
and hydroxy modifications were distinguished
based on their through-bond connectivities in
ADEQUATE experiments, in which the C-2
carbon shifts were dominated by the changes
in orbital hybridization of C-2 carbon. Consist-
ently, in HSQC spectra 2-methyl group chem-
ical shifts were strongly influenced by the ring
current field of the naphthoquinone group.
The 1H and 13C resonances of 2-methyl group
have readily distinguishable positions in dif-
ferent covalently modified forms of vitamin K
(K, KO, 3-OH K and mercapto K), and the 13C
label provided sufficient sensitivity in HSQC
experiments. To observe a mercapto adduct
(an analog of a cysteine adduct in VKOR re-
action), we used a chemically synthesized
racemic dithiothreitol (DTT) adduct of vita-
min K (55).
Reactions with DTT and with VKOR mu-
tants were carried out with a large excess of
KO, and its signal dominated all resulting spec-
tra; however, signals of derivative forms were
readily observed. Compared to the mercapto
adduct generated by using DTT, the enzymatic
mercapto adduct isomers show additional sig-
nals, which result from heterogeneity of the
proteolysis products.
RE FERENCES AND NOTES
1.
J. Goy, M. Crowther, Approaches to diagnosing and managing
anticoagulant-related bleeding. Semin. Thromb. Hemost. 38,
702–710 (2012). doi: 10.1055/s-0032-1326788; pmid: 23034830
2. S. Schulman, R. J. Beyth, C. Kearon, M. N. Levine, Hemorrhagic
complications of anticoagulant and thrombolytic treatment:
American College of Chest Physicians evidence-based clinical
practice guidelines (8th edition). Chest J. 133, 257S–298S
(2008), vol. 133.
3. D. S. Budnitz, M. C. Lovegrove, N. Shehab, C. L. Richards,
Emergency hospitalizations for adverse drug events in older
Americans. N. Engl. J. Med. 365, 2002–2012 (2011).
doi: 10.1056/NEJMsa1103053; pmid: 22111719
4. D. W. Stafford, The vitamin K cycle. J. Thromb. Haemost. 3,
1873–1878 (2005). doi: 10.1111/j.1538-7836.2005.01419.x;
pmid: 16102054
5. W. Li et al., Structure of a bacterial homologue of vitamin K
epoxide reductase. Nature 463, 507–512 (2010). doi: 10.1038/
nature08720; pmid: 20110994
6. M. A. Rishavy, A. Usubalieva, K. W. Hallgren, K. L. Berkner,
Novel insight into the mechanism of the vitamin K
oxidoreductase (VKOR): Electron relay through Cys43 and
Cys51 reduces VKOR to allow vitamin K reduction and
facilitation of vitamin K-dependent protein carboxylation.
J. Biol. Chem. 286, 7267–7278 (2011). doi: 10.1074/
jbc.M110.172213; pmid: 20978134
7. G. Shen et al., Warfarin traps human vitamin K epoxide
reductase in an intermediate state during electron transfer.
Nat. Struct. Mol. Biol. 24, 69–76 (2017). doi: 10.1038/
nsmb.3333; pmid: 27918545
8. S. Liu, W. Cheng, R. Fowle Grider, G. Shen, W. Li, Structures
of an intramembrane vitamin K epoxide reductase homolog
reveal control mechanisms for electron transfer. Nat. Commun.
5, 3110–3110 (2014). doi: 10.1038/ncomms4110;
pmid: 24477003
9. A. Hammed et al., VKORC1L1, an enzyme rescuing the
vitamin K 2,3-epoxide reductase activity in some extrahepatic
tissues during anticoagulation therapy. J. Biol. Chem. 288,
28733–28742 (2013). doi: 10.1074/jbc.M113.457119;
pmid: 23928358
10. C. Vermeer, K. Hamulyák, Vitamin K: Lessons from the past.
J. Thromb. Haemost. 2, 2115–2117 (2004). doi: 10.1111/
j.1538-7836.2004.01002.x; pmid: 15613015
11. L. Goodstadt, C. P. Ponting, Vitamin K epoxide reductase:
Homology, active site and catalytic mechanism.
Trends Biochem. Sci. 29, 289–292 (2004). doi: 10.1016/
j.tibs.2004.04.004; pmid: 15276181
12. J. Ansell et al., Descriptive analysis of the process and quality
of oral anticoagulation management in real-life practice in
patients with chronic non-valvular atrial fibrillation: The
international study of anticoagulation management (ISAM).
J. Thromb. Thrombolysis 23, 83–91 (2007). doi: 10.1007/
s11239-006-9022-7; pmid: 17221328
13. P.-H. Chu, T.-Y. Huang, J. Williams, D. W. Stafford, Purified
vitamin K epoxide reductase alone is sufficient for conversion
of vitamin K epoxide to vitamin K and vitamin K to vitamin
KH2. Proc. Natl. Acad. Sci. U.S.A. 103, 19308–19313 (2006).
doi: 10.1073/pnas.0609401103; pmid: 17164330
14. J. D. Pédelacq, S. Cabantous, T. Tran, T. C. Terwilliger,
G. S. Waldo, Engineering and characterization of a superfolder
green fluorescent protein. Nat. Biotechnol. 24, 79–88 (2006).
doi: 10.1038/nbt1172; pmid: 16369541
15. K. P. Link, The discovery of dicumarol and its sequels.
Circulation 19, 97–107 (1959). doi: 10.1161/01.CIR.19.1.97;
pmid: 13619027
16. M. Gebauer, Synthesis and structure-activity relationships of
novel warfarin derivatives. Bioorg. Med. Chem. 15, 2414–2420
(2007). doi: 10.1016/j.bmc.2007.01.014; pmid: 17275317
17. W. R. Porter, Warfarin: History, tautomerism and activity.
J. Comput. Aided Mol. Des. 24, 553–573 (2010). doi: 10.1007/
s10822-010-9335-7; pmid: 20352297
18. G. M. Cooper et al., A genome-wide scan for common genetic
variants with a large influence on warfarin maintenance dose.
Blood 112, 1022–1027 (2008). doi: 10.1182/blood-2008-01-
134247; pmid: 18535201
19. J. A. Haque, M. G. McDonald, J. D. Kulman, A. E. Rettie, A
cellular system for quantitation of vitamin K cycle activity:
Structure-activity effects on vitamin K antagonism by warfarin
metabolites. Blood 123, 582–589 (2014). doi: 10.1182/
blood-2013-05-505123; pmid: 24297869
20. G. Shen et al., Membrane protein structure in live cells:
Methodology for studying drug interaction by mass
spectrometry-based footprinting. Biochemistry 57, 286–294
(2018). doi: 10.1021/acs.biochem.7b00874; pmid: 29192498
21. G. Shen et al., Stabilization of warfarin-binding pocket of
VKORC1 and VKORL1 by a peripheral region determines their
different sensitivity to warfarin inhibition. J. Thromb.
Haemost. 16, 1164–1175 (2018). doi: 10.1111/jth.14127;
pmid: 29665197
22. B. C. S. Sinhadri, D. Y. Jin, D. W. Stafford, J. K. Tie, Vitamin K
epoxide reductase and its paralogous enzyme have different
structures and functions. Sci. Rep. 7, 17632–17632 (2017).
doi: 10.1038/s41598-017-18008-3; pmid: 29247216
23. T. E. Klein et al., Estimation of the warfarin dose with clinical
and pharmacogenetic data. N. Engl. J. Med. 360, 753–764
(2009). doi: 10.1056/NEJMoa0809329; pmid: 19228618
24. T. J. Wigle, L. E. Jansen, W. A. Teft, R. B. Kim,
Pharmacogenomics guided-personalization of warfarin and
tamoxifen. J. Pers. Med. 7, 20 (2017). doi: 10.3390/
jpm7040020; pmid: 29236081
25. S. Li, S. Liu, Y. Yang, W. Li, Characterization of warfarin
inhibition kinetics requires stabilization of intramembrane
vitamin K epoxide reductases. J. Mol. Biol. 432,
5197–5208 (2020). doi: 10.1016/j.jmb.2020.05.009;
pmid: 32445640
26. K. Inaba, Y. H. Takahashi, K. Ito, S. Hayashi, Critical role of a
thiolate-quinone charge transfer complex and its adduct
form in de novo disulfide bond generation by DsbB.
Proc. Natl. Acad. Sci. U.S.A. 103, 287–292 (2006).
doi: 10.1073/pnas.0507570103; pmid: 16384917
27. S. M. Miller et al., Use of a site-directed triple mutant to
trap intermediates: Demonstration that the flavin C(4a)-thiol
adduct and reduced flavin are kinetically competent
Liu et al., Science 371, eabc5667 (2021)
1 January 2021
9 of 10
RES EARCH | R E S E A R C H A R T I C L E
intermediates in mercuric ion reductase. Biochemistry 29,
2831–2841 (1990). doi: 10.1021/bi00463a028; pmid: 2189497
28. R. B. Silverman, Chemical model studies for the mechanism of
vitamin K epoxide reductase. J. Am. Chem. Soc. 103,
5939–5941 (1981). doi: 10.1021/ja00409a068
29. C. H. Davis, D. Deerfield 2nd, T. Wymore, D. W. Stafford,
L. G. Pedersen, A quantum chemical study of the mechanism
of action of Vitamin K epoxide reductase (VKOR) II. Transition
states. J. Mol. Graph. Model. 26, 401–408 (2007).
doi: 10.1016/j.jmgm.2006.10.005; pmid: 17182266
30. M. J. Fasco, P. C. Preusch, E. Hildebrandt, J. W. Suttie,
Formation of hydroxyvitamin K by vitamin K epoxide reductase
of warfarin-resistant rats. J. Biol. Chem. 258, 4372–4380
(1983). pmid: 6833262
31. B. Matagrin et al., New insights into the catalytic mechanism of
vitamin K epoxide reductase (VKORC1) - The catalytic
properties of the major mutations of rVKORC1 explain the
biological cost associated to mutations. FEBS Open Bio 3,
144–150 (2013). doi: 10.1016/j.fob.2013.02.001;
pmid: 23772386
32. G. B. Lim, Milestone 2: Warfarin: from rat poison to clinical use.
Nat. Rev. Cardiol. (2017). doi: 10.1038/nrcardio.2017.172;
pmid: 29238065
33. S. Schulman, B. Wang, W. Li, T. A. Rapoport, Vitamin K
epoxide reductase prefers ER membrane-anchored
thioredoxin-like redox partners. Proc. Natl. Acad. Sci. U.S.A.
107, 15027–15032 (2010). doi: 10.1073/pnas.1009972107;
pmid: 20696932
34. E. Gross, D. B. Kastner, C. A. Kaiser, D. Fass, Structure of
Ero1p, source of disulfide bonds for oxidative protein folding in
the cell. Cell 117, 601–610 (2004). doi: 10.1016/S0092-8674
(04)00418-0; pmid: 15163408
35. K. Inaba et al., Crystal structure of the DsbB-DsbA complex
reveals a mechanism of disulfide bond generation.
Cell 127, 789–801 (2006). doi: 10.1016/j.cell.2006.10.034;
pmid: 17110337
36. Y. Hui, E. L. Chng, L. P. Chua, W. Z. Liu, R. D. Webster,
Voltammetric method for determining the trace moisture
content of organic solvents based on hydrogen-bonding
interactions with quinones. Anal. Chem. 82, 1928–1934 (2010).
doi: 10.1021/ac9026719; pmid: 20143888
37. M. J. Fasco, L. M. Principe, R- and S-Warfarin inhibition of
vitamin K and vitamin K 2,3-epoxide reductase activities
in the rat. J. Biol. Chem. 257, 4894–4901 (1982).
pmid: 7068669
38. S. Cabantous, T. C. Terwilliger, G. S. Waldo, Protein
tagging and detection with engineered self-assembling
fragments of green fluorescent protein. Nat. Biotechnol.
23, 102–107 (2005). doi: 10.1038/nbt1044;
pmid: 15580262
39. C. Li et al., FastCloning: A highly simplified, purification-free,
sequence- and ligation-independent PCR cloning method. BMC
Biotechnol. 11, 92–92 (2011). doi: 10.1186/1472-6750-11-92;
pmid: 21992524
40. L. Zheng, U. Baumann, J. L. Reymond, An efficient one-step
site-directed and site-saturation mutagenesis protocol. Nucleic
Acids Res. 32, e115–e115 (2004). doi: 10.1093/nar/gnh110;
pmid: 15304544
41. M. Hattori, R. E. Hibbs, E. Gouaux, A fluorescence-detection
size-exclusion chromatography-based thermostability assay
for membrane protein precrystallization screening. Structure
20, 1293–1299 (2012). doi: 10.1016/j.str.2012.06.009;
pmid: 22884106
42. W. Liu, V. Cherezov, Crystallization of membrane proteins in
lipidic mesophases. J. Vis. Exp. 49, 2501–2501 (2011).
doi: 10.3791/2501; pmid: 21490572
43. Z. Otwinowski, W. Minor, Processing of x-ray diffraction data
collected in oscillation mode. Methods Enzymol. 276, 307–326
(1997). doi: 10.1016/S0076-6879(97)76066-X
44. W. Kabsch, XDS. Acta Crystallogr. D Biol. Crystallogr. 66,
125–132 (2010). doi: 10.1107/S0907444909047337;
pmid: 20124692
45. P. Evans, Scaling and assessment of data quality. Acta
Crystallogr. D Biol. Crystallogr. 62, 72–82 (2006). doi: 10.1107/
S0907444905036693; pmid: 16369096
46. A. J. McCoy et al., Phaser crystallographic software.
J. Appl. Crystallogr. 40, 658–674 (2007). doi: 10.1107/
S0021889807021206; pmid: 19461840
47. G. N. Murshudov, A. A. Vagin, E. J. Dodson, Refinement
of macromolecular structures by the maximum-likelihood
method. Acta Crystallogr. D Biol. Crystallogr. 53,
240–255 (1997). doi: 10.1107/S0907444996012255;
pmid: 15299926
48. K. Cowtan, Recent developments in classical density
modification. Acta Crystallogr. D Biol. Crystallogr. 66, 470–478
(2010). doi: 10.1107/S090744490903947X; pmid: 20383000
49. K. D. Cowtan, ‘dm’: An automated procedure for phase
improvement by density modification. Joint CCP4 and
ESF-EACBM Newsletter on Protein Crystallography 31, 34–38
(1994).
50. P. V. Afonine et al., Towards automated crystallographic
structure refinement with phenix.refine. Acta Crystallogr.
D Biol. Crystallogr. 68, 352–367 (2012). doi: 10.1107/
S0907444912001308; pmid: 22505256
51. J. K. Tie, D. Y. Jin, K. Tie, D. W. Stafford, Evaluation of warfarin
resistance using transcription activator-like effector nucleases-
mediated vitamin K epoxide reductase knockout HEK293 cells.
J. Thromb. Haemost. 11, 1556–1564 (2013). doi: 10.1111/
jth.12306; pmid: 23710884
52. J.-K. Tie, D.-Y. Jin, D. L. Straight, D. W. Stafford, Functional
study of the vitamin K cycle in mammalian cells. Blood 117,
2967–2974 (2011). doi: 10.1182/blood-2010-08-304303;
pmid: 21239697
53. C. G. Bevans et al., Determination of the warfarin inhibition
constant Ki for vitamin K 2,3-epoxide reductase complex
subunit-1 (VKORC1) using an in vitro DTT-driven assay.
Biochim. Biophys. Acta 1830, 4202–4210 (2013). doi: 10.1016/
j.bbagen.2013.04.018; pmid: 23618698
54. S. Rost et al., Mutations in VKORC1 cause warfarin resistance
and multiple coagulation factor deficiency type 2. Nature
427, 537–541 (2004). doi: 10.1038/nature02214;
pmid: 14765194
55. P. C. Preusch, J. W. Suttie, A chemical model for the
mechanism of vitamin K epoxide reductase. J. Org. Chem. 48,
3301–3305 (1983). doi: 10.1021/jo00167a028
56. J. Schleucher et al., A general enhancement scheme in
heteronuclear multidimensional NMR employing pulsed field
gradients. J. Biomol. NMR 4, 301–306 (1994). doi: 10.1007/
BF00175254; pmid: 8019138
57. G. E. Martin, R. T. Williamson, P. G. Dormer, W. Bermel,
Inversion of 1J(CC) correlations in 1,n-ADEQUATE spectra.
Magn. Reson. Chem. 50, 563–568 (2012). doi: 10.1002/
mrc.3840; pmid: 22806686
AC KNOWLED GME NTS
We dedicate this paper to the memory of Dr. Evan Sadler, who
has provided W.L. with generous support and invaluable advice as
a mentor. We thank D. Fremont for his critical reading of the
manuscript. Funding: W.L. is supported by National Heart, Lung,
and Blood Institute (NHLBI) (R01 HL121718), W. M. Keck
Foundation (Forefront of Science Award), Children’s Discovery
Institute (MCII 2020-854), National Eye Institute (NEI) (R21
EY028705), and National Institute of General Medical Sciences
(NIGMS) (R01 GM131008). The NECAT beamlines are funded
by NIGMS (P30 GM124165), the Eiger detector is funded by a
NIH-ORIP (HEI S10OD021527), and APS is supported by the U.S.
Department of Energy (DOE) (DE-AC02-06CH11357). Author
contributions: W.L. and S. Liu designed the study; S. Liu purified
and crystallized the proteins; S. Liu collected data with help
from N.S.; W.L. solved the structures; S. Li performed activity
assays; G.S. inspired how to make substrates stably bound;
A.M.K. performed NMR analyses; W.L. oversaw the studies and
analyzed the results; W.L. wrote the manuscript with input from
the other authors. Competing interests: None declared. Data and
materials availability: Atomic coordinates and structure factors
for the reported crystal structures have been deposited in
the Protein Data Bank under accession codes 6WV3 (HsVKOR-
warfarin), 6WV6 (HsVKOR-phenindione), 6WVH (HsVKOR-
Brodifacoum), 6WV7 (HsVKOR-Chlorophacinone), 6WV4 (HsVKOR
Cys43Ser-warfarin), 6WV5 (HsVKOR Cys43Ser-KO), 6WVI (TrVKORL),
6WVB (TrVKORL-warfarin), 6WV9 (TrVKORL-K), 6WVA
(TrVKORL-KO), and 6WV8 (TrVKORL ‘Cys132Ser’-K). All other
data needed to evaluate the conclusions in the paper are present
in the paper or the supplementary materials.
SUPPLEMENTARY MATERIALS
science.sciencemag.org/content/371/6524/eabc5667/suppl/DC1
Figs. S1 to S19
Tables S1 to S3
References (58–64)
Movies S1 to S3
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
10 May 2020; accepted 27 October 2020
Published online 5 November 2020
10.1126/science.abc5667
Liu et al., Science 371, eabc5667 (2021)
1 January 2021
10 of 10
|
10.1126_science.abe9124
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
IMMUNOLOGY
Canonical T cell receptor docking on peptide–MHC
is essential for T cell signaling
Pirooz Zareie, Christopher Szeto†, Carine Farenc, Sachith D. Gunasinghe, Elizabeth M. Kolawole,
Angela Nguyen, Chantelle Blyth, Xavier Y. X. Sng, Jasmine Li, Claerwen M. Jones, Alex J. Fulcher,
Jesica R. Jacobs, Qianru Wei, Lukasz Wojciech, Jan Petersen, Nicholas R.J. Gascoigne,
Brian D. Evavold, Katharina Gaus, Stephanie Gras†‡*, Jamie Rossjohn‡*, Nicole L. La Gruta‡*
APC
INTRODUCTION: T cell receptor (TCR) recogni-
tion of peptide–major histocompatibility com-
plexes (pMHCs) is one of the most diverse
receptor–ligand interactions in biology. Never-
theless, these interactions exhibit a
highly conserved, or canonical, TCR–
pMHC docking polarity in both mice
and humans. Whether this canon-
ical docking polarity is driven by
evolutionarily conserved, germline-
encoded complementarity between
the TCR and MHC or by signaling
constraints imposed by coreceptors
has been a question of enduring de-
bate. Here, we demonstrate that al-
though reversed-polarity TCR–pMHC
recognition is prevalent within a
naïve, viral epitope–specific T cell
repertoire and may exhibit relatively
high pMHCI affinity, such TCRs are
unable to support TCR signaling in
the presence of CD8 coreceptor be-
cause of mislocalization of Lck. These
data support a paradigm in which the
highly conserved TCR–pMHCI dock-
ing polarity is driven by structural
constraints on TCR signaling.
Lck
CD8
RATIONALE: Evidence suggests that
the canonical TCR–pMHC docking
polarity is driven by evolutionary hard-
wiring of complementary germline-
encoded motifs at the TCR and MHC
interface. An alternate model sug-
gests that TCR recognition of pMHC
is driven during thymic selection by the need
for the CD4 or CD8 coreceptors to bind MHC
and deliver coreceptor-associated Lck to the
CD3 signaling complex. We previously iden-
tified reversed-polarity TRBV17+ TCRs from
the preimmune influenza A virus (IAV)–specific
repertoire that bound pMHCI (H-2DbNP366)
with a moderate affinity but were unable to
support robust T cell recruitment. Here, using
a range of canonical and reversed TCRs spe-
cific for the same cognate pMHCI, we tested
the hypothesis that the TCR–pMHCI dock-
ing polarity precedes binding strength as
a key determinant of T cell activation. We
hypothesized that the underlying driver of
the canonical docking polarity is the colocal-
ization of signaling molecules central to the
TCR signal transduction pathway.
TCR-pMHC docking
Canonical
Reversed
APC
CD8
Lck
pMHC
TCR
CD3
Signal
transduction
>10 nm
CD8+ T cell
CD8+ T cell
The canonical polarity of TCR–pMHC docking is essential for colo-
calization of CD3 and coreceptor-associated Lck and for produc-
tive TCR signaling. Schematic shows how canonical TCR–pMHC
recognition colocalizes Lck and CD3, driving TCR-mediated signaling.
By contrast, a reversed TCR–pMHC recognition polarity
mislocalizes Lck and CD3, impeding signaling.
RESULTS: In this study, we demonstrate that
reversed TCRs are prevalent in a naïve virus–
specific repertoire but are poorly represented
in the immune response after virus challenge.
We identified antigen-specific TCRab clono-
types that were either poorly recruited or
clonally expanded and found an overriding
association between immune prevalence and
canonical TCR–pMHCI docking. This was
irrespective of pMHCI affinity, catch or slip
bond formation, or TCR clustering, demon-
strating that a canonical docking polarity is
required for T cell activation. This finding
was verified after viral challenge of adoptively
transferred retrogenic T cells expressing re-
versed or canonical docking TCRs of varying
affinities. The inability of T cells expressing
reversed-docking TCRs to be recruited into the
antiviral immune response demonstrates that
TCR–pMHCI docking topology supersedes
TCR–pMHCI affinity as the primary determi-
nant for effective in vivo immune recruitment.
Using fluorescence lifetime imaging micros-
copy (FLIM)–Förster resonance energy trans-
fer (FRET) analyses, we show that canonical
TCR–pMHCI docking is essential for the
colocalization of CD8–Lck with CD3z, which
is impaired when the TCR engages pMHCI
with reversed polarity. The requirement for
canonical TCR–pMHCI docking can be cir-
cumvented by the removal of the CD8 core-
ceptor or by dissociation of Lck from CD8,
suggesting that sequestration of Lck by the
CD8 coreceptor has a dual role: po-
tentiating signaling arising from ca-
nonical TCR–pMHCI interactions and
impeding reversed-polarity TCR–
pMHCI signaling.
pMHC
TCR
CD3
CONCLUSION: The inability of reversed-
polarity TCRs to participate in the im-
mune response occurs independently
of TCR–pMHCI binding affinity and
instead is a direct consequence of re-
versed TCR–pMHCI engagement. Most
TCR–pMHC complexes that have been
solved to date, upon which the canoni-
cal TCR–pMHCI docking paradigm
has been established, were derived
from expanded immune repertoires.
Thus, we conclude that the highly
conserved docking polarity is driven
predominantly by the structural con-
straints imposed on TCR signaling and
recruitment into an immune response.
In addition to the well-recognized
augmentation of signaling resulting
from canonical TCR–pMHCI engage-
ment, our findings suggest a role for
coreceptor–Lck association in pre-
venting signaling by noncanonical
TCR–pMHC recognition. Such neg-
ative regulation would serve to limit
the extent of functional TCR cross-
reactivity and constrain the number of signaling-
competent TCR-binding modalities.▪
The list of author affiliations is available in the full article online.
*Corresponding author. Email: nicole.la.gruta@monash.edu
(N.L.L.G.); jamie.rossjohn@monash.edu (J.R.); s.gras@
latrobe.edu.au (S.G.)
†Present address: Department of Biochemistry and Genetics,
La Trobe Institute for Molecular Science, La Trobe University,
Melbourne, Victoria, Australia.
‡Joint senior authors.
Cite this article as P. Zareie et al., Science 372, eabe9124
(2021). DOI: 10.1126/science.abe9124
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.abe9124
Zareie et al., Science 372, 1056 (2021)
4 June 2021
1 of 1
RES EARCH
R E S E A R C H A R T I C L E
◥
IMMUNOLOGY
Canonical T cell receptor docking on peptide–MHC
is essential for T cell signaling
Pirooz Zareie1, Christopher Szeto1†, Carine Farenc1, Sachith D. Gunasinghe2, Elizabeth M. Kolawole3,
Angela Nguyen1, Chantelle Blyth1, Xavier Y. X. Sng1, Jasmine Li4, Claerwen M. Jones1,
Alex J. Fulcher5, Jesica R. Jacobs3, Qianru Wei6, Lukasz Wojciech6, Jan Petersen1,7,
Nicholas R.J. Gascoigne6, Brian D. Evavold3, Katharina Gaus2, Stephanie Gras1,7†‡*,
Jamie Rossjohn1,7,8‡*, Nicole L. La Gruta1‡*
T cell receptor (TCR) recognition of peptide–major histocompatibility complexes (pMHCs) is
characterized by a highly conserved docking polarity. Whether this polarity is driven by recognition or
signaling constraints remains unclear. Using “reversed-docking” TCRb-variable (TRBV) 17+ TCRs from the
naïve mouse CD8+ T cell repertoire that recognizes the H-2Db–NP366 epitope, we demonstrate that their
inability to support T cell activation and in vivo recruitment is a direct consequence of reversed docking
polarity and not TCR–pMHCI binding or clustering characteristics. Canonical TCR–pMHCI docking
optimally localizes CD8/Lck to the CD3 complex, which is prevented by reversed TCR–pMHCI polarity.
The requirement for canonical docking was circumvented by dissociating Lck from CD8. Thus, the
consensus TCR–pMHC docking topology is mandated by T cell signaling constraints.
T cell–mediated immunity to pathogens
and cancers requires activation of T cells
through ab T cell antigen receptor (TCR)
recognition of antigenic peptides pre-
sented by major histocompatibility
complex class I (MHCI) or class II (MHCII)
molecules. The extreme diversity inherent in
both the TCR repertoire and the array of
pMHC ligands is reflected in the substantial
variation at the TCR–pMHC interface (1). De-
spite this variation, nearly all of the TCR–pMHC
ternary complexes solved to date exhibit a highly
consistent docking polarity, with the TCRa
chain sitting over the MHCI a2 or MHCII
b1 helix, and TCRb docking over the MHCI
1Infection and Immunity Program and Department of
Biochemistry and Molecular Biology, Biomedicine Discovery
Institute, Monash University, Clayton, Victoria, Australia.
2European Molecular Biology Laboratory (EMBL) Australia
Node in Single Molecule Science and the ARC Centre of
Excellence in Advanced Molecular Imaging, School of Medical
Sciences, University of New South Wales, New South Wales,
Australia. 3Department of Pathology, University of Utah
School of Medicine, Salt Lake City, UT, USA. 4Infection and
Immunity Program and Department of Microbiology,
Biomedicine Discovery Institute, Monash University, Clayton,
Victoria, Australia. 5Monash Micro Imaging, Monash
University, Clayton, Victoria, Australia. 6Immunology
Translational Research Programme and Department of
Microbiology and Immunology, Yong Loo Lin School
of Medicine, National University of Singapore, Singapore
117545. 7Australian Research Council Centre of Excellence
in Advanced Molecular Imaging, Monash University,
Clayton, Victoria, Australia. 8Institute of Infection and
Immunity, Cardiff University School of Medicine, Heath Park,
Cardiff, UK.
†Present address: Department of Biochemistry and Genetics,
La Trobe Institute for Molecular Science, La Trobe University,
Melbourne, Victoria, Australia.
‡Joint senior authors.
*Corresponding author. Email: nicole.la.gruta@monash.edu
(N.L.L.G.); jamie.rossjohn@monash.edu (J.R.); s.gras@latrobe.
edu.au (S.G.)
and MHCII a1 helix (1). Evidence suggests that
the conserved TCR–pMHC docking polarity
is “hard-wired” by evolutionarily conserved
amino acid motifs in the germline-encoded
regions of TCRs and MHC molecules (2–5). An
alternate model suggests that TCR recognition
of pMHC is driven during thymic selection by
the need for the CD4 or CD8 coreceptors to
bind MHC and deliver coreceptor-associated
Lck to the CD3 signaling complex (5). Because
of the proposed positioning of CD3 in the TCR–
pMHC–CD4/CD8 complex, this model posits
that only canonical polarity TCR–pMHC inter-
actions are conducive to signaling (6–8). The
biological significance of the canonical docking
polarity remains unclear and has not been
tested experimentally because of the rarity of
TCR–pMHC docking polarities outside of this
paradigm (1, 9–11). We recently identified CD8+
T cells expressing “reversed” TCRs that bind
their cognate pMHCI in a 180° reversed orien-
tation, signaled poorly, and drove a weak
antiviral immune response (12). Here, we
investigated the key drivers of the canonical
TCR–pMHC docking polarity and its role in
T cell recognition and activation.
Results
Unconventional TRBV17+ TCRs are prevalent in
the naïve H-2Db–NP366–specific repertoire but
do not contribute to the immune response
We have previously described two naïve TRBV17+
TCRs (NP1-B17, hereafter referred to as B17.R1,
and NP2-B17) that recognize the H-2Db–NP366
epitope in a 180° reversed orientation (12). To
determine whether reversed TCR docking was a
general feature of TRBV17+ H-2Db–NP366–specific
TCRs, we generated H-2Db–NP366 tetramers con-
taining single–amino acid substitutions at
H-2Db Glu18 and/or Ala89, residues that are
uniquely important for binding the reversed
B17.R1 TCRa chain (1, 12). Ala89Glu substitu-
tion (H-2DbA89E–NP366) completely abrogated
tetramer binding to B17.R1 at high TCR expres-
sion levels without affecting B13.C1 TCR bind-
ing (Fig. 1, A and B). The loss of B17.R1 TCR
binding was verified by surface plasmon reso-
nance (SPR) analysis (Fig. 1C and Table 1).
We next used comparative staining with
the H-2DbWT–NP366 and H-2DbA89E–NP366 tet-
ramers to determine the proportion of the
naïve H-2Db–NP366–specific CD8+ T cell pop-
ulation that was affected by this mutation,
suggestive of a reversed TCR–pMHCI docking
polarity (Fig. 1, D and E). Although a similar
number of TRBV13+ cells was detected using
either tetramer (Fig. 1F), the mutant A89E
tetramer detected only ~48% of TRBV17+
cells detected by the wild-type (WT) tetramer
(Fig. 1G). By contrast, these two tetramers
showed equivalent staining of both TRBV13+
and TRBV17+ T cells in mice infected with
influenza A virus (IAV) (Fig. 1, H to K). Thus,
although TRBV17+ TCRs that bind H-2Db–NP366
in a reversed orientation are prevalent in the
naïve repertoire, they are not recruited into the
immune response after IAV infection.
Recruitment into the immune response is
associated with TCR–pMHCI docking topology
independently of TCR–pMHCI affinity
To gain a further understanding of TCR-intrinsic
determinants of recruitment, we analyzed
TRBV17+ H-2Db–NP366–specific TCRab sequences
from uninfected (13) and infected B6 mice. The
naïve TRBV17+ TCRab repertoire was diverse,
comprising a range of TCRa-variable (TRAV)
genes with distinct CDR3a and CDR3b se-
quences (Fig. 2A) (13). By contrast, each im-
mune repertoire was characterized by the
dominance of only one or two clones (Fig. 2B).
Thus, the low prevalence of TRBV17+ TCRs in
the H-2Db–NP366–specific immune repertoire
(13) is caused by an inability of most of these
clones to respond to IAV.
To demonstrate that the key criteria for
immune recruitment from the TRBV17 subset
was a canonical docking polarity, we selected
three TRBV17+ TCRs from the immune rep-
ertoire for further structural and biophysical
analyses. These TCRs were taken from ex-
panded clones (mouse 1: B17.C1; mouse 3:
B17.C2) and from the single TRAV14+ clone
from mouse 3 (B17.R2) (Fig. 2B). Although
the B17.R2 TCR was identified from an in-
fected mouse, the sensitive detection method
and apparent lack of clonal expansion means
that it was likely derived from a naïve T cell.
We performed tetramer staining of 293T cells
expressing these TCRs, along with a TRBV13+
TCR (B13.C1) known to drive robust immune
recruitment (12, 13). Those TCRs that were
Zareie et al., Science 372, eabe9124 (2021)
4 June 2021
1 of 13
RES EARCH | R E S E A R C H A R T I C L E
B
B13.C1
B17.R1
C
A
T
W
f
o
I
F
M
%
D
e
n
u
m
m
i
-
e
r
P
H
e
n
u
m
m
I
150
100
50
0
6
6
3
P
N
–
T
W
b
D
2
-
H
6
6
3
P
N
–
T
W
b
D
2
-
H
B13.C1
B17.R1
6
6
3
P
N
–
b
D
2
-
H
A89V
WT
A89E
A A89F
89R
E18A
E18V
18F
E E118L
E 8W
30.58
0
7
1
V
B
R
T
57.02
12.40
TCR
TRBV13
0.20
0
7
1
V
B
R
T
E
I
6
6
3
P
N
–
E
9
8
A
b
D
2
-
H
6
6
3
P
N
–
E
9
8
A
b
D
2
-
H
79.62
20.18
CD44
TRBV13
CD44
TCR
7
1
V
B
R
T
13.21
0
67.92
18.87
TCR
TRBV13
0.24
0
7
1
V
B
R
T
79.14
20.62
TRBV13
B17.R1
400 WT KD = 37.5 ± 4.4 M
A89E KD = >200 M
300
200
100
0
0
25
50
100
Concentration (µM)
200
WT
A89V
A89E
A89F
A89R
E18A
E18V
E18F
E18L
E18W
Overlay
)
U
R
(
s
t
i
n
U
e
s
n
o
p
s
e
R
F
)
e
g
n
a
h
c
d
o
f
(
l
+
3
1
V
B
R
T
J
)
e
g
n
a
h
c
d
o
f
(
l
+
3
1
V
B
R
T
1.0
0.5
0.0
-0.5
-1.0
1.0
0.5
0.0
-0.5
-1.0
H-2DbA89E –NP366
H-2DbA89E –NP366
G
)
e
g
n
a
h
c
d
o
f
(
l
+
1.0
0.5
0.0
-0.5
7
1
V
B
R
-1.0T
K
)
e
g
n
a
h
c
d
o
f
(
l
+
7
1
V
B
R
T
1.0
0.5
0.0
-0.5
-1.0
H-2DbA89E –NP366
H-2DbA89E –NP366
Fig. 1. Reverse-docking TRBV17+ T cells are not recruited into the immune
response. (A and B) HEK293T cells were transfected with B13.C1 (black) and
B17.R1 (gold) TCRs, and binding of WT or mutant H-2Db–NP366 tetramers was
analyzed 48 hours after transfection. Live GFP+ cells were gated and analyzed for
change in geometric mean fluorescence intensity (geoMFI). Shown is geoMFI as
a percentage change from the WT tetramer (A) and representative dot plots
staggered (top panel) or overlaid (bottom panel) of TCR expression and binding
of various mutant tetramers (B). Data shown in (A) are from two independent
experiments combined. Data shown in (B) are dot plots from one representative
experiment. (C) Binding response of B17.R1 TCR against H-2DbWT–NP366 in black
or H-2DbA89E–NP366 in green. Data presented are from a single experiment
representative of two independent experiments. (D to K) Representative dot
plots and graphs showing the proportion of H-2DbWT–NP366 binding [(D) and (H),
black] or H-2DbA89E–NP366 binding [(E) and (I), green] T cells using either
TRBV13 (Vb8.3) or TRBV17 (Vb9) TCRb chains isolated from naïve mice
[10 mice pooled/data point; (D) to (G)] or immune mice 10 days after
infection [one mouse per data point; (H) to (K)]. Plots represent the percentage
change in TRBV13+ T cells bound by the H-2DbA89E–NP366 tetramer relative to
H-2DbWT–NP366 tetramer from naïve mice (F) and immune mice (J), and the
percentage change in TRBV17+ T cells bound by the H-2DbA89E–NP366
tetramer relative to H-2DbWT–NP366 tetramer binding from naïve mice (G) and
immune mice (K). Data shown from 10 pooled naïve mice in (D) to (E)
represent one sample of n = 3. Summary data shown in (F) and (G) are the
mean and SEM of three independent datasets (each containing 10 pooled
mice). Data from an individual immune mouse shown in (H) and (I) represent
one sample of n = 9. Summary data shown in (J) and (K) are mean and
SEM from a representative three samples (each containing one mouse)
collected on one day.
well represented in the immune response, in-
cluding B13.C1 (Fig. 2C), B17.C1 (Fig. 2D), and
B17.C2 (Fig. 2F), all showed equivalent binding
by the H-2DbWT–NP366 and H-2DbA89E–NP366
tetramers. By contrast, the poorly represented
or unrecruited B17.R2 TCR showed significantly
reduced binding of the H-2DbA89E–NP366
tetramer (Fig. 2E) and a 10-fold reduced af-
finity for H-2DbA89E–NP366, suggesting a
reversed TCR–pMHCI docking polarity.
To determine the role of TCR–pMHCI af-
finity in driving immune recruitment of TRBV17+
cells, we determined TCR affinity by SPR (Table 1).
The canonical immune B17.C1 TCR had an
extremely weak affinity for H-2Db–NP366 (KD
> 200 mM) (Fig. 2D). By contrast, the minor or
naïve B17.R2 TCR had a substantially higher
affinity (KD = 6.34 mM) and tetramer binding
(Fig. 2E), similar to that of the immunodom-
inant B13.C1 TCR (KD = 4.13 mM) (Fig. 2C).
Thus, the prevalence of T cells in the immune
response is primarily associated with canon-
ical TCR–pMHC docking polarity, independent
of TCR–pMHC affinity.
Structural determination of TCR–H-2Db–NP366
docking topologies
We next determined the crystal structures of
B17.R2 and B17.C1 TCRs in complex with H-
2Db–NP366 (Fig. 3 and tables S1 to S4). As
suggested from tetramer binding (Fig. 2), the
B17.R2 TCR adopted a reversed docking po-
larity over H-2Db–NP366, forming a docking
angle of 238° and binding in a similar manner
to the previously determined B17.R1–H-2Db–
NP366 and NP2-B17–H-2Db–NP366 complexes
(Fig. 3, A and B, and table S4) (12). By contrast,
Zareie et al., Science 372, eabe9124 (2021)
4 June 2021
2 of 13
RES EARCH | R E S E A R C H A R T I C L E
Table 1. TCRs, recruitment characteristics, docking polarity, and KD.
TCR
TCRa
TCRb
Immune recruitment
Docking polarity
KD (mM)
H-2DbWT–NP366 H-2DbA89E–NP366
B13.C1 (NP1–B13*) TRAV16; RVSGGSNAKL TRBV13–1; SGGGNTGQL
............................................................................................................................................................................................................................................................................................................................................
TRBV17; SRDLGRDTQ
B17.R1 (NP1-B17*)
............................................................................................................................................................................................................................................................................................................................................
B17.R2
TRBV17; SRDLGRDTQ
............................................................................................................................................................................................................................................................................................................................................
TRBV17; SRGTIHSNTEV
B17.C1
............................................................................................................................................................................................................................................................................................................................................
TRBV17; SRGGLSYEQ
B17.C2
............................................................................................................................................................................................................................................................................................................................................
Immunodominant
Naïve or poorly recruited
Naïve or poorly recruited
Clonally expanded
Clonally expanded
TRAV14; SETSGSWQL
TRAV14; SETSASWQL
TRAV4–4; VTGNTGKL
TRAV14D; SRRGSAKL
4.13 ± 1.55
37.5 ± 4.4
6.34 ± 1.58
>>200
ND
Canonical†
Reversed
Reversed
Canonical
Canonical†
ND
>200
62 ± 30
>200
ND
*Previously reported in (12).
performed at a minimum twice in duplicate.
†Structure undetermined; polarity inferred by MHCI mutational analyses.
ND, not determined.
SPR values are the mean ± SEM of experiments
A
CDR3
CDR3
C
BB1133..CC11
6
6
3
P
N
–
b
D
2
-
H
f
o
r
e
b
m
u
N
s
l
l
e
c
+
7
1
V
B
R
T
c
fi
c
e
p
s
-
i
B
6
6
3
P
N
–
b
D
2
-
H
f
o
r
e
b
m
u
N
s
l
l
e
c
+
7
1
V
B
R
T
c
fi
c
e
p
s
-
i
10
9
8
7
6
5
4
3
2
1
0
60
40
20
0
Pre-Immune
m1
m2 m3 m4 m5 m6
Immune
CDR3
CDR3
m1 m2 m3 m4
D
BB1177..CC11
E
BB1177..RR22
F
6
6
3
P
N
–
b
D
2
-
H
BB1177..CC22
TCR
)
U
R
(
s
t
i
n
U
e
s
n
o
p
s
e
R
)
U
R
(
s
t
i
n
U
e
s
n
o
p
s
e
R
)
U
R
(
s
t
i
n
U
e
s
n
o
p
s
e
R
40
30
20
10
0
0
WT KD = 4.13 ± 1.55 M
25
50
Concentration ( M)
1 00
40 WT KD = >200 M
A89E KD = >200 M
25 50
100
Concentration ( M)
200
WT KD = 6.34 ± 1.58 M
A89E KD = 62.0 ± 30 M
30
20
10
0
0
500
400
300
200
100
0
0
25 50
100
200
Concentration ( M)
H-2DbWT–NP366
H-2DbA89E–NP366
Fig. 2. Prevalence of TCRs in the immune response is unrelated to TCR–
pMHCI affinity. (A and B) TRBV17 TCRb+ H-2Db–NP366–specific TCR clonotypes
presented as bar graphs with corresponding CDR3a/CDR3b sequences from
six individual naïve (m1 to m6) (13) (A) or four individual immune mice (m1 to
m4) 10 days after infection with IAV (B). (C to F) HEK293T cells were transfected
with pMIGII vectors encoding abTCR and CD3gdez and H-2DbWT–NP366 (black)
or H-2DbA89E–NP366 (green) tetramer staining analyzed by flow cytometry
48 hours later. Shown are representative flow cytometry plots of TCRb expression
and tetramer binding from live GFP+ cells (left) and SPR sensorgrams (right)
of B13.C1 (C), B17.C1 (D), B17.R2 (E), and B17.C2 (F) TCRs. Data shown in (C) to
(F) are from one experiment representative of two (SPR) or three (flow cytometry)
independently performed experiments.
the B17.C1 TCR adopted a canonical docking
polarity (Fig. 3, C and D, and table S4). In the
B17.R2 TCR–H-2Db–NP366 complex, the TCRa
chain played a lesser role [26.4% of the buried
surface area (BSA)] in the interaction, with
only the CDR3a loop contributing to binding
(Fig. 3B and table S4). By contrast, the TCRb
chain contributed to 73.6% of the BSA, en-
compassing the framework region of the
b-chain (FWb) region (39.2% of the BSA),
the CDR2b loop (27.3% of the BSA), and the
CDR3b loop (7.1% of the BSA) (Fig. 3B and
Zareie et al., Science 372, eabe9124 (2021)
4 June 2021
3 of 13
RES EARCH | R E S E A R C H A R T I C L E
A
B
C
D
B17.R2
B17.C1
B17.C1
B17.R2
2m
366NP
238°
CDR2
CDR3
H-2Db
CDR3
Fw
366NP
2m
CDR1
CDR1
33°
CDR3
FW
H-2Db
CDR2
CDR3
Fig. 3. B17.R2 TCR and B17.C1 TCR in complex with H-2Db–NP366. (A and C) TCR–pMHC complexes of B17.R2 TCR–H-2Db–NP366 (A) and B17.C1 TCR–H-2Db–
NP366 (C). (B and D) TCR atomic footprints on the surface of each corresponding pMHC complex. Spheres represent the center of mass of Va (pink) and Vb (blue).
Pie charts represent the relative contributions of each TCR segment to the pMHCI interaction. Contacts are colored according to the CDR loop involved. The
TCRa chain is colored pink, the TCRb chain blue, H-2Db white, b2m orange, and peptide black/dark gray. CDR1a, CDR2a, and CDR3a are colored in green, teal, and
purple, respectively, and CDR1b, CDR2b, and CDR3b in red, orange, and yellow, respectively. Framework (FW) is colored pink for the FWa and blue for the FWb.
table S4). Similarly, the B17.C1 TCR–H-2Db–
NP366 complex exhibited an unusually high
contribution of the TCRb chain (75% of the
BSA), with the interactions dominated by
the CDR3b loop (37.8% of the BSA) and the
CDR2b loop (23.2% of the BSA) (Fig. 3D and
table S4). Moreover, this complex structure
presents an unusually low number of con-
tacts between the TCR and the MHCI (table
S2), as well as a poor shape complementarity
(table S4), and was consistent with the
low affinity of this interaction (Table 1 and
table S2).
CD8+ T cell recruitment does not correspond to
two-dimensional TCR–pMHCI affinities nor bond
duration under force
To test the possibility that T cell recruitment
correlated with two-dimensional (2D) TCR–
H-2Db–NP366 affinities, we measured the rel-
ative 2D affinity of B13.C1, B17.R1, B17.R2, and
B17.C1 TCRs for H-2Db–NP366 using the 2D
micropipette adhesion frequency assay (2D–
MP) (14–18). Although the reversed B17.R2
TCR had the second highest 2D affinity after
the B13.C1 TCR (fig. S1A), it was not condu-
cive to robust immune recruitment (Fig. 2B).
By contrast, the canonical docking B17.C1 TCR
with a lower 2D affinity (fig. S1A) was expanded
in the immune repertoire (Fig. 2B). Thus,
the recruitment of the H-2Db–NP366–specific
TRBV17+ T cells occurs independently of TCR–
pMHCI affinity.
We next measured the TCR–H-2Db–NP366
bond lifetime under conditions of force (14–17).
CD8+ TCR transductants were stimulated with
peptide bound to H-2DbWT or to mutant
H-2DbD227K to assess the contribution of co-
receptor binding to bond strength (19). Both
of the canonical docking TCRs, the high-
affinity B13.C1 TCR and the low-affinity B17.C1
TCR, were able to form catch bonds, peaking at
~10 pN (fig. S1, B and C). CD8 binding con-
tributed significantly to bond lifetime only for
the low-affinity B17.C1 TCR (fig. S1, B and C). By
contrast, reversed-polarity TCRs (B17.R1 and
B17.R2) showed the formation of slip bonds
with pMHCI, with a loss of bond lifetime with
increasing force (fig. S1, D and E). However, at
least for the high-affinity B17.R2 TCR, the
bond lifetime generated at ~10 pN, an approx-
imation of the physiological force on a TCR
(20–22), was similar to that observed at the
peak of the catch bond formation (fig. S1F).
Thus, although the reversed TCR–H-2Db–NP366
interaction is characterized by slip-bond for-
mation, it exhibits relatively high bond life-
times at the physiological force of 10 pN.
Only canonical docking TCRs can support
immune recruitment irrespective of
TCR–H-2Db–NP366 affinity
To confirm our earlier observations (Fig. 1, D
to K) that T cell recruitment into the immune
response was primarily dependent on a ca-
nonical TCR–pMHCI docking polarity indepen-
dent of TCR–pMHCI binding strength, we
selected B17.R1, which binds with low to mod-
erate affinity, and the B17.R2 TCR, which binds
with high affinity similar to that of the im-
munodominant B13.C1 TCR, for further inves-
tigation. We then generated retrogenic mice
expressing the canonical polarity B13.C1 or
B17.C2 TCRs and the reversed-polarity B17.R1
or B17.R2 TCRs (23). We were unsuccessful
in expressing the B17.C1 TCR in vivo despite
validating construct fidelity and instead gen-
erated TCR-retrogenic mice expressing the
B17.C2 TCR, which exhibited similar prop-
erties. That is, it expressed TRBV17, had a
moderate to low avidity for H-2Db–NP366,
bound H-2Db–NP366 independently of Ala89
(and thus likely docked in a canonical orien-
tation) (Fig. 2F), and was expanded in the
immune repertoire (Table 1). Consistent with
our previously published data (12), adoptive
transfer of retrogenic B13.C1+ and B17.R1+
T cells, either alone or in combination, fol-
lowed by IAV challenge (Fig. 4A), resulted in
the effective recruitment and expansion of
canonical B13.C1+ T cells but not reversed B17.
R1+ T cells (Fig. 4, B to N, and fig. S2A). Failure
to recruit B17.R1+ T cells was not caused by
green fluorescent protein (GFP) expression be-
cause the same experiment performed with
GFP+ B13.C1+ T cells showed similar recruit-
ment profiles as those coexpressing mCherry
(fig. S2, B to E).
To distinguish the impact of TCR–pMHCI
affinity and docking polarity on recruitment,
we adoptively transferred T cells expressing the
high-affinity reversed B17.R2 TCR either alone
or with B13.C1+ T cells before IAV challenge.
The B17.R2 T cells were not detectable in the
immune response after either single transfer
Zareie et al., Science 372, eabe9124 (2021)
4 June 2021
4 of 13
RES EARCH | R E S E A R C H A R T I C L E
A
CD45.2+
TCR retrogenic donors
B13.C1 /
B17.C2
B17.R1 /
B17.R2
Adoptive transfer of
CD8+ T cells (4000)
CD45.1+ host mice
Intranasal influenza A virus
infection (1x10E4
PFU HKx31)
10 days
Harvest BAL, spleen and mLN
B
L
A
B
n
e
e
p
S
l
N
L
m
Individual
transfers
Co-transfers
s
l
l
e
c
T
+
8
D
C
f
o
%
s
l
l
e
c
T
+
8
D
C
f
o
%
s
l
l
e
c
T
+
8
D
C
f
o
%
C
D
E
25
20
15
10
5
0
15
10
5
0
5
4
3
2
1
0
B13.C1
B17.R1
B17.R2
B17.C2
s
l
l
e
c
T
+
8
D
C
f
o
%
s
l
l
e
c
T
+
8
D
C
f
o
%
s
l
l
e
c
T
+
8
D
C
f
o
%
25
20
15
10
5
0
15
10
5
0
5
4
3
2
1
0
I
J
K
F
G
H
50
40
30
20
10
0
20
15
10
5
0
5
4
3
2
1
0
L
M
N
25
20
15
10
5
15
10
5
5
4
3
2
1
0
B17.R2
B17.C2
B13.C1
B17.R1
B13.C1
B17.R2
Fig. 4. TCR–pMHCI docking orientation is a primary determinant of in
vivo T cell activation and recruitment. (A) Schematic diagram of the
experimental protocol. Retrogenic CD45.2+ GFP/mCherry+ CD4− CD8+ T cells
were sorted from B13.C1 (mCherry), B17.R1 (GFP), B17.R2 (GFP), and
B17.C2 (mCherry) retrogenic mice, and 4 × 103 cells were transferred
individually or cotransferred into naïve C57BL/6 mice that were infected
intranasally with 1 × 104 PFUs of HKx31 IAV the following day. Mice were
euthanized for analysis 10 days after infection. (B) Dot plots from the BAL of
mice that received cotransfers of retrogenic T cells. (C to N) Percentage
retrogenic CD8+ T cells of total CD8+ T cells isolated from the BAL [(C), (F),
(I), (L)], spleen [(D), (G), (J), (M)], or mLNs [(E), (H), (K), (N)] from single
adoptive transfers [(C) to (E)] or cotransfers [(F) to (N)] at day 10 after
infection. Each point represents data from an individual mouse (n = 2 to 8),
and the combined dataset was collected over 4 separate days. Each sample
testing cotransferred retrogenic T cells was paired with individual transfers as
experimental controls.
(Fig. 4, C to E) or cotransfers (Fig. 4, B and I to
K). Finally, we adoptively transferred the low-
to-moderate avidity, canonical B17.C2+ T cells
and the high-affinity, reversed B17.R2+ T cells
into B6 mice, which were then challenged with
IAV. Retrogenic B17.C2+ T cells were readily
recovered from bronchoalveolar lavage (BAL)
(Fig. 4, B, C, and L), spleen (Fig. 4, D and M),
and mediastinal lymph nodes (mLNs) (Fig. 4, E
and N). By contrast, the B17.R2 TCR did not
support detectable immune expansion into any
tissue (Fig. 4, L to N). Thus, TCR–pMHCI dock-
ing topology supersedes TCR–pMHCI affinity
as the primary determinant for effective in vivo
immune recruitment.
Reversed-polarity TCRs do not prevent
TCR clustering
To determine whether the reversed TCR–
pMHCI docking prevents the formation of
signaling-competent multimers (10), 5KC
T cell hybridoma cells (TCRab−CD4−CD8−)
(24) expressing either the B13.C1 or B17.R2
TCRs were placed on a supported lipid bi-
layer (SLB) containing ICAM-1 (unstimulated)
or ICAM-1 and H-2Db–NP366 (stimulated) (Fig. 5,
A and B) for analysis of TCR clustering by
dSTORM (Fig. 5A). For both unstimulated
and stimulated T cells, TCRs exhibited a non-
random clustered spatial distribution on the
cell membrane, as indicated by a significantly
larger L(r)-r value relative to complete spatial
randomness (Fig. 5B). The peak of molecular
TCR clustering [max L(r)-r], was higher after
stimulation (ICAM + pMHCI) compared with
unstimulated T cells (ICAM), indicative of
antigen-driven TCR clustering (Fig. 5, B and C).
This antigen-driven TCR clustering was similar
for both the canonical (B13.C1) and reversed
(B17.R2) docking TCRs. Thus, reversed TCR–
pMHCI docking does not impede the forma-
tion of multimeric TCR–CD3 structures.
Reversed TCR recognition of H-2Db–NP366
impedes the localization of CD3 and CD8
Assuming that a similar arch-like structure is
formed after pMHCI recognition by the TCR/
CD3 and CD8 as has been observed for TCR–
pMHCII–CD4 (25), the canonical TCR–pMHC
docking polarity may be essential for coreceptor-
associated Lck to be situated proximally to CD3
for the initiation of signal transduction (25).
Using our current structural understanding
of the interactions between TCR–pMHCI (1),
MHCI, and CD8 (26) and TCRab and the CD3
chains (27, 28), we modeled the quaternary
TCR–pMHC–CD8–CD3 structure for canonical-
polarity B17.C1 TCR–H-2Db–NP366 (Fig. 5D)
and reversed-docking B17.R1 TCR–H-2Db–
NP366 interactions (Fig. 5E). The 180° re-
versal of the B17.R1 TCR over the H-2Db–NP366
substantially altered the position of CD8
relative to CD3.
To experimentally determine whether re-
versed TCR–pMHCI docking affected the local-
ization of CD8-associated Lck to the CD3
complex, we used fluorescence lifetime imag-
ing microscopy (FLIM)–Förster resonance
energy transfer (FRET) microscopy to deter-
mine the close (<10 nm) molecular associa-
tion between CD8b–mCherry and CD3z-GFP
fusion proteins in live TCR+ hybridoma cells
after epitope-specific stimulation (29–32) because
this was not readily feasible in primary T cells.
We used the FLIM–FRET approach over con-
ventional superresolution microscopy (20-
to 30-nm resolution) to allow us to resolve
protein–protein interactions (<10 nm).
Zareie et al., Science 372, eabe9124 (2021)
4 June 2021
5 of 13
RES EARCH | R E S E A R C H A R T I C L E
A
s
e
g
a
m
i
M
R
O
T
S
d
p
a
m
l
r
e
t
s
u
C
R
C
T
B
1200
1000
r
-
)
r
(
L
800
600
400
200
0
0
D
CD8
1C.31B
2R.71B
F
CTV (DC2.4)
-GFP
CD3
-mCherry
CD8
Composite
FLIM
1
C
.
3
1
B
1
R
.
7
1
B
2
R
.
7
1
B
1
C
.
7
1
B
2
C
.
7
1
B
G
5
0
-5
B13.C1
B17.R1
B17.R2
B17.C1
B17.C2
)
%
(
p
m
A
v
A
-10
-15
n
y
S
-20
NP366-374
***
***
***
- + - + - + - + - +
B13.C1
B17.R2
ICAM-1 + pMHCI
ICAM-1 + pMHCI
ICAM-1
Random distribution
ICAM-1
Random distribution
100
200
300
400
500
0
100
200
300
400
500
r (nm)
r (nm)
B13.C1
ns
B17.R2
ns
C
1200
1000
800
600
400
200
r
-
)
r
(
L
.
x
a
M
0
B13.C1
B17.R2
B13.C1
B17.R2
ICAM-1 ICAM-1 + pMHCI
B17.C1-H-2Db–NP366
E
B17.R1-H-2Db–NP366
2m
MHC-I
2m
MHC-I
TCR
CD3
CD3
CD3
CD8
TCR
CD3
CD3
CD3
Fig. 5. Reversed TCR–pMHCI docking does not impede TCR clustering but
mislocalizes the CD8 coreceptor and CD3 complex. (A) Single-molecule
images of 5KC TCR transductants expressing canonical docking B13.C1 or B17.R2
TCRs on supported lipid bilayers decorated with either ICAM-1 only (ICAM-1) or
ICAM-1 + H-2Db–NP366 (ICAM-1 + pMHC) at 5 min. Scale bar, 5 mm. Close-up view
of single-molecule localization microscopy image (2 × 2 mm) as TCR cluster maps
(lower panels) from representative regions (boxed, top panel), with TCRb
molecules in clusters shown in green and molecules outside clusters shown in blue.
Cluster contours are highlighted in red lines. (B) Ripley’s K analysis of TCR
clustering [L(r)-r] against radii (r). Complete spatial randomness is shown as a
solid gray line where L(r)-r = 0. Positive L(r)-r values indicate molecular clustering
relative to the random distribution, shown as the mean (solid line) ± 95%
confidence interval (dashed lines) for TCRs under each condition. Dotted lines
indicate ± SEM. (C) The maximum L(r)-r value derived from the peak of the graph
in (B) corresponds to the spatial scale (r) at which the highest degree of clustering
of localizations is being observed (n = 20). Statistical analysis performed by
one-way ANOVA. (D and E) Representation of the TCR–CD3 complex (PDB 6JXR),
CD8 coreceptor (PDB 3DMM), and ternary complex of B17.C1 TCR–H-2Db–NP366 (D)
and B17.R1 TCR–H2Db–NP366 (PDB 5SWZ) (E). The different chains are shown in
black for CD3zz, red for the CD3ge, pink for the CD3de, and blue for the TCRab
chains. The MHC is shown in white, and the CD8 coreceptor is shown in orange.
(F) DC2.4 cells labeled with CTV were pulsed with 10 mM NP366 peptide for 1 hour
before coculture with 5KCzGFP.CD8bmCherry–expressing TCRs as indicated. T cell
hybrids interacting with a DC2.4 cell were imaged up to 20 min after coculture
by confocal microscopy and subsequently analyzed by FLIM to measure GFP
lifetime 10 to 20 s later (fig. S3). Scale bar, 10 mm. (G) Amplitude weighted lifetime
of GFP (TavAMP) of B13.C1, B17.R1, B17.R2, B17.C1, and B17.C2 TCR+ T cells (±NP366
peptide) measured as percentage change at the synapse versus nonsynapse
(SynDTavAMP). Cells were observed on three different days. Statistical analysis was
performed by two-way ANOVA to examine the effect of stimulation and the day on
which each cell was observed on the SynDTavAMP for each cell line. Significant
effects on the SynDTavAMP by peptide stimulation are indicated by ***P ≤ 0.001 as
indicated. For each cell line, no significant effect on the SynDTavAMP was found for
the day each cell was observed.
Expression of the FRET pair constructs (fig. S3A)
did not negatively affect TCR signaling because
phosphorylated extracellular signal-regulated
kinase (pERK) could be detected after stimula-
tion of the B13.C1 TCR+ 5KCCD3zGFP.CD8abmCherry
cells, similar to B13.C1 TCR+ 5KC T cells ex-
pressing WT CD8ab and CD3z (Fig. 6F and fig.
S3, B to D). Stimulation of B13.C1 TCR+ cells
resulted in FRET, as measured by a substantial
reduction in the amplitude weighted lifetime of
the donor (GFP) at the synapse (fig. S3, E and
F), indicating colocalization of the CD8bmCherry
and CD3zGFP molecules (Fig. 5, F and G). We
also observed FRET after stimulation of cells
expressing the canonical B17.C1 and B17.C2
TCRs (Fig. 5, F and G). However, stimulation
of cells expressing the reversed B17.R1 and
B17.R2 TCRs resulted in negligible FRET (Fig. 5,
F and G). Thus, a reversed TCR–pMHCI dock-
ing topology results in improper localization
of CD8bmCherry and CD3zGFP in a manner that
Zareie et al., Science 372, eabe9124 (2021)
4 June 2021
6 of 13
F
180
160
140
120
100
4000
3000
e
d
u
t
i
n
g
a
m
l
i
a
n
g
s
K
R
E
p
)
C
U
A
(
2000
1000
0
RES EARCH | R E S E A R C H A R T I C L E
B
C
CD8NULL
CD8CxC
CD8WT
B13.C1
B17.R2
0
3
10
4
10
5
10
3
-10
0
3
10
4
10
5
10
3
-10
0
3
10
4
10
5
10
)
l
o
r
t
n
o
c
f
o
%
(
I
F
M
K
R
E
p
CD8WT
CD8NULL
G
180
160
140
120
100
H
180
160
140
120
100
CD8CxC
B13.C1
B17.R2
0
20
40
60
0
20
40
60
0
20
40
60
Min
A
100
80
60
40
20
0
3
-10
TCR
8
D
C
CD8
D
B13.C1
E
B17.R2
IP : CD8α
IB : Lck
IP : CD8α
IP : CD8α
WT
CD8
CD8
NULL
CxC
CD8
WT
CD8
CD8
NULL
CxC
CD8
)
m
i
t
s
b
A
f
o
%
(
]
2
-
L
I
[
**
4000
3000
2000
1000
0
4000
3000
2000
1000
0
Unstim
NP366
60
40
20
Unstim
NP366
60
Unstim
NP366
60
*
*
*
1
2 10
40
20
0
Unstim
40
20
0
Unstim
1
2 10
0
Unstim
1 2 10
[Peptide] ( M)
Fig. 6. Signaling constraints that mandate TCR–pMHCI docking orientation
are driven by Lck sequestration and localization by the CD8 coreceptor.
(A to C) Representative histograms of TCRb expression (top panel) and dot plots
of CD8ab expression (bottom panel) by 5KC TCR transductants expressing
B13.C1 (black) and B17.R2 (red). (D and E) Immunoblots of CD8-associated Lck
in CD8WT CD8NULL and CD8CxC transductants expressing B13.C1 (D) or B17.R2 (E)
TCRs. (F to H) Time-resolved pERK induction (top panels) up to 60 min after
coincubation with peptide-pulsed DC2.4 cells, presented as a percent change
from a no-peptide control (measured at corresponding time point) using cells
that express CD8WT (F), CD8NULL (G), and CD8CxC (H). Middle panel: pERK signal
magnitude analyzed by area under the curve (AUC) analysis of 0 to 60 min of
stimulation. Samples were tested in duplicate (n = 3), and the mean and SEM of
all datasets is shown. Statistical analyses were performed using a paired-samples
t test to pair by dataset. Statistical significance is indicated by **P ≤ 0.01 as
indicated. IL-2 secretion (bottom panels) into the supernatant by CD8WT (F),
CD8NULL (G), and CD8CxC (H) transductants was measured by enzyme-linked
immunosorbent assay (ELISA) after 16 hours of coincubation with peptide-pulsed
DC2.4 cells and is presented as a percentage of the plate-bound anti-CD3
antibody stimulation controls. Samples were tested in duplicate (n = 3), and the
mean and SEM of all datasets is shown. Statistical analyses were performed
using a paired-samples t test to pair by dataset. Statistical significance is
indicated by *P ≤ 0.05 as indicated.
is independent of the strength of TCR–pMHCI
binding.
The CD8 coreceptor inhibits TCR signaling by
reversed-polarity TCRs
We hypothesized that when TCR–pMHCI po-
larity is reversed, the association of Lck with
CD8 prevents, rather than promotes, effective
Lck localization to CD3. To test this, 5KC
T cells were transduced with either the high-
affinity canonical B13.C1 TCR or the high-
affinity reversed B17.R2 TCR (Fig. 6, A to C)
because neither of these TCRs is dependent
on CD8 for binding to H-2Db–NP366 (fig. S1,
B, E, and F). Each TCR was expressed (i) with
WT CD8ab (CD8WT) (Fig. 6A), (ii) in the ab-
sence of CD8 (CD8NULL) (Fig. 6B), or (iii) with
mutant CD8ab containing C227A and C229A
substitutions in the cytoplasmic tail of
CD8a to abrogate Lck binding (CD8CxC)
(33) (Fig. 6, C to E). All cell lines showed a
similar sensitivity to phorbol 12-myristate
13-acetate–ionomycin stimulation or anti-
body-mediated polyclonal stimulation, as
measured by pERK or interleukin-2 (IL-2)
production (fig. S4).
The CD8WT B13.C1 TCR+ transductants me-
diated robust signaling throughout the time
course, inducing a significantly higher pERK
signal magnitude and IL-2 secretion com-
pared with B17.R2 (Fig. 6F). For B13.C1 pERK
was induced as early as 10 min and main-
tained for 60 min after stimulation, and there
was substantial IL-2 production at 16 hours
(Fig. 6F). By contrast, the high-affinity re-
versed B17.R2 TCR showed negligible signal
transduction when coexpressed with CD8WT,
as evidenced by minimal pERK and no de-
tectable IL-2 (Fig. 6F). Whereas the loss of
Zareie et al., Science 372, eabe9124 (2021)
4 June 2021
7 of 13
RES EARCH | R E S E A R C H A R T I C L E
CD8 (CD8NULL) severely attenuated the sig-
nal transduction capacity of the B13.C1 TCR,
detectable, low-level pERK and IL-2 secre-
tion was evident after stimulation of B17.R2
TCR+ cells (Fig. 6G) and was statistically in-
distinguishable from B13.C1. To distinguish
the contribution of CD8 to MHCI binding
versus Lck delivery, we stimulated cells ex-
pressing the mutant CD8CxC. Again, both the
B13.C1 TCR+ and B17.R2 TCR+ cells transduced
a signal of similar kinetics and magnitude,
with increased pERK compared with that ob-
served in the absence of CD8 (Fig. 6H). This
signaling again corresponded to detectable
and equivalent levels of IL-2 production.
Similar findings were made upon analysis
of 5KCzGFP.CD8bmCherry T cells (fig S3, B to D)
and independently generated B13.C1, B17.R1,
and B17.R2 TCR+ cells (fig. S5A). These find-
ings were also in agreement with TCRb down-
regulation analysis after stimulation (fig. S5B).
Signaling mediated by canonical TCR–
pMHCI docking is augmented slightly by
CD8 binding and substantially by CD8 de-
livery of Lck. However, reversed-polarity
TCR–pMHCI recognition prevents signal-
ing caused by CD8 sequestration and mis-
localization of Lck. We present evidence that
the highly conserved TCR–pMHCI docking
polarity is mandated not by binding require-
ments, but instead by the need to colocal-
ize key signaling molecules to enable signal
transduction.
Discussion
Although a topic of much speculation, there has
been no definitive demonstration of whether
the canonical TCR–pMHC docking polarity po-
tentiates effective TCR binding or signaling, nor
of the mechanism by which it does so. Although
evolutionarily conserved pairwise interactions
between TCR and MHC molecules can pre-
dispose TCRs to MHC recognition in a canonical
orientation (5), our data revealed that the in-
ability of reversed TCR–pMHC recognition
to support T cell activation was unrelated
to binding affinity. Instead, these findings
support a paradigm in which the polarity
of TCR recognition of pMHCI is a primary
determinant of T cell signaling through the
colocalization of molecules critical for TCR
signal transduction.
Our data align with current knowledge of
the structural organization of TCR-signaling
molecules. The complete ternary structure of a
TCR–pMHCII–CD4 complex (25) revealed the
formation of a 70-Å wide “arch” between the
TCR and the CD4 coreceptor, within which
the asymmetrically arranged CD3 signaling
complex (28) is positioned. It was postulated
that extreme docking polarities (such as a
reversed polarity) would place the bulk of the
CD3 complex outside of the arch, impeding
optimal Lck delivery (25) and T cell signaling.
Our current work provides clear support for
this model by demonstrating that the inability
of reversed TCRs to signal (i) occurs indepen-
dently of binding strength, (ii) is dependent on
CD8 binding of Lck, and (iii) is characterized
by an inability to colocalize CD8 and CD3 after
antigen stimulation. Moreover, our demon-
stration of robust TCR–CD3 cluster formation
suggest that unusual docking topologies do not
preclude signaling by inhibiting TCR multi-
merization (10, 34), which is consistent with
previous work showing that dense TCR–CD3
cluster formation can occur independently
of signaling and better reflects TCR binding
(35). Our observation that the reversed TCR–
pMHCI interaction formed slip bonds and yet
remained inherently capable of signal trans-
duction, supports the notion that catch bonds
are not essential (although likely still opti-
mal) for signaling.
The observation that CD8 was an impedi-
ment to signaling by reversed TCR–pMHCI
recognition through binding of Lck demon-
strated that this polarity resulted in a mis-
localization of coreceptor-associated Lck and
CD3 after pMHCI ligation. Previous studies
have shown that preventing coreceptor seques-
tration of Lck in vivo, either by deletion of
coreceptors (36) or by mutation of corecep-
tor-binding sites on Lck (37), facilitates TCR
signaling after recognition of non-MHC ligands.
Thus, the association of Lck with the CD8
coreceptor, in addition to dictating the MHC
ligand (5), also dictates the manner in which
the MHC ligand must be recognized.
Other deviations from the typical TCR–
pMHC docking angle exist. Most notably, two
identified human induced regulatory T cell
TCRs were found to dock on pMHCII in a
reversed orientation (9). Although broadly
maintaining the canonical docking polar-
ity, extreme docking angles over the pMHC
have been observed in autoreactive human
CD4+ T cells (38, 39) and in a nonsignaling
mouse H-2Ld–restricted TCR (10). The cur-
rent study provides a potential mechanism by
which such unconventional pMHC-docking
polarities may diminish TCR signaling to
prevent negative selection or abrogate TCR-
mediated signaling. Such exceptions to the
canonical TCR–pMHC docking “rule” should
be explored to further advance our understand-
ing of T cell signaling requirements.
Although Lck association with coreceptors
can have a marked effect on TCR signaling,
some intracellular Lck untethered to corecep-
tors is present, highly active (40, 41), and able
to support in vivo signaling (30, 36, 37). Why,
then, does “free” Lck not allow for signaling by
reversed TCRs? Critically, Lck can be found
associated with the TCR–CD3 complex in mice
lacking CD4 and CD8 coreceptors but not in
mice expressing coreceptors (36). We propose
that when coreceptors are expressed, Lck is
preferentially sequestered away from the TCR–
CD3 complex to impair TCR signaling in the
absence of MHC ligands. In the presence of
MHC ligands, the coreceptor delivers Lck and
promotes MHC-restricted TCR signaling. This
may be exacerbated in the case of CD4, which
binds Lck with higher affinity than CD8 (42, 43).
Although small amounts of residual free Lck
may be able to initiate some early phosphory-
lation events (32), it is insufficient to support full
activation.
Although a useful tool for determining the
mechanism constraining signaling-competent
modalities of TCR recognition, the drivers of
reversed TCR docking in this instance are
unclear. It has been suggested that a reversed
TCR–pMHC orientation may be a consequence
of positively charged residues and/or proline
within the CDR3b loop, preventing interaction
with a conserved cluster of positively charged
residues on the CDR3b contact regions of
MHCI a2 helices (44). A comparison of the
CDR3b loops of naïve TRBV17+ (enriched for
reversed TCRs) and immune H-2Db–NP366–
specific repertoires (13) (enriched for canon-
ical TCRs) revealed a similar frequency of His,
Lys, Arg, and Pro usage. However, given that
the TRBV17 gene element encodes an Arg at
position 108, all but one of the naïve TRBV17+
TCRs contained at least one of these residues
but they were present in only ~40% of the
immune CDR3b sequences. Thus, the relevance
of these residues in driving noncanonical dock-
ing requires further investigation.
The inability of the reversed TCRs to sup-
port signaling would appear to preclude their
ability to support thymic selection. It is possi-
ble, given the reduced threshold for thymic
selection compared with peripheral activation
(45–48), that an attenuated signal may be
sufficient for positive selection. Alternatively,
such TCRs may mediate selection through
canonical TCR–pMHC recognition and exhibit
unconventional pMHC recognition only in the
periphery.
In summary, the current study demonstrates
a dual role for coreceptor association of Lck:
augmenting signaling mediated by canonical
TCR–pMHC interactions and preventing sig-
naling by unconventional modes of recogni-
tion. We hypothesize that, in this way, excessive
TCR cross-reactivity is constrained by the num-
ber of signaling-competent binding modalities,
thereby enhancing the exquisite functional spe-
cificity of the TCR–pMHC interaction.
Materials and Methods
TCR transfection of HEK293T cells
Human embryonic kidney (HEK) 293T cells
(ATCC, #CRL-3216) were maintained in a
humidified incubator at 37°C and 10% CO2.
HEK293T cells were plated at 3.5 × 105 cells/
well of a six-well plate in 3.5 ml of complete
medium [Dulbecco’s modified Eagle’s medium
Zareie et al., Science 372, eabe9124 (2021)
4 June 2021
8 of 13
RES EARCH | R E S E A R C H A R T I C L E
(DMEM; Invitrogen, #11960), 10% fetal bovine
serum (FBS), Hepes, L-glut, and penicillin-
streptomycin (PenStrep)]. The following day,
4.2 ml of FuGene 6 HD (Promega) was added to
171 ml of OptiMEM (Invitrogen) in an Eppendorf
tube and incubated for 10 min at room tem-
perature (RT). The FuGene:OptiMEM mixture
was then added dropwise to 700 ng of pMIGII
encoding an abTCR sequence and 700 ng of
pMIGII encoding CD3gde and z subunits and
incubated for a further 15 min at RT. The
FuGene-OptiMEM-DNA mixture was then
added dropwise to each well of a six-well plate
and swirled to mix gently before returning
to the incubator. After 48 hours, the culture
medium was aspirated and cells were de-
tached from the plate by repeated washing
with fluorescence-activated cell sorting (FACS)
buffer [phosphate-buffered saline (PBS) +
0.1% bovine serum albumin]. Transfected
cells were labeled with indicated tetramers for
1 hour at RT, followed by staining for TCRb
and viability. Tetramer binding was analyzed
by flow cytometry using a Fortessa X20 (BD
Biosciences).
Mice and influenza A virus infection
Female C57BL/6J (CD45.2) mice were bred
and housed at the Monash Animal Research
Platform (MARP; Monash University, Victoria,
Australia). B6.SJptprca (CD45.1) and Rag1−/−
(CD45.2) mice were purchased from the Walter
and Eliza Hall Institute and housed at MARP.
Naïve female C57BL/6J mice aged 6 to
10 weeks were briefly anesthetized by iso-
flurane inhalation and infected intranasally
with 1 × 104 plaque-forming units (PFUs) of
HKx31 (H3N2) influenza A virus in 30 ml of
saline. All animal experimentation was reviewed
and approved by the Monash University Animal
Ethics Committee (AEC8585, AEC14182, and
AEC17693).
Tetramer-based magnetic enrichment for
H-2Db–NP366–specific T cells
Tetramer-based magnetic enrichment of epitope-
specific T cells was performed largely as de-
scribed previously (49, 50). Spleens and all
easily dissected lymph nodes were harvested
from 10 naïve female C57BL/6J mice (pre-
immune repertoire) or individual mice 10 days
after infection with HKx31 IAV intranasally
(immune repertoire). One female C57BL/6J
mouse infected with HKx31 between 10 and
60 days after infection was also harvested to
be used as a positive control for tetramer
staining of naïve samples. For analysis of
the preimmune repertoire, single-cell sus-
pensions from 10 naïve mice were pooled
and then evenly divided into eight separate
50-ml conical centrifuge tubes, each for en-
richment through an LS column (Miltenyi
Biotec). For analysis of immune repertoires,
samples from individual mice were split evenly
into two matched 50-ml conical centrifuge
tubes. Each experimental sample was first
blocked using Fc block (2.4G2 supernatant
+ 1% normal mouse serum + 1% normal rat
serum). Each pair of tubes was labeled with
either H-2DbWT–NP366 or H-2DbA89E–NP366
tetramers conjugated to allophycocyanin (APC)
for 1 hour at RT. Tetramer-labeled cells were
then incubated with anti-APC microbeads
(Miltenyi Biotec) for 30 min at 4°C. Tetramer-
bound cells were positively enriched by passage
through an LS column (Miltenyi Biotec)
mounted on a QuadroMACS (Miltenyi Biotec)
magnetic separator. Enriched samples were
then labeled with antibodies (as listed in
table S5) against cell surface antigens, includ-
ing Vb9 (TRBV17) and Vb8.3 (TRBV13-1), and
stained for viability before analyzing en-
tire samples on a Fortessa X20 (BD Bio-
sciences) or Symphony A3 (BD Biosciences)
flow cytometer.
Single-cell TCR sequencing and T cell cloning of
immune TRBV17+ H-2Db–NP366–specific T cells
Tetramer-bound cells were enriched and iso-
lated as described above for immune reper-
toires. Samples enriched for tetramer-bound
cells were run on a FACSAria III (BD Bio-
sciences) cell sorter, and live CD19−CD4−
CD8+TCRb+CD44+Vb9+ H-2Db–NP366–specific
T cells were single-cell sorted using a FACS
AriaIII Fusion (BD Biosciences) into 96-well
polymerase chain reaction (PCR) plates (Eppendorf)
and stored at −80°C until use. Single-cell multi-
plex reverse transcription (RT)–PCR of abTCR
was performed as previously described (13).
PCR product was sequenced by Sanger se-
quencing at the Monash Micromon Genomics
Facility (Monash University, Victoria, Australia).
Antigen-specific, P2A-linked TCRab gene con-
structs were custom ordered from Genscript
and cloned into pMIGII (RRID: Addgene_52107;
a gift from D.A.A. Vignali) vector and sequenced
to confirm the correct TCR sequence. Plasmids
encoding antigen specific P2A-linked TCRab
were prepared and propagated for retroviral
transduction using 10-beta competent Escherichia
coli (E. coli) (New England Biolabs, #CR019H),
and plasmids were isolated using the EndoFree
Maxi Prep Kit (Qiagen, #12362).
Generation of TCR-retrogenic mice
Plasmids encoding TCRa and TCRb genes of
interest linked by the P2A peptide were or-
dered from Genscript and cloned into the
pMIGII or pMIC vector expressing GFP or
mCherry, respectively (RRID: Addgene_52107,
RRID: Addgene_52114; a gift from D.A.A.
Vignali). TCR-retrogenic mice were generated
as previously described (12, 23) but with the use
of congenically distinct female Rag1−/− (CD45.2)
mice as bone marrow donors and female
B6.SJptprca (CD45.1) as recipients to aid the
identification of donor-derived cells.
Adoptive transfer of retrogenic CD8+ T cells for
IAV challenge
CD45.1−CD45.2+CD4−CD8+ GFP/mCherry+
T cells were isolated by FACS from female
TCR-retrogenic mice using a BD FACSAria III
Fusion or BD Influx cell sorter (BD Bioscien-
ces). Retrogenic T cells (4 × 103) from each
line were resuspended in 200 ml of PBS + 2%
FBS and injected intravenously into naïve
female B6.SJptprca (CD45.1) mice. The next
day, mice were infected with 1 × 104 PFUs of
HKx31 IAV as described above. At the peak of
the CD8 T cell response (10 days after infection),
mice were euthanized and BAL, spleen, and
mLNs were harvested for flow cytometry analy-
sis. The gating strategy for identifying donor-
derived retrogenic T cells is described in fig. S3.
In vitro TCR expression by
retroviral transduction
TCRnull 5KC cells (TCRab−CD4−CD8−) (gift from
P. Marrack) were maintained in a humidified
incubator at 37°C and 10% CO2. 5KC cells were
sorted for loss of CD4 to establish a CD4−CD8−
TCR− cell line. HEK293T cells were plated 1 ×
106 cells/dish in a 15-cm tissue culture dish
(Sarstedt) in 10 ml of complete DMEM (cDMEM)
containing DMEM, 10% FBS, Hepes, L-glut, and
PenStrep). The following day, 30 ml of FuGene
6 HD (Promega, #E2691) was added to 470 ml
of OptiMEM (Invitrogen, #31985) in a micro-
centrifuge tube and incubated for 10 min at
RT. The FuGene:OptiMEM mixture was then
added dropwise to 4 mg of pMIGII encoding
an abTCR sequence, along with 4 mg of pPAM-E
and 2 mg of pVSVg, and incubated for a further
15 min at RT. The FuGene-OptiMEM-DNA
mixture was then added dropwise to the
HEK293T cell culture and swirled to mix
gently before returning to the incubator.
The next day, medium containing FuGene:
OptiMEM:DNA was replaced with fresh cDMEM
and incubated for 12 hours. Supernatant was
removed approximately every 12 hours five to six
times and filtered through a 0.45-mm syringe-
driven filter. Polybrene (Sigma-Aldrich, #H9268)
(6 mg/ml) was added to the supernatant before
resuspending 5KC cells in filtered, retrovirus-
containing supernatant. After five to six virus
transfers, 5KC cells were grown to confluency
in fresh cDMEM and sorted for similar TCRab
and CD8ab expression. For CD8 transductions,
only cells expressing an endogenous ratio of
CD8a:CD8b were sorted.
pERK detection by phospho-flow cytometry
For a positive control, 96-well U-bottom plates
were coated overnight with 100 ml of anti-
mouse CD3 antibody diluted to 10 mg/ml in
PBS overnight. DC2.4 cells (gift from K. Rock)
were cultured in cDMEM and maintained in a
humidified incubator at 37°C and 10% CO2.
DC2.4 cells were stained with Aqua Blue Fixable
viability stain (Life Technologies), seeded at 1 ×
Zareie et al., Science 372, eabe9124 (2021)
4 June 2021
9 of 13
RES EARCH | R E S E A R C H A R T I C L E
105 cells/well in 50 ml of a 96-well plate (Nunc),
and allowed to adhere for 1 hour in the incu-
bator. Transduced 5KC cell lines were labeled
with Aqua Blue Fixable viability stain (Life Tech-
nologies), plated at 50,000 cells/well, and allowed
to equilibrate in the incubator for at 1 hour. NP366
peptide (Genscript) was then added to the DC2.4
cultures at the indicated concentrations and
incubated for a further 1 hour. 5KC cells (5 ×
104) were added to peptide-pulsed DC2.4 cul-
tures at a final culture volume of 100 ml, and
briefly centrifuged at 600g for 1 min to en-
courage contact. At the indicated time points,
100 ml of prewarmed, 2× Lyse/Fix buffer (BD
Biosciences, #558049) was added to each well,
and the mixture was incubated at 37°C for
10 min. Fixed cells were washed twice in 200 ml
of PBS. Fixed cell pellets were permeabilized by
the addition of 100 ml of −20°C Perm Buffer III
(BD Biosciences) and then stored overnight at
−20°C. The following day, fixed and permeabi-
lized cells were washed with 200 ml of FACS
buffer and stained with cell surface anti-
bodies and a rabbit anti-phospho p44 MAPK
(Cell Signaling Technology) for 1 hour on ice.
Cells were washed with FACS buffer and then
stained with an anti-rabbit PE F(ab′)2 frag-
ment (Cell Signaling Technology) for 30 min.
Cells were washed twice in FACS buffer
before running the samples on a BD Fortessa
X20 or BD Symphony A3 (BD Biosciences) flow
cytometer.
IL-2 ELISA
DC2.4 cells were seeded at 1 × 105 cells/well
in 100 ml of a flat-bottom 96-well plate (Nunc)
and allowed to adhere for 1 hour in the in-
cubator. Transduced 5KC cell lines were plated
at a concentration of 1 × 106 cells/ml and al-
lowed to equilibrate in the incubator for at
least 1 hour. NP366 peptide was then added to
the DC2.4 cultures at the indicated concen-
trations and incubated for a further 1 hour.
5KC cells (5 × 104) were added to peptide-
pulsed DC2.4 cultures or to the anti-CD3–coated
wells at a final culture volume of 200 ml and
briefly centrifuged at 600g for 1 min to en-
courage contact. After 16 hours of coincuba-
tion, plates were centrifuged at 935g for 3 min
to pellet cells and supernatant was aspirated,
transferred to a new plate, and stored at −20°C
until required. IL-2 secreted in the supernatant
was measured using the BD IL-2 mouse ELISA
kit (BD Biosciences, #555148) according to the
manufacturer’s instructions. Absorbance was
measured at 450 nm using a CLARIOstar plate
reader (BMG LabTech).
Immunoprecipitation and immunoblotting
5KC T cells (1.5 to 2 × 107) were lysed for
60 min at 4°C in 300 ml of Pierce IP Lysis/Wash
Buffer (25 mM Tris-HCl pH 7.4, 150 mM NaCl,
1% NP-40, 1 mM EDTA, and 5% glycerol) and
1× Halt Protease Inhibitor Cocktail (Pierce).
Lysates were centrifuged at 13,000g for 20 min,
and 30 ml of supernatant was kept as whole-cell
lysate/input. The remaining lysate was pre-
cleared with 20 ml washed Protein G Sepharose
for 60 min at 4°C. Protein G Sepharose (20 ml)
was incubated with 10 mg of anti-CD8a (53-6.7)
for 60 min at 4°C. Antibody-conjugated Protein
G Sepharose was washed three times with
Pierce® IP Lysis/Wash Buffer. Immunoprecipi-
tation was performed by addition of precleared
lysate to antibody-conjugated Protein G Sepharose
and incubation for 3 hours at 4°C. Samples were
centrifuged, washed five times in Pierce IP Lysis/
Wash Buffer and once in 50 mM Tris-HCl (pH
7.4). Elution was performed by boiling in 3×
Laemmli buffer with 50 mM dithiothreitol
(95°C, 5 min). Samples were centrifuged and
supernatant containing immunoprecipitated
CD8 was collected. Samples were resolved on
10 to 14% SDS–polyacrylamide gel electropho-
resis under reducing conditions at 100 V for 3 to
4 hours. Proteins were wet transferred onto
polyvinylidene difluoride membranes at con-
stant 300 mA for 2 hours. Membranes were
blocked for 60 min at RT in 5% skim milk (w/v)
in 0.1% (v/v) TBS-Tween before incubation
with specific antibody (1:3000) overnight at
4°C. Membranes were washed three times in
0.1% (v/v) TBS-Tween and probed with rel-
evant horseradish peroxidase–conjugated sec-
ondary antibody as indicated for 60 min at RT.
After three washes in 0.1% (v/v) TBS-Tween,
enhanced chemiluminesence substrate was
added to membranes for 2 min. Blots were
visualized on a ChemiDoc XRS+ (Bio-Rad
Laboratories).
Confocal and FLIM for analysis of FRET
DC2.4 cells were maintained in phenol-free
cDMEM (Invitrogen 31053), seeded in 35-mm
Fluorodish (World Precision Instruments) cell
culture dishes at 1 × 105 cells/dish in 1 ml of
culture, and incubated overnight at 37°C and
10% CO2. The following day, DC2.4 cells were
washed three times with prewarmed PBS and
then stained with 5 mM CellTrace Violet (CTV;
Invitrogen C34557) for 30 min. CTV was then
aspirated from the dish and the labeled cells
were washed three times with prewarmed,
phenol-free cDMEM. Labeled DC2.4 cells were
incubated with 10 mM NP366 peptide for 1 hour.
5KC hybridoma cells expressing TCR and the
FRET pairs CD3z-GFP and CD8b-mCherry
were plated at a density of 1 × 106 cells/ml in a
six-well dish and equilibrated in the incuba-
tor for at least 1 hour before use. For imaging,
DC2.4 cells were brought into focus and then
100 ml of T cells (~1 × 105 cells) was added to
the culture dish containing labeled DC2.4 cells
and imaged by confocal microscopy up to
20 min later using an Olympus FV1000 run-
ning Fluoview software (Olympus). The fluo-
rescent lifetime of the donor molecule GFP was
measured by time-correlated single-photon
counting using a PicoHarp 300 (PicoQuant)
running Symphotime 64 (PicoQuant) fitted
to an Olympus FV1000 laser scanning con-
focal microscope. A 485-nm pulsed laser was
used. Time-correlated single-photon counting
decay curves were fitted to a biexponential
reconvolution decay model in SymphoTime
64 to determine donor (GFP) lifetime. A good
biexponential reconvolution decay curve fit
was characterized by c2 values close to 1. c2
values that deviated by ±1 were uncommon in
our dataset but were excluded from the analysis.
To detect FRET, the amplitude weighted donor
average lifetime (tAvAmp) was used because this
reflects the quenching of the donor caused by
the FRET process. To determine FRET between
CD8b-mCherry and CD3z-GFP, we measured
tAvAmp at the immunological synapse where the
T cell interacted with a dendritic cell by selecting
a region of interest in the FLIM image. Non-
synaptic tAvAmp was also measured from the
same cell as an internal control and was used to
calculate the percentage change in tAvAmp of
GFP at the synapse (SynDtAvAmp). For analysis,
n-exponential reconvolution using the n = 2
model parameter was used for donor curve
fitting.
2D micropipette adhesion frequency
assay (2D-MP)
The relative 2D affinity of the H-2Db–restricted
nucleoprotein epitope (NP366; ASNENMETM)
TCRs expressed in 5KC hybridoma cell lines was
measured by the previously described 2D-MP
(14–18). In short, human red blood cells (hRBCs)
coated with Biotin-LC-NHS (BioVision) strepta-
vidin (Thermo Fisher Scientific), coated with
biotinylated H-2Db–NP366
D227K, and mounted
on a glass micropipette. 5KC hybridomas ex-
pressing B13.C1, B17.R1, B17.R2, or B17.C1 TCRs
were mounted on opposing glass micropipettes.
The adhesion frequency between the TCR of
interest and pMHC aspirated on opposing
micropipettes was observed using an inverted
microscope. An electronically controlled piezo-
electric actuator brought the opposing cells into
contact and repeated a T cell contact and
separation cycle with the pMHC-coated RBCs
50 times while keeping the contact area (Ac)
and time (t) constant. Upon retraction of the
T cell, adhesion (binding of TCR–pMHC) was
observed as a distention of the RBC membrane,
allowing for quantification of the adhesion
frequency (Pa) at equilibrium. Surface pMHC
(ml) and TCRb (mr) densities were determined
by flow cytometry using an anti-TCRb PE anti-
body (BD Biosciences, #H57-597) and an anti-
H2Db antibody (clone 28-14-8; eBioscience),
both at saturating concentrations, along with
BD QuantiBRITE PE beads for standardization
(BD Biosciences). The calculation of molecules
per area was determined by dividing the num-
ber of TCRs and pMHCs per cell by the respective
surface areas. The relative 2D affinities were
Zareie et al., Science 372, eabe9124 (2021)
4 June 2021
10 of 13
RES EARCH | R E S E A R C H A R T I C L E
calculated using the following equation: AcKa =
–ln[1 – Pa(1)]/mrml.
Biomembrane force probe assay
Bond lifetime measurements under force were
captured using the biomembrane force probe
(BFP) assay. Procedures for coupling pMHC to
glass beads have been described previously
(21). Briefly, hRBCs were first biotinylated
with EZ-link NHS-PEG-Biotin (Thermo Fisher
Scientific) and then reacted to streptavidin.
Borosilicate beads were cleaned, silanized, and
reacted to streptavidin–maleimide (Sigma-
Aldrich, St. Louis, MO). Streptavidin beads
were then coated with H-2Db–NP366 or H-
2Db–D227K-NP366. pMHC monomer–coated
beads (which serve as a force probe) were
then placed on the apex of an hRBC that was
aspirated onto a glass micropipette. The posi-
tion of the edge of the bead was tracked by a
high-resolution camera (1600 frames/s) with
< 3-nm displacement precision. The position
of the edge of the bead was tracked by a high-
resolution camera (1600 frames/s) with <3 nm
displacement precision using a Zeiss micro-
scope. The cell was brought into contact with
the pMHC-coated bead:RBC; the cell was then
retracted and held at the desired force by the
computer-controlled piezoelectric actuator until
bond dissociation occurred. If adhesion was
present, it was detected by tensile force caused
by stretching of the hRBC and tracked by
displacement of the pMHC-coated bead. The
bond lifetime was measured from the time the
desired force was reached to the time it took
the cell to disengage with the bead, which was
visualized as the RBC retracted and the bead
returned to its starting position before the
start of the next cycle. Repeated cycles (known
as force-clamp cycles) could be performed.
Multiple forces were collected for each ligand
(pMHC-coated beads) and were shown in five
to eight bins as the mean ± SEM. For an op-
timal response to antigen, bond lifetimes in-
crease with increasing force before reaching a
peak and then decrease, which is referred to as
a catch bond. When increasing force leads
to decreasing bond lifetime, this is called a
slip bond.
SLB preparation
Glass coverslips of a 0.17 mm thickness were
first cleaned with 1 M KOH for 10 min and
then rinsed with Milli-Q water and placed in
100% ethanol for 20 min. These glass cover-
slips were then plasma cleaned for 5 min.
Afterward, the coverslips were adhered to
eight-well silicon chambers (ibidi, #80841). A
liposome solution of 1 mg/ml with a lipid
ratio of 96.5% DOPC (1,2-dioleoyl-sn-glycero-3-
phosphocholine), 2% DGS-NTA(Ni) (1,2-dioleoyl-sn-
glycero-3-[(N-(5-amino-1-carboxypentyl)iminodiacetic
acid)succinyl] (nickel salt)), 1% Biotinyl-Cap-PE
(1,2-dioleoyl-sn-glycero-3-phosphoethanolamine-
N-(cap biotinyl) (sodium salt)), and 0.5%
PEG5000-PE (1,2-distearoyl-sn-glycero-3-
phosphoethanolamine-N-[methoxy(poly-
ethylene glycol)-5000] (ammonium salt)
(mol%; all available from Avanti Polar Lipids
(DOPC, 850375C), (DGS-NTA(Ni), 790404C),
(Biotinyl-Cap-PE, 870273C), (PEG5000-PE,
880220C) was created by vesicle extrusion, as
described previously (35). Extruded liposomes
were added to eight-well chambers at a ratio
of 1:5 with Milli-Q water (with 10 mM of
CaCl2) and incubated for 30 min at RT before
gently rinsing with PBS repeatedly. During
washing steps, the disruption of the lipid
bilayer was minimized by retaining ~200 ml
of PBS in each well. Lateral mobility of the
freshly prepared SLB was confirmed by adding
10 mg/ml of fluorescently labeled streptavidin
(Invitrogen, #S11223) and monitoring fluores-
cence recovery after photobleaching (FRAP) as
described elsewhere (35). Excess Ca2+ ions in
the system were removed by adding 0.5 mM of
EDTA, followed by gently rinsing with Milli-Q
water. The NTA groups in DGS-NTA(Ni) lip-
ids were then recharged by adding 1 mM of
NiCl2 for 15 min and gently rinsed with PBS
repeatedly. Finally, SLB was blocked with
5% bovine serum albumin in PBS for 15 min
at RT.
T cell activation on SLB
To decorate the SLB with biotinylated pMHC
and His-tagged ICAM-1, 100 mg/ml of strepta-
vidin (Life Technologies) was incubated for
20 min and rinsed with PBS. Afterward, 500 ng/ml
of biotinylated H-2Db–NP366 (pMHCI) (12) and
200 ng/ml of His-tagged ICAM-1 (Sino Biological)
prepared in 5% bovine serum albumin in PBS was
added to the lipid bilayer and incubated for 1 hour
at RT. SLB was gently rinsed with PBS for several
times to remove excess unbound proteins. Before
adding T cell hybridomas, SLB was incubated with
warm DMEM culture medium (37°C) for 30 min.
T cell hybridomas were then allowed to activate on
the lipid bilayer for 5 min at 37°C, followed by
immediate cell fixation with 4% paraformaldehyde
(vol/vol) in PBS for 15 min at RT. Excess fixatives
were removed by rinsing with PBS.
Immunostaining of 5KC T cells
Before immunostaining 5KC T cells were per-
meabilized with 0.1% Triton X-100 (vol/vol)
(Sigma-Aldrich) for 15 min and then rinsed
with PBS. Cells were then blocked with 5%
bovine serum albumin in PBS for 1 hour. T cells
were stained with primary antibodies reactive
against the TCRb subunit and conjugated to
Alexa Fluor 647 fluorophore (BioLegend). Cells
were probed with primary antibodies for 1 hour
at RT. After antibody staining, samples were
repeatedly rinsed with PBS to remove excess
unbound antibodies and fluorophores. Postfixa-
tion was performed using 4% paraformaldehyde
(vol/vol) in PBS for 15 min. Before imaging, 0.1-mm
TetraSpeck microspheres (Invitrogen) were
embedded onto the lipid bilayer.
Single-molecule localization microscopy
with dSTORM
For single-molecule imaging, an imaging buf-
fer consisting of TN buffer [50 mM Tris-HCl
pH 8.0, 10 mM NaCl), oxygen scavenger system
GLOX [0.5 mg/ml glucose oxidase (Sigma-
Aldrich, #G2133); 40 mg/ml catalase (Sigma-
Aldrich, #C-100); and 10% w/v glucose], and
10 mM 2-aminoethanethiol (MEA; Sigma-
Aldrich, #M6500) was used. dSTORM image
sequences were acquired on a total internal
reflection fluorescence microscope (commer-
cial setup, Zeiss Elyra) with a 100× oil-immersion
objective (numerical aperture = 1.46). For Alexa
Fluor 647, 633 nm (15 mW) laser illumination
was used, along with a 405-nm activator laser
(15 mW) for imaging. Time series of 10,000 frames
were acquired per sample with a cooled, electron-
multiplying charge-coupled device camera (iXon
DU-897D; Andor) with an exposure time of
50 ms. Image processing, including fiducial
marker–based drift correction and generation
of x–y particle coordinates for each molecule
detected in the acquisition, were performed
using Zeiss Zen software (Zen Black 2012
version).
Expression, refolding, purification,
crystallization, and structure determination
DNA fragments encoding the TRAV and TRBV
segments of the B17.C1 TCR were purchased,
codon optimized (Genscript), amplified, and
cloned separately into a previously reported
expression vector fused to human Ca and Cb,
respectively (12). The B17.R2 TCR was gener-
ated by mutagenesis of the B17.R1 TCR. The
B13.C1 TCRab construct was purchased, codon
optimized for mammalian cell expression
(Genscript), and cloned into a pHLsec vector.
Each TCR is a chimeric construct of mouse
variable and human constant domain. The
plasmids constructs were confirmed by DNA
sequencing. The B13.C1 TCR was produced in
HEK293S cells as a soluble protein and pu-
rified through its His tag over an affinity
column and size exclusion chromatography.
Soluble H-2Db WT or mutant heavy chain
(generated by site-directed mutagenesis), the
human and mouse b2m, the B17.R1, B17.R2,
and the B17.C1 TCRs a and b chains were ex-
pressed separately in E. coli (Novagen, #70236)
as inclusion bodies, then subsequently solubi-
lized, refolded, and purified as previously
reported (12).
Crystals of the ternary B17.R2–H-2Db–NP366
complex were grown by the hanging-drop,
vapor-diffusion method at 20°C with a protein/
reservoir drop ratio of 1:1 at 3 mg/ml in 10 mM
Tris-HCl, pH 8.0, 150 mM NaCl using 20% PEG
3350, 0.2 M K/Na/tartrate, and 0.1 M Bis-tris-
propane buffer, pH 6.5. Crystals of the ternary
Zareie et al., Science 372, eabe9124 (2021)
4 June 2021
11 of 13
RES EARCH | R E S E A R C H A R T I C L E
B17.C1–H-2Db–NP366 complex were grown by
the hanging-drop, vapor-diffusion method at
20°C with a protein/reservoir drop ratio of 1:1
at 8 mg/ml in 10 mM Tris-HCl, pH 8.0, 150 mM
NaCl using 16% PEG 3350, 0.2 M potassium
thiocyanide, 4% ethylene glycol, and 0.1 M Bis-
tris-propane buffer, pH 7.6. The crystals were
soaked in a cryoprotectant solution containing
the mother liquor solution with the PEG con-
centration increased to 30% (w/v) and then
flash-frozen in liquid nitrogen.
For the B17.C1–H-2Db–NP366 structure, de-
spite successfully reproducing and testing nu-
merous crystals, only a single crystal gave a
diffraction at high resolution (i.e., <5 Å) and
was rapidly destroyed by radiation damage.
Datasets were collected on the MX1 (51) and
MX2 (52) beamline at the Australian Synchro-
tron (Clayton, Victoria, Australia), processed
using XDS software (53), and scaled using
Aimless software (54) from the CCP4 suite
(55). The data cutoff used was CC1/2 > 0.5%
and I/s(I) > 1.5 (56). The structures were
determined by molecular replacement using
the PHASER program (57), with the B17.R1
TCR from the previous B17.R1–H-2Db–NP366
complex used as the search model for the
TCR [PDB accession code 5SWZ (12)]. Manual
model building was conducted using Coot soft-
ware (58), followed by maximum-likelihood
refinement with the Buster program (59).
The final model has been validated using the
PDB validation website, and the final refine-
ment statistics are summarized in table S1.
The electron density at the interface was well
defined despite slightly above average R factors.
The high R factors are caused by poor electron
density for some parts of H2Db a3 domain, the
b2m, as well as the C terminus of TCRa con-
stant domain. However, these regions are distal
from the ligand interface. All molecular graphics
representations were created using PyMol (The
PyMOL Molecular Graphics System, version 2.0;
Schrödinger, LLC). The structures have been
deposited into the PDB database (B17.R2 TCR–
H-2Db–NP366, PDB 7JWI; B17.C1 TCR–H-2Db–
NP366, PDB 7JWJ).
SPR experiments
SPR experiments were conducted at 25°C on
the BIAcore T200 and BIAcore 3000 instru-
ment (GE Healthcare, Buckinghamshire, UK)
with 10 mM Tris-HCl, pH 8, 150 mM NaCl,
0.005% surfactant P20, and 0.5% bovine serum
albumin buffer. The 12H8 antibody was bound
to all flow cells of a CM5 sensor chip through
amine coupling (60), and all TCRs subsequently
bound to the antibody. A negative control (LC13
TCR) (61) was used on each SPR chip bound to
flow cell 1. Each cycle of TCR injection and
pMHC injection was regenerated with Actisep
(Sterogene). pMHC was flowed over the surface
with a concentration range of 0.78 to 200 mM at
a flow rate of 5 or 30 ml/min. A minimum of two
independent experiments were conducted (n =
2) in duplicate. GraphPad Prism software was
used for data analysis with the 1:1 Langmuir
binding model.
Statistical analyses
Statistical analysis was performed with one-way
ANOVA or two-way ANOVA when comparing
multiple groups as indicated. For data obtained
over multiple days, we considered the possibility
of day-to-day variation, which we accounted for
as a “nuisance factor” in the two-way ANOVA.
We did not find a statistically significant effect
of day-to-day variation in our analysis. Where
appropriate, we also performed paired-samples
t tests, as indicated in the figure legends, which
pair the dataset by the day in which they were
obtained. In the figures, P-values are denoted as
*P ≤ 0.05, **P ≤ 0.01, ***P ≤ 0.001, and ****P ≤
0.0001. Ripley’s K analysis was used to quantify
the degree of clustering in a population of
molecules compared with a complete spatial
randomness (62). For each molecule registered
as a localization event, Ripley’s (K) calculates
the number of neighboring localizations with-
in a given radius (r) corrected by the total
density of localizations, providing information
on the degree of spatial clustering of molecules
within a region of interest. In this study, we
performed Ripley’s K analysis on single-molecule
images using a previously published algorithm
(27). To determine the average clustering value
within a ROI, this algorithm used a linearized
form of Ripley’s (K), the L(r)−r. Here, r is defined
as the spatial scale radius. In a complete spatial
randomness, L(r)−r value = 0. Similarly, a posi-
tive value for L(r)−r at a given r radius indicated
a clustering of localization events, whereas a
negative value represented a dispersed spatial
organization (negative clustering). The start
(0 nm), end (500 nm), and step size (10 nm)
for r in the algorithm were user defined. The
maximum L(r)−r value derived from L(r)−r
versus r graph corresponded to the spatial
scale (r) at which the highest degree of clus-
tering of localizations was observed.
RE FERENCES AND NOTES
1.
J. Rossjohn et al., T cell antigen receptor recognition of
antigen-presenting molecules. Annu. Rev. Immunol. 33,
169–200 (2015). doi: 10.1146/annurev-immunol-032414-
112334; pmid: 25493333
2. S. Dai et al., Crossreactive T cells spotlight the germline rules
for ab T cell-receptor interactions with MHC molecules.
Immunity 28, 324–334 (2008). doi: 10.1016/
j.immuni.2008.01.008; pmid: 18308592
3. D. Feng, C. J. Bond, L. K. Ely, J. Maynard, K. C. Garcia,
Structural evidence for a germline-encoded T cell receptor-
major histocompatibility complex interaction ‘codon’.
Nat. Immunol. 8, 975–983 (2007). doi: 10.1038/ni1502;
pmid: 17694060
4. B. D. Stadinski et al., A role for differential variable gene pairing
in creating T cell receptors specific for unique major
histocompatibility ligands. Immunity 35, 694–704 (2011).
doi: 10.1016/j.immuni.2011.10.012; pmid: 22101158
5. F. Van Laethem, A. N. Tikhonova, A. Singer, MHC restriction is
imposed on a diverse T cell receptor repertoire by CD4 and
CD8 co-receptors during thymic selection. Trends Immunol.
33, 437–441 (2012). doi: 10.1016/j.it.2012.05.006;
pmid: 22771139
6. S. Rangarajan, R. A. Mariuzza, T cell receptor bias for MHC: Co-
evolution or co-receptors? Cell. Mol. Life Sci. 71, 3059–3068
(2014). doi: 10.1007/s00018-014-1600-9; pmid: 24633202
7. N. L. La Gruta, S. Gras, S. R. Daley, P. G. Thomas, J. Rossjohn,
Understanding the drivers of MHC restriction of T cell
receptors. Nat. Rev. Immunol. 18, 467–478 (2018).
doi: 10.1038/s41577-018-0007-5; pmid: 29636542
8. P. Zareie, C. Farenc, N. L. La Gruta, MHC restriction: Where are
we now? Viral Immunol. 33, 179–187 (2020). doi: 10.1089/
vim.2019.0195; pmid: 32286180
9. D. X. Beringer et al., T cell receptor reversed polarity
recognition of a self-antigen major histocompatibility complex.
Nat. Immunol. 16, 1153–1161 (2015). doi: 10.1038/ni.3271;
pmid: 26437244
10. J. J. Adams et al., T cell receptor signaling is limited by docking
geometry to peptide-major histocompatibility complex.
Immunity 35, 681–693 (2011). doi: 10.1016/j.
immuni.2011.09.013; pmid: 22101157
11. B. C. Sim, L. Zerva, M. I. Greene, N. R. J. Gascoigne, Control of
MHC restriction by TCR Valpha CDR1 and CDR2. Science 273,
963–966 (1996). doi: 10.1126/science.273.5277.963;
pmid: 8688082
12. S. Gras et al., Reversed T cell receptor docking on a major
histocompatibility class I complex limits involvement in the
immune response. Immunity 45, 749–760 (2016).
doi: 10.1016/j.immuni.2016.09.007; pmid: 27717799
13. T. Cukalac et al., Paired TCRab analysis of virus-specific CD8
(+) T cells exposes diversity in a previously defined ‘narrow’
repertoire. Immunol. Cell Biol. 93, 804–814 (2015).
doi: 10.1038/icb.2015.44; pmid: 25804828
14. J. Huang et al., The kinetics of two-dimensional TCR and
pMHC interactions determine T-cell responsiveness.
Nature 464, 932–936 (2010). doi: 10.1038/nature08944;
pmid: 20357766
15. J. J. Sabatino Jr.., J. Huang, C. Zhu, B. D. Evavold, High
prevalence of low affinity peptide-MHC II tetramer-negative
effectors during polyclonal CD4+ T cell responses. J. Exp. Med.
208, 81–90 (2011). doi: 10.1084/jem.20101574;
pmid: 21220453
16. R. J. Martinez, R. Andargachew, H. A. Martinez, B. D. Evavold,
Low-affinity CD4+ T cells are major responders in the primary
immune response. Nat. Commun. 7, 13848 (2016).
doi: 10.1038/ncomms13848; pmid: 27976744
17. E. M. Kolawole, R. Andargachew, B. Liu, J. R. Jacobs,
B. D. Evavold, 2D kinetic analysis of TCR and CD8 coreceptor
for LCMV GP33 epitopes. Front. Immunol. 9, 2348 (2018).
doi: 10.3389/fimmu.2018.02348; pmid: 30374353
18. R. Andargachew, R. J. Martinez, E. M. Kolawole, B. D. Evavold,
CD4 T cell affinity diversity is equally maintained during acute
and chronic infection. J. Immunol. 201, 19–30 (2018).
doi: 10.4049/jimmunol.1800295; pmid: 29777029
19. T. A. Potter, T. V. Rajan, R. F. Dick 2nd, J. A. Bluestone,
Substitution at residue 227 of H-2 class I molecules abrogates
recognition by CD8-dependent, but not CD8-independent,
cytotoxic T lymphocytes. Nature 337, 73–75 (1989).
doi: 10.1038/337073a0; pmid: 2462676
20. L. Limozin et al., TCR-pMHC kinetics under force in a cell-free
system show no intrinsic catch bond, but a minimal encounter
duration before binding. Proc. Natl. Acad. Sci. U.S.A. 116,
16943–16948 (2019). doi: 10.1073/pnas.1902141116;
pmid: 31315981
21. B. Liu, W. Chen, B. D. Evavold, C. Zhu, Accumulation of
dynamic catch bonds between TCR and agonist peptide-MHC
triggers T cell signaling. Cell 157, 357–368 (2014).
doi: 10.1016/j.cell.2014.02.053; pmid: 24725404
22. J. Hong et al., Force-regulated in situ TCR–peptide-bound MHC
class II kinetics determine functions of CD4+ T cells.
J. Immunol. 195, 3557–3564 (2015). doi: 10.4049/
jimmunol.1501407; pmid: 26336148
23. J. Holst et al., Generation of T-cell receptor retrogenic mice.
Nat. Protoc. 1, 406–417 (2006). doi: 10.1038/nprot.2006.61;
pmid: 17406263
24. J. White, A. Pullen, K. Choi, P. Marrack, J. W. Kappler, Antigen
recognition properties of mutant V beta 3+ T cell receptors are
consistent with an immunoglobulin-like structure for the
receptor. J. Exp. Med. 177, 119–125 (1993). doi: 10.1084/
jem.177.1.119; pmid: 8380294
25. Y. Yin, X. X. Wang, R. A. Mariuzza, Crystal structure of a
complete ternary complex of T-cell receptor, peptide-MHC, and
CD4. Proc. Natl. Acad. Sci. U.S.A. 109, 5405–5410 (2012).
doi: 10.1073/pnas.1118801109; pmid: 22431638
Zareie et al., Science 372, eabe9124 (2021)
4 June 2021
12 of 13
RES EARCH | R E S E A R C H A R T I C L E
26. R. Wang, K. Natarajan, D. H. Margulies, Structural basis of the
CD8 a b/MHC class I interaction: Focused recognition orients
CD8 b to a T cell proximal position. J. Immunol. 183,
2554–2564 (2009). doi: 10.4049/jimmunol.0901276;
pmid: 19625641
27. S. V. Pageon, P. R. Nicovich, M. Mollazade, T. Tabarin, K. Gaus,
Clus-DoC: A combined cluster detection and colocalization
analysis for single-molecule localization microscopy data.
Mol. Biol. Cell 27, 3627–3636 (2016). doi: 10.1091/mbc.e16-07-
0478; pmid: 27582387
28. D. Dong et al., Structural basis of assembly of the human T cell
receptor-CD3 complex. Nature 573, 546–552 (2019).
doi: 10.1038/s41586-019-1537-0; pmid: 31461748
29. T. Żal, M. A. Żal, N. R. J. Gascoigne, Inhibition of T cell
receptor-coreceptor interactions by antagonist ligands
visualized by live FRET imaging of the T-hybridoma
immunological synapse. Immunity 16, 521–534 (2002).
doi: 10.1016/S1074-7613(02)00301-1; pmid: 11970876
30. A. N. Tikhonova et al., ab T cell receptors that do not undergo
major histocompatibility complex-specific thymic selection
possess antibody-like recognition specificities. Immunity 36,
79–91 (2012). doi: 10.1016/j.immuni.2011.11.013;
pmid: 22209676
31. P. P. Yachi, J. Ampudia, N. R. J. Gascoigne, T. Zal,
Nonstimulatory peptides contribute to antigen-induced CD8-T
cell receptor interaction at the immunological synapse.
Nat. Immunol. 6, 785–792 (2005). doi: 10.1038/ni1220;
pmid: 15980863
32. J. Casas et al., Ligand-engaged TCR is triggered by Lck not
associated with CD8 coreceptor. Nat. Commun. 5, 5624 (2014).
doi: 10.1038/ncomms6624; pmid: 25427562
33. J. M. Turner et al., Interaction of the unique N-terminal region
of tyrosine kinase p56lck with cytoplasmic domains of CD4
and CD8 is mediated by cysteine motifs. Cell 60, 755–765
(1990). doi: 10.1016/0092-8674(90)90090-2; pmid: 2107025
34. M. E. Birnbaum et al., Molecular architecture of the ab T cell
receptor-CD3 complex. Proc. Natl. Acad. Sci. U.S.A. 111,
17576–17581 (2014). doi: 10.1073/pnas.1420936111;
pmid: 25422432
35. S. V. Pageon et al., Functional role of T-cell receptor
nanoclusters in signal initiation and antigen discrimination.
Proc. Natl. Acad. Sci. U.S.A. 113, E5454–E5463 (2016).
doi: 10.1073/pnas.1607436113; pmid: 27573839
36. F. Van Laethem et al., Deletion of CD4 and CD8 coreceptors
permits generation of alphabetaT cells that recognize antigens
independently of the MHC. Immunity 27, 735–750 (2007).
doi: 10.1016/j.immuni.2007.10.007; pmid: 18023370
37. F. Van Laethem et al., Lck availability during thymic selection
determines the recognition specificity of the T cell repertoire.
Cell 154, 1326–1341 (2013). doi: 10.1016/j.cell.2013.08.009;
pmid: 24034254
38. M. Hahn, M. J. Nicholson, J. Pyrdol, K. W. Wucherpfennig,
1504–1514.e7 (2020). doi: 10.1016/j.celrep.2020.01.008;
pmid: 32023465
42. D. L. Wiest et al., Regulation of T cell receptor expression in
immature CD4+CD8+ thymocytes by p56lck tyrosine kinase:
Basis for differential signaling by CD4 and CD8 in immature
thymocytes expressing both coreceptor molecules. J. Exp.
Med. 178, 1701–1712 (1993). doi: 10.1084/jem.178.5.1701;
pmid: 8228817
43. H. Sanabria et al., Resolving dynamics and function of transient
states in single enzyme molecules. Nat. Commun. 11, 1231
(2020). doi: 10.1038/s41467-020-14886-w; pmid: 32144241
44. J. Lu et al., Molecular constraints on CDR3 for thymic selection
of MHC-restricted TCRs from a random pre-selection
repertoire. Nat. Commun. 10, 1019 (2019). doi: 10.1038/
s41467-019-08906-7; pmid: 30833553
45. G. M. Davey et al., Preselection thymocytes are more sensitive
to T cell receptor stimulation than mature T cells. J. Exp. Med.
188, 1867–1874 (1998). doi: 10.1084/jem.188.10.1867;
pmid: 9815264
46. K. A. Hogquist et al., T cell receptor antagonist peptides induce
positive selection. Cell 76, 17–27 (1994). doi: 10.1016/0092-
8674(94)90169-4; pmid: 8287475
47. K. A. Hogquist, S. C. Jameson, The self-obsession of T cells:
How TCR signaling thresholds affect fate ‘decisions’ and
effector function. Nat. Immunol. 15, 815–823 (2014).
doi: 10.1038/ni.2938; pmid: 25137456
48. B. Lucas, I. Stefanová, K. Yasutomo, N. Dautigny,
R. N. Germain, Divergent changes in the sensitivity of maturing
T cells to structurally related ligands underlies formation of a
useful T cell repertoire. Immunity 10, 367–376 (1999).
doi: 10.1016/S1074-7613(00)80036-9; pmid: 10204492
49. N. L. La Gruta et al., Primary CTL response magnitude in mice
is determined by the extent of naive T cell recruitment and
subsequent clonal expansion. J. Clin. Invest. 120, 1885–1894
(2010). doi: 10.1172/JCI41538; pmid: 20440073
50. J. J. Moon et al., Naive CD4(+) T cell frequency varies for
different epitopes and predicts repertoire diversity and
response magnitude. Immunity 27, 203–213 (2007).
doi: 10.1016/j.immuni.2007.07.007; pmid: 17707129
51. N. P. Cowieson et al., MX1: A bending-magnet crystallography
beamline serving both chemical and macromolecular
crystallography communities at the Australian Synchrotron.
J. Synchrotron Radiat. 22, 187–190 (2015). doi: 10.1107/
S1600577514021717; pmid: 25537608
52. D. Aragão et al., MX2: A high-flux undulator microfocus
beamline serving both the chemical and macromolecular
crystallography communities at the Australian Synchrotron.
J. Synchrotron Radiat. 25, 885–891 (2018). doi: 10.1107/
S1600577518003120; pmid: 29714201
53. W. Kabsch, XDS. Acta Crystallogr. D Biol. Crystallogr. 66,
125–132 (2010). doi: 10.1107/S0907444909047337;
pmid: 20124692
Unconventional topology of self peptide-major
histocompatibility complex binding by a human autoimmune
T cell receptor. Nat. Immunol. 6, 490–496 (2005).
doi: 10.1038/ni1187; pmid: 15821740
54. P. R. Evans, G. N. Murshudov, How good are my data and what
is the resolution? Acta Crystallogr. D Biol. Crystallogr. 69,
1204–1214 (2013). doi: 10.1107/S0907444913000061;
pmid: 23793146
39. D. K. Sethi et al., A highly tilted binding mode by a self-reactive
T cell receptor results in altered engagement of peptide and
MHC. J. Exp. Med. 208, 91–102 (2011). doi: 10.1084/
jem.20100725; pmid: 21199956
55. M. D. Winn et al., Overview of the CCP4 suite and current
developments. Acta Crystallogr. D Biol. Crystallogr. 67,
235–242 (2011). doi: 10.1107/S0907444910045749;
pmid: 21460441
40. Q. Wei et al., Lck bound to coreceptor is less active than free
Lck. Proc. Natl. Acad. Sci. U.S.A. 117, 15809–15817 (2020).
doi: 10.1073/pnas.1913334117; pmid: 32571924
41. V. Horkova et al., Dynamics of the Coreceptor-LCK
Interactions during T Cell Development Shape the Self-
Reactivity of Peripheral CD4 and CD8 T Cells. Cell Rep. 30,
56. P. A. Karplus, K. Diederichs, Linking crystallographic model and
data quality. Science 336, 1030–1033 (2012). doi: 10.1126/
science.1218231; pmid: 22628654
57. A. J. McCoy et al., Phaser crystallographic software. J. Appl.
Crystallogr. 40, 658–674 (2007). doi: 10.1107/
S0021889807021206; pmid: 19461840
58. P. Emsley, B. Lohkamp, W. G. Scott, K. Cowtan, Features and
development of Coot. Acta Crystallogr. D Biol. Crystallogr. 66,
486–501 (2010). doi: 10.1107/S0907444910007493;
pmid: 20383002
59. G. Bricogne et al., “BUSTER version 1.10.0” (The Debian
Project, 2011).
60. N. A. Borg et al., The CDR3 regions of an immunodominant
T cell receptor dictate the ‘energetic landscape’ of peptide-
MHC recognition. Nat. Immunol. 6, 171–180 (2005).
doi: 10.1038/ni1155; pmid: 15640805
61. L. Kjer-Nielsen et al., A structural basis for the selection of
dominant alphabeta T cell receptors in antiviral immunity.
Immunity 18, 53–64 (2003). doi: 10.1016/S1074-7613(02)
00513-7; pmid: 12530975
62. B. D. Ripley, Tests of “randomness” for spatial point patterns.
J. R. Stat. Soc. B 41, 368–374 (1979). doi: 10.1111/j.2517-
6161.1979.tb01091.x
AC KNOWLED GME NTS
We wish to dedicate this work to the memory of our coauthor
and close colleague, Professor Katharina Gaus. We thank
P. Marrack (National Jewish Health, Colorado) and D. Vignali
(University of Pittsburgh) for provision of reagents;
D. Jayasinghe and F. Weide for technical assistance; P. Harrison
from the Monash Bioinformatics Platform for assistance with
statistical analyses; staff at Monash FlowCore, Monash Animal
Resource Platform, Monash Macromolecular Crystallization
Facility, Monash Micro Imaging Platforms; and the ANSTO
Australian Synchrotron MX1/MX2 beamline scientists.
Funding: This work was supported by the Australian Research
Council (ARC) (DP170103631 to N.L.L. and S.G., DP201102776 to
N.L.L., CE140100011 to J.R. and K.G.); the National Health and
Medical Research Council of Australia (NHMRC) (APP1182086
to N.L.L. and APP1155162 to K.G.); the NSW Cancer Council
(APP1128488 to K.G.); the Singapore Ministry of Health’s
National Medical Research Council (CBRG/0097/2015 to
N.R.J.G.); and the Singapore Ministry of Education (2014–T2–1–
136 to N.R.J.G.). N.L.L. is supported by an ARC Future
Fellowship, S.G. by an NHMRC Senior Research Fellowship, and
J.R. by an ARC Laureate Fellowship. Author contributions:
Conceptual and experimental design: N.L.L., P.Z., J.R., S.G.,
K.G., B.D.E., and N.R.J.G.; recombinant protein expression,
structure determination, and SPR analyses: C.S., C.F., J.P.,
S.G., and J.R.; generation of in vitro cell lines, retrogenic mice,
FLIM-FRET imaging, and assays of T cell function: P.Z. with help
from C.M.J., Q.W., L.W., X.Y.X.S., A.N., C.B., and A.J.F.; 2D
measures of binding: J.R.J., E.M.K., and B.D.E.; SMLM: S.D.G.
and K.G.; data analysis: all authors; manuscript writing: P.Z.,
N.L.L., S.G., and J.R. with contributions from all other authors.
Competing interests: The authors declare no competing
interests. Data and materials availability: Structures for the
B17.R2 TCR–H-2Db–NP366 and B17.C1 TCR–H-2Db–NP366
complexes have been deposited into the PDB database (B17.R2
TCR–H-2Db–NP366, PDB 7JWI; B17.C1 TCR–H-2Db–NP366, PDB
7JWJ). All other data are available in the main text or the
supplementary materials.
SUPPLEMENTARY MATERIALS
science.sciencemag.org/content/372/6546/eabe9124/suppl/DC1
Figs. S1 to S5
Tables S1 to S5
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
22 September 2020; accepted 23 April 2021
10.1126/science.abe9124
Zareie et al., Science 372, eabe9124 (2021)
4 June 2021
13 of 13
|
10.1126_science.abd2638
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
HIV
Recapitulation of HIV-1 Env-antibody coevolution in
macaques leading to neutralization breadth
Ryan S. Roark*, Hui Li*, Wilton B. Williams*, Hema Chug*, Rosemarie D. Mason*, Jason Gorman*,
Shuyi Wang, Fang-Hua Lee, Juliette Rando, Mattia Bonsignori, Kwan-Ki Hwang, Kevin O. Saunders,
Kevin Wiehe, M. Anthony Moody, Peter T. Hraber, Kshitij Wagh, Elena E. Giorgi, Ronnie M. Russell,
Frederic Bibollet-Ruche, Weimin Liu, Jesse Connell, Andrew G. Smith, Julia DeVoto,
Alexander I. Murphy, Jessica Smith, Wenge Ding, Chengyan Zhao, Neha Chohan, Maho Okumura,
Christina Rosario, Yu Ding, Emily Lindemuth, Anya M. Bauer, Katharine J. Bar, David Ambrozak,
Cara W. Chao, Gwo-Yu Chuang, Hui Geng, Bob C. Lin, Mark K. Louder, Richard Nguyen,
Baoshan Zhang, Mark G. Lewis, Donald D. Raymond, Nicole A. Doria-Rose, Chaim A. Schramm,
Daniel C. Douek, Mario Roederer, Thomas B. Kepler, Garnett Kelsoe, John R. Mascola, Peter D. Kwong,
Bette T. Korber, Stephen C. Harrison, Barton F. Haynes, Beatrice H. Hahn, George M. Shaw†
INTRODUCTION: It is widely believed that the
development of an effective neutralizing
antibody–based HIV-1 vaccine will require
consistent activation of multiple germline
precursor B cells that express immunoglobulin
receptors specific for one or more of the canonical
broadly neutralizing antibody (bNAb) epitope
clusters, followed by efficient antigen-driven
selection for antibody affinity maturation. How
to accomplish this feat by immunization has
proven to be a daunting scientific challenge.
One roadblock to rational HIV-1 vaccine design
is the lack of a suitable outbred primate model
in which bNAbs can be commonly induced,
thereby enabling the molecular, biological, and
immunological mechanisms responsible for such
responses to be studied in a reproducible and
iterative fashion.
RATIONALE: Given that most HIV-1 bNAbs
identified to date have come from humans
chronically infected by HIV-1, we hypothesized
that one means to elicit such antibodies in
primates might be by infecting rhesus ma-
caques (RMs) with simian-human immu-
nodeficiency virus (SHIV) strains that bear
primary HIV-1 envelope proteins (Envs), includ-
ing those that induced bNAbs in humans.
SHIV-infected RMs could then be used to
assess the potential of particular HIV-1 Envs
to elicit bNAbs and to characterize the co-
evolutionary pathways of bNAb lineages and
the cognate Env intermediates that elicited
them, thus serving as a molecular guide for
rational vaccine design. Recent innovations
in SHIV design make this experimental strat-
egy testable.
Envelope-antibody coevolution in monkey and human. Env-antibody coevolution in SHIV-infected rhesus
macaques mirrors that in HIV-1–infected humans, including the elicitation of broadly neutralizing antibodies.
The illustration depicts a mirror with images, left to right, of organism, longitudinal changes in viral Env
sequence, Env trimer recognition by virus-elicited broadly neutralizing antibodies, and atomic-level details of
paratope–epitope interactions.
RESULTS: Neutralizing antibodies elicited by
HIV-1 in naturally infected humans coevolve
with viral Envs in distinctive molecular patterns,
in some cases acquiring substantial breadth.
We constructed SHIVs bearing primary trans-
mitted/founder Envs from three HIV–1 infected
humans who developed bNAbs and used these
SHIVs to infect 22 RMs. Seven monkeys de-
veloped bNAbs that exhibited a wide range
of breadth and potency. Unexpectedly, SHIV
infections elicited molecular patterns of Env-
antibody coevolution in monkeys that mir-
rored what was seen in humans infected by
HIV-1 strains bearing homologous Envs. Sim-
ilarities included conserved immunogenetic,
structural, and chemical solutions to epitope
recognition and precise Env–amino acid sub-
stitutions, insertions, and deletions leading to
virus persistence. One rhesus antibody, capa-
ble of neutralizing 49% of a 208-strain global
virus panel, contained a 24–amino acid heavy
chain complementarity-determining region 3
(HCDR3) with a sulfated tyrosine at its tip;
this rhesus bNAb exhibited a V2 apex mode of
recognition similar to human bNAbs PGT145
and PCT64-35S, with critical interactions in-
volving lysine or arginine residues at Env posi-
tions 121, 166, and 169 and an N-linked glycan
at position 160. Another rhesus antibody bound
the CD4 binding site by CD4 mimicry mirror-
ing human bNAbs 8ANC131, CH235, and VRC01.
In other SHIV-infected animals, bNAb responses
targeted a canonical V3 high mannose patch
epitope cluster that included Env residues
324GDIR327 and N332. Molecular patterns of
epitope evolution enabling virus escape, and
at the same time promoting bNAb affinity
maturation, were similar in SHIV-infected
RMs and HIV-1–infected humans.
CONCLUSION: SHIV infection of RMs is the
only model system other than naturally in-
fected humans where the immunogen (Env)
coevolves with neutralizing antibodies. The
high mutability and dynamic replication of
HIV-1 and SHIV in vivo result in a constantly
evolving virus quasispecies, which means
that Envs with binding affinities sufficient to
drive bNAb lineage affinity maturation are
continuously being generated. SHIV-infected
macaques may thus provide insights for vac-
cine design by enabling the identification of
Env intermediates that can guide the evolution
of precursor B cells through stages of affinity
maturation leading to breadth and potency.▪
The list of author affiliations is available in the full article online.
*These authors contributed equally to this work.
†Corresponding author. Email: shawg@pennmedicine.
upenn.edu
Cite this article as R. S. Roark et al., Science 371, eabd2638
(2021). DOI: 10.1126/science.abd2638
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.abd2638
I
H
N
,
N
A
M
R
O
G
.
J
:
E
G
A
M
I
Roark et al., Science 371, 142 (2021)
8 January 2021
1 of 1
RES EARCH
R E S E A R C H A R T I C L E
◥
HIV
Recapitulation of HIV-1 Env-antibody coevolution in
macaques leading to neutralization breadth
Ryan S. Roark1*, Hui Li1*, Wilton B. Williams2,3*, Hema Chug4*, Rosemarie D. Mason5*,
Jason Gorman5*, Shuyi Wang1, Fang-Hua Lee1, Juliette Rando1, Mattia Bonsignori2,3, Kwan-Ki Hwang2,
Kevin O. Saunders2,6, Kevin Wiehe2,3, M. Anthony Moody2,7, Peter T. Hraber8, Kshitij Wagh8,
Elena E. Giorgi8, Ronnie M. Russell1, Frederic Bibollet-Ruche1, Weimin Liu1, Jesse Connell1,
Andrew G. Smith1, Julia DeVoto1, Alexander I. Murphy1, Jessica Smith1, Wenge Ding1, Chengyan Zhao1,
Neha Chohan1, Maho Okumura1, Christina Rosario1, Yu Ding1, Emily Lindemuth1, Anya M. Bauer1,
Katharine J. Bar1, David Ambrozak5, Cara W. Chao5, Gwo-Yu Chuang5, Hui Geng5, Bob C. Lin5,
Mark K. Louder5, Richard Nguyen5, Baoshan Zhang5, Mark G. Lewis9, Donald D. Raymond4,
Nicole A. Doria-Rose5, Chaim A. Schramm5, Daniel C. Douek5, Mario Roederer5, Thomas B. Kepler10,11,
Garnett Kelsoe2,6, John R. Mascola5, Peter D. Kwong5, Bette T. Korber8, Stephen C. Harrison4,12,
Barton F. Haynes2,3, Beatrice H. Hahn1, George M. Shaw1†
Neutralizing antibodies elicited by HIV-1 coevolve with viral envelope proteins (Env) in distinctive
patterns, in some cases acquiring substantial breadth. We report that primary HIV-1 envelope proteins—
when expressed by simian-human immunodeficiency viruses in rhesus macaques—elicited patterns of
Env-antibody coevolution very similar to those in humans, including conserved immunogenetic,
structural, and chemical solutions to epitope recognition and precise Env–amino acid substitutions,
insertions, and deletions leading to virus persistence. The structure of one rhesus antibody, capable of
neutralizing 49% of a 208-strain panel, revealed a V2 apex mode of recognition like that of human
broadly neutralizing antibodies (bNAbs) PGT145 and PCT64-35S. Another rhesus antibody bound the
CD4 binding site by CD4 mimicry, mirroring human bNAbs 8ANC131, CH235, and VRC01. Virus-antibody
coevolution in macaques can thus recapitulate developmental features of human bNAbs, thereby
guiding HIV-1 immunogen design.
A major roadblock to rational HIV-1 vaccine
design is the lack of a suitable primate
model in which broadly neutralizing anti-
bodies (bNAbs) can be commonly induced
and the molecular, biological, and immuno-
logical mechanisms responsible for such re-
sponses studied in a reproducible and iterative
fashion. Because most HIV-1 bNAbs identified
to date have come from humans chronically
infected by HIV-1, we hypothesized that one
means to elicit such antibodies in primates
might be by infecting Indian rhesus macaques
1Departments of Medicine and Microbiology, Perelman School
of Medicine, University of Pennsylvania, Philadelphia, PA 19104,
USA. 2Duke Human Vaccine Institute, Duke University School of
Medicine, Durham, NC 27710, USA. 3Department of Medicine,
Duke University School of Medicine, Durham, NC 27710, USA.
4Laboratory of Molecular Medicine, Boston Children’s Hospital
and Harvard Medical School, Boston, MA 02115, USA. 5Vaccine
Research Center, National Institute of Allergy and Infectious
Diseases, National Institutes of Health, Bethesda, MD 20892,
USA. 6Departments of Immunology and Surgery, Duke
University School of Medicine, Durham, NC 27710, USA.
7Departments of Pediatrics and Immunology, Duke University
School of Medicine, Durham, NC 27710, USA. 8Theoretical
Biology and Biophysics, Los Alamos National Laboratory, Los
Alamos, NM 87545, USA. 9Bioqual, Inc., Rockville, MD 20850,
USA. 10Department of Microbiology, Boston University School of
Medicine, Boston, MA 02118, USA. 11Department of
Mathematics and Statistics, Boston University, Boston, MA
02215, USA. 12Howard Hughes Medical Institute, Harvard
Medical School, Boston, MA 02115, USA.
*These authors contributed equally to this work.
†Corresponding author. Email: shawg@pennmedicine.upenn.edu
(RMs) with simian-human immunodeficien-
cy virus (SHIV) strains that bear primary or
transmitted/founder (T/F) HIV-1 envelope
proteins (Envs) that induced bNAbs in hu-
mans (1–7). SHIV-infected RMs could then be
used to assess the potential of particular HIV-1
Envs to elicit bNAbs and to characterize the
coevolutionary pathways of bNAb lineages
and the cognate Env intermediates that eli-
cited them (2–8), thereby serving as a molec-
ular guide for rational vaccine design. Recent
innovations in SHIV design (9) now make this
experimental strategy testable.
HIV-1 bNAbs target one of several conserved
regions on the native Env trimer, including the
CD4 binding site (CD4bs), V2 apex, V3 high-
mannose patch, glycoprotein gp120–gp41 in-
terface, fusion peptide, glycosylated silent
face, and membrane-proximal external region
(10, 11). Such antibodies generally share cer-
tain features, such as exceptional heavy chain
complementarity-determining region 3 (HCDR3)
length, extensive somatic hypermutation, auto-
reactivity, or unusual mechanisms for binding
glycans or glycopeptides (10–13). Long HCDR3s
result from initial germline immunoglobulin
gene rearrangements, whereas the character-
istic high frequencies of V(D)J mutations that
lead to affinity maturation and neutralizing
breadth result from persistent virus replication,
epitope variation, and continued Env-Ab co-
evolution. In some subjects, cooperating anti-
body lineages that target the same epitope have
been identified, and together they contribute
to Env diversification and bNAb development
(4, 6). A critical question in the HIV-1 vaccine
field is whether the relatively rare examples of
high-titer bNAbs that have been identified in a
subset of chronically infected humans repre-
sent stochastic accidents of nature, not likely
to be replicated by a vaccine, or whether there
are special properties of particular HIV-1 Envs
that, along with deterministic features of Env-
Ab coevolution, make HIV-1 vaccine develop-
ment more plausible.
This study is based on the premise that while
Env diversity within HIV-1 group M is extra-
ordinarily high, primary HIV-1 Envs [that is,
native Env trimers on infectious virions that
allow for persistent replication and clinical
transmission in humans (1)] are nonetheless
constrained in their immediate evolutionary
potential (14–16). This paradox—extreme Env
diversity globally but constraints on immedi-
ate or near-term evolution in an individual—
can be explained by competing evolutionary
forces: positive selection imposed by immune
evasion, balanced by purifying (negative) se-
lection acting to retain viral fitness. At least
four interrelated Env properties, acting in con-
cert, facilitate immune escape—occlusion of
trimer-interface epitopes (17), glycan shielding
(18), epitope variation (19), and conformational
masking (20)—while conservation of structure
and cell-entry properties and retention of pro-
tein stability constrain each primary Env’s im-
mediate and short-term evolutionary trajectory.
When introduced into a human as HIV-1 in-
fection or into RMs as SHIV infection, each Env
will have distinctive restrictions in its exposure
of antigenic sites, thus favoring distinct mo-
lecular pathways of escape from neutralizing
antibodies with each pathway exacting a differ-
ent fitness cost from the virus. These considera-
tions led us to hypothesize that SHIV infection
of RMs will recapitulate HIV-1 infection of hu-
mans with respect to molecular patterns of
Env-Ab coevolution, within the bounds set by
any particular Env sequence and the respec-
tive rhesus and human immunoglobulin gene
repertoires that respond to it (21).
Some of the best-studied examples of HIV-1
Env-Ab coevolution have come from the anal-
ysis of HIV-1–infected human subjects CH505,
CH848, and CAP256, who developed bNAb re-
sponses targeting the CD4bs, V3 high-mannose
patch, and V2 apex of HIV-1 Env, respectively
(2–7). Subjects CH505 and CH848 were each
productively infected by a single T/F virus, and
we previously constructed SHIVs containing
each of these T/F Envs (9). Subject CAP256 was
infected by a single virus (CAP256PI, primary
infection) and then became superinfected by a
second genetically divergent virus (CAP256SU,
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
1 of 19
RES EARCH | R E S E A R C H A R T I C L E
superinfection) (22). The CAP256SU variant
was thought to trigger the V2 bNAb lineage in
this individual (5), so we constructed and ana-
lyzed a SHIV containing this Env (fig. S1). SHIVs
containing CH505, CH848, and CAP256SU
Envs were modified at gp120 residue 375 to
enhance binding and entry into rhesus CD4
T cells. The Env375 substitutions resulted in
no discernible change in antigenicity or sen-
sitivity to anti-HIV-1 NAbs compared with
wild-type virus, but they were essential for
replication in primary RM CD4+ T cells (fig.
S1) (9). We inoculated 22 RMs (table S1) with
SHIV.CH505 (n = 10), SHIV.CH848 (n = 6), or
SHIV.CAP256SU (n = 6) by the intravenous
route and followed them for an average pe-
riod of 2 years (mean of 103 weeks, range 36
to 184 weeks) for viral replication kinetics,
development of strain-specific (autologous)
and broadly reactive (heterologous) NAbs, and
patterns of Env sequence evolution. Eight of
the animals were treated with anti-CD8 mono-
clonal antibody (mAb) at or near the time of
SHIV inoculation to transiently deplete CD8+
cells and increase peak and set point plasma
virus titers. One animal (RM5695), which was
repurposed from an earlier study of HIV-1
gp120 protein immunization (23), was inocu-
lated with plasma from three animals recently
infected by SHIV.CH505 to increase early vi-
rus diversity.
SHIV replication in vivo and elicitation of NAbs
Figure S2 depicts the kinetics of virus repli-
cation and the development of autologous,
strain-specific tier 2 NAbs as measured by
plasma viral RNA (vRNA) and inhibition of
virus entry into TZM-bl cells (18). Acute virus
replication kinetics were consistent for all
SHIVs in all animals with peak viremia oc-
curring about 14 days after inoculation and
reaching titers in the 105 to 108 vRNA/ml range
[geometric mean (GM) = 4.2 × 106]. Peak vRNA
titers were higher in anti-CD8 treated animals
than in untreated animals [geometric mean
titers (GMT) of 2.6 × 107 versus 1.5 × 106, re-
spectively; P < 0.0001]. Plasma virus load set
points at 24 weeks after infection varied
widely, between 100 and 988,110 vRNA/ml,
with a mean for anti-CD8 treated animals of
248,313 ± 136,660 vRNA/ml compared with
8313 ± 3509 vRNA/ml for untreated animals
(P = 0.002). Set point vRNA levels in human
subjects infected with the corresponding HIV-1
CH505, CH848, or CAP256 strains were 81,968,
132,481, and 750,000 vRNA/ml, respectively.
Compared with two human cohort studies of
acute and early HIV-1 infection (24, 25), peak
and set point vRNA titers in SHIV-infected
animals were comparable, although several
macaques developed viral loads near the lower
limit of detection (100 vRNA/ml), which is
rare in humans. Six macaques developed AIDS-
defining events and were euthanized at 36, 40,
60, 65, 89, or 136 weeks after SHIV infection;
four of these animals had been treated with
anti-CD8 mAb (table S1).
Autologous tier 2 NAb responses to the
three SHIVs were detected as early as 8 weeks
after infection and peaked between 24 and
80 weeks, with 50% inhibitory dilutions (ID50)
between 0.05 (1:20 dilution) and 0.000125
(1:8000 dilution) (fig. S2). The kinetics of ap-
pearance and titers of autologous tier 2 NAbs
that developed in response to SHIV infections
were generally comparable to those observed
in humans infected by viruses bearing the ho-
mologous Envs (2, 7, 22). Among all animals,
there was a significant association between
higher set point virus titers and higher autolo-
gous tier 2 NAb titers (Spearman correlation
rank correlation coefficient rs = 0.74, P < 0.0001).
Heterologous plasma neutralizing activity was
assessed against a diverse 22-member global
panel of tier 1a (n = 3) or tier 2 (n = 19) HIV-1
strains (26–29) (table S2). All animals devel-
oped potent neutralizing responses to the three
tier 1a viruses (GMT ID50 = 0.0004). Tier 1a
viruses have “open” Env trimers that sponta-
neously expose linear V3 and conformational
CD4-induced (CD4i) bridging sheet epitopes,
thus explaining their extreme sensitivity to what
are otherwise non-neutralizing antibodies.
Such non-neutralizing antibodies that target
linear V3 and CD4i epitopes are elicited in vir-
tually all HIV-1–infected humans, thereby select-
ing for Envs with an open-closed equilibrium
that strongly favors the closed configuration
(20, 30–32). Tier 2 viruses typify primary or
T/F viruses that are generally resistant to neu-
tralization by heterologous plasma antibodies
except for those that target one of the canon-
ical bNAb epitope clusters (1, 10, 11, 26). In this
study, we used a reciprocal neutralization titer
cutoff of ID50 ≤ 0.05 against ≥33% of heterolo-
gous tier 2 viruses as an indication of neutral-
ization breadth. This is a conservative threshold
consistent with other studies that have charac-
terized bNAb prevalence in human cohorts
(28, 33, 34).
Seven of 22 RMs developed antibody re-
sponses that neutralized between 6 and 18 of
18 heterologous HIV-1 tier 2 viruses in our test
panel (Fig. 1A and table S2). Heterologous neu-
tralization was first detected as early as 8 to
16 weeks after SHIV infection in two animals
and as late as 88 weeks after SHIV infection in
others. Heterologous NAbs reached ID50 titers
as high as 0.0001, with GMTs in the seven ani-
mals ranging from 0.018 (1:55) to 0.004 (1:271)
(Fig. 1A). Peak and set point plasma virus titers
were higher in the seven animals that developed
bNAbs (GMT = 18,150,586 and 47,471 vRNA/ml,
respectively) than in animals that did not
(GMT = 2,101,814 and 1945 vRNA/ml, respec-
tively; P < 0.015 for both). Two animals with
bNAbs (RM5695 and RM6070) were infected
by SHIV.CH505, two (RM6163 and RM6167) by
SHIV.CH848, and three (RM40591, RM42056,
and RM6727) by SHIV.CAP256SU. The remain-
ing 15 animals in the study showed either no or
very limited, low-titer neutralization of heter-
ologous tier 2 viruses (table S2). Altogether, the
findings show that SHIVs bearing primary
T/F Envs reproduce in RMs key features of
virus replication dynamics and NAb elicitation
that are characteristic of HIV-1 infection in hu-
mans, including the potential to elicit bNAbs.
Figure 1B highlights the kinetics, potency,
and breadth of neutralization by plasma from
the SHIV.CH505-infected animal RM5695 and
identifies immunoglobulin G (IgG) as the active
component. By 16 weeks after infection, an
autologous NAb response to SHIV.CH505 was
detectable at an ID50 titer of 0.02 along with
heterologous responses to viruses bearing HIV-1
25710, X1632, Q23, ZM233, T250, and WITO
Envs at titers of 0.05 to 0.01. NAbs increased
rapidly in breadth and titer thereafter. By week
48, bNAbs targeting Q23 and T250 jumped in
titer to between 0.0002 and 0.0005 and against
X1632, 246F3, ZM233, and WITO to titers of
0.005. By week 56, bNAbs in the plasma of
RM5695 neutralized 17 of 18 viruses in the heter-
ologous HIV-1 test panel at titers ranging from
0.05 to 0.0002, along with a divergent tier 2
simian immunodeficiency virus strain (SIVcpz.
MT145.Q171K) that shares selective antigenic
cross-reactivity with HIV-1 in the V2 apex bNAb
epitope cluster (35) at a titer of 0.006. IgG was
purified from RM5695 week 56 plasma and
assayed for neutralizing activity against the
same 19 heterologous viruses: IgG concentra-
tions between 0.002 mg/ml (corresponding to
an ~1:10,000 dilution of rhesus plasma) and
4 mg/ml neutralized 18 of the 19 viruses in a
rank order similar to the polyclonal plasma (Fig.
1B, far right panels). Neither plasma nor purified
IgG neutralized control viruses pseudotyped
with the murine leukemia virus (MLV) Env.
Thus, anti-HIV-1–specific IgG accounted for
all of the autologous and heterologous neu-
tralizing activity in the RM5695 plasma. Of
note, RM5695 plasma and plasma IgG from
week 56 reached ID90 or ID95 thresholds against
most viruses and exhibited steep inflections at
the ID50 midpoint, indicating potent neutral-
ization. Neutralization breadth detected as
early as 16 weeks after infection is unusual in
HIV-1 infection but not unprecedented (36) and
occurs most often with V2 apex bNAbs, likely
because their activity depends more on long
HCDR3s than on extensive somatic hypermuta-
tion. We show below that the bNAb activity in
RM5695 plasma and its isolated IgG fraction
as well as monoclonal bNAbs derived from
RM5695 all targeted a bNAb epitope in the
V2 apex that included the conserved lysine rich
C-strand and N160 glycan. The kinetics of ap-
pearance, breadth, titers, and potency of bNAbs
elicited in the six other SHIV-infected RMs,
including three animals (RM6070, RM40591,
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
2 of 19
RES EARCH | R E S E A R C H A R T I C L E
A
)
0
5
D
I
l
a
c
o
r
p
c
e
R
i
(
r
e
t
i
T
b
A
N
104
103
102
20
<20
104
103
102
20
<20
RM5695 (CH505)
RM6070 (CH505)
RM6163 (CH848)
RM6167 (CH848)
Breadth: 94%
GMT: 185
Breadth: 44%
GMT: 66
104
103
102
20
<20
Breadth: 33%
GMT: 271
104
103
102
20
<20
Breadth: 100%
GMT: 68
104
103
102
20
<20
0
8
16
24
32
48
56
0
8
16 24 36
0 64 80 88 104 112 128
0
64
80
88 104 112 128
RM40591 (CAP256SU)
RM42056 (CAP256SU)
RM6727 (CAP256SU)
Breadth: 67%
GMT: 119
Breadth: 44%
GMT: 72
104
103
102
20
<20
Breadth: 50%
GMT: 55
104
103
102
20
<20
0 72 88 96 104 112 120 128
0
48 56 64 72 80 88 104
Time (Weeks)
0
56
60
Q23
T250
WITO
ZM233
CAP256SU
CH505
CH848
BG505 332N
TRO11
25710
CNE8
X2278
BJOX010000
X1632
Ce1176
246F3
CH119
Ce0217
CNE55
w0
w8
w16
w24
w32
w48
w56
w56
B
y
t
i
v
i
t
c
e
f
n
i
t
n
e
c
r
e
P
125
100
75
50
25
0
125
100
75
50
25
0
0.0001
0.00001
0.001
0.01
0.0001
0.00001
0.001
0.1
0.01
0.01
0.1
0.0001
0.001
0.1
0.00001
0.001
0.0001
0.00001
0.1
0.0001
0.001
0.00001
RM5695 Plasma Dilution
0.01
0.01
0.00001
0.0001
0.1
0.001
0.01
0.0001
0.1
0.001
0.00001
0.01
0.1110
0.1
0.01
0.001
0.0001
RM5695
IgG (mg/ml)
MLV
TRO11
25710
CNE8
X2278
BJOX002000
X1632
CE1176
246F3
CH119
CE0217
CNE55
MLV
T250
Q23
WITO
ZM233
CAP256SU
CH848
BG505.332N
MT145.Q171K
CH505
Fig. 1. Broadly neutralizing antibodies in seven rhesus macaques. (A) RMs
5695 and 6070 were infected by SHIV.CH505; RMs 6163 and 6167 by SHIV.
CH848; and RMs 40591, 42056, and 6727 by SHIV.CAP256SU. Neutralizing
antibody titers (ID50, 50% inhibitory dilution) from longitudinal plasma speci-
mens against 19 global tier 2 HIV-1 strains (26, 27, 84) are depicted. Full
designations of target viruses are provided in fig. S22 and table S3. Maximum
neutralization breadth across all time points and maximum geometric mean
titer (GMT) of neutralization at any one time point are indicated for each
animal. (B) Neutralization curves for longitudinal plasma specimens and
purified plasma IgG (highlighted in bold) from RM5695 show the development
of broad and potent neutralization. MT145K.Q171K is a chimpanzee-derived
SIVcpz strain that shares antigenic cross-reactivity with HIV-1 in the V2 apex
(35, 131). Dashed lines indicate 50% reduction in virus infectivity. MLV, murine
leukemia virus.
and RM42056) whose bNAbs also targeted the
V2 apex C-strand, are summarized in Fig. 1A
and table S2.
Env evolution in SHIV-infected macaques and
HIV-1–infected humans
To explore whether SHIV Env evolution in
RMs recapitulates that of HIV-1 in humans in
a strain-specific manner, we analyzed Env se-
quences in the 22 SHIV-infected animals over
1 to 3 years of follow-up and compared them
with the evolution of the homologous Envs
in humans infected by HIV-1. We used single
genome sequencing for this analysis, because
it allows for retention of genetic linkage across
intact viral genes and enables precise molec-
ular tracking of sequence diversification from
unambiguous T/F genomes (1, 37–39). We ana-
lyzed cDNA sequences derived from plasma
virion RNA, because the half-life of circulating
plasma virus is <1 hour (40) and that of cells
producing most of this virus is <1 day (40).
This short composite half-life (<1 day) reflects
the life span of >99.9% of circulating plasma
virus in individuals not receiving antiretrovi-
ral therapy (41, 42), making the genetic com-
position of the plasma virus quasi-species
an exquisitely sensitive real-time indicator of
in vivo selection pressures acting on virus and
virus-producing cells, including that exerted
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
3 of 19
RES EARCH | R E S E A R C H A R T I C L E
TF
SP
V1 V2
LD V3
V4
V5
FP
MPER
A
n
a
m
u
H
2
7
0
6
M
R
w004
w007
w010
w014
w020
w030
w053
w078
w100
w136
w004
w010
w020
w032
w036
w044
w052
w068
w084
w092
w104
w112
w010
w020
w032
9
6
0
6
M
w040R
w012
w016
w032
w056
w080
w012
w024
w048
7
9
6
6
M
R
3
0
7
6
M
R
w010
w020
w032
0
7
0
6
M
w036R
w002
w004
w012
w016
w024
w048
w064
5
9
6
5
M
R
C
N234
hole
1
130
130
1
130
1
130
1
130
1
130
130
1
CH505 TF
N332
hole
N234
filled
o
PNGS
10A neighbors of PNGS
>80% glycan shielded in M-group
50-80% glycan shielded in M-group
N234
filled
640
620
620
640
620
640
620
640
279
281
234
334
300
302 330
355
417
413
2
334
417
279
302
330
281
234
300
2
413
355
334
417
279 302 330
355
330
334
302
417
234
334
413
334
417
355
234
279 302 330
334
417
234
279 302
330
355
3
3
3
3
3
N332
filled
D
CH505
w053
N332
filled
l
s
e
d
n
i
1
V
RM6072
w044
B
TF
w004
w007
w010
w014
w020
w030
w053
w078
w100
w136
w004
w010
w020
w032
w036
w044
w052
w068
w084
w092
w104
w112
w010
w020
w032
w040
w012
w016
w032
w056
w080
w012
w024
w048
w010
w020
w032
w036
w002
w004
w012
w016
w024
w048
w064
0
3
1
4
3
2
9
7
2
1
8
2
0
0
3
2
0
3
0
3
3
2
3
3
4
3
3
7
4
3
5
5
3
3
1
4
7
1
4
1
7
4
0
2
6
0
4
6
n
a
m
u
H
2
7
0
6
M
R
9
6
0
6
M
R
7
9
6
6
M
R
3
0
7
6
M
R
0
7
0
6
M
R
5
9
6
5
M
R
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
4 of 19
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Env evolution in SHIV.CH505-infected rhesus macaques recapitu-
lates HIV-1 Env evolution in human subject CH505. (A) A pixel plot
(www.hiv.lanl.gov/content/sequence/pixel/pixel.html; www.hiv.lanl.gov/content/
sequence/HIV/HIVTools.html) of amino acid alignments of longitudinal HIV-1
Env sequences obtained by single genome sequencing of plasma virion RNA.
Between 15 and 60 individual sequences are grouped by time point, and amino
acids are colored red to highlight mutations relative to the infecting CH505
transmitted/founder strain illustrated schematically at the top. The image is
highly compressed. Each row within a time block represents a single sequence,
and each column represents an amino acid position in the alignment; thus each
pixel represents a single amino acid that is colored gray if it matches the T/F
sequence, red if it is a mutated residue, and black if it is an insertion or deletion
(indel) relative to the T/F virus. All SHIV sequences differ from the T/F at
position 375, reflecting the SHIV design strategy that enables replication in
rhesus macaques (9). Green tags indicate amino positions (HXB2 numbering)
that are mutated in both human and rhesus. Yellow tags indicate three sites of
identical indels observed in both human and rhesus. (B) LASSIE analysis (44)
of the same longitudinal Env sequences was used to characterize mutations
under positive selection. If a T/F amino acid was retained in <25% of the sequences
in human CH505 infection, the site was considered to be under selective pressure
and tracked in all hosts. The height of each amino acid mutation is proportional
to its frequency at the respective time point. Red, dark blue, and black indicate
acidic, basic, and neutral residues, respectively. “O” indicates asparagine (N)
embedded in an N-linked glycosylation motif. Numbers at the bottom indicate
residue positions (HXB2 numbering). Green numbers indicate mutations that
reached 75% frequency in the human and in at least one animal. (C) Side projection
of the CH505 Env trimer with potential N-linked glycans (PNGS) indicated in blue,
10-Å neighbors of PNGs shown in green, and “glycan holes” that are typically
covered by glycans in >80% and 50 to 80% of group M HIV-1 strains in magenta
and pink, respectively. The light gray area in the central region of the trimer is
the interprotomer surface that forms a cleft with low glycan coverage (45). Two
glycan holes in the CH505 T/F at positions 234 and 332 were filled by the addition
of NXS/T motifs over time in human subject CH505 and in RM6072 as well as in
all other monkeys (see fig. S3). (D) The V1 sequence of CH505 env is shown at the
top of the panel with the hypervariable region indicated in red. Indels in V1 that
arose in the first year of infection in the human host are illustrated in the sequences
below the reference T/F sequence. Indels occurred in nucleotide lengths divisible
by three so as to maintain a viable Env open reading frame, and in the case of
insertions, consisted of direct repeats. Indels that were replicated identically in one
or more rhesus macaques are indicated by color-matched boxes to the right.
Identical indels in human and rhesus CH505 sequences were also found in V4 and
V5 (fig. S9). Single-letter abbreviations for the amino acid residues are as follows:
A, Ala; C, Cys; D, Asp; E, Glu; F, Phe; G, Gly; H, His; I, Ile; K, Lys; L, Leu; M, Met;
N, Asn; P, Pro; Q, Gln; R, Arg; S, Ser; T, Thr; V, Val; W, Trp; and Y, Tyr.
by NAbs, cytotoxic T cells (CTLs), or anti-
retroviral drugs (18, 37, 39, 42, 43). Figure 2A
illustrates Env amino acid substitutions in
virus from sequential plasma specimens from
human subject CH505 and from six SHIV.
CH505-infected RMs. By inspection and the
algorithm-based statistical methods previ-
ously described [(37, 38) and supplementary
materials], amino acid substitutions in Env
were found to be nonrandomly distributed in
patterns that evolved over time beginning as
early as 4 weeks after infection. Using LASSIE
(Longitudinal Antigenic Sequences and Sites
from Intra-Host Evolution) (44), a computa-
tional tool that systematically identifies amino
acid differences in a set of evolved sequences
that reach a predetermined frequency thresh-
old compared with a founder sequence, we
identified in human subject CH505 15 Env
residues where mutations altered the encoded
amino acid in at least 75% of sequences at one
or more subsequent time points. Thirteen of
these 15 sites varied by >75% in one or more
monkeys, and 10 sites varied in at least half of
the animals (Fig. 2B). A highly parallel Env
evolutionary trajectory was observed between
the human host and monkey RM6072. Unex-
pectedly, the substituted amino acids at many
of the variable sites were identical in human
and rhesus sequences, including mutations
that created potential N-linked glycans (PNGs)
that filled “glycan holes” (Fig. 2C and fig. S3)
(45). The conserved patterns of amino acid
substitutions in CH505 Envs were very differ-
ent from mutational patterns of Env sequen-
ces observed in the human subject CH848 and
in RMs infected by SHIVs bearing the CH848
Env (fig. S4). Twenty-nine variable sites were
identified in Envs from the human subject
CH848, and 16 of these were shared in SHIV.
CH848-infected RMs, often with identical
amino acid substitutions, including some that
altered the distribution of PNGs and glycan
holes (fig. S5). As a control, we compared pat-
terns of Env evolution in CH848-infected RMs
against selected sites in the CH505-infected
human and vice versa. Many fewer sites of
shared evolution were observed in both dis-
cordant Env pairings (P < 0.01 for both com-
parisons) (fig. S6). The Env strain-specific
mutational patterns observed in humans and
rhesus animals infected by CH505 and CH848
viruses were different still from the CAP256SU
infected RMs in which 45 selected sites were
identified, including 25 that were shared in
more than one animal (figs. S7, A and B, and
S8). A comparable Env-wide analysis in the
human subject CAP256 could not be per-
formed because this individual was infected
and then superinfected by two widely diver-
gent viruses whose progeny underwent ex-
tensive recombination (22) (fig. S7C). Still, we
could identify common sites of positive se-
lection in human and rhesus sequences, in-
cluding a key mutation at residue 169 that
led to virus escape from V2 apex C-strand tar-
geted bNAbs (fig. S7D). Overall, we identified
89 selected Env residues in CH505, CH848, or
CAP256SU T/F viruses, and 50 of these sites
were shared among humans and RMs that
were infected by viruses bearing the same Env
strain. This result was markedly different for
individuals infected by different virus strains
where no more than 6 of these 89 variable sites
were shared by any CH505, CH848, or CAP256SU
pairing (P < 0.0001), and only 14 sites were
shared overall (P < 0.0001). These findings
thus show distinctive examples of conserved
strain-specific evolution of HIV-1 Env sequences
in humans and RMs.
Another type of Env variation common in
HIV-1 infection that can affect immune recog-
nition results from insertions and deletions
(indels). Indels occur as a consequence of
polymerase slippage and template switching
during reverse transcription of the retroviral
genome (46, 47). Indels are most commonly
observed in the surface-exposed variable loops
1, 2, 4, and 5, which can better accommodate
changes in sequence length than structural or
enzymatic elements of the viral proteome; this
was the case for human subject CH505 and
macaques infected by SHIV.CH505 (Fig. 2A).
Notably, we found multiple examples of iden-
tical indels between human and rhesus CH505
sequences and among CH505 sequences from
different RMs (Fig. 2D and fig. S9). A total of
six distinct V1 indels, three V4 indels, and
one V5 indel were identified in sequences from
human subject CH505 that were replicated
exactly in sequences from one or more mon-
keys. Two of these indels were found in the
human and in all six monkeys. Additional iden-
tically replicated indels in V1 and V5 were
identified only in monkeys. Indels were also ob-
served in sequences from human subject CH848
(figs. S4 and S10). An exact replica of all but
one of these indels was present in one or more
macaques. Additional indels were replicated
in multiple animals (fig. S10). None of the in-
dels in CH505 sequences were found in CH848,
and neither set was found in CAP256SU se-
quences. In both human and rhesus, indels
generally lengthened or shortened variable
loops in a direction that approached the me-
dian lengths of globally circulating viruses
(figs. S9 and S10). We performed statistical
analyses to determine the likelihood that the
conserved sets of Env-specific indels that we
observed in different individuals could occur
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
5 of 19
RES EARCH | R E S E A R C H A R T I C L E
by chance and that likelihood was estimated
to be vanishingly small (see “Extended Dis-
cussion” in the supplementary materials).
NAbs select for mutations in Env
CH505
To examine whether NAb-mediated selection
was a primary driving force in the Env-specific
evolutionary patterns that we observed, early
strain-specific NAb responses and later bNAb
responses were mapped using site-directed
mutagenesis to introduce observed muta-
tions into neutralization-sensitive autologous
and heterologous Envs. We then tested these
mutant Envs, comparing against their wild-
type counterparts, for sensitivity to polyclonal
plasma and mAbs isolated from the infected
RMs or human subjects. Variable sites in the
CH505 Env (Fig. 2, A and B) were interrogated
alone or in combination for their effects on
Env sensitivity to autologous antibodies. Aside
from variable residues in the leader sequence
of gp160 that generally represent CTL escape
mutations (37, 43), CTL epitope reversions to
global consensus at or near residue 417 (4),
and mutations at residues 300, 620, and 640
that occurred inconsistently and late in the
course of infection, most of the variable res-
idues were found to represent escape muta-
tions from autologous NAbs (fig. S11A). These
included residues 234 and 334, where muta-
tions restored PNGs that filled T/F glycan
holes (fig. S3), loop D residues 279 and 281
involved in CD4 binding, CD4 contact residues
in V5 (460/DV5), and residue 130. The tempo-
ral appearance of these NAbs coincided with
the appearance of phenotypically demonstra-
ble NAb escape mutations in Env (Fig. 2, A and
B). These results were corroborated by neu-
tralization patterns of V3-targeted mAbs DH647
and DH648 and CD4bs-targeted mAbs DH650.
UCA (UCA, unmutated common ancestor) and
DH650 (all isolated from SHIV.CH505-infected
RM6072) and the human CD4bs-targeted mAbs
CH235.UCA, CH235.IA3, and CH235.9 isolated
from human subject CH505 (6) (fig. S11, B and
C). Human subject CH505 developed two co-
operative lineages of CD4bs antibodies that
targeted epitopes that included loop D res-
idues 279 to 281 and residues in V5, eventually
leading to neutralization breadth by both anti-
body lineages (4, 6). In RM6072, a mAb lineage,
called DH650, was isolated that targeted these
same epitopes, but it never developed neutrali-
zation breadth, a finding for which we found a
structural explanation (see below).
In addition to the strain-specific NAb re-
sponses that we identified in SHIV.CH505-
infected RMs, we found in two CH505 animals
(RM5695 and RM6070) neutralizing anti-
bodies that targeted heterologous tier 2 viruses
(Fig. 1 and table S2). These bNAbs were first
detected at weeks 8 (RM6070) and 16 (RM5695)
after infection, and their development was tem-
porally associated with the appearance of
mutations in the Env V2 apex, including res-
idues 166 or 169 (Fig. 3A). These are common
contact residues of human V2 apex bNAbs
(5, 8, 48). A minor fraction of sequences lost
PNGs at 156 and 160. By 24 weeks after infec-
tion in RM5695 and 32 weeks after infection
in RM6070, most circulating virus contained
mutations at residues 166 or 169; by 36 to
48 weeks after infection, all sequences were
mutant at one or the other of these positions.
We corroborated this single genome sequence
evidence of strong virus selection in the cen-
tral cavity of the V2 apex of RM5695 by next-
generation sequencing, which revealed that
99.3% of 10,000 sequences sampled between
weeks 48 and 64 after infection contained
mutations at residues 166 or 169. When these
residues were mutated in the wild-type ver-
sions of heterologous primary virus strains
T250, Q23, MT145K, 246F, or BG505 that
were otherwise neutralized by RM5695 and
RM6070 plasma, neutralization was abrogated
(Fig. 3C and fig. S12). Neutralization of these
heterologous strains was variably dependent
on the glycan at N160 (fig. S13), similar to what
has been reported for V2 apex bNAbs in sub-
ject CAP256SU (49). These findings indicated
the presence of potent V2 apex C-strand tar-
geted bNAbs in RM5695 and RM6070. In
summary, the mapping of autologous and het-
erologous NAb responses in RMs infected by
SHIV.CH505 indicated that most mutations in
Env that could not be ascribed to CTL selec-
tion were the result of NAb selection.
CH848
We also examined Env mutations shared be-
tween SHIV.CH848-infected RMs and the
HIV-1–infected CH848 human subject as po-
tential sites targeted by NAbs. The earliest
Env substitution in the human subject that
was also mutated in sequences from all six
RMs was at position R336, which is surface-
exposed on the HIV-1 trimer. The R366G sub-
stitution in RM6167 occurred at week 10 after
SHIV infection when autologous NAb titers to
the T/F SHIV were 1:220. NAb titers to a site-
directed mutant of the T/F virus that con-
tained the R366G substitution fell to 1:72
(fig. S14A), indicative of an early immuno-
dominant epitope-focused NAb response.
Autologous NAb titers in RM6167 increased
to 1:435 by week 36, and this was accompa-
nied by strong selection on the evolving virus
quasispecies resulting in deletions in V1 and
V5 (figs. S4 and S10) and amino acid substi-
tutions at positions indicated in fig. S4. When
these consensus mutations were introduced
into the CH848 T/F Env and tested for neu-
tralization sensitivity to week 10 and 36 plas-
ma, the mutant virus showed complete escape
(fig. S14A). In human subject CH848, an early
strain-specific NAb response targeted many of
these same sites including V1 (7). The virus
quasispecies in subject CH848 escaped by
deleting ~10 amino acids in V1, which was fol-
lowed by the development of the V3 glycan–
focused bNAb lineage DH270 (7, 50). Notably,
the same sequence of events occurred in SHIV.
CH848-infected animals RM6163 and RM6167
(fig. S10). V1 was first targeted by an early
autologous NAb response (fig. S14), followed
by deletions in V1, which was in turn followed
by the development of V3 glycan–targeted
bNAbs (Fig. 1A and table S2). We mapped the
epitopes of these bNAb responses to the glycan
at residue N332 and the Gly-Asp-Ile-Arg (GDIR)
motif at positions 324 to 327 (Fig. 4B). Con-
sistent with this epitope mapping, we identified,
in the evolving viral quasispecies of both RMs,
mutations at residues 332 to 334 and GDIR
residues 324 to 327 (Fig. 4A). Like the proto-
typic human V3 glycan bNAbs DH270, PGT121,
and PGT128, the bNAbs that we identified in
RMs 6163 and 6167 were strictly N332 depen-
dent, and mutations in spatially associated sur-
face residues V295, N301, N138, and N133 had
similar effects on neutralization potency of both
rhesus and human bNAbs (Fig. 4, B and C).
CAP256
In human subject CAP256, recombination be-
tween PI and SU lineage sequences (fig. S7C)
precluded a gp160-wide Env analysis. We fo-
cused instead on mutations in and near the V2
C-strand, because the human subject and two
of six RMs (40591 and 42056) developed V2
apex–targeted bNAbs (Fig. 1A and table S2). In
both human and RMs, very similar patterns
of Env V2 sequence evolution occurred (Fig. 3B).
These mutations included substitutions at the
canonical residues 160, 166, 169, and 171, shown
to be contact residues for other prototypical
human V2 apex bNAbs (5, 8, 51–53). When in-
troduced into heterologous neutralization-
sensitive Envs, mutations at 166, 169, and 171
abrogated neutralization by plasma from RMs
40591 and 42056 as it did for control human
bNAbs PGT145 and VRC26.25, the latter having
been isolated from subject CAP256 (49, 53)
(Fig. 3C). In RMs 40591 and 42056, we ob-
served that V2 apex–targeted antibodies were
variably dependent on binding to the glycan
at Env residue 160 for neutralizing activity, a
pattern that was similar to antibodies from
animals RM5695 and RM6070 and human
subject CAP256 (49). This variable dependence
on N160 for bNAb activity is further shown in
fig. S13 and is discussed in the supplemen-
tary materials. Overall, the molecular and
temporal patterns of Env sequence evolution
reported here for SHIV-infected macaques,
and previously in humans (18, 37–39, 43), high-
light the utility of dynamic measurements of
localized sequence variation as a highly sen-
sitive indicator of epitope-specific adaptive
immune responses.
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
6 of 19
RES EARCH | R E S E A R C H A R T I C L E
A
B
RM5695 (CH505)
156
166
160
169
171
Human (CAP256)
156
160
166
169
171
RM6070 (CH505)
166
156
160
169
171
D
RM40591 (CAP256SU)
RM42056 (CAP256SU)
156
160
166
169
171
156
160
166
169
171
C
N156
N160
164
166
169
170
171
172
Strand-C
Relative ID50
Relative IC50
RM
5695
RM
6070
V2 apex bNAbs
Virus Mutation
w56
plasma
w36
plasma
VRC26
.25
PGT
145
WT
1
N160K
>10
Q23.17
R166S
<0.1
K169E
<0.1
1
>10
<0.1
<0.1
K171A
<0.1
0.2
1
1
<0.1
<0.1
0.2
1
<0.1
<0.1
<0.1
<0.1
Relative ID50
Relative IC50
RM
40591
RM
42056
V2 apex bNAbs
Virus Mutation
w112
plasma
w88
plasma
VRC26
.25
PGT
145
WT
N160K
T250-4
R166S
K169E
K171E
1
2.1
0.2
0.2
0.1
1
>10
<0.1
<0.1
1
1
<0.1
<0.1
1
<0.1
<0.1
<0.1
<0.1
<0.1
<0.1
>3 fold sensitive >3 fold resistant
Fig. 3. Broadly neutralizing antibody responses in four rhesus macaques
map to the V2 apex. (A) Single genome sequencing (N = number of sequences;
w = weeks after SHIV infection) in SHIV.CH505-infected animals RM5695 and RM6070
reveals selection and fixation of mutations in and immediately proximal to strand C
and additional mutations that eliminate PNG sites at 156 and 160. (B) Sequential Env
sequences from human subject CAP256 and SHIV.CAP256SU-infected RMs 40591
and 42056 showed selection and fixation of mutations in and proximal to strand C,
with fewer mutations eliminating the PNG site at 160. (C) Heterologous neutralization
of Q23.17 and T250-4 Env-pseudotyped viruses by rhesus plasma is drastically
reduced (expressed as fold-change from wild type) by mutations at residues 166, 169,
and 171, similar to human V2 apex bNAbs. Enhanced neutralization against the
N160K mutants is illustrated in fig. S13 and further described in the supplementary
materials. (D) Escape mutations from rhesus V2 apex bNAbs are displayed on the
5FYL structure of BG505.N332 SOSIP (top view).
Homologous B cell responses in humans
and macaques
The observation that homologous Envs evolved
in similar molecular patterns in humans and
macaques could be explained by limited num-
bers of antigenic sites accessible for antibody
binding and restricted pathways of virus es-
cape. In addition, homologous human and
rhesus germline B cell receptors could favor
binding to common HIV-1 Env epitopes and
follow similar patterns of Ab-Env coevolution.
We explored the latter possibility by isolating
and characterizing neutralizing mAbs from
two RMs that were infected by SHIV.CH505.
We selected these animals for study because
both exhibited favorable virus replication ki-
netics, comparable early NAb responses, and
an overall pattern of Env evolution similar to
that in the human subject CH505; nonetheless,
one animal (RM5695) developed bNAbs and
the other (RM6072) did not. We asked what
might be the genetic and structural similar-
ities and dissimilarities between the NAbs
elicited in these animals.
CD4bs-directed antibodies
Two broadly neutralizing mAb lineages (CH235
and CH103) targeting the HIV-1 CD4bs were
previously isolated from human subject CH505,
and their evolutionary pathways from germline
receptors to mature antibodies were determined
(2, 4, 6). In animal RM6072, serological analysis
of early plasma samples suggested the presence
of CD4bs-targeted Abs on the basis of selective
Ab binding to CH505 T/F gp120 and resurfaced
Env core but not to their isogenic D371I mu-
tants, which do not bind CD4 (fig. S15A). Ad-
ditional evidence of CD4bs-targeted Abs in
the plasma of this animal included competi-
tive blocking of soluble CD4 (sCD4) and CH235
and CH106 mAbs (fig. S15B). We sorted mem-
ory B cells from RM6072 from longitudinal
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
7 of 19
RES EARCH | R E S E A R C H A R T I C L E
A
Human (CH848)
V3
RM6167 (CH848)
V3
295
301
324
327
332
295
301
324
327
332
C
B
325
326
328
N133
N137
N156
N295
N301
N332
N339
N386
N392
RM6163 (CH848)
V3
295
301
324 327
332
Relative ID50
Relative IC50
RM6163
RM6167
V3 glycan bNAbs
Virus Region Mutation
w80 plasma w80 plasma DH270
PGT128
PGT121
WT
N133D
N138D
N156K
V295N
N301D
N332D
D325N
I326T
N339D
N386D
N392D
V1
V3
6
7
1
1
e
C
V3 GDIR
C3
V4
1.0
2.5
8.4
2.7
0.3
0.3
<0.1
0.2
0.3
1.1
1.9
1.0
1.0
1.8
4.6
3.9
0.3
0.8
0.2
0.5
0.5
2.1
1.9
2.8
1.0
0.4
0.3
0.5
5.8
5.8
>50
>50
1.0
0.6
0.8
0.6
1.0
0.3
0.2
0.7
2.0
27.5
1.0
1.0
1.5
0.5
2.0
2.0
>300
>1000
0.7
0.8
0.6
0.7
0.7
1.0
1.5
1.0
1.5
1.0
>3 fold resistant
>3 fold sensitive
Fig. 4. Broadly neutralizing antibody responses in two rhesus macaques map
to the V3 glycan high-mannose patch. (A) Sequential Env sequences from human
subject CH848 and SHIV.CH848-infected RMs 6167 and 6163 showed selection
and fixation of mutations in the GDIR motif. Serine-to-asparagine substitutions at
position 334 shift the PNG from residue 332 to 334 in the human subject and
in RM6167. (B) Heterologous neutralization of Ce1176 Env-pseudotyped virus by
RMs 6163 and 6167 plasma is reduced (expressed as fold-change from wild type)
by mutations at N332D and V295N, consistent with human V3 glycan bNAbs.
Neutralization of GDIR mutants is reduced by two- to fivefold. Elimination of a glycan
at position N138 enhances neutralization in RMs 6163 and 6167 and for some human V3
glycan bNAbs. (C) V3 glycan bNAb escape mutations are displayed on the 5FYL
structure of BG505.N332 SOSIP (side view) highlighting their close spatial proximity.
time points 20, 24, and 32 weeks after SHIV
infection and selected cells that reacted with
CH505 T/F gp120 but not the D371I mutant.
We isolated a 15-member B cell clonal lineage
designated DH650 (fig. S15, C and D) that se-
lectively bound the autologous CH505 T/F Env
gp120 but not the D371I mutant (fig. S15E).
Some of these mAbs competed with sCD4 and
CH235 and CH106 mAbs for binding CH505
T/F gp120 (fig. S15F). Mature DH650 lineage
mAbs, but not the inferred germline UCA,
bound CH505 T/F gp140 SOSIPv4.1 trimer (fig.
S16A). Most members of the DH650 lineage
neutralized a glycan-deficient mutant of CH505
Env (gly4), and two-thirds of them neutralized
the wild-type CH505 T/F strain (fig. S16B). The
DH650 UCA neutralized neither. None of the
DH650 lineage mAbs neutralized heterologous
viruses (fig. S16C and table S3).
The immunogenetic features of the DH650
lineage mAbs suggest how they recognize HIV-1
Env. The lineage comes from V(D)J recombi-
nation of the macaque VH1-h gene (VH, var-
iable region of immunoglobulin heavy chain;
fig. S15D), which is 91% similar to an ortho-
logous human gene VH1-46 used by the CD4bs
bNAbs CH235 and 8ANC131 (6, 54). DH650
antibodies share key VH residues with CH235
(fig. S17A), which were shown previously to
be contact sites with gp120 Env (6, 55). These
included residue 57, which in both the CH235
UCA and DH650 UCA underwent affinity mat-
uration to R57, which is important for CH235
bNAb activity and is shared among CD4bs
bNAbs using VH1-46 (6, 55). We found this
N57R substitution in DH650 to be essential for
binding to CH505 T/F Env (fig. S17B). Other
DH650 VH1-h gene residues that we found to
be important for Env binding included N35,
Q62, and R72 (Fig. 5 and fig. S17, C and D). A
distinguishing feature of DH650 lineage anti-
bodies was the IGVk2 light chain, which has
an exceptionally long LCDR1 of 17 amino acids
(fig. S17E), which we explored by structural
studies.
The crystal structure of DH650 bound to
the gp120 Env core of the CH505 T/F virus
showed that its interactions with the gp120
CD4bs closely resembled those of the human
CD4bs mAbs CH235, 8ANC131, and VRC01
(Fig. 5, A to D, and table S4). This similarity
included conserved HCDR2-mediated CD4
mimicry and coordination of Env Asp368 by
Arg72. An important difference between the
rhesus and human antibody lineages was in
the light chains (DH650, macaque IGVk2;
CH235, human IGVl3), in which the LCDR1 of
the DH650 light chain was six residues longer
than its CH235 counterpart (fig. S17E). The
structure showed that in the Env-Ab complex,
the CH505 gp120 loop D had undergone a con-
formational change to accommodate the longer
DH650 LCDR1. We inferred from the structure
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
8 of 19
RES EARCH | R E S E A R C H A R T I C L E
that this shift could occur only because of the
absence of a commonly found glycan at gp120
position 234 in the CH505 T/F virus. Moreover,
addition of that glycan, which occurred in
both the human donor and RM6072 by 30 to
36 weeks after infection, conferred resistance to
DH650 (fig. S11C) and likely eliminated selective
pressure in the monkey to enforce deletions in
LCDR1. Thus, infection of RM6072 with SHIV.
CH505 expanded a B cell clone bearing an
antigen receptor encoded by the RM VH1-h
gene segment that is orthologous to the hu-
man VH1-46 gene. This B cell lineage under-
went affinity maturation, including selection
for a critical R57 VH mutation that is also
found in the human CH235 bNAb lineage.
It has been reported that the maturation of
VRC01-class CD4bs bNAbs generally includes
deletions in LCDR1 or mutations to glycine
that confer flexibility (56). Evolution of the
DH605 lineage in RM6072 failed to include
deletions or flexibility in LCDR1 and hence
neutralization breadth did not develop.
V2 apex–directed antibodies
RM5695, infected with an early SHIV.CH505
plasma virus quasispecies (see supplementary
materials), developed bNAbs that, on the basis
of Env escape patterns in vivo and neutraliza-
tion phenotypes of site-directed Env mutants
in vitro, targeted the V2 apex (Fig. 3). We used
an unbiased FACS strategy to isolate 20,000
individual memory B cells from peripheral
blood mononuclear cells (PBMCs) 65 weeks
after SHIV infection and expanded these cells
in 384-well plates. Culture supernatants were
screened for neutralizing activity against two
heterologous virus strains (T250-4 and Q23.17).
Five wells scored positive, and paired heavy
and light chain immunoglobulin genes were
successfully amplified from cellular mRNA
from four of them (Fig. 6A and fig. S18). All
four rhesus mAbs belonged to a single lineage
that we designated RHA1.V2 (rhesus HIV anti-
body 1 targeting a V2 epitope, with lineage
members RHA1.V2.01-04). The IGVH and IGVL
genes of the four RHA1.V2.01-04 mAbs were
closely related, with maximum nucleotide di-
versities of 5.3 and 3.9%, respectively. We used
NextGen sequencing to characterize immu-
noglobulin transcripts expressed by naïve
IgD+IgM+IgG− B cells from RM5695 and used
IgDiscover (57) and a recent database of RM
Ig alleles (21) to identify a personalized im-
munoglobulin gene repertoire (table S5). From
this analysis, we determined the germline origins
of the mature RHA1 mAbs to include a novel
IGVH4 allele, IGHV4-ABB-S*01_S8200 (fig.
S18A), and IGlV1-ACN*02 (fig. S18B). Somatic
hypermutation within the lineage was mod-
est, with nucleotide divergence from germline
of 7.1 to 8.5% for VH and 6.2 to 6.6% for VL
(Fig. 6A). These values are comparable to some
human V2 apex bNAbs, which typically have
lower frequencies of VH mutations than mem-
bers of other human bNAb classes. The mature
rhesus bNAb heavy chains contained a two–
amino acid insertion within HCDR1 (figs.
S18A and S19A) and a 24-residue-long HCDR3
(IMGT numbering) that was derived from re-
arrangement of the VH4-ABB-S*01_S8200,
DH3-9, and JH2-P genes plus six nontemplated
amino acids (figs. S18A and S19B). This HCDR3
was rich in aromatic and negatively charged
residues, like HCDR3s of human V2 apex bNAbs
(fig. S19C). Despite this long HCDR3, the ma-
ture rhesus bNAb RHA1.V2.01 was not auto-
or polyreactive (fig. S20).
The HCDR3 of the RHA1 lineage antibodies
was similar in length to the human V2 apex
bNAb PCT64-35S (24 versus 25 amino acids,
respectively), and the two broadly neutralizing
mAbs contained conserved motifs within their
respective HCDR3s including a negatively
charged “DDY” segment (fig. S19C), which in
PCT64-35S was shown to be tyrosine-sulfated
and to interact with positively charged V2 apex-
hole residues of Env (8, 58). All four rhesus
A
DH650HC DH650LC
B
o
90
CDRH2
CD4bl
CH505 T/F
gp120 core
C
D
CD4bl
VRC01
8ANC131
CH235
DH650
DH650
Heavy chain
V-domain
Fig. 5. Structure of Fab DH650 bound to CH505 gp120 core mimics human CD4bs antibodies.
(A) Ribbon representations of Fab DH650 heavy chain (purple) and light chain (pink) and CH505 gp120
core (blue). The heavy-chain CDRH2 is in magenta, the light-chain CDRL1 is in orange, and the CD4 binding
loop (CD4bl) is in red. Loop D of gp120, which shifts from its position in other complexes to accommodate
the long CDRL1 of DH650, is in yellow; the first N-acetyl glucosamine of glycan 276, which also shifts to
accommodate CDRL1, is in stick representation (carbon, yellow; nitrogen, blue; oxygen, red). (B) 2D class
averages after negative stain EM analysis of DH650-CH505 DS-SOSIP trimer showing three Fabs bound per
trimer. (C) Superposition of Fab DH650-CH505 gp120 complex with other CD4 binding site Fab-gp120
complexes, with the essentially identical structures of gp120 as the common reference. The view corresponds
to the orientation in the right-hand part of (A). Only the Fv modules of the Fabs are shown. Complex of
DH650 (this study) is shown in gray, VRC01 (PDB ID 3NGB) in cyan, CH235.12 (PDB ID 5F96) in magenta,
and 8ANC131 (PDB ID 4RWY) in green. DH650-bound gp120 core Cas have root mean square deviations
of 0.78, 0.82, and 0.71 Å from those of the VRC01-, CH235.12-, and 8AC131-bound gp120 structures,
respectively. (D) Left panel: Recognition of CD4 binding loop of CH505 T/F gp120 by variable heavy domain
of DH650 (top). Coordination of Asp368 by Arg72 of DH650 (bottom). Right panel: Comparison with
other CD4 binding site antibodies (6).
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
9 of 19
RES EARCH | R E S E A R C H A R T I C L E
rhesus
mAb
VH gene
RHA1.V2.01
IGHV4-ABB*0
1_S8200
RHA1.V2.02
IGHV4-ABB*0
1_S8200
RHA1.V2.03
IGHV4-ABB*0
1_S8200
RHA1.V2.04
IGHV4-ABB*0
1_S8200
%mut
(nt)
8.5%
8.5%
7.5%
7.1%
Indels
HCDR1 +2
HCDR1 +2
HCDR1 +2
HCDR1 +2
HCDR
lengths
10.8.24
10.8.24
10.8.24
10.8.24
HCDR3
V gene
C ARKGEDFYEDDYGQYFTAGWFFDLW
C ARKGEDFYEDDYGQYFTAGWYFDLW
C ARKGEDFYEDDYGQYFTAGWYFDLW
C ARKGEDFYEDDYGQYFTAGWFFDLW
IG V1-ACN*02
IG V1-ACN*02
IG V1-ACN*02
IG V1-ACN*02
%mut
(nt)
6.2%
6.6%
6.6%
6.2%
LCDR
lengths
8.3.8
8.3.8
8.3.8
8.3.8
LCDR3
CQCYDSSVLF
CQCYDTTVLF
CQCYDSNVLF
CQCYDTTVLF
A
B
C
D
RHA1.V2.01
Clades A/ACD/AD
Clade AG
Clade G
Clades
C/BC
Clade AE
Clades
D/CD
Clade B
Breadth: 49%
Median IC50: 0.35 ug/mL
Clade AC
E
F
F
<0.10
0.10 -1.00 1.00-10.00 10.00-50.00
IC50 (µg/ml)
RHA1
RHA1
Light
Light
HIV-T250.4
wild type N160K R166S K169E K171E
RHA1.V2.01
PGT145
PCT64-35S
VRC26.25
PG9
CH01
>50
>50
>50
0.012
0.003
0.009
0.0008 0.0008
0.004
0.032
>50
>50
>50
>50
>50
>50
0.008
0.052
>50
>50
>50
>50
19
26
15
0.59
0.88
0.001
1.4
>50
SIV-MT145K
wild type N160D R166S K169E K171Q
RHA1.V2.01
PGT145
PCT64-35S
VRC26.25
PG9
CH01
0.081
0.024
0.20
>50
>50
>50
0.0008 0.0008
0.045
0.26
>50
>50
>50
>50
>50
>50
0.047
0.23
>50
>50
>50
>50
>50
5.8
0.48
0.24
0.55
0.001
0.24
47
HIV-CH505 in vivo immunotype
wild type R166K R166G R166S R169K
R169T
RHA1.V2.01
RHA1.V2.02
RHA1.V2.03
RHA1.V2.04
0.29
0.13
0.63
0.52
1.4
0.59
2.5
2.4
>50
>50
>50
>50
>50
>50
>50
>50
1.5
0.24
2.0
2.5
>50
>50
>50
>50
RHA1.V2.01
RHA1
RHA1
Heavy
Heavy
D98
Side view
K169
K169
D100b
R166
E100a
R166
R166
R166R16666
Y100d
D100c
K121
Top view
G
G
Side view
Side view
R166K/
R169K
>50
>50
>50
>50
RHA1 Heavy
PGT145 Heavy
RHA1 Heavy
PCT64 Heavy
Fig. 6. Rhesus bNAb lineage RHA1 targets the V2 apex, is broadly reactive, and
contains a sulfated tyrosine in HCDR3 that shows precise chemical mimicry to
human bNAbs PCT64-35S and PGT145. (A) Immunogenetic characteristics of four
RHA1 lineage bNAbs. A key feature is the 24–amino acid–long HCDR3 that
contains an acidic EDDY core motif. (B) Neutralization breadth and potency of RHA1.
V2.01 against a 208-strain global virus panel. The dendrogram depicts phylogenetic
relatedness of the HIV-1 Envs tested. (C) Neutralization expressed as IC50 (micrograms
per milliliter) of wild-type heterologous viruses and their V2 apex mutants by RHA1.V2.01
and by prototypic human V2 apex bNAbs. Like most human V2 apex bNAbs, RHA1.
V2.01 is strictly dependent on N160 and positively charged residues at 166 and 169.
(D) Neutralization of CH505 T/F (wild-type) virus and C-strand variants, or “immuno-
types,” that evolved in vivo in RM5695. Predominant mutations at 24 weeks after SHIV
infection in RM5695 were R166K or R169K; at 48 weeks, R166S, R166G, R169T, R166K,
and R169K were prevalent; at 64 weeks R166S or R169T became fixed (see Fig. 3A).
Panel (D) shows progressive loss in neutralization sensitivity to RHA1 bNAbs by
the evolving CH505 Envs, beginning with CH505 T/F wild type (most sensitive),
CH505.R166K or R169K (intermediately sensitive), and ending with CH505.R166G,
R166S, R169T, or R166K+R169K (all resistant). Results are expressed as IC50
(micrograms per milliliter). (E) Neutralization fingerprint for RHA1.V2.01 shows it to
cluster within the PGT145 class. (F) Cryo-EM structure (side view) of RHA1.V2.01
in complex with BG505 DS-SOSIP at 4-Å resolution. Inset (top) highlights
electrostatic contacts of the HCDR3 with Env protomers, including interactions of
the tyrosine-sulfated 100d residue with Env K121. Inset (bottom) shows the trimer
apex cavity highlighting glycans at N160 and the C-strands. (G) Alignment of
gp120 from the complex trimer structure with RHA1.V2.01 Fab to trimer complexes
with human Fabs PGT145 (PDB-5V8L) and PCT64-35S (modeled with PDB-6CA6
fit to EMD-7865) reveals alignment of tyrosine sulfated residues within the
respective HCDR3 tips, which insert into the hole at the V2 apex of the Env trimer.
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
10 of 19
RES EARCH | R E S E A R C H A R T I C L E
mAbs were tested for neutralization against
the 19-member global panel of tier 2 viruses
and showed similar patterns of reactivity,
neutralizing 15 to 17 strains (fig. S21). One
antibody (RHA1.V2.01) was tested for neutrali-
zation against a 208-member global virus panel
and was found to neutralize 102 heterologous
virus strains, or 49%, at a maximum concen-
tration threshold of 50 mg/ml (Fig. 6B and
fig. S22). Neutralization of heterologous virus
strains depended on Env residues N160, R/
K166, and R/K169, with partial dependence
on K171 (Fig. 6C and fig. S23). This precise
pattern of neutralization sensitivity to N160,
R/K166, and C-strand residue R/K169 and
K171 mutations was shared by the human
V2 apex bNAbs PCT64-35S and PGT145 but
was different from that of PG9, VRC26.25,
and CH01. CH505 Envs that evolved in
RM5695 in vivo coincident with the develop-
ment and maturation of RHA1 lineage anti-
bodies showed evidence of strong selection at
residues 166 and 169 (Fig. 3A). Introduction
of these mutated residues into the CH505 T/
F Env resulted in loss of neutralization
sensitivity to RHA1 mAbs (Fig. 6D).
Bioinformatical comparisons of human and RM
V2 apex bNAbs
Using a neutralization “fingerprint” analysis
(59), which compares the potency of individ-
ual bNAbs against a large set of HIV-1 strains
(fig. S22), we observed clustering of RHA1.V2.01
within the PGT145 class of V2 apex bNAbs that
includes PCT64-35M and PGDM1400 (Fig. 6E
and fig. S19, D and E). In a hierarchical clus-
tering analysis of the neutralization profiles
of RHA1.V2.01 and other prototypic human
V2 apex bNAbs measured against the 208-virus
panel, RHA1.V2.01 grouped most closely with
PCT64-35M (fig. S24A). This finding was strongly
supported statistically by the overlap of vi-
ruses that were sensitive or resistant to those
antibodies [Fisher’s exact test P = 2 × 10−16;
odds ratio = 13.14; accuracy ratio = 0.78 (all
concordant = 1; none = 0)] (fig. S24B) and by
correlation of median inhibitory concentra-
tion (IC50) titers of RHA1.V2.01 and PCT64-
35M (coefficient of determination R2 = 0.4467;
P = 1.4 × 10−10) compared with other V2 apex
bNAbs (fig. S24C). We next examined neutral-
ization profiles across different HIV-1 group
M subtypes (fig. S25). We found that RHA1.
V2.01 was more subtype-independent than any
of the human V2 apex bNAbs but was other-
wise most similar to PCT64-35M and CAP256.
VRC26. Finally, we performed a neutralization
signature analysis using GenSig (www.hiv.lanl.
gov/content/sequence/GENETICSIGNATURES/
gs.html) (fig. S26). Neutralization “signatures”
identify individual Env residues that contrib-
ute directly or indirectly to antibody binding,
including those potentially involved in select-
ing for affinity maturation (60). The signa-
ture analysis was performed using Fisher’s
test with a binary IC50 threshold greater or
less than 50 mg/ml and a Wilcoxon test that
measures the difference in IC50 distributions
with and without a given amino acid residue.
High-confidence signature sites were defined
as those meeting at least two of three criteria:
(i) contact sites; (ii) at least one phylogenet-
ically corrected signature at the site; or (iii)
at least one signature at the site that had a
false discovery rate q < 0.1. For the rhesus mAb
RHA1.V2.01, statistically robust signatures were
identified at residues 130, 160, 166, 167, and
169. Among all of the V2 apex bNAbs analyzed,
only PCT64-35M shared all five of these signa-
ture sites.
Structure of V2 apex bNAb RHA1.V2.01
The structure of mAb RHA1.V2.01 in complex
with the BG505 DS-SOSIP Env trimer, deter-
mined by cryo–electron microscopy (cryo-EM)
at 3.9-Å resolution, showed notable similarity
to PGT145 and PCT64-35S (Fig. 6, F and G,
and table S6). These antibodies bind Env with
a 1:1 stoichiometry near the trimer threefold
axis and are surrounded by the three N160
glycans. Their respective HCDR3s adopt a
needle-like antiparallel b-hairpin conforma-
tion that extends from the combining surface
of the Fab and inserts into a cationic hole at
the trimer apex. The N-terminal ends of each
of the three C-strands abut the apex hole and
are oriented perpendicular to the inserting
HCDR3. Like PCT64-35S and PGT145, the
acidic EDDY motif of RHA1.V2.01 was tyrosine-
sulfated and made key contacts with Env
residues 121, 166, and 169. (Fig. 6F, boxed in-
sert). When the Env-bound structures of RHA1.
V2.01, PCT64-35S, and PGT145 were overlaid,
the respective EDDY motifs aligned at the tips
of their respective HCDR3 loops around the
b-hairpin turn (Fig. 6G). Otherwise, the over-
all Fab orientations differed, indicating the
HCDR3 tip structural mimicry to be the main
source of the neutralization similarity among
these antibodies. The HCDR1 of RHA1.V2.01,
which contained a nontemplated two–amino
acid insertion in addition to other strongly
selected mutations, was sandwiched between
the Env N160 glycans of two protomers and
proximal to the C-strand of one, with buried
surface area of 52 and 49 Å2, respectively, for
the two glycans and a key electrostatic inter-
action between D29 and K171 (Fig. 6F and
figs. S19A and S27F). Thus, the V2 apex bNAb
lineage in RM5695 exhibits genetic, chemical,
and structural solutions to epitope recognition
that are shared with human V2 apex targeted
bNAbs, especially PCT64-35S and PGT145.
Discussion
A principal finding of this study is that SHIVs
bearing primary T/F HIV-1 Envs elicit strain-
specific and heterologous NAbs in RMs that
can replicate to a considerable degree responses
to HIV-1 in humans. This mimicry includes
the frequency, kinetics, titers, immunogenetics,
structures, and target epitopes of elicited anti-
bodies; structural and chemical features of
epitope recognition; and coevolutionary path-
ways of antibody maturation and Env escape.
All are key features to be considered in vaccine
design. Our findings add substantially to earlier
reports of sporadic neutralization of heterolo-
gous tier 2 viruses elicited in RMs by SHIVs
bearing lab-adapted or animal-passaged HIV-1
Envs (61–64) or Env immunogens (65–72). The
current results, together with a recent report
by Wang and colleagues (73), show how closely
neutralizing antibody responses in RMs can
mirror responses in humans and indicate the
extent to which protective responses elicited by
reverse engineered or lineage-based vaccines in
RMs might be expected to predict human re-
sponses to candidate vaccines (10, 11, 74, 75).
While bNAbs from animal RM5695 were
exceptional, the breadth and potency of bNAbs
from the other six animals were more in keep-
ing with a large cross section of chronically
infected humans who were studied for neu-
tralization breadth generally after many years
of infection (28, 33, 34). In those studies, “elite
neutralizers”—those with the most broadly
reactive antibodies and from whom many
of the best bNAb mAbs have been isolated—
actually represented a very small fraction
(<5%) of infected individuals. What is most
notable about our study is not the overall breadth
and potency of rhesus bNAb responses but
the fact that we observed bNAbs arising in 7
of 22 prospectively studied animals after less
than 1 or 2 years of SHIV infection and that
the epitope specificities of these antibodies
could be mapped to canonical HIV-1 bNAb sites
in six. These findings suggest that the rhesus
macaque is a favorable model for HIV-1 vaccine
studies where the key variables of priming,
boosting, affinity maturation, and antibody du-
rability can be explored rapidly and iteratively.
An unexpected observation was the extent
to which Env evolution in macaques recapitu-
lated evolution of homologous Envs in human
infections. Similarities included site-specific
and amino acid–specific mutations and iden-
tities or near-identities of indels. These sim-
ilarities likely resulted from: (i) the highly
evolved and functionally constrained nature
of primary T/F Env trimers, (ii) limited sites
of antibody accessibility and variable fitness
costs of escape mutations, and (iii) homolo-
gous germline B cell responses in different
animals and humans to conserved Env epi-
topes. Equally surprising were the genetic
and structural similarities between rhesus
and human antibodies that targeted CD4bs
or V2 apex epitopes and their conserved mech-
anisms of epitope recognition. This included
the HCDR2-mediated CD4 mimicry of the
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
11 of 19
RES EARCH | R E S E A R C H A R T I C L E
rhesus antibody DH650 and the tyrosine-
sulfated, 24-residue-long HCDR3 of the rhesus
antibody RHA1.V2.01, which bound N160 gly-
cans and positively charged Env residues at
positions 121, 166, 169, and 171. Together, the
conserved patterns of Env-specific sequence
variation and the homologous and ortholo-
gous B cell responses in humans and RMs
represent notable examples of convergent
evolution (14) that may aid in the design and
testing of novel HIV-1 vaccines.
Our findings suggest that HIV-1 Envs are
not equal in their propensity for eliciting
epitope-specific bNAb responses. For example,
we found that CAP256SU Env, which elicited
V2 apex bNAbs in human subject CAP256
(3, 5, 22), induced bNAbs of the same specif-
icity in two of six SHIV-infected RMs. CH848
Env, which elicited V3 glycan targeted bNAbs
in a human subject (7), did the same in 2 of 6
SHIV infected RMs. And CH505 Env, which
elicited CD4bs targeted bNAbs in a human
subject (2, 4, 6), induced homologous strain-
specific CD4bs targeted NAbs in RM6072.
CH505 Env also elicited V2 apex targeted NAbs
with variable breadth in vaccinated CH03 heavy
chain Ig knock-in mice and an immunized RM
(69). In the present study, it did so as well in 2
of 10 SHIV infected RMs. In still other SHIV
infected RMs, we have identified fusion pep-
tide targeted bNAbs and broadly reactive and
potent CD4bs bNAbs. These findings highlight
the potential for RMs to develop antibody re-
sponses targeting an array of different canonical
bNAb epitope clusters and suggest a tendency
for certain Envs to preferentially elicit bNAb
responses targeting particular epitopes.
SHIV replication in RMs is the only model
system other than naturally infected humans
where the immunogen (Env) coevolves with anti-
bodies. The high mutability and dynamic rep-
lication of HIV-1 and SHIV result in a constantly
evolving virus quasispecies (9, 18, 38, 39, 42, 43),
which means that Envs with binding affinities
sufficient to drive bNAb lineage affinity mat-
uration are constantly being generated. The
SHIV-infected macaque can therefore be par-
ticularly informative for vaccine design by
enabling the identification and then rapid
testing of Env intermediates that guide the
evolution of germline bNAb precursor B cells
through stages of affinity maturation to ac-
quire breadth and potency. In the CAP256 (5)
and PC64 (8) infected human subjects, viral
sequences showed very similar patterns of Env
evolution at residues 166 and 169, which in
turn were similar to the pattern in the SHIV.
CH505-infected RM5695. Deep sequencing of
RM5695 plasma vRNA covering the V1V2 re-
gion revealed selection focused primarily on
residues 166 and 169 with mutations at these
two sites rapidly and completely replacing the
infecting virus strain. The infecting SHIV.
CH505 virus had an Arg at these two positions,
which evolved progressively to R166K or R169K
and then to R166G, R166S, or R166T (Fig. 3).
The earliest mutants, R166K and R169K, were
approximately fivefold more resistant to the
mature rhesus bNAb mAbs than the infecting
virus, whereas the subsequent R166G, R166S,
and R166T mutants were >100-fold more re-
sistant (Fig. 6D). Thus, sequential Envs that
varied at residues 166 and 169 in animal RM5695
showed progressive phenotypic escape from
V2 apex bNAb antibodies, closely resembling
the viral Env-bNAb coevolution observed in
humans CAP256 and PC64. Cryo-EM analy-
sis of the RHA1.V2.01 mAb provided a struc-
tural explanation for this loss of antibody
recognition by showing that Env residues 166
and 169 were primary electrostatic contacts
with the antibody. Mutations in these two res-
idues in the V2 apex appear to be largely or
solely responsible for driving affinity matura-
tion of diverse antibody lineages to breadth in
multiple rhesus animals and humans within a
relatively short time (<1 year), suggesting that
Env intermediates or “immunotypes” (5) re-
quired for V2 apex bNAb elicitation may be
few and simple. This hypothesis has important
implications for V2 apex–targeted HIV-1 vac-
cine design, which can be tested rapidly in Ig
knock-in mice and outbred macaques using
immunogens designed from V2 apex variants
of CH505 and other primary Envs. The goal of
such research would be to learn the rules
governing consistent bNAb induction in RMs
and then translate these findings to human
studies using SOSIP, mRNA, or other non-
SHIV-based vaccine platforms.
V3 high-mannose patch glycopeptides are
also commonly targeted by bNAbs in HIV-1–
infected humans (33) and are of high interest
for HIV-1 vaccine development (50, 76, 77).
Site-directed mutagenesis coupled with anti-
body neutralization showed that polyclonal
bNAb responses in SHIV.CH848-infected RMs
6167 and 6163 targeted canonical N332 and
324GDIR327 motifs, similar to human V3 glycan
bNAbs (78, 79). Deletions in V1 of SHIV.CH848
sequences preceded the development of V3
glycan bNAbs in both monkeys and human
subject CH848. Long, glycosylated (N133 and
N138) V1 segments obstruct access of V3 gly-
can bNAbs (7, 80), and germline-targeted Env
immunogens with shortened V1 segments de-
pleted of glycans enhance Ab access to V3 high-
mannose patch epitopes (50, 76). Because Env-Ab
coevolution leading to V3 glycan bNAbs gen-
erally requires more extensive somatic hyper-
mutation compared with V2 apex bNAbs (10, 11),
a rhesus model in which vaccinations with
germline-targeted Envs (50, 76, 77) is followed
by infection with SHIVs whose Envs are similarly
targeted, offers a novel strategy for identifying
the “finishing” or “polishing” immunogens nec-
essary for bNAb affinity maturation and an
outbred primate model system to test them.
It is generally believed that the develop-
ment of an effective neutralizing antibody–
based HIV-1 vaccine will require consistent
activation of multiple germline precursor B
cells that express immunoglobulin receptors
specific for one or more of the canonical bNAb
epitope clusters, followed by efficient antigen-
driven selection for antibody affinity matu-
ration (10, 11, 50, 74–77, 81). The present study
demonstrates that the SHIV-infected rhesus
model can inform both of these critical steps
in bNAb elicitation. The fact that only a minority
of SHIV infected animals in the current study
developed bNAbs is a faithful reflection of the
natural prevalence of bNAb responses in HIV-1–
infected humans (28, 33, 34, 82) and further
argues for the relevance of the rhesus model.
A clear limitation of our study is that SHIV
infection is not a viable vaccine strategy for
humans nor is CD8 depletion, which was em-
ployed to increase peak and set point virus
loads. Nonetheless, it should be possible to
combine established immunization platforms
such as Env trimers, outer domain scaffolds,
virus-like particles, or DNA and RNA expres-
sion followed by SHIV infection to identify op-
timized priming and boosting immunogens
that elicit broad neutralization in macaques
as a molecular guide for HIV-1 vaccine design in
humans.
Materials and methods
Nonhuman primates and SHIV inocula
Indian RMs were housed at Bioqual, Inc.,
Rockville, MD, according to guidelines of the
Association for Assessment and Accreditation
of Laboratory Animal Care standards. Exper-
iments were approved by the University of
Pennsylvania and Bioqual Institutional Ani-
mal Care and Use Committees. RMs were se-
dated for blood draws or excisional biopsies of
lymph nodes, anti-CD8 mAb infusions, and
SHIV inoculations. A subset of animals received
an intravenous infusion of either 25 to 50 mg/kg
of anti-CD8a mAb (MT807R1) or anti-CD8b
mAb (CD8beta255R1) 1 week before or at the
time of SHIV inoculation. The intent of admin-
istering anti-CD8 mAbs was to transiently re-
duce CD8+ cells, thereby allowing for higher
peak and set point viral loads. This was done
in only a subset of animals because it was not
known a priori whether this might accelerate
disease progression to an extent that would
preclude long-term follow-up of animals for
bNAb induction. In humans, higher plasma
virus loads and lower CD4+ T cell counts are
correlated with the development of bNAbs
(33, 34). SHIV infections were done intrave-
nously. All animals in this study, with the
exception of RM5695, were inoculated with
molecularly cloned virus [50 or 500 ng p27Ag in
Dulbecco's modified Eagle's medium (DMEM)
or RPMI 1640 with 10% heat-inactivated fetal
bovine serum (FBS)] containing the preferred
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
12 of 19
RES EARCH | R E S E A R C H A R T I C L E
Env375 variant or a mixture of Env375 Env
variants. Animals RM6069, 6070, 6072, 6163,
and 6167 were repurposed from a previous study
of SHIV replication dynamics and immunopa-
thogenesis (9). Animal RM5695 was repur-
posed from a previous study of HIV-1 CH505
Env gp120 protein immunization (23). This
animal was infected with 0.2 ml of SHIV-
positive plasma from monkeys RM6069 (weeks
10 and 20), RM6070 (weeks 10 and 20), and
RM6072 (weeks 4, 10, and 20), for a total plasma
inoculum of 1.4 ml. The rationale for this exper-
iment was to increase early viral diversity and
potentially include antigen-antibody immune
complexes in the inoculum. The genetic com-
position of this plasma virus inoculum, and the
particular viral genomes that were successfully
transmitted to RM5695, can be seen in the
Highlighter and LASSIE panels in Fig. 2, A and
B, and the sequences can be accessed from
GenBank (accession # MT484881-MT484977).
Processing and storage of clinical specimens
All blood samples were collected in sterile
vacutainers containing acid citrate dextrose for-
mula A (ACD-A) anticoagulant. ACD-A anti-
coagulated blood (40 ml) was combined in a
sterile 50-ml polypropylene conical tube, cen-
trifuged at 2100 revolutions per minute (rpm)
(1000g) for 10 min at 20°C, and the plasma
collected in a fresh 50-ml conical tube without
disturbing the buffy coat white blood cell (WBC)
layer and large red cell pellet. The plasma was
centrifuged again at 2500 rpm (~1500g) for
15 min at 20°C to remove all platelets and cells.
Plasma was collected and aliquoted into 1-ml
cryovials and stored at −80°C. The red blood cell
(RBC)/WBC pellet was resuspended in an equal
volume of Hanks balanced salt solution (HBSS)
without Ca++ or Mg++ and containing 2 mM
EDTA and then divided into four 50-ml conical
tubes. Additional HBSS-EDTA (2 mM) buffer was
added to bring the volume of the RBC/WBC
mixture to 30 ml in each tube. The cell suspen-
sion was then carefully underlayered with 14 ml
96% Ficoll-Paque and centrifuged at 1800 rpm
(725g) for 20 min at 20°C in a swinging bucket
tabletop centrifuge with slow acceleration and
braking so as not to disrupt the Ficoll-cell
interface. Mononuclear cells at the Ficoll interface
were collected and transferred to a new 50-ml
centrifuge tube containing HBSS-EDTA (without
Ca++ or Mg++) and centrifuged at 1000 rpm
(~200g) for 15 min at 20°C. This pellets PBMCs
and leaves most of the platelets in the super-
natant. The supernatant was removed, and the
cell pellet was resuspended in 40 ml HBSS
(with Mg++ and Ca++ and without EDTA) + 1%
FBS. To remove additional contaminating plate-
lets, the cell suspension was centrifuged again
at 1000 rpm (~200g) for 15 min at 20°C and
the supernatant discarded. The cell pellet was
tap-resuspended in the residual 0.1 to 0.3 ml of
media and then brought to a volume of 10 ml
HBSS (with Mg++ and Ca++) + 1% FBS. Cells were
counted and viability assessed by trypan
blue exclusion. Cells were centrifuged again
at 1200 rpm (300g) for 10 min at 20°C, the
supernatant discarded, and the cells resus-
pended at a concentration of 5 to 10 × 106
cells/ml in CryoStor cell cryopreservation me-
dia (Sigma cat. no. C2999) and aliquoted into
1-ml cryovials (CryoClear cryovials; Globe Sci-
entific Inc., cat. no. 3010). Cells were stored in
a Corning CoolCell LX cell freezing container
at −80°C overnight and then transferred to
vapor-phase liquid N2 for long-term storage.
Mononuclear cells collected from lymph nodes
(LN) and spleen were processed in a similar
manner to blood mononuclear cells. LN and
spleen were excised and placed immediately
into RPMI1640 medium on wet ice. LNs were
diced with a sterile scalpel, and spleen was
homogenized and the material passed through
a sterile mesh grid. Cells were collected from
the pass-through and subjected to Ficoll den-
sity gradient purification as described above.
SHIV construction and characterization
The experimental design for constructing SHIVs
bearing primary or T/F Envs with allelic var-
iation at gp120 residue 375 was previously
described, including SHIV.CH505 and SHIV.
CH848 (9). For the construction of SHIV.
CAP256SU, we synthesized (GenScript) se-
quence KF996583.1 from GenBank (GenBank:
KF996583.1) and cloned it into the pCRXL-
TOPO-SIVmac766 backbone (9) by recombi-
nant polymerase chain reaction (PCR). The
QuikChange II XL Site-Directed Mutagenesis
kit (Agilent Technologies) was used to create
allelic variants (M, Y, F, W, or H) of the wild-
type Env375S codon. Wild-type and mutant
plasmids were transformed into MAX Effi-
ciency Stbl2 Competent Cells (Invitrogen)
for maxi-DNA preparations. Each 10-kb viral
genome was sequenced in its entirety to au-
thenticate its identity. Infectious SHIV stocks
were generated in 293T cells as previously de-
scribed (9). On day 0, five million 293T cells
were plated in 100-mm tissue culture dishes in
10 ml of complete MEM growth media with
10% FBS. On day 1, 6 ug of SHIV plasmid DNA
was combined with 18 ml of FuGENE 6 (Promega)
in 500 ml of DMEM was added dropwise to
tissue culture dishes. Media containing virus
was harvested on day 3 and aliquoted for stor-
age at −80°C. Virus concentration was estimated
by p27 antigen (p27Ag) ELISA (Zeptometrix)
and infectious particle concentration was de-
termined by entry into TZM-bl cells in the
presence of DEAE-Dextran, as previously de-
scribed (18). Typically, 293T-derived SHIV stocks
contained >1000 ng/ml p27Ag and >1,000,000
IU/ml on TZM-bl cells. The replication kinetics
of each of the SHIV.CAP256SU Env375 var-
iants in primary, activated human and rhe-
sus CD4 T cells were determined as previously
described (9). 293T supernatants containing
300 ng p27Ag of each variant were added to
2 × 106 purified human or rhesus CD4 T cells
in complete RPMI growth medium [RPMI1640
with 15% FBS (Hyclone), 100 U/ml penicillin-
streptomycin (Gibco), 30 U/ml IL-2 (aldesleukin,
Prometheus Laboratories), and 30 mg/ml DEAE-
Dextran]. As 300 ng p27Ag is equal to ~3 ×
109 virions, ~3 × 105 IU on TZM cells, or ~3 ×
104 IU on primary CD4 T cells, the estimated
MOI of this titration was estimated to be
about 0.01. The cell and virus mixtures were
incubated for 2 hours under constant rotation
at 37°C to facilitate infection, washed three
times with RPMI1640, and resuspended in
complete RPMI1640 medium lacking DEAE-
Dextran. Cells were plated into 24-well plates
at 2 million cells in 1 ml and cultured for 11
days, with sampling of 0.2 ml supernatant
and media replacement every 2 to 3 days for
11 days. Supernatants were assayed fr p27Ag
concentration by ELISA (Zeptometrix).
Plasma vRNA quantification
Plasma viral load measurements were per-
formed by the NIH/NIAID-sponsored Non-
human Primate Virology Core Laboratory at
the Duke Human Vaccine Institute. This core
facility is CLIA certified and operates a highly
standardized, quality-controlled Applied Bio-
systems Real-Time SIV and HIV vRNA PCR
assays. QIAsymphony SP and QIAgility auto-
mated platforms (QIAGEN) are used for high-
throughput sample processing and PCR setup.
Viral RNA is extracted and purified from plas-
ma, annealed to a target specific primer and
reverse transcribed into cDNA. The cDNA is
treated with RNase and added to a custom
real-time PCR master mix containing target
specific primers and a fluorescently labeled
hydrolysis probe. Thermal cycling is performed
on a QuantStudio3 (ThermoFisher Scientific)
real-time quantitative PCR (qPCR) instrument.
Viral RNA copies (cp) per reaction is interpo-
lated using quantification cycle data. Raw data
are quality-controlled, positive and negative
controls are checked, and the mean viral RNA
cp per milliliter is calculated. Over the course
of this study, the sensitivity limits for accurate
vRNA quantification using 0.5 ml of NHP plasma
improved from 250 RNA cp/ml to 62 RNA cp/ml.
We chose a conservative threshold of 100
RNA cp/ml for a limit of detection and 250
RNA cp/ml for the limit of quantification.
Viral sequencing, pixel plots, and LASSIE analysis
Single genome sequencing of SHIV 3′ half ge-
nomes was performed as previously described
(1, 9). Geneious R7 was used for alignments
and sequence analysis and sequences were
visualized using the LANL Highlighter and
Pixel tools (www.hiv.lanl.gov/content/sequence/
pixel/pixel.html and www.hiv.lanl.gov/content/
sequence/HIV/HIVTools.html; 44). The specific
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
13 of 19
RES EARCH | R E S E A R C H A R T I C L E
implementation of this software for this project
is described in the figure legends.
IgG isolation from plasma
Total polyclonal IgG was isolated from rhesus
plasma using the Protein A/Protein G GraviTrap
kit (GE Healthcare). Plasma was heat-inactivated
(1 hour at 56°C), clarified by centrifugation at
21,000g for 4 min, and applied to the Protein
A/G column. The sample was washed and eluted
per the manufacturer’s instructions and then
buffer-exchanged with phosphate-buffered sa-
line (PBS). The concentration of purified IgG
sample was quantified using the Pierce BCA
Protein Assay Kit (ThermoFisher).
Neutralizing antibody assay
Assays for neutralizing antibodies were per-
formed using TZM-bl indicator cells, as previ-
ously described (9, 18). This assay is essentially
identical to that used by Montefiori, Seaman,
and colleagues (83) (www.hiv.lanl.gov/content/
nab-reference-strains/html/home.htm), the only
difference being that in our assay we plate virus
and test plasma onto adherent TZM-bl cells and
hold the concentration of test plasma constant
(5% v/v) constant across all wells, which con-
tain 10% heat-inactivated fetal bovine serum
in the complete RPMI1640 culture medium.
Target viruses express HIV-1 Envs whose com-
plete designations and subtype classifications
are included in fig. S22 and table S3 and re-
ported elsewhere (26, 27, 84). In Fig. 1A, we
calculated neutralization breadth (ID50 ≤ 0.05)
across all time points for each animal from
titers displayed in table S2 and plotted in Fig. 1A.
We calculated the maximum geometric mean
titer (GMT) of neutralization at any one time
point from values with ID50 ≤ 0.05, as displayed
in table S2 and plotted in Fig. 1A.
Binding antibody assays
HIV-1 Env binding by recombinant mAbs and
plasma or sera were tested in ELISA, as pre-
viously described (50, 85). In brief, recombi-
nant Envs were coated directly to Nunc-absorb
(ThermoFisher) plates overnight at 4°C or
captured using AbC- mAb (AVIDITY, Colorado,
USA) that was directly coated to Nunc-absorb
plates overnight at 4°C. Antibody binding was
detected with goat anti-human or goat anti-
rhesus HRP-labeled anti-IgG Fc antibodies
(Jackson ImmunoResearch Laboratories), and
HRP detection was subsequently quantified
with 3,3′,5,5′-tetramethylbenzidine (TMB). Com-
petitive ELISA to assess cross-blocking of
recombinant mAbs or plasma antibodies were
previously described (7, 86). We biotinylated
the antibodies using BIOTIN-X-NHS (Cayman
Chemicals, CAT# 13316). Competitive inhibition
of biotinylated mAbs was measured as a per-
cent of binding in the presence of a competing
nonbiotinylated mAb relative to binding in
the absence of this competing mAb. MAbs
were also tested for binding HIV-1 Envs using
Biolayer interferometry (BLI) as described
(87). Here, antibody binding was measured
using mAb-captured sensors that were placed
into solutions of CH505 gp120 or SOSIP trimers
at 50 mg/ml for 1000 s. MAbs were captured
using anti-human IgG Fc sensors, and nonspecific
or background binding was subtracted using
binding levels by anti-influenza HA mAb (CH65).
Rhesus B cell staining and sorting
of strain-specific mAbs
CH505 gp120 T/F CD4bs-specific antibodies
were isolated from memory B cells in PBMCs,
lymph node, or bone marrow collected at weeks
20, 24, 32, and 52 using two approaches: direct
single-cell sorting into PCR plates (weeks 20,
32, and 52) (23, 88) and memory B cell cultures
(week 24) (89, 90). For direct sorting, we per-
formed single-cell isolation of memory B cells
decorated with AlexaFluor 647 (AF647) or Bril-
liant Violet 421 (BV421)–tagged HIV-1 CH505
TF gp120 using a BD FACSAria or a BD FACSAria
II (BD Biosciences, San Jose, CA), as previously
described (23). The flow cytometry data were
analyzed using FlowJo (Treestar, Ashland, OR)
(88, 91, 92). For our sort strategy, we isolated
antigen-specific IgD-negative, CD27-All mem-
ory B cells that bound BV421-tagged CH505 T/F
gp120 but not AF647-tagged CH505 T/F gp120
D371 mutant protein; antibodies isolated from
these B cells were referred to as CH505 differ-
ential binders or CD4BS antibodies (23). For
memory B cell cultures, from 8 million PBMCs,
we sorted 30,792 CH505 T/F gp120-specific B
cells, defined as CD3-negative, CD14-negative,
CD16-negative, IgD-negative, CD27-All, CD20-
positive, AF647-tagged CH505 T/F gp120-positive
and BV421-tagged CH505 T/F gp120-positive.
As previously described (90), cells were flow
sorted in bulk into wells containing 5000
MS40L feeder cells, RPMI-1640 supplemented
with 15% FBS, 1 mM sodium pyruvate, 1% non-
essential amino acids, 25 mM HEPES buffer,
2.5 mg/ml ODN2006 (Invivogen, TLRL-2006-5),
5 mM CHK2-inhibitor (Calbiochem, 220486),
100 ng/ml recombinant human interleukin (IL)–
21 (Peprotech, cat. no. 2001-21), 10 ng/ml recom-
binant Human BAFF (Peprotech, cat. no. 310-13),
200 U/ml IL-2 (from the myeloma IL-2 producing
cell line IL2-t6, provided by A. Lanzavecchia, IRB,
Bellinzona, Switzerland), and 100 ml superna-
tant of the Herpesvirus papio (HVP)–infected
Baboon cell line S594 (NHP Reagent Resource).
The concentration of each supplement was pre-
viously determined to achieve optimal in vitro
stimulation. After overnight incubation at 37 °C
in 5% CO2, memory B cells were transferred at
limiting dilution into 96-well round-bottom tis-
sue culture plates containing 5000 MS40L feeder
cells. Culture medium was refreshed 7 days
after plating and harvested 2 weeks after plating
to test for binding to CH505 T/F gp120, CH505
T/F gp120 D371, RSC3 (93), and RSC3 D371 P363N,
as well as for neutralization of pseudotyped the
CH505 T/F HIV-1 strain in the TZM-bl-based
neutralization assay using a single dilution of
supernatant (89, 94). CD4bs DH650 lineage
autologous tier 2 NAbs were isolated from
CH505 differential-binding memory B cells in
PBMCs from week 20, 24, and 32 after SHIV
CH505 infection of RM6072. In capturing max-
imum numbers of B cells bearing candidate
CD4bs antibodies, we used a less stringent gating
strategy for capturing CH505 T/F gp120 (+) and
CH505 T/F gp120 D371 (−) B cells. In so doing, we
also captured non-CD4 binding site autologous
tier 2 NAbs, including DH647 and DH648 that
were isolated from memory B cells in PBMCs
at week 20 after SHIV infection of RM6072.
Rhesus B cell staining, culture, and microneutralization
screening for V2 apex bNAbs
Cryopreserved PBMCs from RM5695 at week 65
after SHIV infection were thawed and stained
with LIVE/DEAD Fixable Aqua Dead Cell Stain
(Life Technologies), as previously described
(95, 96). Cells were washed and stained with
an antibody cocktail of CD3 (clone SP34-2,
BD Biosciences), CD4 (clone OKT4, BioLegend),
CD8 (clone RPA-T8, BioLegend), CD14 (clone
M5E2, BioLegend), CD20 (clone 2H7, BioLegend),
IgG (clone G18-145, BD Biosciences), IgD (poly-
clonal, Dako), and IgM (clone G20-127, BD Bio-
sciences) at room temperature in the dark for
20 min. The stained cells were washed three
times with PBS, resuspended in 1 ml of PBS and
passed through a 70-mm cell mesh (BD Biosci-
ences). Total memory B cells (CD3-CD4-CD8-
CD14-CD20+IgD-IgM-IgG+) were sorted with a
modified three-laser FACSAria cell sorter using
the FACSDiva software (BD Biosciences) and
flow cytometric data were subsequently ana-
lyzed using FlowJo (v9.7.5). B cells were sorted
at 1 cell per well of a 384-well plate containing B
cell culture media following a human B cell
culture protocol (97) that was optimized for the
expansion of rhesus B cells. Briefly, sorted B
cells were expanded for 14 days in B cell cul-
ture medium consisting of Iscove’s modified
Dulbecco’s medium (IMDM) with GlutaMAX
supplemented with 10% heat-inactivated FBS,
1X MycoZap Plus-PR, 100 U/ml IL-2, 0.05 mg/ml
IL-4, 0.05 mg/ml IL-21, 0.05 mg/ml BAFF, 2 mg/ml
CpG ODN2006, and 3T3-msCD40L feeder cells
at a density of 5000 cells per well. Supernatants
from ~14,000 individual wells were evaluated
for neutralization of HIV-1 Q23.17 and T250-4
Env-pseudotyped viruses using the high through-
put NVITAL automated microneutralization
assay, as previously described (98). Wells were
selected for RT-PCR on the basis of 50% reduc-
tion in infectivity of at least one virus. Out of
nearly 14,000 wells tested, five met these neu-
tralization criteria. Paired heavy and light Ig
chains were successfully amplified from four of
these wells. These amplicons were sequenced,
cloned, and expressed as IgG1 (RHA1.V2.01-04),
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
14 of 19
RES EARCH | R E S E A R C H A R T I C L E
and all four mAbs exhibited broad and potent
neutralization.
Rhesus B cell cloning and expression
Heavy (IGHV) and light (IGKV, IGLV) chain
genes were isolated via single cell PCR ap-
proaches (88, 99), and the gene sequences
were computationally analyzed using rhesus
Cloanalyst program (100–102). Antibody im-
munogenetics were reported as gene families
and segments, mutation frequencies, and CDR3
lengths using the rhesus Cloanalyst database of
reference genes (21). Using the rhesus cloanalyst
program, we identified B cell clonal lineages for
antibodies with the same inferred IGHV VDJ
rearrangement and CDR3 length, and paired
with the same light chain (Ig VJ segments). For
DH650 lineage, the UCA and intermediate (IA)
genes were inferred computationally using the
Cloanalyst program. The automated inference
of antibody clonality was followed up by vi-
sual inspection of the DNA sequence align-
ments for confirmation. The heavy and light
chain gene sequences from the sorted B cells
or inferred UCA and IAs were commercially
generated and used to express purified recom-
binant mAbs as described (88). Heavy and
light chain immunoglobulin repertoire next-
generation sequencing of monkey RM6072
was performed with the Illumina MiSeq plat-
form using primers targeting the VH1 and Vk2
families to identify DH650 clonal members
using a previously described protocol (23, 103).
For mAb isolation from RM5695, bulk cDNA
was synthesized from the five neutralization-
positive B cell culture wells using random
hexamers as previously described (104). Sub-
sequently, immunoglobulin heavy chain (IgG)
and light chain (IgK and IgL) variable regions
were separately amplified by nested PCR cycles
using pools of rhesus macaque primers as pre-
viously described (105). Sequences were ana-
lyzed using Cloanalyst (102) to infer putative
variable region germline genes and identify
clonal lineages. The heavy and lambda chain
variable regions of the RM5695 lineage were
codon optimized, synthesized with a murine im-
munoglobulin signal peptide (tripeptide sequence
VHS) immediately preceding the 5′ end of FRW1,
and cloned into rhesus IgG1 (RhCMV-H) and IgL
(RhCMV-L) expression vectors directly upstream
of the respective constant regions using AgeI/
NheI and AgeI/ScaI restriction sites, respectively
(105). Recombinant mAbs were expressed by
cotransfection of paired heavy and lambda chain
plasmids into 293Freestyle cells as previously
described (93), purified from cell supernatant
using the Protein A/Protein G GraviTrap kit (GE
Healthcare), and buffer-exchanged into PBS.
Next-generation sequencing of naïve B cells
and IgDiscover analysis
PBMCs isolated from RM5695 plasma obtained
at week 48 after infection were stained with
LIVE/DEAD Aqua, CD3-PerCP-Cy55, CD4-BV785,
CD8-BV711, CD14-PE-Cy7, CD20-BV605, IgD-
FITC, IgG-Ax680, and IgM-BV650. Approxi-
mately 300,000 naïve B cells (CD20+, IgG−,
IgD+, IgM+) were bulk sorted into RPMI with
10% FBS and 1% Pen-Strep using a BD FACSAria
II. Total RNA was extracted using RNAzol RT
per the manufacturer’s guidelines (Molecular
Research Center, Inc). Reverse transcription
of mRNA transcripts, IgM and IgL variable
region library preparation, and next-generation
sequencing were performed as previously de-
scribed (106), as these methods can be efficiently
used for both humans and RMs. Both heavy
and lambda immunoglobulin libraries were
sequenced on Illumina Miseq with 2 × 300 bp
runs using the MiSeq Reagent V3 kit (600-cycle).
Filtered, quality-controlled IgM and IgL se-
quences were analyzed using IgDiscover (57)
to curate a personalized immunoglobulin
repertoire library for RM5695. A naïve IgK
variable region library was not prepared, be-
cause the RM5695 lineage uses a lambda chain.
IgDiscover can identify functional VH, VL,
VK, JH, JL, and JK genes, and denotes any
novel genes and/or alleles with respect to the
provided reference database. For this analysis,
a recently published RM database was used as
the template (21). The resulting personalized
database was then used in SONAR (107) to
accurately assign germline immunoglobulin
genes and precisely determine somatic hyper-
mutation frequencies for the RHA1.V2 lineage.
Auto- and polyreactivity analysis
RHA1.V2.01 mAb reactivity to nine autoanti-
gens was measured using the AtheNA Multi-
Lyte ANA kit (Zeus scientific, Inc, #A21101).
Antibodies were twofold serially diluted start-
ing at 50 mg/ml, and kit methods were followed
per manufacturer’s instructions. Samples were
analyzed using AtheNA software. Indirect im-
munofluorescence binding of RHA1.V2.01 mAb
to human epithelial (HEp-2) cells (Zeus Scien-
tific, Somerville, NJ) was also performed per
manufacturer’s instructions. Briefly, 20 ml of
diluted antibody (50 mg/ml) was added to
antinuclear antibody (ANA) test slides. Slides
were incubated for 20 min at room temper-
ature in a humid chamber and then washed
with 1X PBS. Goat-anti-rhesus Ig-FITC (Southern
Biotech, Birmingham, AL) secondary antibody
was added at a concentration of 30 mg/ml to
each well. The slides were incubated for 20 min,
washed twice, dried, fixed with 33% glycerol
and cover-slipped. Slides were imaged using
an Olympus AX70 microscope with a SpotFlex
FX1520 camera. Images were acquired on a
40× objective using the FITC fluorescence chan-
nel. Positivity was determined by comparison
with positive and negative control nonhuman
primate mAbs DH1037 and DH570.30, respec-
tively. Staining patterns were identified using
the Zeus Scientific pattern guide.
Hierarchical clustering of RHA1.V2.01 and other
V2 apex bNAb profiles
We used the Heatmap webtool at the Los
Alamos HIV database (www.hiv.lanl.gov/
content/sequence/HEATMAP/heatmap.html)
(93, 106, 107). Input data were log10 trans-
formed IC50 titers for RHA1.V2.01 and other
V2 apex bNAbs for the 208 virus panel. For
clustering, Euclidean distances and the Ward
algorithm (108) were used. Bootstraps were
calculated using 1000 iterations.
Neutralization signature analyses
These analyses were performed using the
webtool GenSig (www.hiv.lanl.gov/content/
sequence/GENETICSIGNATURES/gs.html)
(59) as described in (60). Briefly, IC50 titers
for RM5695 RHA1.V2.01 and other V2 apex
bNAbs, together with Env sequences from
the 208-strain virus panel, were used as in-
puts. Two statistical strategies were used:
(i) Fisher’s test with a binary phenotype, IC50
titer above or below the threshold (resistant
and sensitive, respectively), and (ii) Wilcoxon
test that compares distribution of IC50 titers
with and without a feature. The latter analyses
were conducted two ways, including and ex-
cluding resistant viruses. In both strategies,
every amino acid and glycan at each site were
tested for associations, with and without a
phylogenetic correction. A relaxed multiple
test false discovery rate (FDR) (q-value < 0.2)
was used to first filter all possible hypothetical
associations. At this high threshold, several
phylogenetically uncorrected signatures were
found, raising the suspicion that phylogenetic
artifacts could be at play. However, some of
these signatures could be relevant and under-
lie the clade-specific sensitivity and/or resistance
of the bNAb in question. Thus, to down-select
the most statistically robust and relevant signa-
ture sites, each selected site was required to
meet at least two of the following three criteria:
(i) contact site (as defined below), (ii) at least
one phylogenetically corrected signature, and
(iii) at least one strong signature at q < 0.1. All
associations with q < 0.2 at selected sites were
used to identify sequence features associated
with sensitivity or resistance to the bNAb. The
Wilcoxon signatures did not identify any new
sites or associations. Env contact sites were
defined using the cryo-EM structure of PGT145
in complex with Env trimer [Protein Data Bank
(PDB) ID 5V8L] and a cutoff of 8.5 Å from any
antibody heavy atom (HXB2 positions: 120, 121,
122, 123, 124, 125, 126, 127, 128, 129, 130, 156, 157,
158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168,
169, 170, 171, 172, 173, 174, 175, 184, 191, 192, 200,
201, 305, 306, 307, 308, 309, 312, 313, 314, 315,
and 321).
Soluble HIV-1 envelope trimer generation
HIV Env trimer BG505 DS-SOSIP.664 was
produced in transiently transfected 293F cells
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
15 of 19
RES EARCH | R E S E A R C H A R T I C L E
as previously described (109, 110). Briefly, the
plasmid encoding the SOSIP trimer and a
plasmid encoding furin were mixed at 4:1
ratio and transfected into 293F cells at 0.75 mg
plasmid per 1 L cells (1.5 × 106/ml) using
293Fectin (ThermoFisher) or Turbo293 trans-
fection reagent (Speed BioSystems). Cells were
then incubated in a shaker at 120 rpm, 37°C,
9% CO2. The following day, 80 ml HyClone
SFM4HEK293 medium and 20 ml FreeStyle
293 Expression Medium were added to each
liter of cells. Env trimer protein was purified
from the day-7 supernatant with VRC01 af-
finity chromatography, followed by gel filtra-
tion on a Sephadex200 16/60HL column and
V3-exposed trimers were removed with nega-
tive selection on a 447-52D affinity column.
The antigenicity of the trimers was confirmed
with a panel of antibodies in the Meso Scale
Discovery (MSD) platform.
Antibody Fab preparation
Variable regions of the RHA1.V2.01 antibody
heavy and light chain genes were synthesized
(Genscript) and subcloned into the pVRC8400
vector, in which a HRV3C cleavage site was
inserted in the heavy-chain hinge region.
The heavy and light chain pair was cotrans-
fected in Expi293F cells (ThermoFisher) using
Turbo293 transfection reagent (Speed BioSys-
tems) as described previously (111). The culture
supernatant was harvested 6 days after trans-
fection and loaded on a protein A column, the
column was washed with PBS, and IgG proteins
were eluted with a low pH buffer. The eluted
IgG proteins were cleaved by HRV3C, and the
cleavage mixture was passed through a pro-
tein A column.
Cryo-EM data collection and processing
Antibody Fab fragments of RHA1.V2.01 were
incubated with BG505 DS-SOSIP with the Fab
in molar excess. At a 1 mg/ml concentration,
2.3 ml of the complex was deposited on a C-flat
grid (www.protochips.com). An FEI Vitrobot
Mark IV was used to vitrify the grid with a
wait time of 30 s, blot time of 3 s, and blot
force of 1. Automated data collection was per-
formed with Leginon (112) on a Titan Krios
electron microscope equipped with a Gatan
K2 Summit direct detection device. Exposures
were taken in movie mode for 10 s with the
total dose of 71.06 e–/Å2 fractionated over 50
raw frames. Images were preprocessed through
Appion (113, 114); MotionCor2 (115) was used
for frame alignment and dose weighting. The
CTF was determined using CTFFind4 (116, 117).
Initial particle picking was done with DoG
Picker (113, 114). RELION (118) was then used
for particle extraction. CryoSPARC 2.12 (119)
was used for two-dimensional (2D) classifica-
tions, ab initio 3D reconstruction, homogeneous
refinement, and nonuniform 3D refinement.
Initial 3D reconstruction was performed using
C1 symmetry, confirming one Fab per trimer,
and C1 symmetry was applied for the final
reconstruction and refinement. Coordinates
from PDB ID 6NNF (120) and 6CA6 (58) were
used for initial fit to the reconstructed map.
Simulated annealing was performed on the
first refinement followed by iterative manual
fitting and real space refinement in Phenix
(121) and Coot (122). Geometry and map
fitting were evaluated throughout the process
using Molprobity (123) and EMRinger (124).
PyMOL (www.pymol.org) was used to generate
figures.
Expression and crystallization of CH505
gp120:DH650 Fab complex
CH505 gp120 core (residues 44 to 492, DV1-V2
and DV3) (125) was expressed in HEK293S
GnTI− cells and purified by affinity chroma-
tography on Galanthus nivalis lectin (Vector
Laboratories) followed by gel filtration on
Superdex 200 column (GE). Deglycosylation
was carried out with endoglycosidase H (endoH;
New England Biolabs) in deglycosylation buffer
(50 mM sodium acetate, pH 6, 5 mM EDTA,
500 mM NaCl, 10 ml endoH, 1 mg/ml leupeptin,
1 mg/ml aprotinin) at 37°C overnight, followed
by buffer exchange to 40 mM Tris HCl, pH 7.4,
1 M NaCl, 2 mM MnCl2, 2 mM CaCl2, and
passage through a concanavalin A column
(Sigma) to remove any gp120 that had not been
fully deglycosylated by the endoH treatment.
The eluate was buffer exchanged by passage
over a Superdex200 column (GE) into 2.5mM
Tris-HCl pH7.5, 350mM NaCl, followed by
concentration to 4 mg/ml. The Fab fragment
of mAb DH650 was expressed in HEK293T
cells, as described in (125). DH650-gp120 core
complex was formed by incubating gp120 core
with DH650 in 1:1.3 molar ratio followed by
gel filtration on Superdex 200 (GE) column in
buffer 2.5 mM Tris-HCl pH 7.5, 350 mM NaCl.
The complex was concentrated to 8.5 mg/ml
and crystallized by hanging-drop vapor dif-
fusion in 20% PEG 8K, 100 mM Tris pH 8,
500 mM NaCl.
Determination of CH505 gp120:DH650 Fab
crystal structure
X-ray diffraction data to 2.8-Å resolution were
collected at the Advanced Photon Source
(Argonne National Laboratories) on NE-CAT
beamline 24-ID-C. Intensity data were integrated,
scaled, and merged with XDS and XSCALE
(126) (see table S4). An initial model, obtained
by molecular replacement with Phaser (127)
using the ZM176.6 gp120 core (PDB ID 4LST),
was rebuilt and refined with COOT (128) and
Buster (Global Phasing, Ltd., Cambridge, UK),
respectively. The final model (PDB ID 6XCJ)
includes residues 52 to 128, 207 to 299, 327 to
396, and 409 to 488 of CH505 gp120, heavy
chain residues 1 to 223, and light chain residues
1 to 219 of Fab DH650.
Negative-stain electron microscopy of CH505
SOSIP:DH650 Fab complex
CH505 DS-SOSIP, prepared as described in (7),
was mixed with a DH650 Fab in a 1:4 molar
ratio. After incubation for 1 hour, the complex
was loaded on a Superose 6 column (GE). The
SOSIP-Fab complex was diluted to 0.1 mg/ml
and applied to freshly glow-discharged, carbon-
coated EM grids and negatively stained with
1% uranyl formate. Images were recorded on
an FEI Tecnai T12 electron microscope, oper-
ated at 120 kV, and equipped with a charge-
coupled detector. Particles were selected and
class averages computed with EMAN2 (129).
Neutralization fingerprint
The neutralization fingerprint of a monoclonal
antibody is defined as the potency pattern with
which the antibody neutralizes a set of diverse
viral strains. The neutralization fingerprints
of V2 targeting bNAbs, including VRC26,
PGT145, PG9, and VRC38 classes, as well as a
set of other HIV-1 bNAbs were compared and
clustered according to fingerprint similarity,
as described previously (59), using a panel of
208 HIV-1 viral strains.
Statistical analyses
Statistical tests were calculated by using Graph-
Pad Prism 7 software. The Mann-Whitney test
was used to determine whether the viral loads
(peak and set point) of anti-CD8 treated animals
were significantly different from untreated
animals and whether viral loads (peak and set
point) were greater in animals that developed
bNAbs compared with those that did not. We
chose a nonparametric rank-based test because
both peak and set point viral loads of the
untreated group and the bNAb and non-bNAb
groups failed the D’Agostino and Pearson nor-
mality test (P < 0.05). The Spearman’s rank
correlation test was used to determine whether
there is a significant correlation between set
point viral loads and autologous tier 2 NAb
titers. The geometric means were calculated
using the Column statistics function of GraphPad
Prism. The Chi-squared test was used on 2 × 2
contingency tables of shared and nonshared
mutations within or between groups of humans
and animals infected by viruses bearing the
same (homologous) or different (heterologous)
Envs to determine whether Env residue muta-
tions demonstrated significant strain specific-
ity. Separate tests were run to analyze total
shared and nonshared mutations for the groups
overall and for any discordant CH505, CH848,
or CAP256SU group pairing.
Data and software availability
GenBank accession numbers for all HIV-1 env
sequences analyzed in this study are as follows:
MN471655-MN472016, MT484339-MT487491,
MT580365-MT580445, MT509359, GQ999989,
HQ625604, KC863461-KC863464, KF241776,
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
16 of 19
RES EARCH | R E S E A R C H A R T I C L E
KF996577-KF996601, KF996604, KF996606,
KF996610-KF996630, KF996632-KF996662,
KF996664-KF996678, KF996680, KF996682-
KF996683, KF996685-KF996716, KT698223-
KT698227, MF572809-MF572829, EF203980-
EF203981, and MK205498-MK205507. HIV-
1 Env sequences from the human subject CAP256
were previously published (3, 5, 130). GenBank
accession numbers for immunoglobulin genes
are: MT581213-MT581268, MT610888-MT610895,
and MT656172-MT656253. Cryo-EM maps and
fitted coordinates have been deposited with
database codes EMDB-22295 and PDB ID 6XRT,
respectively. DH650 bound to the gp120 core
was deposited with database code PDB ID 6XCJ.
RE FE RENCES AND N OT ES
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
B. F. Keele et al., Identification and characterization of
transmitted and early founder virus envelopes in primary
HIV-1 infection. Proc. Natl. Acad. Sci. U.S.A. 105, 7552–7557
(2008). doi: 10.1073/pnas.0802203105; pmid: 18490657
H. X. Liao et al., Co-evolution of a broadly neutralizing
HIV-1 antibody and founder virus. Nature 496, 469–476
(2013). doi: 10.1038/nature12053; pmid: 23552890
N. A. Doria-Rose et al., Developmental pathway for potent
V1V2-directed HIV-neutralizing antibodies. Nature 509,
55–62 (2014). doi: 10.1038/nature13036; pmid: 24590074
F. Gao et al., Cooperation of B cell lineages in induction of
HIV-1-broadly neutralizing antibodies. Cell 158, 481–491 (2014).
doi: 10.1016/j.cell.2014.06.022; pmid: 25065977
J. N. Bhiman et al., Viral variants that initiate and drive
maturation of V1V2-directed HIV-1 broadly neutralizing
antibodies. Nat. Med. 21, 1332–1336 (2015). doi: 10.1038/
nm.3963; pmid: 26457756
M. Bonsignori et al., Maturation pathway from germline to
broad HIV-1 neutralizer of a CD4-mimic antibody. Cell 165,
449–463 (2016). doi: 10.1016/j.cell.2016.02.022;
pmid: 26949186
M. Bonsignori et al., Staged induction of HIV-1 glycan-
dependent broadly neutralizing antibodies. Sci. Transl. Med.
9, eaai7514 (2017). doi: 10.1126/scitranslmed.aai7514;
pmid: 28298420
E. Landais et al., HIV envelope glycoform heterogeneity and
localized diversity govern the initiation and maturation of a
V2 apex broadly neutralizing antibody lineage. Immunity 47,
990–1003.e9 (2017). doi: 10.1016/j.immuni.2017.11.002;
pmid: 29166592
H. Li et al., Envelope residue 375 substitutions in simian-
human immunodeficiency viruses enhance CD4 binding and
replication in rhesus macaques. Proc. Natl. Acad. Sci. U.S.A.
113, E3413–E3422 (2016). doi: 10.1073/pnas.1606636113;
pmid: 27247400
P. D. Kwong, J. R. Mascola, HIV-1 vaccines based on antibody
identification, B cell ontogeny, and epitope structure.
Immunity 48, 855–871 (2018). doi: 10.1016/j.
immuni.2018.04.029; pmid: 29768174
D. Sok, D. R. Burton, Recent progress in broadly neutralizing
antibodies to HIV. Nat. Immunol. 19, 1179–1188 (2018).
doi: 10.1038/s41590-018-0235-7; pmid: 30333615
B. F. Haynes et al., Cardiolipin polyspecific autoreactivity in
two broadly neutralizing HIV-1 antibodies. Science 308,
1906–1908 (2005). doi: 10.1126/science.1111781;
pmid: 15860590
14.
13. G. Y. Chuang et al., Structural survey of broadly neutralizing
antibodies targeting the HIV-1 Env trimer delineates epitope
categories and characteristics of recognition. Structure 27,
196–206.e6 (2019). doi: 10.1016/j.str.2018.10.007;
pmid: 30471922
E. C. Holmes, Error thresholds and the constraints to RNA
virus evolution. Trends Microbiol. 11, 543–546 (2003).
doi: 10.1016/j.tim.2003.10.006; pmid: 14659685
J. Woo, D. L. Robertson, S. C. Lovell, Constraints on
HIV-1 diversity from protein structure. J. Virol. 84, 12995–13003
(2010). doi: 10.1128/JVI.00702-10; pmid: 20881050
16. H. K. Haddox, A. S. Dingens, J. D. Bloom, Experimental
15.
estimation of the effects of all amino-acid mutations to HIV’s
envelope protein on viral replication in cell culture. PLOS
Pathog. 12, e1006114 (2016). doi: 10.1371/journal.
ppat.1006114; pmid: 27959955
17.
18.
19.
R. Wyatt et al., The antigenic structure of the HIV gp120
envelope glycoprotein. Nature 393, 705–711 (1998).
doi: 10.1038/31514; pmid: 9641684
X. Wei et al., Antibody neutralization and escape by HIV-1.
Nature 422, 307–312 (2003). doi: 10.1038/nature01470;
pmid: 12646921
B. R. Starcich et al., Identification and characterization of
conserved and variable regions in the envelope gene of
HTLV-III/LAV, the retrovirus of AIDS. Cell 45, 637–648
(1986). doi: 10.1016/0092-8674(86)90778-6; pmid: 2423250
20. P. D. Kwong et al., HIV-1 evades antibody-mediated
21.
neutralization through conformational masking of
receptor-binding sites. Nature 420, 678–682 (2002).
doi: 10.1038/nature01188; pmid: 12478295
A. Ramesh et al., Structure and diversity of the rhesus
macaque immunoglobulin loci through multiple de novo
genome assemblies. Front. Immunol. 8, 1407 (2017).
doi: 10.3389/fimmu.2017.01407; pmid: 29163486
22. P. L. Moore et al., Multiple pathways of escape from
HIV broadly cross-neutralizing V2-dependent antibodies.
J. Virol. 87, 4882–4894 (2013). doi: 10.1128/JVI.03424-12;
pmid: 23408621
23. W. B. Williams et al., Initiation of HIV neutralizing B cell
lineages with sequential envelope immunizations.
Nat. Commun. 8, 1732 (2017). doi: 10.1038/s41467-017-
01336-3; pmid: 29170366
24. Z. M. Ndhlovu et al., Magnitude and kinetics of CD8+ T Cell
activation during hyperacute HIV infection impact viral set
point. Immunity 43, 591–604 (2015). doi: 10.1016/
j.immuni.2015.08.012; pmid: 26362266
25. M. L. Robb et al., Prospective study of acute HIV-1 infection in
adults in east Africa and Thailand. N. Engl. J. Med. 374,
2120–2130 (2016). doi: 10.1056/NEJMoa1508952;
pmid: 27192360
26. M. S. Seaman et al., Tiered categorization of a diverse panel
of HIV-1 Env pseudoviruses for assessment of neutralizing
antibodies. J. Virol. 84, 1439–1452 (2010). doi: 10.1128/
JVI.02108-09; pmid: 19939925
27. A. deCamp et al., Global panel of HIV-1 Env reference strains
for standardized assessments of vaccine-elicited neutralizing
antibodies. J. Virol. 88, 2489–2507 (2014). doi: 10.1128/
JVI.02853-13; pmid: 24352443
28. P. Hraber et al., Prevalence of broadly neutralizing
antibody responses during chronic HIV-1 infection. AIDS 28,
163–169 (2014). doi: 10.1097/QAD.0000000000000106;
pmid: 24361678
29. HIV-1 strains 25710 and Q23.17 have been variably reported
as tier 1b or tier 2 in neutralization sensitivity to different
panels of heterologous HIV-1 positive plasma. In all studies,
these two virus strains have been shown to be much closer
to tier 2 viruses than to lab-adapted tier 1A viruses in their
neutralization sensitivity patterns. See (26–28).
J. M. Decker et al., Antigenic conservation and
immunogenicity of the HIV coreceptor binding site. J. Exp.
Med. 201, 1407–1419 (2005). doi: 10.1084/jem.20042510;
pmid: 15867093
K. L. Davis et al., Human immunodeficiency virus type 2
(HIV-2)/HIV-1 envelope chimeras detect high titers of broadly
reactive HIV-1 V3-specific antibodies in human plasma.
J. Virol. 83, 1240–1259 (2009). doi: 10.1128/JVI.01743-08;
pmid: 19019969
J. Chen et al., Effect of the cytoplasmic domain on antigenic
characteristics of HIV-1 envelope glycoprotein. Science 349,
191–195 (2015). doi: 10.1126/science.aaa9804;
pmid: 26113642
E. Landais et al., Broadly neutralizing antibody responses in a
large longitudinal sub-Saharan HIV primary infection cohort.
PLOS Pathog. 12, e1005369 (2016). doi: 10.1371/journal.
ppat.1005369; pmid: 26766578
E. Landais, P. L. Moore, Development of broadly neutralizing
antibodies in HIV-1 infected elite neutralizers. Retrovirology
15, 61 (2018). doi: 10.1186/s12977-018-0443-0;
pmid: 30185183
30.
31.
32.
33.
34.
38.
J. F. Salazar-Gonzalez et al., Genetic identity, biological
phenotype, and evolutionary pathways of transmitted/
founder viruses in acute and early HIV-1 infection. J. Exp.
Med. 206, 1273–1289 (2009). doi: 10.1084/jem.20090378;
pmid: 19487424
39. K. J. Bar et al., Early low-titer neutralizing antibodies impede
HIV-1 replication and select for virus escape. PLOS Pathog. 8,
e1002721 (2012). doi: 10.1371/journal.ppat.1002721;
pmid: 22693447
40. M. Markowitz et al., A novel antiviral intervention results in
more accurate assessment of human immunodeficiency
virus type 1 replication dynamics and T-cell decay in vivo.
J. Virol. 77, 5037–5038 (2003). doi: 10.1128/JVI.77.8.5037-
5038.2003; pmid: 12663814
41. D. D. Ho et al., Rapid turnover of plasma virions and CD4
lymphocytes in HIV-1 infection. Nature 373, 123–126 (1995).
doi: 10.1038/373123a0; pmid: 7816094
42. X. Wei et al., Viral dynamics in human immunodeficiency
virus type 1 infection. Nature 373, 117–122 (1995).
doi: 10.1038/373117a0; pmid: 7529365
43. P. Borrow et al., Antiviral pressure exerted by HIV-1-specific
cytotoxic T lymphocytes (CTLs) during primary infection
demonstrated by rapid selection of CTL escape virus.
Nat. Med. 3, 205–211 (1997). doi: 10.1038/nm0297-205;
pmid: 9018240
44. P. Hraber et al., Longitudinal antigenic sequences and sites
from intra-host evolution (LASSIE) identifies immune-
selected HIV variants. Viruses 7, 5443–5475 (2015).
doi: 10.3390/v7102881; pmid: 26506369
45. K. Wagh et al., Completeness of HIV-1 envelope glycan shield
at transmission determines neutralization breadth. Cell Rep.
25, 893–908.e7 (2018). doi: 10.1016/j.celrep.2018.09.087;
pmid: 30355496
J. M. Coffin, Structure, replication, and recombination of
retrovirus genomes: Some unifying hypotheses. J. Gen. Virol.
42, 1–26 (1979). doi: 10.1099/0022-1317-42-1-1;
pmid: 215703
46.
47. N. Wood et al., HIV evolution in early infection: Selection
pressures, patterns of insertion and deletion, and the impact
of APOBEC. PLOS Pathog. 5, e1000414 (2009). doi: 10.1371/
journal.ppat.1000414; pmid: 19424423
48. B. M. Mabvakure et al., Positive selection at key residues in
the HIV envelope distinguishes broad and strain-specific
plasma neutralizing antibodies. J. Virol. 93, e01685-18
(2019). doi: 10.1128/JVI.01685-18; pmid: 30567996
49. N. A. Doria-Rose et al., New member of the V1V2-directed
CAP256-VRC26 lineage that shows increased breadth and
exceptional potency. J. Virol. 90, 76–91 (2015). doi: 10.1128/
JVI.01791-15; pmid: 26468542
50. K. O. Saunders et al., Targeted selection of HIV-specific
51.
52.
53.
54.
antibody mutations by engineering B cell maturation.
Science 366, eaay7199 (2019). doi: 10.1126/science.aay7199;
pmid: 31806786
R. Andrabi et al., Identification of common features in
prototype broadly neutralizing antibodies to HIV envelope V2
apex to facilitate vaccine design. Immunity 43, 959–973
(2015). doi: 10.1016/j.immuni.2015.10.014; pmid: 26588781
J. Gorman et al., Structures of HIV-1 Env V1V2 with broadly
neutralizing antibodies reveal commonalities that enable
vaccine design. Nat. Struct. Mol. Biol. 23, 81–90 (2016).
doi: 10.1038/nsmb.3144; pmid: 26689967
J. Gorman et al., Structure of super-potent antibody CAP256-
VRC26.25 in complex with HIV-1 envelope reveals a combined
mode of trimer-apex recognition. Cell Rep. 31, 107488
(2020). doi: 10.1016/j.celrep.2020.03.052; pmid: 32268107
J. F. Scheid et al., Sequence and structural convergence of
broad and potent HIV antibodies that mimic CD4 binding.
Science 333, 1633–1637 (2011). doi: 10.1126/
science.1207227; pmid: 21764753
55. K. Wiehe et al., Functional relevance of improbable antibody
mutations for HIV broadly neutralizing antibody development.
Cell Host Microbe 23, 759–765.e6 (2018). doi: 10.1016/
j.chom.2018.04.018; pmid: 29861171
35. H. J. Barbian et al., Neutralization properties of simian
56. T. Zhou et al., Multidonor analysis reveals structural
immunodeficiency viruses infecting chimpanzees and gorillas.
mBio 6, e00296-15 (2015). doi: 10.1128/mBio.00296-15;
pmid: 25900654
36. V. Sanchez-Merino et al., Detection of broadly neutralizing
activity within the first months of HIV-1 infection. J. Virol. 90,
5231–5245 (2016). doi: 10.1128/JVI.00049-16; pmid: 26984721
37. N. Goonetilleke et al., The first T cell response to
transmitted/founder virus contributes to the control of acute
viremia in HIV-1 infection. J. Exp. Med. 206, 1253–1272
(2009). doi: 10.1084/jem.20090365; pmid: 19487423
elements, genetic determinants, and maturation pathway for
HIV-1 neutralization by VRC01-class antibodies. Immunity 39,
245–258 (2013). doi: 10.1016/j.immuni.2013.04.012; pmid:
23911655
57. M. M. Corcoran et al., Production of individualized V gene
databases reveals high levels of immunoglobulin genetic
diversity. Nat. Commun. 7, 13642 (2016). doi: 10.1038/
ncomms13642; pmid: 27995928
58. K. Rantalainen et al., Co-evolution of HIV envelope and
apex-targeting neutralizing antibody lineage provides
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
17 of 19
RES EARCH | R E S E A R C H A R T I C L E
59.
benchmarks for vaccine design. Cell Rep. 23, 3249–3261
(2018). doi: 10.1016/j.celrep.2018.05.046; pmid: 29898396
I. S. Georgiev et al., Delineating antibody recognition in
polyclonal sera from patterns of HIV-1 isolate neutralization.
Science 340, 751–756 (2013). doi: 10.1126/science.1233989;
pmid: 23661761
60. C. A. Bricault et al., HIV-1 neutralizing antibody signatures
61.
and application to epitope-targeted vaccine design. Cell Host
Microbe 25, 59–72.e8 (2019). doi: 10.1016/j.
chom.2018.12.001; pmid: 30629920
L. M. Walker et al., Rapid development of glycan-specific,
broad, and potent anti-HIV-1 gp120 neutralizing antibodies in
an R5 SIV/HIV chimeric virus infected macaque. Proc. Natl.
Acad. Sci. U.S.A. 108, 20125–20129 (2011). doi: 10.1073/
pnas.1117531108; pmid: 22123961
62. M. Jia, H. Lu, M. Markowitz, C. Cheng-Mayer, X. Wu,
Development of broadly neutralizing antibodies and their
mapping by monomeric gp120 in human immunodeficiency
virus type 1-infected humans and simian-human
immunodeficiency virus SHIVSF162P3N-infected macaques.
J. Virol. 90, 4017–4031 (2016). doi: 10.1128/JVI.02898-15;
pmid: 26842476
63. N. Gao et al., Development of broad neutralization activity in
simian/human immunodeficiency virus-infected rhesus
macaques after long-term infection. AIDS 32, 555–563
(2018). doi: 10.1097/QAD.0000000000001724;
pmid: 29239895
64. N. Gao et al., Development of antibodies with broad
neutralization specificities against HIV-1 after long term SHIV
infection in macaques. Viruses 12, 163 (2020). doi: 10.3390/
v12020163; pmid: 32023860
65. H. X. Liao et al., Induction of antibodies in guinea pigs and
rhesus monkeys against the human immunodeficiency virus
type 1 envelope: Neutralization of nonpathogenic and
pathogenic primary isolate simian/human immunodeficiency
virus strains. J. Virol. 74, 254–263 (2000). doi: 10.1128/
JVI.74.1.254-263.2000; pmid: 10590113
66. B. F. Haynes et al., Analysis of HIV-1 subtype B third variable
region peptide motifs for induction of neutralizing antibodies
against HIV-1 primary isolates. Virology 345, 44–55
(2006). doi: 10.1016/j.virol.2005.08.042; pmid: 16242749
67. S. Wang et al., Polyvalent HIV-1 Env vaccine formulations
delivered by the DNA priming plus protein boosting approach
are effective in generating neutralizing antibodies against
primary human immunodeficiency virus type 1 isolates from
subtypes A, B, C, D and E. Virology 350, 34–47 (2006).
doi: 10.1016/j.virol.2006.02.032; pmid: 16616287
68. M. Pauthner et al., Elicitation of robust tier 2 neutralizing
antibody responses in nonhuman primates by HIV
envelope trimer immunization using optimized approaches.
Immunity 46, 1073–1088.e6 (2017). doi: 10.1016/
j.immuni.2017.05.007; pmid: 28636956
69. K. O. Saunders et al., Vaccine induction of heterologous
tier 2 HIV-1 neutralizing antibodies in animal models. Cell Rep.
21, 3681–3690 (2017). doi: 10.1016/j.celrep.2017.12.028;
pmid: 29281818
70. A. Torrents de la Peña et al., Immunogenicity in rabbits of
HIV-1 SOSIP trimers from clades A, B, and C, given
individually, sequentially, or in combination. J. Virol. 92,
e01957-17 (2018). doi: 10.1128/JVI.01957-17;
pmid: 29367243
R. Kong et al., Antibody lineages with vaccine-induced
antigen-binding hotspots develop broad HIV neutralization.
Cell 178, 567–584.e19 (2019). doi: 10.1016/j.
cell.2019.06.030; pmid: 31348886
71.
72. P. J. Klasse, G. Ozorowski, R. W. Sanders, J. P. Moore, Env
exceptionalism: Why are HIV-1 env glycoproteins atypical
immunogens? Cell Host Microbe 27, 507–518 (2020).
doi: 10.1016/j.chom.2020.03.018; pmid: 32272076
Z. Wang et al., A broadly neutralizing macaque monoclonal
antibody against the HIV-1 V3-Glycan patch. eLife 9, e61991
(2020). doi: 10.7554/eLife.61991; pmid: 33084569
73.
74. B. F. Haynes, D. R. Burton, J. R. Mascola, Multiple roles for
HIV broadly neutralizing antibodies. Sci. Transl. Med. 11,
eaaz2686 (2019). doi: 10.1126/scitranslmed.aaz2686;
pmid: 31666399
75. K. E. Stephenson, K. Wagh, B. Korber, D. H. Barouch, Vaccines
and broadly neutralizing antibodies for HIV-1 prevention.
Annu. Rev. Immunol. 38, 673–703 (2020). doi: 10.1146/
annurev-immunol-080219-023629; pmid: 32340576
76. A. Escolano et al., Immunization expands B cells specific to
HIV-1 V3 glycan in mice and macaques. Nature 570, 468–473
(2019). doi: 10.1038/s41586-019-1250-z; pmid: 31142836
77.
J. M. Steichen et al., A generalized HIV vaccine design
strategy for priming of broadly neutralizing antibody
responses. Science 366, eaax4380 (2019). doi: 10.1126/
science.aax4380; pmid: 31672916
78. R. Pejchal et al., A potent and broad neutralizing antibody
recognizes and penetrates the HIV glycan shield. Science
334, 1097–1103 (2011). doi: 10.1126/science.1213256;
pmid: 21998254
79. D. Sok et al., A prominent site of antibody vulnerability on
80.
HIV envelope incorporates a motif associated with CCR5
binding and its camouflaging glycans. Immunity 45, 31–45
(2016). doi: 10.1016/j.immuni.2016.06.026; pmid: 27438765
J. P. Julien et al., Broadly neutralizing antibody PGT121
allosterically modulates CD4 binding via recognition of the
HIV-1 gp120 V3 base and multiple surrounding glycans. PLOS
Pathog. 9, e1003342 (2013). doi: 10.1371/journal.
ppat.1003342; pmid: 23658524
81. B. F. Haynes, G. Kelsoe, S. C. Harrison, T. B. Kepler,
B-cell-lineage immunogen design in vaccine development
with HIV-1 as a case study. Nat. Biotechnol. 30, 423–433
(2012). doi: 10.1038/nbt.2197; pmid: 22565972
82. P. L. Moore, The neutralizing antibody response to the
HIV-1 env protein. Curr. HIV Res. 16, 21–28 (2018).
doi: 10.2174/1570162X15666171124122044; pmid: 29173180
83. M. Sarzotti-Kelsoe et al., Optimization and validation of
the TZM-bl assay for standardized assessments of
neutralizing antibodies against HIV-1. J. Immunol. Methods
409, 131–146 (2014). doi: 10.1016/j.jim.2013.11.022;
pmid: 24291345
84. Q. Han et al., Difficult-to-neutralize global HIV-1 isolates
85.
are neutralized by antibodies targeting open envelope
conformations. Nat. Commun. 10, 2898 (2019). doi: 10.1038/
s41467-019-10899-2; pmid: 31263112
L. D. Williams et al., Potent and broad HIV-neutralizing
antibodies in memory B cells and plasma. Sci. Immunol. 2,
eaal2200 (2017). doi: 10.1126/sciimmunol.aal2200;
pmid: 28783671
97.
J. Huang et al., Isolation of human monoclonal antibodies
from peripheral blood B cells. Nat. Protoc. 8, 1907–1915
(2013). doi: 10.1038/nprot.2013.117; pmid: 24030440
98. N. Doria-Rose et al., High throughput HIV-1
microneutralization assay. Protoc. Exch. 10.1038/
protex.2013.069 (2013). doi: 10.1038/protex.2013.069
99. K. Smith et al., Rapid generation of fully human monoclonal
antibodies specific to a vaccinating antigen. Nat. Protoc. 4,
372–384 (2009). doi: 10.1038/nprot.2009.3; pmid: 19247287
100. B. Ewing, L. Hillier, M. C. Wendl, P. Green, Base-calling of
automated sequencer traces using phred. I. Accuracy
assessment. Genome Res. 8, 175–185 (1998). doi: 10.1101/
gr.8.3.175; pmid: 9521921
101. J. M. Volpe, L. G. Cowell, T. B. Kepler, SoDA: Implementation
of a 3D alignment algorithm for inference of antigen receptor
recombinations. Bioinformatics 22, 438–444 (2006).
doi: 10.1093/bioinformatics/btk004; pmid: 16357034
102. T. B. Kepler et al., Reconstructing a B-cell clonal lineage. II.
Mutation, selection, and affinity maturation. Front. Immunol.
5, 170 (2014). doi: 10.3389/fimmu.2014.00170;
pmid: 24795717
103. R. Zhang et al., Initiation of immune tolerance-controlled HIV
gp41 neutralizing B cell lineages. Sci. Transl. Med. 8, 336ra62
(2016). doi: 10.1126/scitranslmed.aaf0618; pmid: 27122615
104. T. Tiller et al., Efficient generation of monoclonal antibodies
from single human B cells by single cell RT-PCR and
expression vector cloning. J. Immunol. Methods 329, 112–124
(2008). doi: 10.1016/j.jim.2007.09.017; pmid: 17996249
105. D. Mason et al., Targeted isolation of antibodies directed
against major sites of SIV Env vulnerability. PLOS Pathog. 12,
e1005537 (2016). doi: 10.1371/journal.ppat.1005537;
pmid: 27064278
106. S. J. Krebs et al., Longitudinal analysis reveals early
development of three MPER-directed neutralizing antibody
lineages from an HIV-1-Infected individual. Immunity 50,
677–691.e13 (2019). doi: 10.1016/j.immuni.2019.02.008;
pmid: 30876875
86. S. M. Alam et al., Antigenicity and immunogenicity of RV144
vaccine AIDSVAX clade E envelope immunogen is enhanced
by a gp120 N-terminal deletion. J. Virol. 87, 1554–1568
(2013). doi: 10.1128/JVI.00718-12; pmid: 23175357
87. S. M. Alam et al., Mimicry of an HIV broadly neutralizing
107. C. A. Schramm et al., SONAR: A high-throughput pipeline for
inferring antibody ontogenies from longitudinal sequencing of
B cell transcripts. Front. Immunol. 7, 372 (2016).
doi: 10.3389/fimmu.2016.00372; pmid: 27708645
108. J. H. Ward Jr.., Hierarchical grouping to optimize an objective
antibody epitope with a synthetic glycopeptide. Sci. Transl.
Med. 9, eaai7521 (2017). doi: 10.1126/scitranslmed.aai7521;
pmid: 28298421
88. H. X. Liao et al., High-throughput isolation of immunoglobulin
genes from single human B cells and expression as
monoclonal antibodies. J. Virol. Methods 158, 171–179
(2009). doi: 10.1016/j.jviromet.2009.02.014; pmid: 19428587
function. J. Am. Stat. Assoc. 58, 236–244 (1963).
doi: 10.1080/01621459.1963.10500845
109. R. W. Sanders et al., A next-generation cleaved, soluble HIV-
1 Env trimer, BG505 SOSIP.664 gp140, expresses multiple
epitopes for broadly neutralizing but not non-neutralizing
antibodies. PLOS Pathog. 9, e1003618 (2013). doi: 10.1371/
journal.ppat.1003618; pmid: 24068931
89. M. Bonsignori et al., Analysis of a clonal lineage of
110. M. Pancera et al., Structure and immune recognition of
HIV-1 envelope V2/V3 conformational epitope-specific
broadly neutralizing antibodies and their inferred unmutated
common ancestors. J. Virol. 85, 9998–10009 (2011).
doi: 10.1128/JVI.05045-11; pmid: 21795340
90. M. Blasi et al., IDLV-HIV-1 Env vaccination in non-human
primates induces affinity maturation of antigen-specific
memory B cells. Commun. Biol. 1, 134 (2018). doi: 10.1038/
s42003-018-0131-6; pmid: 30272013
91. M. A. Moody et al., HIV-1 gp120 vaccine induces affinity
maturation in both new and persistent antibody clonal
lineages. J. Virol. 86, 7496–7507 (2012). doi: 10.1128/
JVI.00426-12; pmid: 22553329
92. K. Wiehe et al., Antibody light-chain-restricted recognition of
the site of immune pressure in the RV144 HIV-1 vaccine trial
is phylogenetically conserved. Immunity 41, 909–918 (2014).
doi: 10.1016/j.immuni.2014.11.014; pmid: 25526306
93. X. Wu et al., Rational design of envelope identifies broadly
neutralizing human monoclonal antibodies to HIV-1.
Science 329, 856–861 (2010). doi: 10.1126/
science.1187659; pmid: 20616233
94. D. C. Montefiori, Evaluating neutralizing antibodies against
HIV, SIV, and SHIV in luciferase reporter gene assays.
Curr. Protoc. Immunol. 64, 12.11.1–12.11.17 (2005).
pmid: 18432938
95. S. P. Perfetto et al., Amine-reactive dyes for dead cell
discrimination in fixed samples. Curr. Protoc. Cytom. 53,
9.34.1–9.34.14 (2010). doi: 10.1002/0471142956.cy0934s53;
pmid: 20578108
96. M. M. Donaldson et al., Optimization and qualification of an
8-color intracellular cytokine staining assay for quantifying
T cell responses in rhesus macaques for pre-clinical vaccine
studies. J. Immunol. Methods 386, 10–21 (2012).
doi: 10.1016/j.jim.2012.08.011; pmid: 22955212
trimeric pre-fusion HIV-1 Env. Nature 514, 455–461 (2014).
doi: 10.1038/nature13808; pmid: 25296255
111. Y. D. Kwon et al., Surface-matrix screening identifies
semi-specific interactions that improve potency of a near pan-
reactive HIV-1-neutralizing antibody. Cell Rep. 22, 1798–1809
(2018). doi: 10.1016/j.celrep.2018.01.023; pmid: 29444432
112. C. Suloway et al., Automated molecular microscopy: The new
Leginon system. J. Struct. Biol. 151, 41–60 (2005).
doi: 10.1016/j.jsb.2005.03.010; pmid: 15890530
113. N. R. Voss, C. K. Yoshioka, M. Radermacher, C. S. Potter,
B. Carragher, DoG Picker and TiltPicker: Software tools to
facilitate particle selection in single particle electron
microscopy. J. Struct. Biol. 166, 205–213 (2009).
doi: 10.1016/j.jsb.2009.01.004; pmid: 19374019
114. G. C. Lander et al., Appion: An integrated, database-driven
pipeline to facilitate EM image processing. J. Struct. Biol. 166,
95–102 (2009). doi: 10.1016/j.jsb.2009.01.002;
pmid: 19263523
115. S. Q. Zheng et al., MotionCor2: Anisotropic correction of
beam-induced motion for improved cryo-electron
microscopy. Nat. Methods 14, 331–332 (2017). doi: 10.1038/
nmeth.4193; pmid: 28250466
116. A. Rohou, N. Grigorieff, CTFFIND4: Fast and accurate defocus
estimation from electron micrographs. J. Struct. Biol. 192,
216–221 (2015). doi: 10.1016/j.jsb.2015.08.008;
pmid: 26278980
117. K. Zhang, Gctf: Real-time CTF determination and correction.
J. Struct. Biol. 193, 1–12 (2016). doi: 10.1016/
j.jsb.2015.11.003; pmid: 26592709
118. S. H. Scheres, RELION: Implementation of a Bayesian
approach to cryo-EM structure determination. J. Struct. Biol.
180, 519–530 (2012). doi: 10.1016/j.jsb.2012.09.006;
pmid: 23000701
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
18 of 19
RES EARCH | R E S E A R C H A R T I C L E
119. A. Punjani, J. L. Rubinstein, D. J. Fleet, M. A. Brubaker,
cryoSPARC: Algorithms for rapid unsupervised cryo-EM
structure determination. Nat. Methods 14, 290–296 (2017).
doi: 10.1038/nmeth.4169; pmid: 28165473
120. Q. Liu et al., Improvement of antibody functionality by
structure-guided paratope engraftment. Nat. Commun. 10,
721 (2019). doi: 10.1038/s41467-019-08658-4;
pmid: 30760721
121. P. D. Adams et al., Recent developments in the PHENIX
software for automated crystallographic structure
determination. J. Synchrotron Radiat. 11, 53–55 (2004).
doi: 10.1107/S0909049503024130; pmid: 14646133
122. P. Emsley, K. Cowtan, Coot: Model-building tools for
123.
molecular graphics. Acta Crystallogr. D 60, 2126–2132
(2004). doi: 10.1107/S0907444904019158; pmid: 15572765
I. W. Davis, L. W. Murray, J. S. Richardson, D. C. Richardson,
MOLPROBITY: Structure validation and all-atom contact
analysis for nucleic acids and their complexes. Nucleic Acids
Res. 32 (suppl. 2), W615–W619 (2004). doi: 10.1093/nar/
gkh398; pmid: 15215462
124. B. A. Barad et al., EMRinger: Side chain-directed model
and map validation for 3D cryo-electron microscopy.
Nat. Methods 12, 943–946 (2015). doi: 10.1038/nmeth.3541;
pmid: 26280328
125. D. Fera et al., Affinity maturation in an HIV broadly
neutralizing B-cell lineage through reorientation of variable
domains. Proc. Natl. Acad. Sci. U.S.A. 111, 10275–10280
(2014). doi: 10.1073/pnas.1409954111; pmid: 24982157
126. W. Kabsch, XDS. Acta Crystallogr. D 66, 125–132 (2010).
doi: 10.1107/S0907444909047337; pmid: 20124692
127. A. J. McCoy et al., Phaser crystallographic software.
J. Appl. Crystallogr. 40, 658–674 (2007). doi: 10.1107/
S0021889807021206; pmid: 19461840
128. P. Emsley, B. Lohkamp, W. G. Scott, K. Cowtan, Features and
development of Coot. Acta Crystallogr. D 66, 486–501
(2010). doi: 10.1107/S0907444910007493; pmid: 20383002
129. G. Tang et al., EMAN2: An extensible image processing suite
for electron microscopy. J. Struct. Biol. 157, 38–46 (2007).
doi: 10.1016/j.jsb.2006.05.009; pmid: 16859925
130. F. K. Treurnicht et al., Adaptive changes in HIV-1 subtype
C proteins during early infection are driven by changes in
HLA-associated immune pressure. Virology 396, 213–225
(2010). doi: 10.1016/j.virol.2009.10.002; pmid: 19913270
131. R. Andrabi et al., The chimpanzee SIV envelope trimer:
Structure and deployment as an HIV vaccine template.
Cell Rep. 27, 2426–2441.e6 (2019). doi: 10.1016/
j.celrep.2019.04.082; pmid: 31116986
ACKN OWLED GMEN TS
We thank A. Abuahmad, H. Chen, A. Eaton, A. Foulger, M. Gladden,
T. Gurley, G. Hernandez, A. Huang, C. Jones, J. Kim, J. Meyer,
A. Monroe, R. Parks, A. Ransier, J. Rathmann, P. Rawls, A. Sanzone,
J. Sprenz, K. Tilahun, R. Verardi, T. Von Holle, A. Wang, and
R. Zhang for technical assistance and the flow cytometry core staff
at the Duke Human Vaccine Institute and NIAID/NIH Vaccine
Research Center. We thank T. Denny, T. Demarco, and N. DeNaeyer
and members of the Nonhuman Primate Virology Core Laboratory
at the Duke Human Vaccine Institute for SHIV plasma vRNA
measurements, and J. Baalwa, D. Ellenberger, F. Gao, K. Hong,
F. McCutchan, D. Montefiori, L. Morris, J. Overbaugh, E. Sanders-Buell,
R. Swanstrom, M. Thomson, S. Tovanabutra, C. Williamson, and
L. Zhang for contributing HIV-1 Env plasmids. We thank K. McKee,
C. Moore, S. O’Dell, G. Padilla, S. D. Schmidt, C. Whittaker,
A. B. McDermott, and M. Seaman for assistance with neutralization
assays, D. Fera for helpful advice and discussion, and the staff at
Bioqual for exceptional care and assistance with nonhuman
primates. Cryo-EM was performed at the Simons Electron
Microscopy Center and National Resource for Automated
Molecular Microscopy located at the New York Structural Biology
Center and was supported by grants from the Simons Foundation
(SF349247), NYSTAR, and the NIH National Institute of General
Medical Sciences GM103310. X-ray diffraction data were collected
on the Northeastern Collaborative Access Team beamline 24 ID-C
(Advanced Photon Source). Funding: This work was funded by
the Bill and Melinda Gates Foundation (OPP1145046 and
OPP1206647/INV-007939); by the Intramural Research Program
of the NIAID/NIH Vaccine Research Center; and by the Division of
AIDS (NIAID/NIH) through support of the Duke Center for HIV/
AIDS Vaccine Immunology-Immunogen Discovery (UM1 AI100645),
the Duke Consortium for HIV/AIDS Vaccine Development (UM1
AI144371), the Penn Center for AIDS Research (P30 AI045008),
and grants AI131251, AI131331, AI128832, AI050529, AI150590, and
AI140897. R.S.R. was supported by an NIH training grant in HIV
pathogenesis (T32-AI007632). X-ray diffraction data were funded
by NIH grant P41 GM103403. Author contributions:
Conceptualization: G.M.S., B.H.H., S.C.H., B.T.K., P.D.K., J.R.M.,
G.K., M.R., D.C.D., H.L., and R.S.R. Investigation and data analysis:
R.S.R., H.L., W.B.W., H.C., R.D.M., J.G., S.W., F.-H.L., J.R., M.B.,
K.-K.H., K.O.S., K.Wi., M.A.M., P.T.H., K.Wa., E.E.G., R.M.R., F.B.-R.,
W.L., J.C., A.G.S., J.D., A.I.M., J.S., W.D., C.Z., N.C., M.O., C.R.,
Y.D., E.L., A.M.B., K.J.B., D.A., C.W.C., G.-Y.C., H.G., B.C.L., M.K.L.,
R.N., B.Z., M.G.L., D.D.R., N.A.D.-R., C.A.S., D.C.D., M.R., T.B.K.,
G.K., J.R.M., P.D.K., B.T.K., S.C.H., B.F.H., B.H.H., and G.M.S.
Writing – original draft: G.M.S., R.S.R., W.B.W., B.T.K., H.C., S.C.H.,
and P.D.K. Writing – review, editing, and approval: all authors.
Competing interests: The authors do not have any competing
interests. Data and materials availability: Sequencing data are
available from GenBank: MN471655-MN472016, MT484339-
MT487491, MT580365-MT580445, MT509359, GQ999989,
HQ625604, KC863461-KC863464, KF241776, KF996577-
KF996601, KF996604, KF996606, KF996610-KF996630,
KF996632-KF996662, KF996664-KF996678, KF996680,
KF996682-KF996683, KF996685-KF996716, KT698223-KT698227,
MF572809-MF572829, EF203980-EF203981, MK205498-
MK205507, MT581213-MT581268, MT610888-MT610895, and
MT656172-MT656253. Cryo-EM maps and fitted coordinates have
been deposited with database codes EMDB-22295 and PDB ID
6XRT, respectively. DH650 bound to the gp120 core was deposited
with database code PDB ID 6XCJ.
SUPPLEMENTARY MATERIALS
science.sciencemag.org/content/371/6525/eabd2638/suppl/DC1
Supplementary Text
Figs. S1 to S27
Tables S1 to S6
References (132–136)
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
10 June 2020; accepted 9 November 2020
Published online 19 November 2020
10.1126/science.abd2638
Roark et al., Science 371, eabd2638 (2021)
8 January 2021
19 of 19
|
10.1126_science.abe0511
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
BACTERIAL PHYLOGENY
A rooted phylogeny resolves early bacterial evolution
Gareth A. Coleman†, Adrián A. Davín†, Tara A. Mahendrarajah, Lénárd L. Szánthó, Anja Spang,
Philip Hugenholtz‡*, Gergely J. Szöllősi‡*, Tom A. Williams‡*
INTRODUCTION: Bacteria are the most diverse
and abundant cellular organisms on Earth, and
in recent years environmental genomics has
revealed the existence of an enormous diver-
sity of previously unknown lineages. Despite
the abundance of genomic sequence data, the
root of the bacterial tree and the nature of the
most recent common ancestor of Bacteria have
remained elusive. The problem is that even
with the help of new data, tracing billions of
years of bacterial evolution back to the root
has remained challenging because standard
phylogenetic models do not account for the
full range of evolutionary processes that shape
bacterial genomes. In particular, standard mod-
els treat horizontal gene transfer as an imped-
iment to the reconstruction of the tree of life
that must be removed from analyses. But if hori-
zontal gene transfer is modeled appropriately, it
can provide information about the deep past
that is unavailable to standard methods.
RATIONALE: We reconstructed and rooted the
bacterial tree by applying a hierarchical phylo-
genomic approach that explicitly uses informa-
tion from gene duplications and losses within
a genome as well as gene transfers between
genomes. This approach allowed us to root the
tree without including an archaeal outgroup.
Outgroup-free rooting is a promising approach
for Bacteria, both because the position of the
universal root is uncertain and because the
long branch separating Bacteria from Archaea
has the potential to distort the reconstruction
of within-Bacteria relationships. Outgroup-free
gene tree-species tree reconciliation allowed us
to quantitatively model both the vertical and
horizontal components of bacterial evolution
and integrate information from 11,272 gene
families to resolve the root of the bacterial
tree. Notably, these analyses also provided
estimates of the gene content of the last bac-
terial common ancestor.
A rooted phylogeny of Bacteria. The reconciliation of bacterial gene phylogenies places the root between
the major clades of Gracilicutes (including Proteobacteria and Bacteroidota) and Terrabacteria (including
Firmicutes and Cyanobacteria). On the basis of this hypothesis, ancestral genome reconstruction predicts
that the last bacterial common ancestor (LBCA) was a complex, double-membraned cell and that, on
average, two-thirds of gene transmissions have been vertically inherited along this rooted tree.
RESULTS: Our analyses place the root between
two major bacterial clades, the Gracilicutes
and Terrabacteria. We found no support for
a root between the Candidate Phyla Radiation
(CPR), a lineage comprising putative sym-
bionts and parasites with small genomes,
and all other Bacteria. Instead, the CPR was
inferred to be a member of the Terrabacteria
and formed a sister lineage to the Chloro-
flexota and Dormibacterota. This suggests
that the CPR evolved by reductive genome
evolution from free-living ancestors. Gene
families inferred to have been present at the
root indicate that the last bacterial common
ancestor was already a complex double-
membraned cell capable of motility and chem-
otaxis that possessed a CRISPR-Cas system.
Although ~92% of gene families have ex-
perienced horizontal transfers during their
history, tracing their evolution along the most
likely rooted tree revealed that about two-thirds
of gene transmissions have been vertical. Thus,
bacterial evolution has a major vertical com-
ponent, despite a profound impact of hori-
zontal gene transfer through time. Horizontal
gene flows can also provide insight into the
temporal sequence of events during bacterial
diversification, because donor lineages must
be at least as old as recipients. Analysis of
gene transfers in our dataset suggests that the
diversification of the Firmicutes, CPR, Acid-
obacteriota, and Proteobacteria is the oldest
among extant bacterial phyla.
CONCLUSION: The vertical and horizontal com-
ponents of genome evolution provide com-
plementary sources of information about
bacterial phylogeny. The vertical component
provides a robust framework for interpreting
species diversity and allows us to reconstruct
ancestral states, while the horizontal compo-
nent helps to root the vertical tree and orient it
in time. The inferred Gracilicutes-Terrabacteria
root will be useful for investigating the tem-
po and mode of bacterial diversification,
metabolic innovation, and changes in cell ar-
chitecture such as the evolutionary transitions
between double (diderm) and single (mono-
derm) membranes. Future development of
methods that harness the complementarity of
vertical and horizontal gene transmission will
continue to further our understanding of life
on Earth.▪
The list of author affiliations is available in the full article online.
†These authors contributed equally to this work.
‡These authors contributed equally to this work.
*Corresponding author. Email: p.hugenholtz@uq.edu.au
(P.H.); ssolo@elte.hu (G.J.Sz.); tom.a.williams@bristol.ac.
uk (T.A.W.)
Cite this article as G. A. Coleman et al., Science 372,
eabe0511 (2021). DOI: 10.1126/science.abe0511
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.abe0511
Coleman et al., Science 372, 588 (2021)
7 May 2021
1 of 1
Corrected 30 June 2021. See full text. RES EARCH
R E S E A R C H A R T I C L E
◥
BACTERIAL PHYLOGENY
A rooted phylogeny resolves early bacterial evolution
Gareth A. Coleman1†, Adrián A. Davín2†, Tara A. Mahendrarajah3, Lénárd L. Szánthó4,5, Anja Spang3,6,
Philip Hugenholtz2‡*, Gergely J. Szöllősi4,5,7‡*, Tom A. Williams1‡*
A rooted bacterial tree is necessary to understand early evolution, but the position of the root is contested.
Here, we model the evolution of 11,272 gene families to identify the root, extent of horizontal gene
transfer (HGT), and the nature of the last bacterial common ancestor (LBCA). Our analyses root the tree
between the major clades Terrabacteria and Gracilicutes and suggest that LBCA was a free-living
flagellated, rod-shaped double-membraned organism. Contrary to recent proposals, our analyses reject
a basal placement of the Candidate Phyla Radiation, which instead branches sister to Chloroflexota
within Terrabacteria. While most gene families (92%) have evidence of HGT, overall, two-thirds of gene
transmissions have been vertical, suggesting that a rooted tree provides a meaningful frame of reference
for interpreting bacterial evolution.
A species tree captures the relationships
among organisms but requires a root to
provide the direction of evolution. Root-
ing deep radiations (1) is among the
greatest challenges in phylogenetics, and
there is no consensus on the root of the bac-
terial tree. On the basis of evidence (2–5) that
the root of the tree of life lies between Bacteria
and Archaea, early analyses with an archaeal
outgroup placed the bacterial root near Aquifi-
cales and Thermotogales (6, 7) or Planctomycetes
(8). Alternative approaches, including analyses
of gene flows and the evolution of multimeric
protein complexes as well as other complex
characters (9), have instead suggested roots
within the monoderm (single-membrane) Bac-
teria (10) or between Chloroflexi and all other
cellular life (9). The development of techniques
for sequencing microbes directly from environ-
mental samples, without the need for labor-
atory cultivation, has greatly expanded the
genomic representation of natural prokaryotic
diversity (11–14). Recent phylogenomic analy-
ses of expanded sets of diverse bacteria have
placed the root between one of the recently
identified groups, the Candidate Phyla Radia-
tion [CPR, also known as Patescibacteria (15, 16)]
and all other Bacteria (11, 16, 17). The CPR is
characterized by small cells and genomes that
1School of Biological Sciences, University of Bristol, Bristol
BS8 1TQ, UK. 2Australian Centre for Ecogenomics, School of
Chemistry and Molecular Biosciences, The University of
Queensland, Brisbane, Queensland 4072, Australia.
3Department of Marine Microbiology and Biogeochemistry,
NIOZ, Royal Netherlands Institute for Sea Research, 1790 AB
Den Burg, Netherlands. 4Department of Biological Physics,
Eötvös Loránd University, 1117 Budapest, Hungary.
5MTA-ELTE “Lendület” Evolutionary Genomics Research
Group, 1117 Budapest, Hungary. 6Department of Cell- and
Molecular Biology, Uppsala University, SE-75123 Uppsala,
Sweden. 7Institute of Evolution, Centre for Ecological
Research, 1121 Budapest, Hungary.
†These authors contributed equally to this work.
‡These authors contributed equally to this work.
*Corresponding author. Email: p.hugenholtz@uq.edu.au (P.H.);
ssolo@elte.hu (G.J.Sz.); tom.a.williams@bristol.ac.uk (T.A.W.)
appear to have predominantly symbiotic or
parasitic lifestyles, but much remains to be
learned about their ecology and physiology
(15, 17–19). If correct, the early divergence of
the CPR has important implications for our
understanding of the earliest period of cellular
evolution. Along with evidence that the root of
the archaeal domain lies between the reduced
and predominantly host-associated DPANN
superphylum (originally named after Diapher-
otrites, Parvarchaeota, Aenigmarchaeota, Nano-
archaeota, and Nanohaloarchaeota) and all
other Archaea (1, 20), the CPR root implies
that streamlined, metabolically minimalist pro-
karyotes have coexisted with the more familiar,
self-sufficient lineages throughout the history
of cellular life (19, 21).
Improved taxon sampling can help to re-
solve evolutionary relationships (22, 23), and
the quantity and diversity of genome sequence
data now available presents an opportunity
to address long-standing questions about the
origins and diversification of Bacteria. How-
ever, deep phylogenetic divergences are diffi-
cult to resolve, both because the phylogenetic
signal for such relationships is overwritten by
new changes over time, and also because the
process of sequence evolution is more complex
than the best-fitting models currently availa-
ble. In particular, variation in nucleotide or
amino acid composition across the sites of the
alignment and the branches of the tree can
induce long branch attraction (LBA) artifacts
in which deep-branching, fast-evolving, poorly
sampled or compositionally biased lineages
group together irrespective of their evolution-
ary history (24). These issues are widely ap-
preciated (11) but are challenging to address
adequately, particularly when sequences from
thousands of taxa (11, 13, 14, 16, 17) are used to
estimate trees of global prokaryotic diversity,
which precludes the use of the best-fitting
phylogenetic methods available.
Archaeal outgroup rooting does not
unambiguously establish the root
of the bacterial tree
The standard approach to rooting is to include
an outgroup in the analysis, and all published
phylogenies in which CPR forms a sister line-
age to the rest of the Bacteria (11, 16, 17) have
made use of an archaeal outgroup. Outgroup
rooting on the bacterial tree, however, has
three serious limitations. First, interpretation
of the results requires the assumption that the
root of the tree of life lies between Bacteria
and Archaea. While this is certainly the con-
sensus view, the available evidence is limited
and difficult to interpret (2–5, 25), and alter-
native hypotheses in which the universal root
is placed within Bacteria have been proposed
on the basis of indels (26, 27) or the analysis
of slow-evolving characters (9). Second, the
long branch leading to the archaeal outgroup
has the potential to distort within-Bacteria
relationships because of LBA. Third, joint analy-
ses of Archaea and Bacteria use a smaller num-
ber of genes that are widely conserved and have
evolved vertically since the divergence of the
two lineages, and sequence alignment is more
difficult owing to the low sequence identity
between homologs of the two domains.
We evaluated the performance of outgroup
rooting on a bacterial tree using 265 Bacteria
(see below) and 149 Archaea from a shared
subset of 29 phylogenetic markers (table S1).
Using this archaeal outgroup, the maximum
likelihood (ML) phylogeny under the best-
fitting model (LG+C60+R8+F, which accounts
for site heterogeneity in the substitution pro-
cess) placed the bacterial root between a clade
comprising Cyanobacteria, Margulisbacteria,
CPR, Chloroflexota, and Dormibacterota on
one side of the root and all other taxa on the
other (fig. S1). However, bootstrap support
for this root, and indeed many other deep
branches in both the bacterial and archaeal
subtrees, was low (50 to 80%). We therefore
used approximately unbiased (AU) tests (28)
to determine whether a range of published
alternative rooting hypotheses (table S2)
could be rejected, given the model and data.
The AU test asks whether the optimal trees
that are consistent with these other hypothe-
ses have a significantly worse likelihood score
than the ML tree. In this case, the likelihoods
of all tested trees were statistically indisting-
uishable (AU test, P > 0.05, table S2), indicat-
ing that outgroup rooting cannot resolve the
bacterial root on this alignment.
An alternative to outgroup rooting for deep
microbial phylogeny
Given the limitations of using a remote archaeal
outgroup to establish the root of the bacterial
tree, we explored outgroup-free rooting using
gene tree-species tree reconciliation (1, 29–31).
We recently applied this approach to root the
Coleman et al., Science 372, eabe0511 (2021)
7 May 2021
1 of 9
Corrected 30 June 2021. See full text. RES EARCH | R E S E A R C H A R T I C L E
archaeal tree (1), and similar approaches have
been used to investigate the root of eukaryotes
(32, 33) and to map and characterize whole-
genome duplications in plants (34). Gene tree-
species tree reconciliation methods work by
adding a layer to the standard framework for
inferring trees from molecular data. This ad-
ditional step models the way in which gene
trees can differ from each other and the over-
arching rooted species tree. Substitution models
[such as LG (35)] describe how the constituent
sequences of a gene family evolve along a gene
tree via a series of amino acid substitutions
that allow us to infer the most likely gene
tree. Reconciliation models describe how a
gene tree evolves along the rooted species tree,
beginning with gene birth (origination) and
followed by a combination of vertical descent
and events such as gene duplications, transfers,
and losses (this series of events is known as a
DTL reconciliation). Combining the substitution-
based modeling of sequences along the gene
tree with the reconciliation-based modeling of
gene trees along a rooted species tree allows
us to infer the most likely rooted species tree
from the constituent gene families. In other
words, reconciliation methods aggregate phy-
logenetic signal across gene families and, be-
cause the likelihood of reconciliations depends
on the position of the root, can be used to test
the support for competing root positions (1, 29),
providing a genome-wide (and gene transfer–
aware) extension of the classical approach used
to root the tree of life on the basis of ancient
gene duplications (3, 4).
Our method, amalgamated likelihood esti-
mation (ALE), improves on earlier approaches
by explicitly accounting for uncertainty in the
gene tree topologies and in the events leading
to those topologies while simultaneously esti-
mating rates of gene duplication, transfer, and
loss directly from the data (31). Simulations
suggest that root inferences under ALE are
robust to variation in taxon sampling and the
proportion of extinct lineages (fig. S2), that the
method finds the correct root even under high
levels of gene transfer (1, 29), and that the num-
bers of D, T, and L events are accurately re-
covered from the data (figs. S3 to S8). These
results suggest that ALE is appropriate for
the problem at hand (36).
Rooting Bacteria without an outgroup
To obtain an unrooted species tree for ALE
analysis, we selected a focal dataset of 265
genomes representative of bacterial diversity
according to the Genome Taxonomy Database
(GTDB) (13). We inferred the tree from a con-
catenation of 62 conserved single-copy markers
(table S1) using the LG+C60+R8+F model in
IQ-Tree 1.6.10 (Fig. 1), which was chosen as the
best-fitting model using the Bayesian informa-
tion criterion (BIC) (37). This yielded highly
congruent trees when removing 20 to 80% of
the most compositionally heterogeneous sites
from the alignment (fig. S9), suggesting that
the key features of the topology are not
composition-driven LBA artifacts. One excep-
tion was the position of the Fusobacteriota,
which was recovered as a sister lineage to a
clade comprising Deinococcota, Synergistota,
and Thermotogota (DST) when 20% of the
most heterogeneous sites were removed (fig.
S9A) but was recovered as a single lineage be-
tween Terrabacteria plus DST and Gracilicutes
in all other trees.
We used ALE to test the support for 62 root
positions (tables S3 and S4) on the unrooted
topology by reconciling gene trees for 11,272
homologous gene families [inferred using MCL
(38)] from the 265 bacterial genomes. Note
that this method does not assume that the root
lies between Bacteria and Archaea. In addition
to testing root positions corresponding to pub-
lished hypotheses, we exhaustively tested all
inner nodes of the tree above the phylum level.
The ALE analysis rejected all of the root posi-
tions tested (AU test, P < 0.05) except for three
adjacent branches, lying between the two
major clades of Gracilicutes (comprising most
diderm lineages) and Terrabacteria (compris-
ing monoderm and atypical diderm line-
ages) (Fig. 1); the difference between the
three root positions was the position of the
Fusobacteriota in relation to these two major
clades (Fig. 1B). Alternative roots were rejected
with increasing confidence as distance from
the optimal root region increased (Fig. 1C and
table S3).
We tested the robustness of the inferred
root region by (i) excluding gene families with
extreme duplication, transfer, or loss rates; (ii)
repeating the analysis using gene families con-
structed with an assignment to families in the
Clusters of Orthologous Genes (COG) (39) onto-
logy; and (iii) repeating the analysis on a sec-
ondary independent sampling of the tree, in
which CPR makes up 40% of the genomes (11)
(figs. S10 to S13 and table S5). These analyses
consistently recovered the root between the
Gracilicutes and Terrabacteria, regardless of
the position of the Fusobacteriota. A Gracilicutes-
Terrabacteria root was previously reported
(40, 41), but these studies did not include the
CPR, which has recently been suggested to
represent the earliest diverging bacterial line-
age (11, 16). Our outgroup-free analysis con-
sistently recovered CPR nested within the
Terrabacteria, as a sister clade to Chloroflexota
and Dormibacterota, even with CPR represent-
ing more than 40% of the taxa included. This
finding implies that the CPR evolved by ge-
nome reduction from a free-living ancestor, a
scenario that has been proposed previously (21).
Transfers contain information about the rel-
ative timing of divergences, because for each
transfer, the donor must be at least as old as
the recipient (42, 43). To establish the relative
ages of the crown groups of different phyla, we
used high-confidence relative age constraints
recovered in at least 95 of 100 bootstrap rep-
licates common to the focal and secondary
datasets (36). Simulations suggest that this
approach accurately recovers relative clade ages
(98.4% accuracy on a simulated dataset the
same size as the focal dataset, fig. S14). Our
analysis (Fig. 2) predicts that the Firmicutes
crown group is the oldest among extant bac-
terial phyla (median rank: 2 ± 1.43 SD) fol-
lowed by the crown groups of the CPR (median
rank: 3 ± 2), Proteobacteria (median rank: 3 ±
1.59), and Acidobacteriota (median rank: 3 ±
1.56), suggesting that these lineages were the
earliest to diversify within Bacteria. The crown
groups of lineages predominantly associated
with animal hosts, Spirochaetota (median rank:
10 ± 0.85) and Elusimicrobiota (median rank:
11 ± 0.62), appear to be the youngest among
extant phyla.
Is bacterial evolution treelike?
How much of bacterial evolution can be ex-
plained by the concept of a rooted species
tree? Horizontal gene transfer (HGT) is fre-
quent in prokaryotes, and published analyses
indicate that most or all prokaryotic gene fam-
ilies have experienced HGT during their his-
tory (1, 44). This implies that there is no single
tree that fully describes the evolution of all
bacterial genes or genomes (45, 46). Extensive
HGT is existentially challenging for concate-
nation, because it greatly curtails the number
of genes that evolve on a single underlying tree
(47). Phylogenetic networks (46, 48) were the
first methods to explicitly acknowledge non-
vertical evolution, but they can be difficult to
interpret biologically. Gene tree-species tree
reconciliation unites tree and network-based
approaches by modeling both the horizontal
components of genome evolution (a fully re-
ticulated network allowing all possible transfers)
and the vertical trace (a common rooted spe-
cies tree). This framework enables us to quan-
tify the contributions of vertical and horizontal
processes to bacterial evolutionary history.
Our analyses (Fig. 3) reveal that most bac-
terial gene families present in two or more
species (9678 of 10,518 MCL families, or 92%)
have experienced at least one gene transfer
during their evolution; only very small fami-
lies have escaped transfer entirely on the time
scales considered here (fig. S15). Consistent
with previous analyses (1, 49), transfer rates
vary across gene functional categories, with
genes encoding proteins involved in defense
mechanisms (such as antibiotic biosynthesis)
and in the production of secondary metabo-
lites being the most frequently transferred,
and those coding for translational and cell
cycle proteins the least (Fig. 3B). Despite this
accumulation of HGT, most gene families
evolve vertically the majority of the time, with
Coleman et al., Science 372, eabe0511 (2021)
7 May 2021
2 of 9
Corrected 30 June 2021. See full text. RES EARCH | R E S E A R C H A R T I C L E
A
B
DST
Cyanobacteria-Margulis
DST
Cyanobacteria-Margulis
Actinobacteriota
Firmicutes
Armati-Eremi
Chloroflexota
CPR
Fusobacteriota
Spirochaetota
Elusimicrobiota
FCB
PVC
Dependentiae
ACD
Acidobacteriota
Nitrospirota
MBDD
Proteobacteria
Actinobacteriota
Firmicutes
Armat-Eremi
Chloroflexota
CPR
Fusobacteriota
Spirochaetota
Elusimicrobiota
FCB
PVC
Dependentiae
ACD
Acidobacteriota
Nitrospirota
MBDD
Proteobacteria
C
Fusobacteriota
DST
Cyanobacteria-Margulis
Actinobacteriota
Firmicutes
Armati-Eremi
Chloroflexota
CPR
Spirochaetota
Elusimicrobiota
FCB
PVC
Dependentiae
ACD
Acidobacteriota
Nitrospirota
MBDD
Proteobacteria
Root 1
Root 2
Root 3
Terrabacteria
Gracilicutes
Fusobacteriota/ DST
Distance from optimal rooting region
Coleman et al., Science 372, eabe0511 (2021)
7 May 2021
3 of 9
Corrected 30 June 2021. See full text. RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. A rooted phylogeny of Bacteria. (A) We used gene tree-species tree
reconciliation to infer the root of the bacterial tree. The unrooted maximum
likelihood phylogeny was inferred from a concatenation of 62 marker genes
under the best-fitting model, LG+C60+R8+F, which accounts for site heteroge-
neity in the substitution process and uses a mixture of eight substitution rates
estimated from the data to model across-site evolutionary rate variation.
Branches are colored according to bootstrap support value. The root falls
between two major clades of Bacteria, the Gracilicutes and the Terrabacteria, on
one of three statistically equivalent adjacent branches indicated by arrows,
shown as rooted trees in (B). All alternative roots tested were rejected (tables S3
and S4), with likelihoods decreasing as a function of distance from the root
region, as shown in (C). Previously proposed root positions, including the CPR
root, are highlighted in red. FCB are the Fibrobacterota, Chlorobia, Bacteroidota,
and related lineages; PVC are the Planctomycetota, Verrucomicrobiota, Chlamydiota,
and related lineages; DST are the Deinococcota, Synergistota, and Thermotogota;
ACD are Aquificota, Campylobacterota, and Deferribacterota; F/A are Firmicutes and
Actinobacteriota; MBDD are Myxococcota, Bdellovibrionota, Desulfomonadota, and
Desulfobacterota. Scale bar, 0.3 amino acid substitutions per site.
A
older
Bacteroidota
Verrucomicrobiota
Acidobacteriota
Proteobacteria
Elusimicrobiota
Spirochaeotota
Actinobacteriota
Firmicutes
CPR
Chloroflexota
Cyanobacteria
Gracilicutes
Terrabacteria
Verrucomicrobiota
Bacteroidota
Acidobacteriota
Elusimicrobiota
Proteobacteria
Spirochaetota
Actinobacteriota
Firmicutes
B
Order of diversification (oldest to youngest)
CPR
Chloroflexota
Cyanobacteria
Fig. 2. Relative crown group ages of major bacterial phyla. Gene transfers
that occurred during the diversification of Bacteria provide a record of the
temporal sequence of events. We used the information provided by directional
(donor-to-recipient) patterns of gene transfer to infer the relative ages of
bacterial crown groups, focusing on phyla represented by at least five genomes in both
of our datasets. To summarize this time information, we sampled 1000 time orders
that were fully compatible with the constraints recovered from both datasets.
(A) Pairwise relative ages of phyla. The proportion of sampled time orders in which
each phylum on the x axis was recovered as younger than each phylum on the y axis.
(B) Relative age distributions of major phyla. For each sampled time order, we
ranked the phyla from oldest (1) to youngest (11) and plotted the distribution of the
ranks. The crown group radiations of Firmicutes, CPR, Proteobacteria, and
Acidobacteriota appear to be the oldest among sampled phyla, while those of
Elusimicrobiota and Spirochaetota are the youngest.
mean verticality estimated to be 64% in the
focal and 68% in the secondary dataset.
Genome-wide reconciliation of gene trees
with the species tree demonstrates that the
optimal rooted species tree provides an apt
summary of much of bacterial evolutionary
history, even for the deepest branches of the
tree (50). From the gene’s eye view, gene fam-
ilies evolve neither entirely vertically nor hori-
zontally; core genes are occasionally transferred,
and even frequently exchanged genes contrib-
ute useful vertical signal; for example, the
median number of genes that evolve vertically
on a branch of the species tree is 998.92 in the
focal analysis (table S6), far greater than the
number of genes that have been concatenated
at the level of all Bacteria. From the perspec-
tive of the genome, constituent genes have
different ages (or residence times), correspond-
ing to the time at which they originated or were
most recently acquired by gene transfer, with-
in the resolution of our taxonomic sampling.
This analysis indicates that, on average, 82%
of all genes from adequately represented
phyla (five or more genomes) were most re-
cently acquired after the diversification of that
phylum, although all genomes retain a smaller
proportion (10 to 27%) of genes that have de-
scended vertically from the stem lineage of
their phylum or even earlier (Fig. 3C). There
are two explanations for this distribution of
gene persistence times: (i) de novo gene orig-
ination within phyla (that is, lineage-specific
gene families) and (ii) the cumulative impact
of gene transfer, which curtails gene persist-
ence times when looking back from the pres-
ent day even though most transmissions are
vertical.
Ancestral proteome of the last bacterial
common ancestor
Reconciliation analyses not only allow us to
infer the acquisition of genes across the tree
but also to estimate the metabolic potential of
the last bacterial common ancestor (LBCA). We
built a second, smaller set of COG-based gene
families better suited for functional annota-
tion and reconciled their gene trees with the
Coleman et al., Science 372, eabe0511 (2021)
7 May 2021
4 of 9
Corrected 30 June 2021. See full text. RES EARCH | R E S E A R C H A R T I C L E
A
B
Acidobacteriota
C
Fig. 3. The verticality of bacterial genome evolution. (A) The rooted bacterial
species tree (Fig. 1), with branches colored according to verticality: the fraction
of genes at the bottom of a branch that descend vertically from the top of that
branch (see inset; V, vertical; O, origination; T, transfer into a branch) (36). Node
heights reflect relative time order consistent with highly supported gene
transfers (Fig. 2). (B) Transfer propensity by COG functional category; that is,
the proportion of gene tree branches that are horizontal T/(V+T) for COG gene
families. Genes involved in information processing, particularly translation (J),
show the lowest transfer propensity (median: 0.31), while genes involved in cell
defense mechanisms (V, such as genes involved in antibiotic defense and
biosynthesis) are most frequently transferred (median transfer propensity: 0.47).
(C) From the genome’s eye view, this combination of vertical and horizontal
processes gives rise to a distribution of gene persistences (residence times),
reflecting the point in evolutionary history [within the Crown group, on the Stem,
or earlier (Before)] at which the gene was most recently acquired. Across all
phyla examined, 82% of genes on sampled genomes were most recently
acquired since the crown group radiation of that phylum. Lineage acronyms are
as in Fig. 1.
species tree (36). In the following reconstruc-
tion, we indicate when gene content inferences
differ between roots (36). Posterior probabil-
ities (PPs) for genes directly relevant to our
reconstruction are provided in table S7, and
all of the pathways we discuss below were con-
firmed in our analysis of the secondary dataset
(36). From the root placement and estimated
rates of gene family extinction in the focal
analysis (1), we predict that LBCA encoded 1293
to 2143 COG family members, the majority of
which (median estimates: 65 to 69.5%; 95%
confidence interval: 57 to 82%) survived to be
sampled in at least one present-day genome.
On the basis of the relationship between COG
family members and genome size for extant
Bacteria (Pearson’s correlation coefficient =
0.96, P = 8 × 10−153), we estimate the genome
size of LBCA to be 2.7 ± 0.4 Mb (SE) for root
1 of the focal analysis (Fusobacteriota with
Terrabacteria) (Fig. 1B), 2.6 ± 0.4 Mb for root 2
(Fusobacteriota with Gracilicutes), and 1.6 ±
0.5 Mb for root 3 (Fusobacteriota root). Under
all three roots, the trend in genome size evo-
lution from LBCA to modern taxa is an ongoing
moderate increase through time in estimated
COG family complements and genome sizes.
The most notable departure from this trend
is a reduction in genome size of between 0.47
and 0.56 Mb on the CPR stem lineage after
Coleman et al., Science 372, eabe0511 (2021)
7 May 2021
5 of 9
Corrected 30 June 2021. See full text. RES EARCH | R E S E A R C H A R T I C L E
divergence from their common ancestor with
Chloroflexota and Dormibacterota (fig. S16).
COG families lost on the CPR stem include
components of the electron transport chain,
carbon metabolism, flagellar biosynthesis and
motor switch proteins, amino acid biosynthe-
sis, the Clp protease subunit ClpX, and RNA
polymerase sigma factor-54 (table S8), consist-
ent with their absence in extant CPR (18).
The inferred ancestral gene set for LBCA
includes most components of the modern bac-
terial transcription, translation, and DNA repli-
cation systems (table S7). This gene set also
includes an FtsZ-based cell division machin-
ery and pathways for signal transduction,
membrane transport, and secretion (Fig. 4)
(36). Further, we identified proteins involved
in bacterial phospholipid biosynthesis, suggest-
ing that LBCA had bacterial-type ester-lipid
membranes (Fig. 4). We also identified most
of the proteins required for flagella and pili
synthesis and those for quorum sensing, sug-
gesting that LBCA was motile (51, 52). Given
that bacterial genes are typically maintained
by strong purifying selection (53), these find-
ings imply that LBCA lived in an environment
in which dispersal, chemotaxis, and surface
attachment were advantageous.
Moderate support for the presence of the
shape-determining proteins MreB (PP = 0.9,
0.71, and 0.49 for roots 1 to 3, respectively, as
depicted in Fig. 1B), MreC (PP = 0.82/0.78/
0.68), and MreD (PP = 0.86/0.84/0.74) at the
root suggests that LBCA was a rod-shaped cell
(52). We also obtained high root PPs for pro-
teins mediating outer cell envelope biosynthe-
sis, including lipopolysaccharides (LPSs), from
which we infer that LBCA had a double mem-
brane with an LPS layer (36). Consistent with
this inference, there was strong support for the
flagellar subunits FlgH, FlgI, and FgA, which
anchor flagella in diderm membranes (54), and
for the type IV pilus subunit PilQ, which among
extant bacteria is specific to diderms (54, 55).
Altogether, this supports hypotheses (9) in
which LBCA was a diderm (54–56) and argues
against scenarios in which the Gram-negative
double membrane originated by endosymbiosis
between monoderms [single-membraned bac-
teria (10)] or via the arrest of sporulation (57)
in a spore-forming monoderm ancestor. Thus,
diderm-to-monoderm transitions must have
occurred subsequently on multiple occasions
within Bacteria (54–56).
We recovered components of several core
pathways for carbohydrate metabolism with
high posterior support, including glycolysis,
the tricarboxylic acid (TCA) cycle, and the pen-
tose phosphate pathway (Fig. 4, figs. S17 and
S18, and table S7) (36). Modern bacteria fix
carbon using several different pathways, in-
cluding the Calvin cycle, the 3-hydroxypropionate
bicycle and variations thereof, the reductive
glycine pathway (58), the Wood-Ljungdahl
pathway (WLP), and the reverse TCA cycle, of
which the latter two have been suggested to
PP >0.75
At least 50% subunits with PP >0.75
PP=0.5-0.74
At least 1 subunit with PP >0.5
PP <0.5
PP <0.5
Type IV pili
PilQ
PilC
6P-D-Glucono-
1,5-lactano
PilB
D-Gluconate-6P
PilT
Inner membrane
Peptidoglycan wall
Xylulose-5P
Outer membrane
Erythrose-4P
Ribulose-5P
Glycolysis
Glucose
Fructose-6-P
Formate
G3P
Formyl-THF
Glycerol
Glycerone-P
Fructose-1,6-P
Glycerone
Sedoheptose-7P
Phosphatic acid
Glyceraldehyde-3P
Ribose-5P
JpsG
DpsG
Bacterial Secretion
System II
Methenyl-THF
1-acylglycerol-3P
Bacterial
Phospholipid
Biosynthesis
1,3-BP-Glycerate
Pentose Phosphate
Pathway
Sec
Lipopolysaccharides
Outer membrane
proteins
OmpH
BamA
LPS attachment
LptDE
LptB
MsbA
crr
Glucose
Reductive Glycine
Pathway
CO2
Wood-Ljungdahl
Pathway
Lipopolysaccharide
Biosynthesis
D-Arabinose-5P
Lauroyl-(KDO)2-lipid
(KDO)2-Lipid A
CMP-3-deoxy-D-manno
octulosonate (KDO)
Lipid A
UDP-GlcNAc
Lipid X
s
r
e
t
r
o
p
s
n
a
r
T
C
B
A
UDP-2,3-
diacylglucosamine
Disacchaide-
1-phosphate
Methylene-THF
Sulfur Metabolism
Sulfate
APS
PAPS
Sulfite
Sulfide
sbp
Sulfate
Sulfate
APS
Sulfite
Sulfide
Methyl-THF
S-amino-
methyldihydro-
lipoylprotien
Glycine
Methyl-CoFeSP
CO2
CO
Acetyl-phosphate
Acetyl-CoA
3-P-Glycerate
2-P-Glycerate
PEP
Pyruvate
ADP+P
ATP
ATP synthase
Cas6
Acetate
CRISPR-Cas
Cas7
Cas7
Cas7
Cas7
Cas7
Cas5
SS
SS
Cas1
Cas2
Electron transport chain
Complex IV
Complex III
Complex II
Complex I
NarGHI
RNF
Oxaloacetate
Citrate
Malate
Fumerate
TCA
cis-Aconitate
Isocitrate
Succinate
2-oxogluterate
Succinyl-CoA
Gram-negative
flagellum
Fig. 4. Ancestral reconstruction of the last bacterial common ancestor (LBCA).
The reconstruction is based on genes that could be mapped to at least one branch
within the root region with a PP > 0.5 (figs. S17 and S18) (36). The presence of a gene
within a pathway is indicated as shown in the key. Our analyses suggest that LBCA was a
rod-shaped, motile, flagellated double-membraned cell. We recover strong support for
central carbon pathways, including glycolysis, the TCA cycle, and the pentose phosphate
pathway. We did not find unequivocal evidence for the presence of a carbon fixation
pathway, but we did find moderate support for components of both the WLP and the
reverse TCA cycle. Although not depicted here, our analyses suggest that the machinery
for core information processing and quorum sensing was also present in LBCA (table
S7). As depicted in the key, arrows and boxes are shaded to indicate presence
probabilities and, for multisubunit complexes, the number of subunits recovered with PP
> 0.5. Note that the recovery of individual subunits for larger complexes does not imply
that the complex was present [see online data supplement (80) for further discussion].
Coleman et al., Science 372, eabe0511 (2021)
7 May 2021
6 of 9
Corrected 30 June 2021. See full text.
RES EARCH | R E S E A R C H A R T I C L E
have emerged early in the history of life
(41, 59–63). Of these, we identified several en-
zymes of the TCA cycle and the reductive gly-
cine pathway, although we did not recover the
key enzymes of either pathway, and the direc-
tionality of the recovered enzymes is difficult
to assess (64) (Fig. 4 and figs. S17 and S18).
Furthermore, we identified several enzymes of
the methyl branch of the WLP for acetate bio-
synthesis and components of a putative Rho-
dobacter nitrogen-fixing (RNF) complex (Fig. 4
and figs. S17 and S18), which together may in-
dicate that LBCA was capable of acetogenic
growth (36, 65). However, the key enzyme of
the WLP, the carbon monoxide dehydrogenase/
acetyl coenzyme A synthase complex (41), had
only moderate root support (PP = 0.5 to 0.75)
for two subunits and low support (PP < 0.5) for
other subunits. Thus, while our analyses sup-
port the antiquity of components of the WLP,
acetogenesis, the TCA cycle, and several other
core metabolic pathways, they do not confi-
dently establish the combination of pathways
used by LBCA (36).
Finally, our reconstruction also indicated
high posterior support for elements of an ad-
aptive immune CRISPR-Cas system (66, 67),
including the universally conserved Cas endo-
nuclease, Cas1 (PP = 0.96/0.93/0.89), essen-
tial for spacer acquisition and insertion into
CRISPR cassettes (68, 69). Among other roles,
CRISPR systems are crucial in antiviral de-
fense and are activated in response to viral
exposure (70); therefore, these findings are
consistent with hypotheses suggesting that
LBCA was already coevolving with parasitic
replicators such as bacteriophages and plas-
mids (71, 72).
Vertical and horizontal evolution
are complementary
Here, we have used reconciliation methods to
model both the vertical and horizontal com-
ponents of bacterial evolution. These compo-
nents are complementary, illuminating different
facets of bacterial evolution, and we show that
the horizontal component can be used to root
and orient the vertical tree. Our analyses root
the Bacteria between two major clades, the
Terrabacteria and Gracilicutes, in contrast to
recent outgroup-rooted analyses that place the
root on the CPR branch. Instead, we predict
that CPR evolved from a common ancestor with
the Chloroflexota and Dormibacterota by reduc-
tive evolution. We infer that the last bacterial
common ancestor was a fully fledged free-living
diderm cell with an LPS layer, a multimeric
flagellum, and a type III CRISPR-Cas system.
Phylogenetic models are necessarily simpli-
fied, and there is much work to be done to
better capture the full heterogeneity of the
evolutionary process in the reconciliation
framework, from varying diversification rates
to endosymbioses. With increased sampling
and improved methods, reconciliation analy-
ses should be able to probe still deeper into the
early evolutionary history of life on Earth.
Methods summary
Phylogenetics
We used two alternative approaches to assem-
ble representative sets of bacterial genomes.
In the focal analysis, we sampled 265 genomes
evenly from across the GTDB taxonomy (13).
In the secondary analysis, we sampled 341 ge-
nomes according to the diversity of major
bacterial lineages reported in a previous study
(11). We used the OMA (73) algorithm to iden-
tify candidate single-copy orthologs and man-
ually inspected initial single gene trees to
identify a set of 62 congruent phylogenetic
markers. Sequences were aligned using MAFFT
7.453 (74) and trimmed using BMGE 1.12 (75)
with the BLOSUM30 matrix. Unrooted species
trees were inferred from a concatenation of
the 62 markers under the LG+C60+R8+F mod-
el in IQ-TREE 1.6.10 (76), which was the best-
fitting model according to the BIC (37). To
perform outgroup rooting analyses, we searched
the genomes of 148 Archaea for orthologs of the
62-marker gene set and identified a subset of
29 genes with congruent single-gene phyloge-
nies. AU tests (28) were performed in IQ-TREE.
Gene tree-species tree reconciliation
To infer gene families, we performed all-versus-
all DIAMOND (77) searches among the input
protein sets and clustered the results using the
MCL algorithm (38) with an inflation param-
eter of 1.2. Gene clusters were aligned and
trimmed as described above, and bootstrap
distributions inferred under the best-fitting
model in IQ-TREE. We used ALEml_undated
(31) to perform gene tree-species tree recon-
ciliation. The relative ages of bacterial crown
groups were estimated with MaxTiC (43) using
only those transfer-based age constraints that
were recovered in both the focal and second-
ary datasets. Estimates of gene family and
lineage verticality were averaged over the re-
conciliations obtained in the focal analysis
when rooting on each of the three candidate
branches in the root region.
Simulations and sensitivity analyses
To evaluate ALE performance, we simulated
gene family evolution using Zombi (78) com-
bined with rejection sampling to obtain sets
of simulated gene families similar to the real
data in terms of inferred DTL events. We then
compared simulated and inferred numbers
of events under a range of conditions (36).
To evaluate the robustness of root inferences,
we ordered gene families by decreasing DTL
rates, rate ratios, and a range of other proxies
for lack of informativeness and potential for
introducing bias (36) and compared the like-
lihoods of competing root hypotheses as in-
creasing proportions of gene families were
excluded from the calculation.
Ancestral metabolic reconstruction
We inferred COG gene families by assigning
the sequences on each sampled genome to
gene families from the COG ontology (39) using
eggNOG-mapper 2 (79), which were then used
to perform gene tree-species tree reconciliation.
Root origination probabilities for each of the
23 COG functional categories were inferred by
maximizing the total reconciliation likelihood
over all gene families in that category. These
category-specific probabilities for origination
at the root were then used to estimate the PP
that each gene family was present at the root
of the tree. To infer the gene family content
and metabolic repertoire of LBCA, functional
annotations of protein sequences were ob-
tained and assigned to COG families present
at the root. The LBCA proteome was recon-
structed taking into account the respective PPs
for key gene families and metabolic pathways.
A detailed account of all analyses is provided
in the supplementary methods (36).
REFERENCES AND NOTES
2.
1. T. A. Williams et al., Integrative modeling of gene and genome
evolution roots the archaeal tree of life. Proc. Natl. Acad. Sci. U.S.A.
114, E4602–E4611 (2017). doi: 10.1073/pnas.1618463114;
pmid: 28533395
J. R. Brown, W. F. Doolittle, Root of the universal tree of life
based on ancient aminoacyl-tRNA synthetase gene
duplications. Proc. Natl. Acad. Sci. U.S.A. 92, 2441–2445
(1995). doi: 10.1073/pnas.92.7.2441; pmid: 7708661
J. P. Gogarten et al., Evolution of the vacuolar H+-ATPase:
Implications for the origin of eukaryotes. Proc. Natl. Acad.
Sci. U.S.A. 86, 6661–6665 (1989). doi: 10.1073/
pnas.86.17.6661; pmid: 2528146
3.
4. N. Iwabe, K. Kuma, M. Hasegawa, S. Osawa, T. Miyata,
Evolutionary relationship of archaebacteria, eubacteria, and
eukaryotes inferred from phylogenetic trees of duplicated
genes. Proc. Natl. Acad. Sci. U.S.A. 86, 9355–9359 (1989).
doi: 10.1073/pnas.86.23.9355; pmid: 2531898
5. O. Zhaxybayeva, P. Lapierre, J. P. Gogarten, Ancient gene
duplications and the root(s) of the tree of life. Protoplasma
227, 53–64 (2005). doi: 10.1007/s00709-005-0135-1;
pmid: 16389494
6. F. U. Battistuzzi, S. B. Hedges, A major clade of prokaryotes
with ancient adaptations to life on land. Mol. Biol. Evol. 26,
335–343 (2009). doi: 10.1093/molbev/msn247;
pmid: 18988685
7. M. Bocchetta, S. Gribaldo, A. Sanangelantoni, P. Cammarano,
Phylogenetic depth of the bacterial genera Aquifex and
Thermotoga inferred from analysis of ribosomal protein,
elongation factor, and RNA polymerase subunit sequences.
J. Mol. Evol. 50, 366–380 (2000). doi: 10.1007/
s002399910040; pmid: 10795828
8. C. Brochier, H. Philippe, A non-hyperthermophilic ancestor
for Bacteria. Nature 417, 244 (2002). doi: 10.1038/417244a;
pmid: 12015592
9. T. Cavalier-Smith, Rooting the tree of life by transition
analyses. Biol. Direct 1, 19 (2006). doi: 10.1186/1745-6150-1-19;
pmid: 16834776
10. J. A. Lake, Evidence for an early prokaryotic endosymbiosis.
Nature 460, 967–971 (2009). doi: 10.1038/nature08183;
pmid: 19693078
11. L. A. Hug et al., A new view of the tree of life. Nat. Microbiol. 1,
16048 (2016). doi: 10.1038/nmicrobiol.2016.48; pmid: 27572647
12. S. Mukherjee et al., 1,003 reference genomes of bacterial
and archaeal isolates expand coverage of the tree of life.
Nat. Biotechnol. 35, 676–683 (2017). doi: 10.1038/nbt.3886;
pmid: 28604660
13. D. H. Parks et al., A standardized bacterial taxonomy based on
genome phylogeny substantially revises the tree of life.
Coleman et al., Science 372, eabe0511 (2021)
7 May 2021
7 of 9
Corrected 30 June 2021. See full text. RES EARCH | R E S E A R C H A R T I C L E
Nat. Biotechnol. 36, 996–1004 (2018). doi: 10.1038/nbt.4229;
pmid: 30148503
14. D. H. Parks et al., Recovery of nearly 8,000 metagenome-
assembled genomes substantially expands the tree of life.
Nat. Microbiol. 2, 1533–1542 (2017). doi: 10.1038/s41564-017-
0012-7; pmid: 28894102
15. C. T. Brown et al., Unusual biology across a group comprising
more than 15% of domain Bacteria. Nature 523, 208–211
(2015). doi: 10.1038/nature14486; pmid: 26083755
16. Q. Zhu et al., Phylogenomics of 10,575 genomes reveals
evolutionary proximity between domains Bacteria and Archaea.
Nat. Commun. 10, 5477 (2019). doi: 10.1038/s41467-019-
13443-4; pmid: 31792218
37. D. Posada, T. R. Buckley, Model selection and model averaging
in phylogenetics: Advantages of Akaike Information Criterion
and Bayesian approaches over likelihood ratio tests. Syst. Biol.
53, 793–808 (2004). doi: 10.1080/10635150490522304;
pmid: 15545256
38. A. J. Enright, S. Van Dongen, C. A. Ouzounis, An efficient
algorithm for large-scale detection of protein families. Nucleic
Acids Res. 30, 1575–1584 (2002). doi: 10.1093/nar/30.7.1575;
pmid: 11917018
39. M. Y. Galperin, D. M. Kristensen, K. S. Makarova, Y. I. Wolf,
E. V. Koonin, Microbial genome analysis: The COG approach.
Brief. Bioinform. 20, 1063–1070 (2019). doi: 10.1093/bib/
bbx117; pmid: 28968633
17. C. J. Castelle, J. F. Banfield, Major new microbial groups
40. K. Raymann, C. Brochier-Armanet, S. Gribaldo, The two-
expand diversity and alter our understanding of the tree of life.
Cell 172, 1181–1197 (2018). doi: 10.1016/j.cell.2018.02.016;
pmid: 29522741
18. C. J. Castelle et al., Biosynthetic capacity, metabolic variety
and unusual biology in the CPR and DPANN radiations.
Nat. Rev. Microbiol. 16, 629–645 (2018). doi: 10.1038/s41579-
018-0076-2; pmid: 30181663
19. J. P. Beam et al., Ancestral absence of electron transport
chains in Patescibacteria and DPANN. Front. Microbiol. 11, 1848
(2020). doi: 10.3389/fmicb.2020.01848; pmid: 33013724
20. C. J. Castelle et al., Genomic expansion of domain archaea
highlights roles for organisms from new phyla in anaerobic
carbon cycling. Curr. Biol. 25, 690–701 (2015). doi: 10.1016/j.
cub.2015.01.014; pmid: 25702576
21. R. Méheust, D. Burstein, C. J. Castelle, J. F. Banfield, The
distinction of CPR bacteria from other bacteria based on
protein family content. Nat. Commun. 10, 4173 (2019).
doi: 10.1038/s41467-019-12171-z; pmid: 31519891
22. A. Graybeal, Is it better to add taxa or characters to a difficult
phylogenetic problem? Syst. Biol. 47, 9–17 (1998).
doi: 10.1080/106351598260996; pmid: 12064243
23. S. M. Hedtke, T. M. Townsend, D. M. Hillis, Resolution of
phylogenetic conflict in large data sets by increased taxon
sampling. Syst. Biol. 55, 522–529 (2006). doi: 10.1080/
10635150600697358; pmid: 16861214
24. J. Bergsten, A review of long-branch attraction. Cladistics 21,
163–193 (2005). doi: 10.1111/j.1096-0031.2005.00059.x
25. R. Gouy, D. Baurain, H. Philippe, Rooting the tree of life:
The phylogenetic jury is still out. Philos. Trans. R. Soc. London
Ser. B 370, 20140329 (2015). doi: 10.1098/rstb.2014.0329;
pmid: 26323760
26. J. A. Lake, R. G. Skophammer, C. W. Herbold, J. A. Servin,
Genome beginnings: Rooting the tree of life. Philos. Trans. R.
Soc. London Ser. B 364, 2177–2185 (2009). doi: 10.1098/
rstb.2009.0035; pmid: 19571238
27. R. G. Skophammer, J. A. Servin, C. W. Herbold, J. A. Lake,
Evidence for a Gram-positive, eubacterial root of the tree of
life. Mol. Biol. Evol. 24, 1761–1768 (2007). doi: 10.1093/
molbev/msm096; pmid: 17513883
28. H. Shimodaira, An approximately unbiased test of phylogenetic
tree selection. Syst. Biol. 51, 492–508 (2002). doi: 10.1080/
10635150290069913; pmid: 12079646
29. G. J. Szöllosi, B. Boussau, S. S. Abby, E. Tannier, V. Daubin,
Phylogenetic modeling of lateral gene transfer reconstructs
the pattern and relative timing of speciations. Proc. Natl. Acad.
Sci. U.S.A. 109, 17513–17518 (2012). doi: 10.1073/
pnas.1202997109; pmid: 23043116
30. L. A. David, E. J. Alm, Rapid evolutionary innovation during an
Archaean genetic expansion. Nature 469, 93–96 (2011).
doi: 10.1038/nature09649; pmid: 21170026
31. G. J. Szöllősi, W. Rosikiewicz, B. Boussau, E. Tannier, V. Daubin,
Efficient exploration of the space of reconciled gene trees.
Syst. Biol. 62, 901–912 (2013). doi: 10.1093/sysbio/syt054;
pmid: 23925510
32. L. A. Katz, J. R. Grant, L. W. Parfrey, J. G. Burleigh, Turning
the crown upside down: Gene tree parsimony roots the
eukaryotic tree of life. Syst. Biol. 61, 653–660 (2012).
doi: 10.1093/sysbio/sys026; pmid: 22334342
33. D. M. Emms, S. Kelly, STRIDE: Species Tree Root Inference
from Gene Duplication Events. Mol. Biol. Evol. 34, 3267–3278
(2017). doi: 10.1093/molbev/msx259; pmid: 29029342
34. A. Zwaenepoel, Y. Van de Peer, Inference of ancient whole-
genome duplications and the evolution of gene duplication
and loss rates. Mol. Biol. Evol. 36, 1384–1404 (2019).
doi: 10.1093/molbev/msz088; pmid: 31004147
35. S. Q. Le, O. Gascuel, An improved general amino acid
replacement matrix. Mol. Biol. Evol. 25, 1307–1320 (2008).
doi: 10.1093/molbev/msn067; pmid: 18367465
domain tree of life is linked to a new root for the Archaea. Proc.
Natl. Acad. Sci. U.S.A. 112, 6670–6675 (2015). doi: 10.1073/
pnas.1420858112; pmid: 25964353
41. P. S. Adam, G. Borrel, S. Gribaldo, Evolutionary history of
carbon monoxide dehydrogenase/acetyl-CoA synthase, one of
the oldest enzymatic complexes. Proc. Natl. Acad. Sci. U.S.A.
115, E1166–E1173 (2018). doi: 10.1073/pnas.1716667115;
pmid: 29358391
42. A. A. Davín et al., Gene transfers can date the tree of life.
Nat. Ecol. Evol. 2, 904–909 (2018). doi: 10.1038/s41559-018-
0525-3; pmid: 29610471
43. C. Chauve et al., MaxTiC: fast ranking of a phylogenetic tree
by maximum time consistency with lateral gene transfers.
bioRxiv 127548 [Preprint]. 6 October 2017. https://doi.org/
10.1101/127548.
44. T. Dagan, W. Martin, Ancestral genome sizes specify the
minimum rate of lateral gene transfer during prokaryote
evolution. Proc. Natl. Acad. Sci. U.S.A. 104, 870–875 (2007).
doi: 10.1073/pnas.0606318104; pmid: 17213324
45. W. F. Doolittle, Phylogenetic classification and the universal
tree. Science 284, 2124–2128 (1999). doi: 10.1126/
science.284.5423.2124; pmid: 10381871
46. W. F. Doolittle, E. Bapteste, Pattern pluralism and the Tree of
Life hypothesis. Proc. Natl. Acad. Sci. U.S.A. 104, 2043–2049
(2007). doi: 10.1073/pnas.0610699104; pmid: 17261804
47. T. Dagan, W. Martin, The tree of one percent. Genome Biol. 7,
118 (2006). doi: 10.1186/gb-2006-7-10-118; pmid: 17081279
48. D. Alvarez-Ponce, P. Lopez, E. Bapteste, J. O. McInerney, Gene
similarity networks provide tools for understanding eukaryote
origins and evolution. Proc. Natl. Acad. Sci. U.S.A. 110,
E1594–E1603 (2013). doi: 10.1073/pnas.1211371110;
pmid: 23576716
49. R. Jain, M. C. Rivera, J. A. Lake, Horizontal gene transfer
among genomes: The complexity hypothesis. Proc. Natl. Acad.
Sci. U.S.A. 96, 3801–3806 (1999). doi: 10.1073/
pnas.96.7.3801; pmid: 10097118
50. P. Puigbò, Y. I. Wolf, E. V. Koonin, The tree and net
components of prokaryote evolution. Genome Biol. Evol.
2, 745–756 (2010). doi: 10.1093/gbe/evq062;
pmid: 20889655
51. R. Liu, H. Ochman, Stepwise formation of the bacterial flagellar
system. Proc. Natl. Acad. Sci. U.S.A. 104, 7116–7121 (2007).
doi: 10.1073/pnas.0700266104; pmid: 17438286
52. F. El Baidouri, C. Venditti, S. Suzuki, A. Meade, S. Humphries,
Phenotypic reconstruction of the last universal common
ancestor reveals a complex cell. bioRxiv 2020.08.20.260398
[Preprint]. 21 August 2020. https://doi.org/10.1101/
2020.08.20.26039.
53. I. Sela, Y. I. Wolf, E. V. Koonin, Theory of prokaryotic genome
evolution. Proc. Natl. Acad. Sci. U.S.A. 113, 11399–11407
(2016). doi: 10.1073/pnas.1614083113; pmid: 27702904
54. L. C. Antunes et al., Phylogenomic analysis supports the
ancestral presence of LPS-outer membranes in the Firmicutes.
eLife 5, e14589 (2016). doi: 10.7554/eLife.14589;
pmid: 27580370
55. D. Megrian, N. Taib, J. Witwinowski, C. Beloin, S. Gribaldo, One
or two membranes? Diderm Firmicutes challenge the Gram-
positive/Gram-negative divide. Mol. Microbiol. 113, 659–671
(2020). doi: 10.1111/mmi.14469; pmid: 31975449
56. N. Taib et al., Genome-wide analysis of the Firmicutes
illuminates the diderm/monoderm transition. Nat. Ecol. Evol. 4,
1661–1672 (2020). doi: 10.1038/s41559-020-01299-7;
pmid: 33077930
57. E. I. Tocheva, D. R. Ortega, G. J. Jensen, Sporulation, bacterial
cell envelopes and the origin of life. Nat. Rev. Microbiol.
14, 535–542 (2016). doi: 10.1038/nrmicro.2016.85;
pmid: 28232669
Commun. 11, 5090 (2020). doi: 10.1038/s41467-020-18906-7;
pmid: 33037220
59. M. C. Weiss et al., The physiology and habitat of the last
universal common ancestor. Nat. Microbiol. 1, 16116 (2016).
doi: 10.1038/nmicrobiol.2016.116; pmid: 27562259
60. F. L. Sousa, W. F. Martin, Biochemical fossils of the
ancient transition from geoenergetics to bioenergetics
in prokaryotic one carbon compound metabolism.
Biochim. Biophys. Acta Bioenerg. 1837, 964–981
(2014). doi: 10.1016/j.bbabio.2014.02.001;
pmid: 24513196
61. F. L. Sousa, S. Nelson-Sathi, W. F. Martin, One step beyond a
ribosome: The ancient anaerobic core. Biochim. Biophys.
Acta Bioenerg. 1857, 1027–1038 (2016). doi: 10.1016/
j.bbabio.2016.04.284; pmid: 27150504
62. G. Borrel, P. S. Adam, S. Gribaldo, Methanogenesis and the
Wood–Ljungdahl pathway: An ancient, versatile, and fragile
association. Genome Biol. Evol. 8, 1706–1711 (2016).
doi: 10.1093/gbe/evw114; pmid: 27189979
63. G. Fuchs, Alternative pathways of carbon dioxide fixation:
Insights into the early evolution of life? Annu. Rev. Microbiol.
65, 631–658 (2011). doi: 10.1146/annurev-micro-090110-
102801; pmid: 21740227
64. T. Nunoura et al., A primordial and reversible TCA cycle in a
facultatively chemolithoautotrophic thermophile. Science
359, 559–563 (2018). doi: 10.1126/science.aao3407;
pmid: 29420286
65. K. Schuchmann, V. Müller, Autotrophy at the thermodynamic
limit of life: A model for energy conservation in acetogenic
bacteria. Nat. Rev. Microbiol. 12, 809–821 (2014). doi: 10.1038/
nrmicro3365; pmid: 25383604
66. K. S. Makarova et al., Evolution and classification of the
CRISPR-Cas systems. Nat. Rev. Microbiol. 9, 467–477 (2011).
doi: 10.1038/nrmicro2577; pmid: 21552286
67. E. V. Koonin, K. S. Makarova, Origins and evolution of
CRISPR-Cas systems. Philos. Trans. R. Soc. London Ser. B 374,
20180087 (2019). doi: 10.1098/rstb.2018.0087;
pmid: 30905284
68. J. K. Nuñez et al., Cas1-Cas2 complex formation mediates
spacer acquisition during CRISPR-Cas adaptive immunity.
Nat. Struct. Mol. Biol. 21, 528–534 (2014). doi: 10.1038/
nsmb.2820; pmid: 24793649
69. K. S. Makarova et al., An updated evolutionary classification of
CRISPR-Cas systems. Nat. Rev. Microbiol. 13, 722–736 (2015).
doi: 10.1038/nrmicro3569; pmid: 26411297
70. R. Barrangou et al., CRISPR provides acquired resistance
against viruses in prokaryotes. Science 315, 1709–1712 (2007).
doi: 10.1126/science.1138140; pmid: 17379808
71. E. V. Koonin, The origins of cellular life. Antonie van
Leeuwenhoek 106, 27–41 (2014). doi: 10.1007/s10482-014-
0169-5; pmid: 24756907
72. M. Krupovic, V. V. Dolja, E. V. Koonin, Origin of viruses:
Primordial replicators recruiting capsids from hosts.
Nat. Rev. Microbiol. 17, 449–458 (2019). doi: 10.1038/s41579-
019-0205-6; pmid: 31142823
73. A. C. J. Roth, G. H. Gonnet, C. Dessimoz, Algorithm of
OMA for large-scale orthology inference. BMC Bioinformatics
9, 518 (2008). doi: 10.1186/1471-2105-9-518;
pmid: 19055798
74. K. Katoh, D. M. Standley, MAFFT multiple sequence alignment
software version 7: Improvements in performance and
usability. Mol. Biol. Evol. 30, 772–780 (2013). doi: 10.1093/
molbev/mst010; pmid: 23329690
75. A. Criscuolo, S. Gribaldo, BMGE (Block Mapping and
Gathering with Entropy): A new software for selection
of phylogenetic informative regions from multiple
sequence alignments. BMC Evol. Biol. 10, 210 (2010).
doi: 10.1186/1471-2148-10-210; pmid: 20626897
76. L. T. Nguyen, H. A. Schmidt, A. von Haeseler, B. Q. Minh,
IQ-TREE: A fast and effective stochastic algorithm for
estimating maximum-likelihood phylogenies. Mol. Biol. Evol.
32, 268–274 (2015). doi: 10.1093/molbev/msu300;
pmid: 25371430
77. B. Buchfink, C. Xie, D. H. Huson, Fast and sensitive protein
alignment using DIAMOND. Nat. Methods 12, 59–60 (2015).
doi: 10.1038/nmeth.3176; pmid: 25402007
78. A. A. Davín, T. Tricou, E. Tannier, D. M. de Vienne, G. J. Szöllősi,
Zombi: A phylogenetic simulator of trees, genomes and
sequences that accounts for dead linages. Bioinformatics 36,
1286–1288 (2020). doi: 10.1093/bioinformatics/btz710;
pmid: 31566657
36. See supplementary materials.
autotrophic growth of Desulfovibrio desulfuricans. Nat.
Annotation through Orthology Assignment by eggNOG-Mapper.
Coleman et al., Science 372, eabe0511 (2021)
7 May 2021
8 of 9
58. I. Sánchez-Andrea et al., The reductive glycine pathway allows
79. J. Huerta-Cepas et al., Fast Genome-Wide Functional
Corrected 30 June 2021. See full text. RES EARCH | R E S E A R C H A R T I C L E
Mol. Biol. Evol. 34, 2115–2122 (2017). doi: 10.1093/molbev/
msx148; pmid: 28460117
80. G. Coleman et al., Extended Data for A rooted phylogeny
resolves early bacterial evolution, Version 9, Figshare (2020);
https://doi.org/10.6084/m9.figshare.12651074.v9.
ACKN OW LEDG MEN TS
Funding: G.A.C. is supported by a Royal Society Research Grant
to T.A.W. T.A.W. is supported by a Royal Society University
Research Fellowship and NERC grant NE/P00251X/1. G.J.Sz.
received funding from the European Research Council under the
European Union’s Horizon 2020 research and innovation program
under grant agreement 714774 and grant GINOP-2.3.2.-15-2016-
00057. A.S. is supported by the Swedish Research Council
(VR starting grant 2016-03559 to A.S.) and the NWO-I foundation
of the Netherlands Organisation for Scientific Research (WISE
fellowship to A.S.). A.A.D. and P.H. are supported by an Australian
Research Council Laureate Fellowship (grant FL150100038).
Author contributions: The project was conceived of by T.A.W.,
G.J.Sz., P.H., A.S., G.A.C., and A.A.D. G.A.C., A.A.D., T.A.W., L.L.Sz.,
and G.J.Sz. performed phylogenomic analyses. G.J.Sz. developed
new analytical methods. T.A.M., G.A.C., A.A.D., and A.S. performed
metabolic annotations and reconstructions. All authors
contributed to interpretation and writing. Competing interests:
The authors have no competing interests to declare. Data
and materials availability: All data and code are provided online
at Figshare (80). New methods are described in detail in the
supplementary methods (36).
SUPPLEMENTARY MATERIALS
science.sciencemag.org/content/372/6542/eabe0511/suppl/DC1
Materials and Methods
Figs. S1 to S18
Tables S1 to S13
References (81–95)
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
28 July 2020; resubmitted 5 November 2020
Accepted 1 April 2021
10.1126/science.abe0511
Coleman et al., Science 372, eabe0511 (2021)
7 May 2021
9 of 9
Corrected 30 June 2021. See full text.
|
10.1126_science.abe5601
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
CRISPR
Toxin-antitoxin RNA pairs safeguard
CRISPR-Cas systems
Ming Li†*, Luyao Gong†, Feiyue Cheng†, Haiying Yu, Dahe Zhao, Rui Wang, Tian Wang,
Shengjie Zhang, Jian Zhou, Sergey A. Shmakov, Eugene V. Koonin, Hua Xiang*
INTRODUCTION: CRISPR-Cas systems efficiently
protect bacteria and archaea from viruses and
other types of foreign DNA, but, characteristi-
cally of defense systems, they also impart non-
negligible fitness costs on the host, for example,
the risk of autoimmunity and the repulsion to
exogenous beneficial genes. Presumably, these
costs result in frequent loss of CRISPR-Cas in
bacteria, which is reflected in its patchy distri-
bution, even among closely related bacterial
strains. Nevertheless, in the current genome
sequence databases, ~40% of bacterial and
~90% of archaeal genomes carry CRISPR-cas
loci, suggesting the possibility that in addition
to the direct benefits of adaptive immunity,
mechanisms might exist that mitigate the costs
of CRISPR systems and prevent their loss.
RATIONALE: We specifically looked into an
archaeal type I-B CRISPR-Cas, where the
genes encoding the subunits of the CRISPR
effector Cascade cannot be deleted individ-
ually but can be readily deleted as a whole,
including a 311–base pair intergenic region.
These observations suggest that the Cascade
gene cassette (cas6-cas8-cas7-cas5) includes a
toxic component that makes it addictive to the
host (elicits cellular toxicity once any of the
cascade genes is deleted). We cloned and exten-
sively analyzed the intergenic region between
cas6 and cas8, which allowed us to identify the
Cascade-repressed toxin gene creT, along with
an associating CRISPR repeat–like sequence
that appears to be required for transcriptional
repression of creT. We hypothesized that the
Sequestering the rare tRNA-Arg
UCU
A
U
creA
creT
CU U
A A A A
G
G G
CreT RNA
cas genes
CRISPR
Cascade-CreA
Cascade-crRNA
5' handle
5' handle
3' handle
Partial matching
Full matching
creA
creT
Gene repression
DNA interference
Toxin-antitoxin RNA pair CreTA safeguards CRISPR-Cas. CRISPR effector (Cascade) is not only guided
by CRISPR RNA to inactivate full-matching foreign nucleic acids but is also co-opted by CreA RNA to
transcriptionally repress the toxin gene creT through partial complementarity between CreA and the creT
promoter. When Cascade is inactivated, the derepressed CreT RNA sequesters the rare tRNAUCU that decodes a
minor arginine codon and arrests cellular growth, thus making the CRISPR effector addictive to the host cell.
repeat-like sequence is part of a CRISPR RNA–
resembling antitoxin (CreA) RNA, which re-
presses the toxin jointly with Cascade. We
reasoned that CreTA would make the cascade
genes addictive for the host.
RESULTS: The intergenic sequence between
cas6 and cas8 caused toxicity in cells lacking
one or more cascade genes. By extensive muta-
tional analysis, we identified the RNA toxin
gene creT and its critical elements, namely, a
combination of a strong Shine-Dalgarno motif,
an efficient start codon, two minor arginine co-
dons (AGA) located immediately downstream,
and a stable stem-loop structure. Overexpres-
sion of tRNAUCU relieved the toxicity of CreT,
supporting a mechanism whereby this RNA
toxin arrests cellular growth by sequestering
the rare arginine tRNAUCU.
Mutational analysis of creT and its neighbor-
ing sequences revealed an adjacent CRISPR
repeat–like sequence that is required to suppress
the toxicity of CreT. This repeat-like sequence is
immediately followed by a spacer-like sequence
and a transcription terminator. By Northern
blotting and RNA sequencing, we validated the
expression of CreA RNA, a CRISPR RNA var-
iant that lacks a 3′ handle. The spacer of CreA
partially matches the promoter of creT (PcreT),
and using a reporter gene, we confirmed that
CreA, as a complex with Cascade, represses
PcreT. Similar to CRISPR interference, repression
of creT requires a protospacer adjacent motif
(PAM) and the PAM-proximal base pairing. In
cells lacking CreTA, the cascade genes become
susceptible to disruption by transposable ele-
ments. Our bioinformatic analysis identified
several CreTA analogs associated with diverse
archaeal and bacterial CRISPR-cas loci and
containing PAMs corresponding to those of
the respective CRISPR systems. Notably, these
CreTA analogs hold little conservation in nu-
cleic acid sequence, suggesting that they have
highly divergently evolved and, conceivably,
exploited different toxicity mechanisms.
CONCLUSION: Our data unearth previously un-
noticed toxin-antitoxin RNA pairs that prevent
the loss of CRISPR-cas loci by making them
addictive to the host cell. The naturally occur-
ring reprogramming of CRISPR effectors for
gene regulation highlights the multifunctional-
ity of CRISPR-Cas in bacteria and archaea and
illuminates the emerging topic of the evolution
of antiviral defense and gene regulation.▪
The list of author affiliations is available in the full article online.
†These authors contributed equally to this work.
*Corresponding author. Email: lim_im@im.ac.cn (M.L.);
xiangh@im.ac.cn (H.X.)
Cite this article as M. Li et al., Science 372, eabe5601
(2021). DOI: 10.1126/science.abe5601
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.abe5601
Li et al., Science 372, 481 (2021)
30 April 2021
1 of 1
RES EARCH
R E S E A R C H A R T I C L E
◥
CRISPR
Toxin-antitoxin RNA pairs safeguard
CRISPR-Cas systems
Ming Li1,2,3†*, Luyao Gong1,3†, Feiyue Cheng1,3†, Haiying Yu1, Dahe Zhao1, Rui Wang1,3‡, Tian Wang4,
Shengjie Zhang1,3, Jian Zhou1, Sergey A. Shmakov5, Eugene V. Koonin5, Hua Xiang1,3,6*
CRISPR-Cas systems provide RNA-guided adaptive immunity in prokaryotes. We report that the
multisubunit CRISPR effector Cascade transcriptionally regulates a toxin-antitoxin RNA pair, CreTA.
CreT (Cascade-repressed toxin) is a bacteriostatic RNA that sequesters the rare arginine tRNAUCU
(transfer RNA with anticodon UCU). CreA is a CRISPR RNA–resembling antitoxin RNA, which requires
Cas6 for maturation. The partial complementarity between CreA and the creT promoter directs Cascade
to repress toxin transcription. Thus, CreA becomes antitoxic only in the presence of Cascade. In
CreTA-deleted cells, cascade genes become susceptible to disruption by transposable elements. We
uncover several CreTA analogs associated with diverse archaeal and bacterial CRISPR-cas loci. Thus,
toxin-antitoxin RNA pairs can safeguard CRISPR immunity by making cells addicted to CRISPR-Cas, which
highlights the multifunctionality of Cas proteins and the intricate mechanisms of CRISPR-Cas regulation.
H ighly diversified CRISPR-Cas systems pro-
vide adaptive immunity in prokaryotes
(1–4). Adaptation complexes incorpo-
rate segments of foreign DNA (spacers)
into CRISPR arrays, and small CRISPR
RNAs (crRNAs) guide a multisubunit effector
complex (class 1) or a single-protein effector
(class 2) to cleave the cognate foreign DNA
or RNA at sequences complementary to the
spacers (5). By disrupting their nucleolytic
activity, DNA-cleaving CRISPR effectors have
been engineered to develop versatile gene
regulators (6, 7). However, regulatory func-
tions of CRISPR-Cas in bacteria and archaea
are poorly understood. A role of the type II
Cas9 effector of Francisella novicida in re-
pressing a virulence-related regulon through
limited complementarity between a non-
canonical RNA guide and the target gene has
been recently demonstrated (8). Furthermore,
recent preliminary results reveal widespread
autoregulation of transcription by the Cas9
effector, which uses a natural single guide
RNA with partial complementarity to the cas9
gene promoter (9).
Here, we report that some type I-B CRISPR
effectors are natural gene regulators that tran-
1State Key Laboratory of Microbial Resources, Institute of
Microbiology, Chinese Academy of Sciences, Beijing, China.
2CAS Key Laboratory of Microbial Physiological and
Metabolic Engineering, Chinese Academy of Sciences,
Beijing, China. 3College of Life Science, University of Chinese
Academy of Sciences, Beijing, China. 4College of Life
Sciences, Sichuan Normal University, Chengdu, China.
5National Center for Biotechnology Information, National
Library of Medicine, National Institutes of Health, Bethesda,
MD, USA. 6Center for Ocean Mega-Science, Chinese
Academy of Sciences, Qingdao, China.
*Corresponding author. Email: lim_im@im.ac.cn (M.L.); xiangh@
im.ac.cn (H.X.)
†These authors contributed equally to this work. ‡Present address:
Non-coding RNA and Drug Discovery Key Laboratory of Sichuan
Province, Chengdu Medical College, Chengdu, Sichuan, China.
scriptionally repress a dormancy-inducing
toxin-antitoxin (TA) system, hereafter CreTA,
after Cascade-repressed toxin-antitoxin. CreTA
is encoded within the respective CRISPR-cas
loci and, unlike previously characterized TA
modules that all encode a protein toxin (10–12),
consists of two RNA molecules. Thus, CreTA
functions as an addiction module that prevents
the loss of the genes encoding the multisubunit
I-B CRISPR effector complex, Cascade.
CreT is a bacteriostatic toxin suppressed
by Cascade
We initially identified a CreTA module within
the 311–base pair (bp) intergenic region be-
tween cas6 and cas8 in Haloarcula hispanica
(Fig. 1A). In our previous study, we failed to
delete any of the cascade genes (cas5, cas6,
cas7, and cas8) individually from the wild-type
(WT) strain but managed to delete them simul-
taneously (Dcas5-8) (13). However, starting from
a strain lacking the intergenic 311 bp (DTA),
mutants DTADcas5, DTADcas6, DTADcas7, and
DTADcas8 were easily obtained (fig. S1). In
these cas mutants and Dcas5-8, a plasmid car-
rying the intergenic 311 bp (pTA) consistently
showed toxicity, i.e., a marked reduction in
transformation efficiency (by ~104-fold), com-
pared with the empty vector (Fig. 1B). By con-
trast, this toxic effect was not observed in WT
or DTA cells that encode a complete set of the
Cascade subunits. We inferred that the Cascade
complex represses a cryptic toxin, which we
named CreT for “Cascade-repressed toxin,”
and the creT gene is embedded within the
cascade gene cassette.
We tested the toxic effect of a series of
truncated variants of pTA and identified a
132-bp region that reduced transformation
efficiency in DTADcas6, DTADcas5, DTADcas7,
and DTADcas8 cells (fig. S2 and pTA07 in Fig.
1C). The 132-bp sequence contained the archaeal
promoter elements BRE (TF-IIB recognition)
and TATA-box (14) (Fig. 1A). As expected, the
toxic effect of pTA in DTADcas6 cells was no
longer observed when the predicted TATA-
box was mutated (pTTm in Fig. 1C). Based
on the positions of the promoter elements,
we predicted a 78–nucleotide (nt) creT tran-
script produced from the 132-bp region in
pTA07 (Fig. 1A). Using a strong promoter (15),
we constructed a plasmid overexpressing this
78-nt RNA (pOE) and observed very low trans-
formation efficiency (~10 CFU/mg; CFU, colony-
forming unit) in both DTA and DTADcas6 cells
(Fig. 1C); by contrast, high transformation ef-
ficiency (104 to 105 CFU/mg) was observed when
this strong promoter was mutated (pOEm).
Thus, the 78-nt creT transcript was clearly
toxic. The 78-nt transcript contained a pair of
inverted repeats (10 nt each) (Fig. 1A), which
allows this RNA to fold into a stem-loop struc-
ture (fig. S3). When we mutated one of the
inverted repeats to disrupt this folding poten-
tial, the transformation efficiency of pTA in
DTADcas6 cells was recovered to ~105 CFU/mg
(pIRm), which was then markedly reduced (to
<10 CFU/mg) when we complementarily mu-
tated the other repeat to restore the stem-
loop (pIRcm) (Fig. 1C). We next expressed an
inactive CreT mutant in DTA and found that
its abundance did not substantially change
as a result of the stem-loop disruption (fig. S3).
The stem-loop structure appears to be critical
for the function rather than the stability of
CreT RNA.
To characterize the toxic effect (bacteriostatic
or bactericidal), we controlled creT expression
using a tryptophan-inducible promoter (16)
and introduced it into another haloarchaeon
Haloferax volcanii. (The inducible promoter
does not work in H. hispanica.) Compared
with the creT– strain (containing the empty
vector), the creT+ strain showed a growth
defect in the inducing medium (Fig. 1D). By
plating the inducing cultures onto noninduc-
ing plates (Fig. 1E), we measured their CFU
curves. Notably, the CFU of the creT+ strain
rose very slowly but did not decline (Fig. 1D),
indicating that CreT suppressed cell multi-
plication. These results indicate that CreT is
a bacteriostatic toxin that inhibits a cellular
process(es) conserved (at least) among differ-
ent species of haloarchaea.
CreT is a small RNA that sequesters tRNAUCU
We noticed that the 5′ end of CreT RNA con-
tains an 8-nt sequence that fully matches the
3′ end of 16S ribosomal RNA (rRNA) (Fig. 2A).
When we modulated this complementarity
to neighboring sequences on 16S rRNA (see
sites 1 and 2 in Fig. 2A), plasmids carrying the
mutated creT were not toxic (high transfor-
mation efficiency was observed in DTADcas6
Li et al., Science 372, eabe5601 (2021)
30 April 2021
1 of 11
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. CreT is a bacteriostatic toxin repressed by Cascade. (A) The
creT-creA operon. BRE and TATA-box are promoter elements. TSS indicates
the transcription start site. Red nucleotides within IRs were mutated to
modulate complementarity. yR and yS of creA are analogous to the repeat
and spacer of CRISPR, respectively. Positions are relative to the TSS of
creT. (B) Transformation of cas mutants by pTA (carrying creTA). Vector is
the empty pWL502. (C) Mutational analysis of pTA. pTA07 carried the
underlined sequence in (A). pTTm contained a mutated TATA box (indicated
with a red star). CreT was overexpressed using PphaR (pOE) or its mutant
P*phaR (pOEm); T8 is a terminator of eight thymine nucleotides. One of
the IRs was mutated in pIRm, and the other was further complementarily
mutated in pIRcm. (D) Effect of CreT on the growth of H. volcanii cells.
(E) Calculation of CFU by dilution plating. Error bars represent mean ±
standard deviation (n = 3).
cells) (Fig. 2B). This suggests that the 8-nt se-
quence interacts with the 3′ terminus of 16S
rRNA and likely acts as an efficient Shine-
Dalgarno (SD) motif, which was observed
to enhance translation efficiency in some
haloarchaea (17, 18). We further noticed that
CreT contains a canonical AUG start codon and
subjected it to saturation mutagenesis. CreT
remained toxic (i.e., markedly impaired the
plasmid transformation efficiency in DTADcas6)
only when AUG was mutated to GUG (Fig. 2C),
another efficient start codon in haloarchaea.
Notably, when AUG was mutated to UUG, a
less efficient start codon in haloarchaea (19),
CreT was inactivated. This suggests that the
toxicity of CreT depends on strong transla-
tion initiation signals.
The translation initiation signals in CreT
are immediately followed by two rare AGA
arginine codons (usage frequency among all
codons is 0.22% in H. hispanica) and then by
an opal stop codon (UGA) (Fig. 2A). We first
mutated the stop codon (UGA to CGA) and
found that the mutated CreT remained toxic
(Fig. 2D). Then, we made synonymous muta-
tions in the two AGA codons and found that
CreT lost toxicity when the two AGA codons
were replaced by the more common arginine
codons CGA (usage frequency among all co-
dons is 0.87%), CGU (0.63%), CGC (2.14%), or
CGG (2.11%) but remained toxic when replaced
by the other rare arginine codon AGG (0.21%)
(Fig. 2D). Thus, it seemed most likely that the
mechanism of CreT toxicity that required the
strong translation initiation signals involved
sequestering and depleting tRNAUCU that de-
codes the rare arginine codons. Indeed, we
showed that overexpression of tRNAUCU, which
decodes both AGA and AGG codons, relieved
the toxicity of the WT (AGA) CreT and the AGG
mutant, whereas overexpression of tRNACCU,
which decodes AGG but not AGA, relieved the
toxicity of the AGG mutant but not of the WT
CreT (Fig. 2E). We further measured the level
of tRNAUCU in the H. volcanii cells with or
without an inducible creT gene but did not
observe detectable differences at 6 or 12 hours
after induction, which indicates that the over-
all amount of charged and uncharged tRNAUCU
was not affected by CreT (fig. S4). We conclude
that CreT acts by sequestering the arginine-
charged tRNAUCU and thus hampering decod-
ing of the rare arginine codons.
Among the 3859 H. hispanica genes, 2025
contain at least one AGA/AGG codon (1075 genes
have a single AGA/AGG codon and 950 genes
have two or more). The speed and/or accuracy of
translation of these genes could be substantially
affected by the availability of tRNAUCU/CCU. In
Escherichia coli, a subset of essential genes
contains the rare AGA/AGG codons preferen-
tially within their first 25 codons, and even a
single AGG codon in this region substantially
reduces the protein expression level (20, 21).
Under the minor codon modulator hypoth-
esis, the availability of the least abundant
tRNAUCU/CCU globally regulates the translation
of these essential genes and hence modulates
key cellular functions. We similarly analyzed
Li et al., Science 372, eabe5601 (2021)
30 April 2021
2 of 11
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. CreT sequesters
tRNAUCU. (A) The
structure of CreT RNA.
(B) Mutational analysis
of the SD motif. SD was
mutated to match the
16S rRNA site 1 or
2 [indicated in (A)].
(C) Saturating mutagen-
esis of the start codon.
(D) Synonymous
mutation of the two rare
Arg (arginine) codons
and mutation of the stop
codon (UGA to CGA).
The usage frequency
among all codons
in H. hispanica is shown
to the right of the plot.
(E) Suppression of CreT
by overexpression of
tRNAs (tRNAUCU or
tRNACCU). The dot indi-
cates a wobble base pair.
Vector is the empty
pWL502. (F) Usage
frequency of AGA/AGG
in H. hispanica genes.
Starting from the
initiation codon, the
number of AGA/AGG
within every 25-codon
window was divided by
the total number of
AGA/AGG within the first 250 codons to calculate the frequency for each window. Error bars represent mean ± standard deviation (n = 3).
the 2025 H. hispanica genes containing AGA/
AGG and found that these genes also prefer-
entially used AGA/AGG within the first 25 co-
dons, and this bias became more prominent
when the 1075 genes containing a single AGA/
AGG codon were specifically investigated (Fig.
2F). We further examined the set of genes with
a single AGA/AGG codon within the first 25 co-
dons and found that these genes are associated
with various key cellular functions (table S1).
Therefore, the minor codon modulator hypoth-
esis is likely to apply also in H. hispanica, so
that sequestration of the rare tRNAUCU by
CreT impairs the translation of some of these
essential genes and thus inhibits cell growth
and division.
CreA is an antitoxin RNA that resembles crRNA
We then asked how Cascade represses CreT.
Canonically, Cascade is guided by crRNAs,
but the only CRISPR array in H. hispanica can
be deleted from the WT cells containing creT
without eliciting toxicity (22), suggesting that
crRNAs are not involved in CreT suppression.
We noticed that pTA showed high transfor-
mation efficiency in DTA cells that retained all
Cascade subunits, but mutants lacking the
sequence immediately downstream of creT
showed very low efficiency (Fig. 1C and fig.
S2B). This finding implied that this sequence
contained uncharacterized elements involved
in the repression of the toxicity of CreT. Within
the region downstream of creT, we detected a
CRISPR repeat–like sequence (hereafter “yR”)
(Fig. 1A). The 30-nt yR shares 21 nucleotides
with the CRISPR repeat from H. hispanica
(Fig. 3A), and its transcript can be processed
by Cas6 (fig. S5). yR is directly followed by a
~33-bp “spacer” (hereafter “yS”) containing a
T-rich sequence that might function as a tran-
scription terminator (14, 23) (Fig. 1A). Although
type I crRNAs typically carry 5′ and 3′ handles
derived from flanking repeat sequences, we
have recently shown that replacing the down-
stream repeat with a transcription termina-
tor produced functionally active noncanonical
crRNA guides without the 3′ handle (24). There-
fore, we predicted that the region downstream
of creT encodes a crRNA-resembling antitoxin
RNA (CreA), which consists of an 8-nt 5′ handle
(remnant of yR), the ~33-nt yS, and no 3′
handle (Fig. 3A). Using a yS-specific probe, we
detected CreA RNA in WT but not in DTA cells
unless these cells were transformed with pTA
derivatives (Fig. 3B). In DTADcas6 cells, larger-
sized precursors instead of mature CreA were
detected, supporting the prediction that the
CreA precursor RNA is processed by Cas6.
When the promoter of creT (PcreT) was inac-
tivated (pTTm), mature CreA and a ~90-nt
precursor was detected in DTA and DTADcas6,
respectively (Fig. 3B), indicating that creA was
expressed from its own promoter. When PcreT
remained active (pIRm), we observed an addi-
tional, longer precursor likely corresponding
to the creT-creA cotranscript (Fig. 3B).
We then used RNA sequencing (RNA-seq)
to explore the production of CreA from pIRm
(fig. S6). Because Cas6 cleavage generates a
hydroxylated 5′ terminus that is inaccessible
to adapter ligation during library construc-
tion, mature CreA molecules need be pre-
treated by polynucleotide kinase to add a 5′
monophosphate (see Materials and methods).
As expected, RNA-seq revealed a high abun-
dance of mature CreA in DTA but not in
DTADcas6 cells (fig. S6), further validating
the Cas6-dependent biogenesis of CreA. In
addition, the predicted Cas6 cleavage site
and an 8-nt 5′ handle of CreA were validated
by RNA-seq.
Li et al., Science 372, eabe5601 (2021)
30 April 2021
3 of 11
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. CreA is analogous to crRNA and is processed by Cas6. (A) The nucleotide identity between the RNA transcripts from CRISPR repeat and CreA YR. A
scheme depicting the Cas6-mediated processing of crRNA and CreA is given. S, CRISPR spacer; YS, the “spacer” of CreA. (B) Northern blot of CreA and its
precursors using the YS-probe. 7S RNA served as the internal control. M is a 100-nt RNA marker.
CreA guides Cascade to transcriptionally
repress creT
We noticed a partial match between the CreA
RNA and the PcreT sequence (Fig. 4A). The
target sequence (protospacer) of type I crRNAs
is typically 5′ preceded by a protospacer
adjacent motif (PAM), and the PAM-proximal
base pairing between the spacer and the
protospacer provide a critical “seed” during
R-loop formation (25, 26). Like this pattern,
the first 11 nucleotides [except the sixth nu-
cleotide, which is known to not be involved
in base paring (26)] of yS are complementary
to a target sequence in PcreT, which is flanked
by 5′-TTC-3′ [the PAM of H. hispanica CRISPR
(27)] (Fig. 4A). Using scanning mutagenesis,
we found that when the conserved 5′ handle
and every seed nucleotide that base pairs with
PcreT were individually mutated, the pTA de-
rivatives transformed DTA cells consistently
with a much lower efficiency (~10 CFU/mg)
than the WT pTA (104-106 CFU/mg) (fig. S7),
indicating that these elements of CreA are
critical for its antitoxin activity; by contrast,
the yS nucleotides outside the seed were not
important (fig. S7B). We then selected three
seed mutants of CreA and verified that com-
plementarily mutating PcreT restored their
antitoxin activity and the high plasmid trans-
formation efficiency in DTA (Fig. 4B). There-
fore, the CreA-PcreT complementarity is critical
for toxin repression.
To validate the PcreT -inhibiting effect of
CreA, we introduced a PcreT-controlled green
fluorescent protein gene (gfp) into WT and
DTA cells (Fig. 4C). As expected, fluorescence
was observed in DTA cells (lacking CreA) but
not in WT cells (producing CreA) unless we
mutated PcreT to disrupt the CreA-PcreT com-
plementarity (Fig. 4C). We also tested the role
of PAM in repressing PcreT. Because the PAM
nucleotides 5′-TTC-3′ are located within the
complement of the purine-rich BRE element
of PcreT (see Fig. 4A), we preserved the purine-
rich character of BRE by mutating PAM to
5′-CCC-3′ (PAM– in Fig. 4D). Nevertheless, only
minimal fluorescence was detected in both WT
and DTA cells, indicating that this mutation
inactivated the BRE element. We next dupli-
cated the BRE sequence (2BRE in Fig. 4D) to
allow mutation of the PAM nucleotides with-
out disrupting the promoter elements. As ex-
pected, the two-BRE PcreT drove fluorescence
production in DTA but not in WT cells (pro-
ducing CreA), indicating that the mutated PcreT
was also repressed by CreA. When the PAM
was further mutated (2BRE/PAM– in Fig. 4D),
fluorescence was observed in both types of
cells with equivalent intensity (which was
weaker than that of the 2BRE mutant in DTA,
presumably because of the direct effects of
mutating the nucleotides next to BRE). Thus,
the PAM motif, as well as the partial comple-
mentarity between CreA and PcreT, is required
for PcreT repression. We then introduced the
PcreT-controlled gfp into cells lacking cas1, cas2,
cas3, or cas4, and notably, minimal fluores-
cence was observed in these cells unless we
mutated PcreT to disrupt the CreA-PcreT com-
plementarity (fig. S8). These results suggest
that these four cas genes, which encode pro-
teins that are not Cascade subunits, are not
required for PcreT repression.
We also performed primer extension to ana-
lyze the activity of PcreT on the pIRm plasmid
(fig. S9). This plasmid carries a mutated creT
(see Fig. 1C), so we could test PcreT activity in
cells lacking Cascade protein(s). In this set-
ting, PcreT-driven transcription was not detected
in DTA cells that encode a complete set of
Cascade proteins but was derepressed in cells
lacking Cas6 (the nuclease processing CreA)
or Cas7 (the backbone subunit of Cascade)
(fig. S9).
Taken together, our results demonstrate that
Cascade transcriptionally represses creT based
on the partial complementarity between CreA
and PcreT. When we programmed the CRISPR
array with a spacer identical to yS (replacing
the WT CRISPR array with a mini-CRISPR
containing yS), creA was no longer required
to repress creT (fig. S10), indicating that ca-
nonical crRNAs were reprogrammed to regu-
late creT transcription and further supporting
the guide RNA–dependent regulatory role of
Cascade.
CreTA safeguards CRISPR immunity
Our previous studies have extensively dem-
onstrated the adaptive CRISPR immunity in
H. hispanica (13, 22, 27). Targeting of com-
plementary DNA by the Cascade-crRNA com-
plex results in the recruitment of the Cas3
helicase-nuclease to elicit interference (cleavage
of the target DNA) and/or primed adaptation
(acquisition of new CRISPR spacers from the
target DNA). We asked how self-interference
and self-adaptation are precluded when CreA
directs Cascade binding to the PcreT DNA. By
modifying the plasmid carrying the PcreT-
controlled gfp, we coexpressed CreA mutants
with differently extended (to 15, 20, 25, or 36 bp)
complementarity to PcreT and tested three possi-
ble outcomes, i.e., plasmid interference, primed
adaptation, or gene repression (fig. S11). All
these self-targeting plasmids showed high
transformation efficiency (~105 CFU/mg) in
DTA cells. Because CreA lacks the 3′ handle,
this result is compatible with our previous
observation that deleting the entire 3′ handle
from a virus-targeting crRNA resulted in no
or little antivirus immunity (24). We then
analyzed the CRISPR array from the trans-
formants, which might have been expanded
by new spacers if primed self-adaptation was
occurring. CRISPR expansion was not observed
Li et al., Science 372, eabe5601 (2021)
30 April 2021
4 of 11
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. CreA directs Cascade to repress PcreT. (A) Scheme illustrating the
CreA-mediated PcreT repression and the partial complementarity between CreA
and PcreT. Substituting any of the red nucleotides inactivated CreA (see fig. S7C).
Positions are relative to the TSS of creT. (B) Nucleotide substitutions to alter
the CreA-PcreT complementarity. (C) Fluorescence from a gfp gene controlled
by PcreT or PcreT-m5 in WT or DTA cells. Vector is the empty pWL502.
Representative microscopy images are provided. (D) Fluorescence from a gfp
gene controlled by the WT PcreT (wt) or its PAM mutant (PAM–). The BRE
element was duplicated (2BRE) to allow PAM mutation without disrupting the
BRE element. Mutated nucleotides are highlighted in red. Error bars represent
mean ± standard deviation (n = 3); two-tailed Student’s t test [****P < 0.0001;
N.S., not significant (P > 0.05)].
until the CreA-PcreT complementarity was in-
creased to 25 or 36 bp (fig. S11C), indicating
that limited complementarity could not prime
adaptation. We then measured the fluores-
cence from the transformant cells and observed
a gradual decrease in fluorescence intensity
as the CreA-PcreT complementarity increased
(fig. S11D). These results indicate that the
mismatches between CreA and its target pre-
clude autoimmunity while allowing tran-
scription regulation and might have been
fine-tuned by selection to optimize the level
of gene repression.
We then sought to investigate the impact of
the CreTA toxicity modulation on CRISPR-Cas
itself. Because cascade genes occupy a large
genomic region that is susceptible to disrup-
tion by transposable element insertion and by
deletion, CreTA that becomes toxic in the ab-
sence of Cascade might stabilize cascade genes
by acting as an addiction module. Given that a
transposition burst of the insertion (IS) ele-
ment ISH27 (1390 bp) has been reported to
occur in H. hispanica cells stored at 4°C (28),
we investigated the WT and DTA strains that
had been stored at 4°C for 2 years. The stored
H. hispanica cultures were resuscitated and
challenged with pTarget, which carries the tar-
get of spacer1 (the first spacer of the CRISPR
array) (Fig. 5A). The survivors that failed to
repel the target plasmid were screened on a
selective medium, and their cascade genes
were analyzed (Fig. 5B). We found that the
cascade genes were interrupted by ISH27 in
~60% (17 of 28) of the DTA survivors but re-
mained intact in nearly all the WT survivors
(one lost the whole CRISPR-cas locus) (table
S2). By contrast, cas3, which is required for
plasmid interference but not for the CreT toxin
repression, was occasionally disrupted in both
the WT and DTA survivors. These findings
show that CreTA is an addiction module that
safeguards the genetic integrity of Cascade in
H. hispanica (Fig. 5C).
Distribution of CreTA homologs and analogs
across prokaryotes and among
CRISPR subtypes
By sequence similarity search of the National
Center for Biotechnology Information (NCBI)
nucleotide genomic database, we identified
only three creTA homologs, all located between
the cas6 and cas8 genes in the I-B CRISPR-cas
locus from haloarchaeal genomes (fig. S12).
Li et al., Science 372, eabe5601 (2021)
30 April 2021
5 of 11
RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. Screening and analysis of cells with inactivated CRISPR-Cas. (A) Scheme illustrating the plasmid challenge assay. Storage at 4°C induced the transposition
“burst” of the IS element (28). pyrF encodes the pyrimidine biosynthetic enzyme orotidine-5′-monophosphate decarboxylase. (B) Example clones surviving the challenge
assay. Ctrl represents the culture before challenge; M is a double-stranded DNA ladder. Larger-sized PCR products indicate the insertion of IS element (ISH27). See table S2
for more information. (C) The model of CreTA-mediated addiction. creT is repressed by Cascade-CreA in WT cells and derepressed when Cascade is disrupted.
This paucity of detectable CreTA homologs
is not surprising, given that small RNAs
often show limited sequence conservation,
and furthermore, yR and yS are likely to have
divergently coevolved with Cas6 and PcreT,
respectively. We examined the intergenic
sequences flanking the cas6 gene in other
haloarchaeal genomes that encode type I-B
CRISPR-Cas and identified five strains carry-
ing creA analogs with limited sequence sim-
ilarity to CreTA from H. hispanica, all of
which were located upstream of cas6 (Fig. 6A).
Each creA analog contained a yR sequence
that was similar (60 to 80% identity) to the
repeat from the co-occurring CRISPR. Using
a probe against the putative yS downstream
of each yR, we detected mature CreA RNA in
four strains that were available in our lab-
oratory (Fig. 6B). We predicted that the se-
quences upstream of each creA encompassed
a creT gene and cloned these regions sepa-
rately into a plasmid, which was then used to
transform H. hispanica cells. These (puta-
tive) creT-containing plasmids consistently
showed much lower transformation efficien-
cy (<102 CFU/mg) than the empty control
(~105 CFU/mg) (Fig. 6C), validating the toxic-
ity of each creT analog. Like the H. hispanica
creT, these toxic genes contained one or more
pairs of inverted repeats and lacked open read-
ing frames larger than 10 codons (fig. S13). The
putative promoter of each creT contained a
target sequence that was partially comple-
mentary to their cognate creA, and each target
sequence was flanked by the 5′-TTC-3′ PAM
(Fig. 6D and fig. S13), strongly suggesting that
the cognate Cascade complexes bind to and
repress these promoters. Thus, these creTA
analogs are probably Cascade-regulated TA
modules that safeguard the accompanying
cascade genes, as in the case of H. hispanica.
However, none of the creT RNAs in the CreTA-
like modules contain a combination of a SD
sequence, a start codon, and immediately fol-
lowing rare codons, suggesting distinct toxicity
mechanisms. The I-B CRISPR loci in several
haloarchaeal genomes closely related to those
encoding CreTA or its analogs lack sufficiently
long intergenic regions flanking cas6 and do
not contain readily detectable yR sequence
in the parts of the CRISPR loci, indicating that
CreTA is a recurrent, but evolutionarily labile,
accessory of CRISPR-Cas systems.
By searching for CRISPR repeat–like se-
quences flanking cas6 genes in other archaea
and bacteria, we predicted additional CreTA
analogs, most of which are associated with a
bacterial type I-B CRISPR (Fig. 7). All of these
cas6-flanking intergenic regions contained
both yR and yS with partial complementarity
to a target sequence flanked by a PAM (Fig. 7).
Type I-B CRISPRs are highly diversified, with
at least 10 distinct subfamilies of cas8b encod-
ing the large subunits of the effector complex
(29). Six I-B CreTA analogs were associated
with cas8b1 and contained the cognate 5′-TCA-
3′ or 5′-TTA-3′ PAM sequences (30), whereas
two analogs were associated with cas8b2 and
contained the corresponding 5′-CCT-3′ PAM
sequence (31). Thus, the CreTA-regulation ap-
parently coevolved with CRISPR immunity ac-
cording to their PAM specificity. Consistently,
we also found a CreTA analog associated with
an archaeal I-D CRISPR and containing the
PAM (5′-GTG-3′) reported for this subtype
(32, 33) (Fig. 7). A putative CreTA was also
predicted for the III-A CRISPR-Cas in a ther-
mophilic bacterium. Like the haloarchaeal
CreTA analogs, inverted repeats were fre-
quently observed between the bacterial repeat-
like sequences and their putative targets (fig.
S14). It remains to be studied experimentally
which of these and other repeat-like sequences
found in CRISPR-cas loci, indeed, are com-
ponents of CreTA-like modules. Nevertheless,
these observations suggest that protection of
cas genes by two-RNA TA modules, conceiv-
ably with different toxicity mechanisms, might
be a widespread phenomenon.
Discussion
In this work, we demonstrated the dual func-
tionality of the multisubunit subtype I-B
CRISPR effectors (Cascades) in haloarchaea.
The Cascade complex of H. hispanica is not
only guided by canonical crRNAs to inactivate
cognate foreign DNA but is also co-opted by a
noncanonical guide (CreA) to down-regulate
the transcription of a toxin RNA gene (creT)
through the partial complementarity between
the spacer-like sequence of CreA and the creT
promoter. CRISPR-Cas systems efficiently pro-
tect bacteria and archaea from viruses and
other types of foreign DNA, but characteris-
tically of defense systems, they also impart
non-negligible fitness costs on the host (34).
Specifically, in the case of CRISPR, this fitness
cost appears to come, primarily, from auto-
immunity (35, 36) or from targeting beneficial
or addictive plasmids (37–39). Presumably, these
costs result in frequent loss of CRISPR-Cas
Li et al., Science 372, eabe5601 (2021)
30 April 2021
6 of 11
RES EARCH | R E S E A R C H A R T I C L E
Fig. 6. CreTA-like modules associating with type I-B CRISPR-Cas. (A) Haloarchaeal CRISPR-cas loci associated with creTA analogs. The homology between
CRISPR repeat (R) and creA YR is shown. (B) Northern blot of CreA analogs. M1 is a 100-nt RNA marker; M2 is a biotin-labeled 64-nt single-stranded DNA.
(C) Transformation of H. hispanica with plasmids carrying a creT analog. Vector is the empty pWL502. Error bars represent mean ± standard deviation (n = 3). (D) The
homology between creA YS and its target. Hmar, Haloarcula marismortui; Hmed, Haloferax mediterranei; Hmuk, Halomicrobium mukohataei; Nsp, Natrinema sp.
J7-2; Hhub, Halobacterium hubeiense.
systems in bacteria, which is reflected in the
patchy distribution of CRISPR-Cas even among
closely related bacterial strains (40). Never-
theless, in the current genome sequence data-
bases, ~40% of bacterial and ~90% of archaeal
genomes carry CRISPR-cas loci, suggesting
the possibility that in addition to the direct
benefits brought about by adaptive immunity,
mechanisms mitigating the costs of CRISPR
systems and preventing their loss might exist.
Here, we reveal one such mechanism whereby
the cas genes encoding the CRISPR effector
subunits become essential to the host, thanks
to their ability to down-regulate the expres-
sion of a toxin, with the help of a natural guide
RNA that serves as an antitoxin. The CreTA-
like elements ensure the preservation of the
effector cas genes but not the adaptation
module of the CRISPR array. The presence of
solo effector gene suites, without the adap-
tation module or CRISPR array, is not un-
common in bacterial and archaeal genomes.
Thus, in the latest genomic census of CRISPR-
Cas systems (29), among the 6254 identified
CRISPR-cas loci, 500 were solo effectors, and
more specifically, among the 669 I-B loci, 75
contained effector genes only. A comprehen-
sive computational and experimental analy-
sis of the solo effector genomic neighborhoods
should show how often these genes are safe-
guarded by CreTA-like modules.
The discovery of the role of CreTA modules
as safeguards of cas genes complements the
recent discovery of the widespread autore-
pression of transcription by Cas9, the type II
CRISPR effector. Such autorepression is a
distinct strategy of CRISPR cost mitigation
in which a noncanonical guide RNA is used,
similar to the case of CreTA (9). The use of
guide RNAs is a general principle for delivery
of protein effectors to distinct sites on DNA
and RNA molecules. Inactivation of foreign
nucleic acids is only one of the RNA-guided
functions, even if this was the primary driving
force in the evolution of CRISPR. It appears
likely that the autoregulation of Cas9 and the
regulation of CreTA-like modules are not the
only cases of dual functionality of CRISPR ef-
fectors and, perhaps, other defense systems that
might possess various regulatory capacities, in
addition to their direct involvement in para-
site inactivation. Because self-targeting CRISPR
spacers that partially match the host DNA are
widespread (35), it is possible that some ca-
nonical crRNAs have been selected to direct
CRISPR effectors to regulate host genes.
Some Cas nucleases can be activated upon
target recognition to degrade host RNA, which
causes cell death and the abortion of phage
infection or arrests cellular growth until the
foreign targets are eliminated (41–43). CRISPR-
Cas and other defense systems also frequently
associate with diverse TA modules in bacterial
and archaeal genomes, and it has been pro-
posed that immune systems interact with the
TA such that the latter induce dormancy or
cell death when the former fail, for example,
in the presence of virus-encoded anti-CRISPR
proteins (44). Here, we describe a different
type of such interaction where the immune
system down-regulates the toxin expression,
rendering the immunity genes addictive to the
host. In such cases, the immune systems and
the TA behave as a pair of symbiotic selfish
genetic elements, in accord with the general
Li et al., Science 372, eabe5601 (2021)
30 April 2021
7 of 11
RES EARCH | R E S E A R C H A R T I C L E
Fig. 7. CreA-like genes associated with other archaeal (I-D) or bacterial (I-B and III-A) CRISPR-cas loci. Alignments of the homologous regions between
CRISPR repeat (R) and creA YR and between creA YS and its putative target are shown. The CRISPR-cas locus of Clostridium sp. BL-8 also encodes a type II
toxin-antitoxin system (relEF).
paradigm of the evolutionary entanglement
between defense mechanisms and mobile ele-
ments (45). However, it remains to be inves-
tigated whether CreTA-like TA modules could
also be bifunctional, serving both as the last line
of defense and the safeguard for the cas genes.
The CreTA module is a previously unknown
type of TA, where both the toxin and the anti-
toxin are represented by small RNA molecules,
although the antitoxin activity strictly depends
on the Cascade protein complex. Given the small
size and poor sequence conservation of both
RNA components, it seems likely that such mod-
ules are more common than is shown in this
work. The CRISPR-Cas systems appear to be
specifically prone to spawn such TA modules
because repeat propagation outside the CRISPR
array is a common phenomenon that is thought
to give rise, in particular, to the tracrRNA of
type II and some type V CRISPR (46). Such
ectopic repeats provide the scaffold for non-
canonical guide RNAs that can evolve into anti-
toxins controlling either an RNA or a protein
toxin. Discovery and structural and functional
dissection of such CRISPR-regulated CreTA-like
modules can be expected to reveal unknown
facets of TA and CRISPR biology and, particu-
larly, their behavior as selfish genetic elements.
Materials and methods
Strains and growth conditions
Haloarchaeal strains were cultivated at 37°C
in AS-168 medium (per liter, 200 g NaCl, 20 g
MgSO4·7H2O, 2 g KCl, 3 g trisodium citrate,
1 g sodium glutamate, 50 mg FeSO4·7H2O,
0.36 mg MnCl2·4H2O, 5 g Bacto Casamino
Acids, and 5 g yeast extract, pH 7.2), unless
specified. H. hispanica ATCC 33960 DpyrF
strain DF60 (47) or its derivatives were culti-
vated in AS-168 medium supplemented with
uracil (at a final concentration of 50 mg/liter).
The strains transformed by the pWL502 de-
rivatives were cultivated in yeast extract–
subtracted AS-168.
The H. volcanii H1424 strain was cultivated
in the Hv-YPC medium (48) with some modifi-
cations [per liter, 144 g NaCl, 30 g MgCl2·6H2O,
33 g MgSO4·7H2O, 4.2 g KCl, 0.333 g CaCl2, 5 g
yeast extract, 1 g peptone (soya), 1 g Bacto
Casamino acids, and 12 ml of 1 M Tris HCl
(pH 7.5)] supplemented with uracil and thy-
midine (at final concentrations of 50 and
40 mg/liter, respectively). The strains trans-
formed by the pTA1228 (16) derivatives were
cultivated in Hv-YPC medium without adding
uracil and thymidine. Tryptophan was added
to a final concentration of 500 mg/liter to
induce CreT expression.
Natrinema sp. J7-2 was cultivated in the
18% MGM medium (per liter, 144 g NaCl, 21 g
MgSO4·7H2O, 0.5 g CaCl2, 18 g MgCl2·6H2O,
4.2 g KCl, 3 g yeast extract, 5 g tryptone, pH 7.5).
E. coli JM109 was used for cloning and
cultivated at 37°C in Luria-Bertani medium.
Ampicillin was added to a final concentration
of 100 mg/liter when needed.
Plasmid engineering and transformation
The cas6-cas8 intergenic sequence (NC_
015943.1: 145387-145697) was amplified from
the H. hispanica genomic DNA using the high-
fidelity KOD-Plus DNA polymerase (TOYOBO,
Osaka, Japan), digested by BamHI and KpnI
(New England Biolabs, MA, USA), and inserted
into the predigested pWL502 (47) with T4 DNA
ligase (New England Biolabs, MA, USA). Using
a series of internal primer pairs (table S3), var-
ious truncated versions were generated for this
fragment. When needed, the sequence of PphaR
(15) was directly designed on the primer. The
point mutation was introduced using the
overlap extension polymerase chain reaction
(PCR) strategy. For example, primer pairs TA-
F/TTm-R and TTm-F/TA-R were separately
used to amplify the two moieties, and the pro-
ducts were mixed and used as the template
for a second round of PCR reaction with the
primer pair TA-F/TA-R. This DNA product con-
tained a TATA box–mutated PcreT, because the
mutation had been designed within the com-
plementary part of the primers TTm-F and
TTm-R. The cloning was performed using
E. coli JM109, and the plasmid was extracted
using the AxyPrep Plasmid Miniprep Kit
(Corning, NY, USA). The insert was validated
by DNA sequencing.
Haloarchaeal cells were transformed ac-
cording to the online Halohandbook (https://
haloarchaea.com/wp-content/uploads/2018/
10/Halohandbook_2009_v7.3mds.pdf), and
Li et al., Science 372, eabe5601 (2021)
30 April 2021
8 of 11
RES EARCH | R E S E A R C H A R T I C L E
the transformants were selected on the yeast
extract–subtracted AS-168 plates or the Hv-YPC
plates without adding uracil and thymidine.
The transformation efficiency (CFU/mg) was
log-transformed, and then the average and
standard deviation were calculated for plotting.
Two-tailed Student’s t test was also performed
based on the log-transformed data. Each trans-
formation assay was double-checked by a dif-
ferent lab member.
Gene knockout and knockin
To knock out creTA or a cas gene, their up-
stream ~500 bp and downstream ~500 bp were
amplified using the corresponding primer pair
UF/UR or DF/DR, and the generated upstream
and downstream fragments were connected by
overlap extension PCR using the primer pair
UF/DR. The final PCR products were digested
and ligated into the suicide vector pHAR (47)
and then introduced into the parental (WT or
DTA) cells. The mutants were obtained through
the previously described two-step screening
(47) and further validated by colony PCR using
the UF/DR primer pair and subsequent DNA
sequencing. In a similar way, the mini-CRISPR
containing the yS spacer was constructed by
overlap extension PCR and used to replace the
WT CRISPR on the DTA chromosome.
Northern blot analysis
Haloarchaeal cells were collected from 3 ml
of early-stationary culture by centrifugation,
and the total RNA was extracted using TRIzol
(Thermo Fisher Scientific, MA, USA) according
to the standard protocol. RNA concentration
was determined by a Nanodrop 1000 spectro-
photometer (Thermo Fisher Scientific, MA,
USA). A total of 8 mg of RNA was mixed with an
equal volume of the RNA loading dye (Takara,
Shiga, Japan), denatured by heating at 65°C
for 10 min, and coelectrophoresed with the
Century-Plus RNA ladder (Thermo Fisher
Scientific, MA, USA) or a biotin-labeled 64-nt
single-stranded DNA on an 8% polyacrylamide
gel (containing 7.6 M urea). The ladder lane
was excised and imaged after ethidium bro-
mide staining. The RNA samples were electro-
transferred onto a Biodyne B nylon membrane
(Pall, NY, USA). The biotin-labeled probe was
used for hybridization, and the signal was de-
tected using the Chemiluminescent Nucleic
Acid Detection Module Kit (Thermo Fisher
Scientific, MA, USA), according to the manu-
facturer’s protocol. The membrane was imaged
using the Tanon 5200 Multi chemiluminescent
imaging system (Tanon Science & Technology,
Shanghai, China). Each blotting assay was per-
formed with two biological replicates, and a
representative result was provided.
RNA-seq analysis
The DTA or DTADcas6 colonies transformed by
pIRm were picked to inoculate 10 ml of yeast
extract–subtracted AS-168. After subinocula-
tion and a 2-day cultivation, the late exponen-
tial culture was collected and total RNA was
extracted. A total of 50 mg of RNA was treated
with polynucleotide kinase (New England
Biolabs, MA, USA) according to the manufac-
turer’s protocol. The kinase was inactivated
by incubating at 65°C for 20 min and then
removed through the phenol-chloroform ex-
traction. After precipitation with the same
volume of isopropanol and 0.1 volume of 3M
sodium acetate, the RNA sample was redis-
solved. RNA molecules ranging from 30 to
300 nt were selected to construct a small RNA
library with the NEXTFLEX Small RNA-Seq
Kit (Bioo Scientific, TX, USA) and then sub-
jected to Illumina HiSeq sequencing (paired-
end, 150-bp reads). The raw data was processed
to remove adapters. The resulting reads were
mapped to the creA sequence using custom
Perl scripts (49).
Primer extension analysis
The 5′-FAM (6-carboxyfluorescein)–labeled
primer+155 (table S3) was ordered from Thermo
Fisher Scientific. A total of 2.5 mg of the labeled
primer was mixed with 5 mg of the total RNA,
and reverse transcription was performed using
200 enzyme units (U) of M-MLV reverse tran-
scriptase (Promega, WI, USA). The extension
products were screened using the ABI3730xl
DNA Analyzer (Thermo Fisher Scientific, MA,
USA), and the results were analyzed using
GeneMapper 4.1.
Heterologous expression of CreT
The tryptophan-inducible promoter p.tnaA
(16) was linked with the 78-bp creT gene using
the overlap extension PCR strategy. The hy-
brid DNA was digested by BamHI and KpnI,
inserted into the predigested expression vector
pTA1228 (16), and then introduced into the
H. volcanii H1424 cells. Single colonies of each
transformant (by empty or modified pTA1228)
were selected and separately inoculated into
100 ml of Hv-YPC medium in triplicate for
growth-curve measurements. The inducing
medium contained 500 mg/liter tryptophan
for toxin induction. Optical density at 600 nm
(OD600) was monitored using the Shimadzu
UV-2550 spectrophotometer. At different time
points (0, 24, 48, 72, and 96 hours) through the
growth curve, the cell culture was sampled,
serially diluted, and plated onto the Hv-YPC
media for CFU calculation. The triplicates were
measured for each treatment to get the mean
and standard deviation.
Fluorescence measurement
The gene of a soluble-modified red-shifted
GFP protein (50) was linked to the PcreT pro-
moter (or its derepressed mutant) using the
overlap extension PCR strategy. The hybrid
DNA was introduced into the H. hispanica
cells using the pWL502 vector. For each trans-
formation assay, three individual colonies were
selected and cultured to the late exponential
phase, and their OD600 and fluorescence
were simultaneously determined using the
Synergy H4 Hybrid multimode microplate
reader (BioTeck, VT, USA). The fluorescence/
OD600 ratio was calculated for each of the
three individual samples, and average and
standard deviation were calculated. Two-tailed
Student’s t test was performed. The transfor-
mant cells were visualized using the Leica TCS
SP8 confocal microscope in combination with
Leica application suite software (Las X).
Plasmid challenge assay
Two complementary oligonucleotides that
included the sequence of spacer1, a 5′-TTC-3′
trinucleotide as the PAM, and two sticky ends
were ordered from Thermo Fisher Scientific.
The oligonucleotides were mixed, denatured,
annealed, and then inserted into pWL502
(predigested by BamHI and KpnI) to gener-
ate the target plasmid pTarget. H. hispanica
strains stored at 4°C for two years were re-
suscitated in fresh medium and then subino-
culated before transformation by pTarget. The
survivors were randomly selected for colony
PCR analyses. The forward primer ~500 bp
upstream of cas6 (cascade-F) and the back-
ward primer ~500 bp downstream of cas5
(cascade-R) were used to amplify the four
cascade genes (table S3). The forward primer
cascade-F and the backward primer ~500 bp
downstream of cas3 (cas3-R) were used to
amplify the genomic fragment containing the
cascade genes and cas3. The PCR products
were further analyzed by DNA sequencing to
get the information of the precise position of
IS insertion events (table S2).
Codon usage analysis
The H. hispanica protein-coding genes were
downloaded from the NCBI ftp site (https://
ftp.ncbi.nlm.nih.gov/genomes/all/GCA/000/
223/905/GCA_000223905.1_ASM22390v1/).
The usage of AGA/AGG codons within these
genes was examined. For each of the 2025
AGA/AGG-containing genes (or the 1075 genes
containing only one AGA or AGG codon), we
calculated the frequency of (AGA+AGG) codon
usage for every 25 codons, starting from its
initiation codon to the 250th codon.
Bioinformatic analysis
The folding potential of RNA was analyzed
using the RNAfold webserver (51). Sequence
alignments were constructed using the T-Coffee
webserver, and the results were visualized using
the GeneDoc software (version 2.6.002). The
base pairings between CreT and 16S rRNA were
analyzed using the IntaRNA web server (52). To
identify the target site of H. hispanica CreA, a
regular expression (“CCTTG.GCTAT”) was used
Li et al., Science 372, eabe5601 (2021)
30 April 2021
9 of 11
RES EARCH | R E S E A R C H A R T I C L E
to search the genome sequence. The targets
of CreA analogs were similarly predicted (see
next section).
Search for creTA homologs and analogs
To find creTA homologs, the BLASTN program
was run with default parameters against the
NCBI nucleotide collection database or against
the RefSeq Genome Database (taxonomy ID:
183963).
To identify potential creTA homologs or ana-
logs in haloarchaea, the intergenic sequences
flanking haloarchaeal cas6 genes (downloaded
from the NCBI database) were searched for a
conserved 5′ handle sequence “NTTGAAGN.”
This motif, together with the upstream 22 bp,
was assumed to represent the YR. The 30 to
40 bp downstream of this motif were assumed
to represent the YS, and the cas6-flanking se-
quences were searched using a regular expres-
sion for a putative target matching the first
1 to 5 and 7 to 11 nucleotides of the YS. Only
when the target was identified, the assumed
YR and YS sequences were defined as a creA
analog, and a creT analog was predicted to be
located between creA and its target.
For the preliminary identification of creTA
analogs in other archaea and bacteria, a data-
base containing ~5000 archaeal and ~40,000
nonredundant bacterial genomes was down-
loaded from the NCBI. The Cas6 proteins were
identified using the corresponding hidden
Markov model profiles (40) and the HMMER
suite (53), and CRISPR arrays were identified
using the minCED tool (https://github.com/
ctSkennerton/minced) with default parame-
ters. For each cas6 gene, the upstream and
downstream sequences (500 bp from each side)
were extracted and then searched for repeat-
like elements (putative YR) with the BLASTN
program (-task blastn-short), using the CRISPR
repeat sequence from the same genome as the
query (49). The sequences containing a single
repeat-like sequence (to exclude from consid-
eration of CRISPR arrays) were manually exam-
ined for putative creTA as described above.
RE FE RENCES AND N OT ES
1. R. Barrangou et al., CRISPR provides acquired resistance
against viruses in prokaryotes. Science 315, 1709–1712 (2007).
doi: 10.1126/science.1138140; pmid: 17379808
2. S. J. J. Brouns et al., Small CRISPR RNAs guide antiviral
defense in prokaryotes. Science 321, 960–964 (2008).
doi: 10.1126/science.1159689; pmid: 18703739
3. B. Wiedenheft, S. H. Sternberg, J. A. Doudna, RNA-guided
genetic silencing systems in bacteria and archaea. Nature 482,
331–338 (2012). doi: 10.1038/nature10886; pmid: 22337052
4. F. Hille et al., The biology of CRISPR-Cas: Backward and
forward. Cell 172, 1239–1259 (2018). doi: 10.1016/
j.cell.2017.11.032; pmid: 29522745
5. E. V. Koonin, K. S. Makarova, F. Zhang, Diversity, classification
and evolution of CRISPR-Cas systems. Curr. Opin. Microbiol.
37, 67–78 (2017). doi: 10.1016/j.mib.2017.05.008;
pmid: 28605718
6. L. S. Qi et al., Repurposing CRISPR as an RNA-guided platform
for sequence-specific control of gene expression. Cell 152,
1173–1183 (2013). doi: 10.1016/j.cell.2013.02.022;
pmid: 23452860
7. M. L. Luo, A. S. Mullis, R. T. Leenay, C. L. Beisel, Repurposing
endogenous type I CRISPR-Cas systems for programmable
gene repression. Nucleic Acids Res. 43, 674–681 (2015).
doi: 10.1093/nar/gku971; pmid: 25326321
8. H. K. Ratner et al., Catalytically active Cas9 mediates
transcriptional interference to facilitate bacterial virulence. Mol.
Cell 75, 498–510.e5 (2019). doi: 10.1016/j.molcel.2019.05.029;
pmid: 31256988
9. R. E. Workman et al., A natural single-guide RNA repurposes
Cas9 to autoregulate CRISPR-Cas expression. Cell 184,
675–688.e19 (2021). doi: 10.1016/j.cell.2020.12.017;
pmid: 33421369
10. R. Page, W. Peti, Toxin-antitoxin systems in bacterial growth
arrest and persistence. Nat. Chem. Biol. 12, 208–214 (2016).
doi: 10.1038/nchembio.2044; pmid: 26991085
11. L. R. Walling, J. S. Butler, Toxins targeting transfer RNAs:
Translation inhibition by bacterial toxin-antitoxin systems.
Wiley Interdiscip. Rev. RNA 10, e1506 (2019). doi: 10.1002/
wrna.1506; pmid: 30296016
12. N. Fraikin, F. Goormaghtigh, L. Van Melderen, Type II toxin-antitoxin
systems: Evolution and revolutions. J. Bacteriol. 202, e00763-e19
(2020). doi: 10.1128/JB.00763-19; pmid: 31932311
13. M. Li, R. Wang, D. Zhao, H. Xiang, Adaptation of the Haloarcula
hispanica CRISPR-Cas system to a purified virus strictly
requires a priming process. Nucleic Acids Res. 42, 2483–2492
(2014). doi: 10.1093/nar/gkt1154; pmid: 24265226
14. M. Brenneis, O. Hering, C. Lange, J. Soppa, Experimental
characterization of Cis-acting elements important for
translation and transcription in halophilic archaea. PLOS Genet.
3, e229 (2007). doi: 10.1371/journal.pgen.0030229;
pmid: 18159946
15. S. Cai et al., A novel DNA-binding protein, PhaR, plays a central
role in the regulation of polyhydroxyalkanoate accumulation
and granule formation in the haloarchaeon Haloferax
mediterranei. Appl. Environ. Microbiol. 81, 373–385 (2015).
doi: 10.1128/AEM.02878-14; pmid: 25344243
16. T. Allers, S. Barak, S. Liddell, K. Wardell, M. Mevarech,
Improved strains and plasmid vectors for conditional
overexpression of His-tagged proteins in Haloferax volcanii.
Appl. Environ. Microbiol. 76, 1759–1769 (2010). doi: 10.1128/
AEM.02670-09; pmid: 20097827
17. S. Sartorius-Neef, F. Pfeifer, In vivo studies on putative
Shine-Dalgarno sequences of the halophilic archaeon
Halobacterium salinarum. Mol. Microbiol. 51, 579–588 (2004).
doi: 10.1046/j.1365-2958.2003.03858.x; pmid: 14756795
18. M. Huber et al., Translational coupling via termination-reinitiation
in archaea and bacteria. Nat. Commun. 10, 4006 (2019).
doi: 10.1038/s41467-019-11999-9; pmid: 31488843
19. O. Hering, M. Brenneis, J. Beer, B. Suess, J. Soppa, A novel
mechanism for translation initiation operates in haloarchaea.
Mol. Microbiol. 71, 1451–1463 (2009). doi: 10.1111/
j.1365-2958.2009.06615.x; pmid: 19210623
20. G. F. T. Chen, M. Inouye, Suppression of the negative effect
of minor arginine codons on gene expression; preferential
usage of minor codons within the first 25 codons of the
Escherichia coli genes. Nucleic Acids Res. 18, 1465–1473
(1990). doi: 10.1093/nar/18.6.1465; pmid: 2109307
21. G. T. Chen, M. Inouye, Role of the AGA/AGG codons, the rarest
codons in global gene expression in Escherichia coli. Genes Dev. 8,
2641–2652 (1994). doi: 10.1101/gad.8.21.2641; pmid: 7958922
22. R. Wang, M. Li, L. Gong, S. Hu, H. Xiang, DNA motifs determining
the accuracy of repeat duplication during CRISPR adaptation
in Haloarcula hispanica. Nucleic Acids Res. 44, 4266–4277
(2016). doi: 10.1093/nar/gkw260; pmid: 27085805
23. D. Dar, D. Prasse, R. A. Schmitz, R. Sorek, Widespread
formation of alternative 3′ UTR isoforms via transcription
termination in archaea. Nat. Microbiol. 1, 16143 (2016).
doi: 10.1038/nmicrobiol.2016.143; pmid: 27670118
24. L. Gong et al., Primed adaptation tolerates extensive structural
and size variations of the CRISPR RNA guide in Haloarcula
hispanica. Nucleic Acids Res. 47, 5880–5891 (2019).
doi: 10.1093/nar/gkz244; pmid: 30957847
25. E. Semenova et al., Interference by clustered regularly
interspaced short palindromic repeat (CRISPR) RNA is
governed by a seed sequence. Proc. Natl. Acad. Sci. U.S.A. 108,
10098–10103 (2011). doi: 10.1073/pnas.1104144108;
pmid: 21646539
26. H. Zhao et al., Crystal structure of the RNA-guided immune
surveillance Cascade complex in Escherichia coli. Nature 515,
147–150 (2014). doi: 10.1038/nature13733; pmid: 25118175
27. M. Li, R. Wang, H. Xiang, Haloarcula hispanica CRISPR
authenticates PAM of a target sequence to prime
discriminative adaptation. Nucleic Acids Res. 42, 7226–7235
(2014). doi: 10.1093/nar/gku389; pmid: 24803673
28. F. Pfeifer, U. Blaseio, Transposition burst of the ISH27 insertion
element family in Halobacterium halobium. Nucleic Acids Res.
18, 6921–6925 (1990). doi: 10.1093/nar/18.23.6921;
pmid: 2175883
29. K. S. Makarova et al., An updated evolutionary classification of
CRISPR-Cas systems. Nat. Rev. Microbiol. 13, 722–736 (2015).
doi: 10.1038/nrmicro3569; pmid: 26411297
30. J. E. Walker et al., Development of both type I-B and type II
CRISPR/Cas genome editing systems in the cellulolytic
bacterium Clostridium thermocellum. Metab. Eng. Commun. 10,
e00116 (2019). doi: 10.1016/j.mec.2019.e00116;
pmid: 31890588
31. P. Boudry et al., Function of the CRISPR-Cas system of the
human pathogen Clostridium difficile. mBio 6, e01112-155
(2015). doi: 10.1128/mBio.01112-15; pmid: 26330515
32. K. Osakabe et al., Genome editing in plants using CRISPR type
I-D nuclease. Commun. Biol. 3, 648 (2020). doi: 10.1038/
s42003-020-01366-6; pmid: 33159140
33. J. Lin et al., DNA targeting by subtype I-D CRISPR-Cas shows
type I and type III features. Nucleic Acids Res. 48, 10470–10478
(2020). doi: 10.1093/nar/gkaa749; pmid: 32960267
34. J. Iranzo, J. A. Cuesta, S. Manrubia, M. I. Katsnelson, E. V. Koonin,
Disentangling the effects of selection and loss bias on gene
dynamics. Proc. Natl. Acad. Sci. U.S.A. 114, E5616–E5624 (2017).
doi: 10.1073/pnas.1704925114; pmid: 28652353
35. A. Stern, L. Keren, O. Wurtzel, G. Amitai, R. Sorek, Self-targeting by
CRISPR: Gene regulation or autoimmunity? Trends Genet. 26,
335–340 (2010). doi: 10.1016/j.tig.2010.05.008; pmid: 20598393
36. R. B. Vercoe et al., Cytotoxic chromosomal targeting by
CRISPR/Cas systems can reshape bacterial genomes and expel or
remodel pathogenicity islands. PLOS Genet. 9, e1003454 (2013).
doi: 10.1371/journal.pgen.1003454; pmid: 23637624
37. D. Bikard, A. Hatoum-Aslan, D. Mucida, L. A. Marraffini, CRISPR
interference can prevent natural transformation and virulence
acquisition during in vivo bacterial infection. Cell Host
Microbe 12, 177–186 (2012). doi: 10.1016/j.chom.2012.06.003;
pmid: 22901538
38. W. Jiang et al., Dealing with the evolutionary downside of
CRISPR immunity: Bacteria and beneficial plasmids. PLOS
Genet. 9, e1003844 (2013). doi: 10.1371/journal.
pgen.1003844; pmid: 24086164
39. L. van Sluijs et al., Addiction systems antagonize bacterial
adaptive immunity. FEMS Microbiol. Lett. 366, fnz047 (2019).
pmid: 30834930
40. K. S. Makarova et al., Evolutionary classification of CRISPR-Cas
systems: A burst of class 2 and derived variants. Nat. Rev.
Microbiol. 18, 67–83 (2020). doi: 10.1038/s41579-019-0299-x;
pmid: 31857715
41. A. J. Meeske, S. Nakandakari-Higa, L. A. Marraffini, Cas13-induced
cellular dormancy prevents the rise of CRISPR-resistant
bacteriophage. Nature 570, 241–245 (2019). doi: 10.1038/
s41586-019-1257-5; pmid: 31142834
42. J. T. Rostøl, L. A. Marraffini, Non-specific degradation
of transcripts promotes plasmid clearance during type III-A
CRISPR-Cas immunity. Nat. Microbiol. 4, 656–662 (2019).
doi: 10.1038/s41564-018-0353-x; pmid: 30692669
43. S. Song, T. K. Wood, A primary physiological role of toxin/
antitoxin systems is phage inhibition. Front. Microbiol. 11, 1895
(2020). doi: 10.3389/fmicb.2020.01895; pmid: 32903830
44. E. V. Koonin, F. Zhang, Coupling immunity and programmed
cell suicide in prokaryotes: Life-or-death choices. BioEssays 39,
e201600186 (2017). doi: 10.1002/bies.201600186;
pmid: 27896818
45. E. V. Koonin, K. S. Makarova, Y. I. Wolf, M. Krupovic, Evolutionary
entanglement of mobile genetic elements and host defence
systems: Guns for hire. Nat. Rev. Genet. 21, 119–131 (2020).
doi: 10.1038/s41576-019-0172-9; pmid: 31611667
46. G. Faure et al., Comparative genomics and evolution of
trans-activating RNAs in class 2 CRISPR-Cas systems. RNA Biol.
16, 435–448 (2019). doi: 10.1080/15476286.2018.1493331;
pmid: 30103650
47. H. Liu, J. Han, X. Liu, J. Zhou, H. Xiang, Development of pyrF-
based gene knockout systems for genome-wide manipulation
of the archaea Haloferax mediterranei and Haloarcula hispanica.
J. Genet. Genomics 38, 261–269 (2011). doi: 10.1016/
j.jgg.2011.05.003; pmid: 21703550
48. T. Allers, H. P. Ngo, M. Mevarech, R. G. Lloyd, Development of
additional selectable markers for the halophilic archaeon
Haloferax volcanii based on the leuB and trpA genes. Appl.
Environ. Microbiol. 70, 943–953 (2004). doi: 10.1128/
AEM.70.2.943-953.2004; pmid: 14766575
49. H. Yu, M. Li, Scripts accompanying “Toxin-antitoxin RNA pairs
safeguard CRISPR-Cas systems,” Version 1.0, Zenodo (2021);
https://doi.org/10.5281/zenodo.4586956.
Li et al., Science 372, eabe5601 (2021)
30 April 2021
10 of 11
RES EARCH | R E S E A R C H A R T I C L E
50. C. J. Reuter, J. A. Maupin-Furlow, Analysis of proteasome-
dependent proteolysis in Haloferax volcanii cells, using
short-lived green fluorescent proteins. Appl. Environ.
Microbiol. 70, 7530–7538 (2004). doi: 10.1128/
AEM.70.12.7530-7538.2004; pmid: 15574956
51. R. Lorenz et al., ViennaRNA Package 2.0. Algorithms Mol. Biol.
6, 26 (2011). doi: 10.1186/1748-7188-6-26; pmid: 22115189
52. M. Mann, P. R. Wright, R. Backofen, IntaRNA 2.0: Enhanced
and customizable prediction of RNA-RNA interactions. Nucleic
Acids Res. 45, W435–W439 (2017). doi: 10.1093/nar/gkx279;
pmid: 28472523
53. R. D. Finn, J. Clements, S. R. Eddy, HMMER web server:
Interactive sequence similarity searching. Nucleic Acids Res. 39,
W29–W37 (2011). doi: 10.1093/nar/gkr367; pmid: 21593126
ACKN OW LEDG MEN TS
We thank X. Chen for providing the Natrinema sp. J7-2 strain,
X. Zhang for help with microscopy, and K. S. Makarova for help with
the bioinformatics analysis. Funding: H.X. is supported by the
Strategic Priority Research Program of the Chinese Academy of
Sciences (XDA24020101), the National Key R&D Program of China
(2020YFA0906800), and the National Natural Science Foundation
of China (91751201). M.L. is supported by the National Natural
Science Foundation of China (31771381, 32022003, and 31970544),
the National Transgenic Science and Technology Program
(2019ZX08010-001 and 2019ZX08010-003), the Youth Innovation
Promotion Association of CAS (2020090), and the Young Elite
Scientists Sponsorship Program by CAST (2017QNRC001). S.A.S.
and E.V.K. are supported by the Intramural Research Program
of the National Institutes of Health (National Library of Medicine).
Author contributions: M.L. and H.X. designed experiments and
supervised the project. M.L., L.G., and F.C. constructed pTA
derivatives and performed the transformation assays with input
from T.W. and J.Z. M.L. and L.G. performed the RNA-seq, Northern
blotting, primer extension, and plasmid challenge assays. F.C.
performed the fluorescence analyses and measured the H. volcanii
growth curves. M.L. and H.Y. analyzed the RNA-seq data. R.W.
constructed the cas knockout mutants. M.L. performed the
bioinformatic analyses with input from H.Y., S.S., D.Z., and S.Z. M.L.,
E.V.K., and H.X. analyzed the data and wrote the manuscript,
which was edited and approved by all authors. Competing interests:
M.L., F.C., and H.X. have filed a related patent. Data and materials
availability: All data are available in the main text or the
supplementary materials. Reagents are available upon request from
H.X. Perl scripts are available on Zenodo (49).
SUPPLEMENTARY MATERIALS
science.sciencemag.org/content/372/6541/eabe5601/suppl/DC1
Figs. S1 to S14
Tables S1 to S3
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
16 September 2020; resubmitted 29 December 2020
Accepted 10 March 2021
10.1126/science.abe5601
Li et al., Science 372, eabe5601 (2021)
30 April 2021
11 of 11
|
10.1126_science.abe2424
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
CORONAVIRUS
Transmission heterogeneities, kinetics, and
controllability of SARS-CoV-2
Kaiyuan Sun*†, Wei Wang†, Lidong Gao†, Yan Wang, Kaiwei Luo, Lingshuang Ren, Zhifei Zhan,
Xinghui Chen, Shanlu Zhao, Yiwei Huang, Qianlai Sun, Ziyan Liu, Maria Litvinova,
Alessandro Vespignani, Marco Ajelli, Cécile Viboud‡, Hongjie Yu*‡
INTRODUCTION: The role of transmission het-
erogeneities in severe acute respiratory syn-
drome coronavirus 2 (SARS-CoV-2) dynamics
remains unclear, particularly those heteroge-
neities driven by demography, behavior, and
interventions. To understand individual het-
erogeneities and their effect on disease con-
trol, we analyze detailed contact-tracing data
from Hunan, a province in China adjacent to
Hubei and one of the first regions to experience
a SARS-CoV-2 outbreak in January to March
2020. The Hunan outbreak was swiftly brought
under control by March 2020 through a com-
bination of nonpharmaceutical interventions
including population-level mobility restriction
(i.e., lockdown), traveler screening, case isola-
tion, contact tracing, and quarantine. In parallel,
highly detailed epidemiological information on
SARS-CoV-2–infected individuals and their close
contacts was collected by the Hunan Provincial
Center for Disease Control and Prevention.
RATIONALE: Contact-tracing data provide infor-
mation to reconstruct transmission chains and
understand outbreak dynamics. These data can
in turn generate valuable intelligence on key
epidemiological parameters and risk factors for
transmission, which paves the way for more-
targeted and cost-effective interventions.
RESULTS: On the basis of epidemiological in-
formation and exposure diaries on 1178 SARS-
CoV-2–infected individuals and their 15,648
close contacts, we developed a series of statisti-
cal and computational models to stochastically
reconstruct transmission chains, identify risk
factors for transmission, and infer the infec-
tiousness profile over the course of a typical
Transmission chains
.5
.5
0
iz
5
_viz
Changsha
Yueyang
Shaoyang
Loudi
Changsha
Yueyang
Changde
Shaoyang
Loudi
Zhuzhou
Changde
Zhuzhou
iOthers
Yiyang
Others
0
1
2
3
8
5
Number of secondary infections
6
4
7
e
s
a
c
e
g
A
e
s
a
c
e
g
A
90
75
60
45
30
15
0
90
75
60
45
30
15
0
Data
Negative
binomial
9
10
Contact matrices
Community
Social
Extended family
Household
0 15 30 45 60 75 90 0 15 30 45 60 75 90
Age contact
Age contact
Transmission kinetics
0.8
0.6
0.4
0.2
0.0
y
t
i
l
i
b
a
b
o
r
P
)
s
y
a
d
(
n
o
i
t
a
o
s
i
l
o
t
t
e
s
n
O
>6
4-6
2-4
<2
0
30
20
10
Generation intervals
(days)
40
-20
-10
0
10
20
30
Serial intervals
(days)
30 40
50 60 70 80 90
Contribution of presymptomatic
transmission (percent)
Transmission chains, contact patterns, and transmission kinetics of SARS-CoV-2 in Hunan, China,
based on case and contact-tracing data from Hunan, China. (Top left) One realization of the
reconstructed transmission chains, with a histogram representing overdispersion in the distribution of
secondary infections. (Top right) Contact matrices of community, social, extended family, and household
contacts reveal distinct age profiles. (Bottom) Earlier isolation of primary infections shortens the generation
and serial intervals while increasing the relative contribution of transmission in the presymptomatic phase.
infection. We observe overdispersion in the
distribution of secondary infections, with 80%
of secondary cases traced back to 15% of in-
fections, which indicates substantial transmis-
sion heterogeneities. We find that SARS-CoV-2
transmission risk scales positively with the
duration of exposure and the closeness of social
interactions, with the highest per-contact risk
estimated in the household. Lockdown inter-
ventions increase transmission risk in families
and households, whereas the timely isolation of
infected individuals reduces risk across all types
of contacts. There is a gradient of increasing
susceptibility with age but no significant dif-
ference in infectivity by age or clinical severity.
Early isolation of SARS-CoV-2–infected indi-
viduals drastically alters transmission kinetics,
leading to shorter generation and serial inter-
vals and a higher fraction of presymptomatic
transmission. After adjusting for the censoring
effects of isolation, we find that the infectious-
ness profile of a typical SARS-CoV-2 patient
peaks just before symptom onset, with 53% of
transmission occurring in the presymptomatic
phase in an uncontrolled setting. We then use
these results to evaluate the effectiveness of
individual-based strategies (case isolation and
contact quarantine) both alone and in combi-
nation with population-level contact reductions.
We find that a plausible parameter space for
SARS-CoV-2 control is restricted to scenarios
where interventions are synergistically com-
bined, owing to the particular transmission
kinetics of this virus.
CONCLUSION: There is considerable heteroge-
neity in SARS-CoV-2 transmission owing to
individual differences in biology and contacts
that is modulated by the effects of interven-
tions. We estimate that about half of secondary
transmission events occur in the presympto-
matic phase of a primary case in uncontrolled
outbreaks. Achieving epidemic control requires
that isolation and contact-tracing interventions
are layered with population-level approaches,
such as mask wearing, increased teleworking,
and restrictions on large gatherings. Our study
also demonstrates the value of conducting high-
quality contact-tracing investigations to advance
our understanding of the transmission dynamics
of an emerging pathogen.▪
The list of author affiliations is available in the full article online.
*Corresponding author. Email: kaiyuan.sun@nih.gov (K.S.);
yhj@fudan.edu.cn (H.Y.)
†These authors contributed equally to this work.
‡These authors contributed equally to this work.
This is an open-access article distributed under the terms
of the Creative Commons Attribution license (https://
creativecommons.org/licenses/by/4.0/), which permits
unrestricted use, distribution, and reproduction in any
medium, provided the original work is properly cited.
Cite this article as K. Sun et al., Science 371, eabe2424
(2021). DOI: 10.1126/science.abe2424
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.abe2424
Sun et al., Science 371, 254 (2021)
15 January 2021
1 of 1
RES EARCH
R E S E A R C H A R T I C L E
◥
CORONAVIRUS
Transmission heterogeneities, kinetics, and
controllability of SARS-CoV-2
Kaiyuan Sun1*†, Wei Wang2†, Lidong Gao3†, Yan Wang2, Kaiwei Luo3, Lingshuang Ren2, Zhifei Zhan3,
Xinghui Chen2, Shanlu Zhao3, Yiwei Huang3, Qianlai Sun3, Ziyan Liu3, Maria Litvinova4,5,
Alessandro Vespignani6,5, Marco Ajelli4,6, Cécile Viboud1‡, Hongjie Yu2*‡
A long-standing question in infectious disease dynamics concerns the role of transmission
heterogeneities, which are driven by demography, behavior, and interventions. On the basis of detailed
patient and contact-tracing data in Hunan, China, we find that 80% of secondary infections traced back
to 15% of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) primary infections, which
indicates substantial transmission heterogeneities. Transmission risk scales positively with the duration
of exposure and the closeness of social interactions and is modulated by demographic and clinical
factors. The lockdown period increases transmission risk in the family and households, whereas isolation
and quarantine reduce risks across all types of contacts. The reconstructed infectiousness profile of
a typical SARS-CoV-2 patient peaks just before symptom presentation. Modeling indicates that
SARS-CoV-2 control requires the synergistic efforts of case isolation, contact quarantine, and
population-level interventions because of the specific transmission kinetics of this virus.
A lthough it has been well documented
that the clinical severity of COVID-19
increases with age (1–5), information is
limited on how transmission risk varies
with demographic factors, clinical pre-
sentation, and contact type (6–12). Individual-
based interventions such as case isolation,
contact tracing, and quarantine have been
shown to accelerate case detection and inter-
rupt transmission chains (13). However, these
interventions are typically implemented in
conjunction with population-level physical
distancing measures, and their effects on con-
tact patterns and transmission risk remains
difficult to separate (14–24). A better under-
standing of the factors driving severe acute
respiratory syndrome coronavirus 2 (SARS-
CoV-2) transmission is key to achieving epi-
demic control while minimizing societal cost,
particularly as countries relax physical dis-
tancing measures.
Hunan, a province in China adjacent to
Hubei, where the COVID-19 pandemic began,
experienced sustained SARS-CoV-2 transmis-
sion in late January and early February 2020,
1Division of International Epidemiology and Population
Studies, Fogarty International Center, National Institutes of
Health, Bethesda, MD, USA. 2School of Public Health, Fudan
University, Key Laboratory of Public Health Safety, Ministry
of Education, Shanghai, China. 3Hunan Provincial Center for
Disease Control and Prevention, Changsha, China.
4Department of Epidemiology and Biostatistics, Indiana
University School of Public Health, Bloomington, IN, USA.
5ISI Foundation, Turin, Italy. 6Laboratory for the Modeling of
Biological and Socio-technical Systems, Northeastern
University, Boston, MA, USA.
*Corresponding author. Email: kaiyuan.sun@nih.gov (K.S.); yhj@
fudan.edu.cn (H.Y.) †These authors contributed equally to this
work. ‡These authors contributed equally to this work.
followed by a quick suppression of the out-
break by March 2020. As in many other provinces
in China, epidemic control was achieved by
layering interventions targeting SARS-CoV-2
cases and their contacts with population-level
physical distancing measures. In this study,
we reconstruct transmission chains among all
identified SARS-CoV-2 infections in Hunan,
as of 3 April 2020, on the basis of granular
epidemiological information collected through
extensive surveillance and contact-tracing ef-
forts. We identify the demographic, clinical,
and behavioral factors that drive transmission
heterogeneities and evaluate how interventions
modulate the topology of the transmission
network. Further, we reconstruct the infec-
tiousness profile of SARS-CoV-2 over the
course of a typical infection and estimate the
feasibility of epidemic control by individual-
and population-based interventions.
We analyze detailed epidemiological records
for 1178 SARS-CoV-2–infected individuals and
their 15,648 close contacts—representing 19,227
separate exposure events—compiled by the
Hunan Provincial Center for Disease Control
and Prevention. Cases were identified be-
tween 16 January and 3 April 2020; primary
cases were captured by passive surveillance,
contact tracing, or traveler screening and then
were laboratory confirmed by reverse tran-
scription polymerase chain reaction (RT-PCR).
Individuals who were close contacts of the
primary cases were followed for at least
2 weeks after the last exposure to the infected
individual. Before 7 February 2020, contacts were
tested only if they developed symptoms during
the quarantine period. After 7 February 2020,
RT-PCR testing was required for all con-
tacts, and specimens were collected at least
once from each contact during quarantine,
regardless of symptoms. Upon positive RT-
PCR test results, infected individuals were
isolated in dedicated hospitals, regardless of
their clinical severity, and their contacts were
quarantined in medical observation facilities.
The case ascertainment process is visualized
in fig. S1.
The dataset includes 210 epidemiological
clusters representing 831 cases, with an ad-
ditional 347 sporadic cases (29%) unlinked to
any cluster (see supplementary materials and
methods for more details). For each cluster, we
stochastically reconstruct transmission chains
and estimate the timing of infection most com-
patible with each patient’s exposure history.
We analyze an ensemble of 100 reconstructed
transmission chains to account for uncertain-
ties in exposure histories (Fig. 1 visualizes
one realization of the transmission chains, and
fig. S2A illustrates variability in the topology of
the aggregation of 100 realizations of trans-
mission chains).
We observe between zero and four gener-
ations of transmission, with the largest cluster
involving 20 SARS-CoV-2–infected individuals.
The number of secondary infections ranges
from 0 to 10, with a distribution of secondary
infections best characterized by a negative
binomial distribution with mean m = 0.40
[95% confidence interval (CI): 0.35 to 0.47]
and variance m(1 + m / k) = 0.96 (95% CI: 0.74
to 1.26), where k = 0.30 (95% CI: 0.23 to 0.39)
is the dispersion parameter (Fig. 1). We find
that 80% of secondary infections can be traced
back to 15% of SARS-CoV-2–infected individ-
uals, which indicates substantial transmission
heterogeneities at the individual level. We can
also assess geographic diffusion within Hunan
province and find that the majority of trans-
mission events occur within the same prefec-
ture (94.3%; 95% CI: 93.7 to 95.0%), with
occasional spread between prefectures (5.7%;
95% CI: 5.0 to 6.3%).
Characterizing SARS-CoV-2 transmission
heterogeneities at the individual level
To dissect the individual transmission heter-
ogeneities and identify predictors of transmis-
sion, we analyze the infection risk among a
subset of 14,622 individuals who were close
contacts of 870 SARS-CoV-2 patients. This
dataset excludes primary cases whose infected
contacts reported a travel history to Wuhan.
The dataset represents 74% of all SARS-CoV-2
cases recorded in the Hunan patient database.
Contacts of these 870 patients have been care-
fully monitored so that 17,750 independent
exposure events have been captured.
We start by characterizing variation in
transmission risk across the diverse set of
17,750 exposures. We study how the per-
contact transmission risk varies with the type
Sun et al., Science 371, eabe2424 (2021)
15 January 2021
1 of 8
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. SARS-CoV-2 transmission chains. (Top)
One realization of the reconstructed transmission
chains among 1178 SARS-CoV-2–infected individ-
uals in Hunan province. Each node in the network
represents a patient infected with SARS-CoV-2,
and each link represents an infector-infectee
relationship. The color of the node denotes the
reporting prefecture of the infected individuals.
(Bottom) Distribution of the number of secondary
infections. Blue bars represent the ensemble
averaged across 100 stochastic samples of the
reconstructed transmission chains. Orange bars
represent the best fit of a negative binomial
distribution to the ensemble average. Vertical
lines indicate 95% CIs across 100 samples (of both
data and the models’ fitting results). Some
confidence intervals are narrow and not visible on
the plot. For sensitivity analysis, we also fit the
distribution with geometric and Poisson distributions.
On the basis of the Akaike information criterion (AIC),
the negative binomial distribution fit the data the
best (average AIC score for negative binomial
distribution: 1902; for geometric distribution: 1981;
and for Poisson distribution: 2259).
of exposures, exposure duration, exposure
timing, and physical distancing intervention,
after adjusting for demographic, clinical, and
travel-related factors. Exposures are grouped
into five categories on the basis of contact
type—i.e., household, extended family, social,
community, and health care (table S2)—with
the duration of exposure approximated by
the time interval between the initial and
final dates of exposure. To gauge the impact
of physical distancing on transmission risks,
we further stratify exposures by the date of
occurrence, with 25 January 2020 marking
the beginning of lockdown in Hunan [based
on Baidu Qianxi mobility index (25); fig. S3A,
insert]. To address putative variation in infec-
tiousness over the course of infection, we
distinguish whether exposures overlap with
the date of symptom onset of a primary case, a
period associated with high viral shedding.
We use a mixed-effects multiple logistic re-
gression model (GLMM-logit) to quantify the
effects of these factors on the per-contact risk
of transmission (see table S3 for a detailed
definition of all risk factors and summary
statistics).
On the basis of the point estimates of the
regression (see fig. S3A for regression results),
we find that household contacts pose the highest
risk of transmission followed by extended
family, social, and community contacts, in
agreement with a prior study (12). Health care
contacts have the lowest risk, which suggests
that adequate protective measures were adop-
ted by patients and health care staff in Hunan.
Notably, the impact of physical distancing
differs by contact type (Table 1): The risk of
transmission in the household increases
during the lockdown period, likely because
of increased contact frequency at home as a
result of physical confinement. By contrast,
the transmission risk decreases for commu-
nity and social contacts during lockdown, pos-
sibly because of the adoption of prudent
behaviors such as mask wearing, hand wash-
ing, and coughing-sneezing etiquette. We find
that longer exposures are riskier, with 1 addi-
tional day of exposure increasing the transmis-
sion risk by 10% (95% CI: 5 to 15%). Further,
transmission risk is higher around the time
of symptom presentation of the primary case
(Table 1). Additionally, susceptibility to infec-
tion (defined as the risk of infection given
contact with a primary case) varies by age:
Children aged 0 to 12 years are significantly
less susceptible than individuals aged 26 to
64 years (odds ratio 0.41; 95% CI: 0.26 to 0.63),
and patients older than 65 years are signif-
icantly more susceptible (odds ratio 1.39; 95%
CI: 1.02 to 1.91). By contrast, we find no statistical
support for age difference in infectivity (fig. S3A).
These results are in agreement with previous
findings (12, 26, 27).
For each of the 17,750 contact exposure
events, we estimate the probability of trans-
mission using the point estimate of the
baseline odds and the odds ratios from the
GLMM-logit regression (fig. S3A). In Fig. 2A,
we plot the distribution of transmission prob-
abilities for household, extended family, social,
and community contacts separately. The aver-
age per-contact transmission probability is
highest for household contacts (7.2%; 95% CI:
1.2 to 19.6%) followed by family (1.7%; 95% CI:
0.4 to 5.6%) and social contacts (0.9%; 95% CI:
0.2 to 2.7%), whereas the risk is lowest for com-
munity contacts (0.4%; 95% CI: 0.1 to 1.1%).
These transmission probabilities reflect the joint
Sun et al., Science 371, eabe2424 (2021)
15 January 2021
2 of 8
RES EARCH | R E S E A R C H A R T I C L E
Table 1. SARS-CoV-2 transmission risk in Hunan by contact type, duration of exposure, and
whether the exposure window contains the date of symptom onset of the primary case—a
period of intense viral shedding. Risk is further stratified by the date of implementation of social
distancing interventions in Hunan, which is 25 January 2020. The regression model is adjusted
for demographic characteristics of the cases and their contacts, clinical symptoms, and travel history.
Details are provided in the materials and methods, and the full results of the regression, including
additional risk factors, are shown in fig. S3.
Risk factors
Odds ratio
95% CI
Social contacts
Household contacts
Extended family contacts
(1.39, 3.49)
Before 25 January 2020
............................................................................................................................................
(2.47, 5.79)
After 25 January 2020
............................................................................................................................................
Reference
Before 25 January 2020
............................................................................................................................................
(0.60, 1.46)
After 25 January 2020
............................................................................................................................................
(0.37, 1.06)
Before 25 January 2020
............................................................................................................................................
(0.21, 0.78)
After 25 January 2020
............................................................................................................................................
(0.19, 0.74)
Before 25 January 2020
............................................................................................................................................
(0.05, 0.71)
After 25 January 2020
............................................................................................................................................
(0.03, 0.68)
Before 25 January 2020
............................................................................................................................................
(0.01, 0.90)
After 25 January 2020
Health care contacts
............................................................................................................................................
(1.05, 1.15)
Duration of exposure (days)
.....................................................................................................................................................................................................................
(1.09, 2.04)
Symptom onset within exposure window (yes)
.....................................................................................................................................................................................................................
2.20***
3.79***
1.00
0.94
0.63
0.41**
0.37**
0.20*
0.15*
0.10*
1.10***
1.49*
Community contacts
*P < 0.05; **P < 0.01; ***P < 0.001.
effect of duration of exposure (Fig. 2B) super-
imposed on differences in transmission risk
by type of contact (fig. S3A). Although con-
fidence intervals on risk estimates are broad,
there is statistical support for separating out
contacts in five categories and including a
time covariate to capture the effect of the
lockdown rather than collapsing the contact
data into fewer categories (table S4). By con-
trast, there is no statistical support for a more
complex model that considers a different ef-
fect of contact duration by type of contact
(table S4). It is worth noting that the per-
contact transmission probabilities were esti-
mated in a situation of intense interventions
and high population awareness of the di-
sease, and thus, they may be not generaliz-
able elsewhere.
The number of contacts is also a key driver
of individual transmission potential and varies
by contact type. Figure 2C presents the contact
degree distribution, defined as the number of
distinct contacts per individual. We find that
the distributions of individual contact degree
are overdispersed with dispersion parameter
0 < k < 1 across all contact types. Further-
more, household (k = 0.72) and extended
family (k = 0.64) contacts are less dispersed
than social (k = 0.19) and community (k = 0.14)
contacts, which suggests that contact hetero-
geneities are inversely correlated with the close-
ness of social interactions. Figure 2D visualizes
the age-specific contact patterns between the
primary cases and their contacts, demonstrat-
ing diverse mixing patterns across different
types of contact. Specifically, household con-
tacts present the canonical three-bands pattern,
where the diagonal illustrates age-assortative
interactions, and the two off-diagonals repre-
sent intergenerational mixing (28, 29). Other
contact types display more diffusive mixing
patterns by age. We also observe that among
all primary cases, young and middle-aged adults
have the most social contacts (Fig. 2E).
Next, we summarize the overall transmis-
sion potential of an individual by calculating
the cumulative contact rate (CCR) of all primary
cases. The CCR captures how contact oppor-
tunities vary with demography, temporal varia-
tion in the infectiousness profile, an individual’s
contact degree, and interventions (see section
4.3 in the materials and methods for detailed
definition). After adjusting for age, sex, clinical
presentation, and travel history to Wuhan,
we find that physical distancing measures
increase CCRs for household and extended
family contacts and decrease (although not
statistically significantly) CCRs for social and
community contacts (Fig. 2E). By contrast,
faster case isolation universally reduces CCRs,
decreasing transmission opportunities across
all contact types (Fig. 2E).
Characterizing the natural history of SARS-
CoV-2 infection by strength of interventions
We have characterized SARS-CoV-2 transmis-
sion risk factors and have shown that individual-
and population-based interventions have a
differential impact on contact patterns and
transmission potential. Next, we use our proba-
bilistic reconstruction of infector-infectee pairs
to further dissect transmission kinetics and
project the impact of interventions on SARS-
CoV-2 dynamics. On the basis of the recon-
structed transmission chains, we estimate a
median serial interval of 5.3 days, with an
interquartile range (IQR) of 2.7 to 8.3 days,
which represents the time interval between
symptom onset of an infector and that of his
or her infectee (fig. S7, B and D). The median
generation interval—defined as the interval
between the infection times of an infector and
his or her infectee—is 5.3 days, with an IQR of
3.1 to 8.7 days (fig. S7, A and C). We estimate
that 63.4% (95% CI: 60.2 to 67.2%) of all
transmission events occur before symptom
onset, which is comparable to findings from
other studies (6–8, 10–13, 18, 30, 31). However,
these estimates are affected by the intensity of
interventions; in Hunan, isolation and quar-
antine were in place throughout the epidemic.
Case isolation and contact quarantine are
meant to prevent potentially infectious indivi-
duals from contacting susceptible individuals,
effectively shortening the infectious period. As
a result, we would expect right censoring of
the generation and serial interval distribu-
tions (32). Symptomatic cases represent 86.5%
of all SARS-CoV-2 infections in our data; among
these patients, we observe longer generation
intervals for cases isolated later in the course
of their infection (Fig. 3A). The median gene-
ration interval increases from 4.0 days (IQR,
1.9 to 7.3 days) for cases isolated 2 days since
symptom onset to 7.0 days (IQR, 3.6 to 11.3 days)
for those isolated >6 days after symptom
onset (P < 0.001; Mann-Whitney U test). We
observe similar trends for the serial interval
distribution (Fig. 3B). The median serial in-
terval increases from 1.7 days (IQR, −1.6 to
4.8 days) for cases isolated <2 days after symp-
tom onset to 7.3 days (IQR, 3.4 to10.8 days)
for those isolated >6 days after symptom onset
(P < 0.001; Mann-Whitney U test).
Faster case isolation restricts transmission
to the earlier stages of infection, thus inflating
the contribution of presymptomatic transmis-
sion (Fig. 3C). The proportion of presympto-
matic transmission is estimated at 87.3% (95%
CI: 79.8 to 93.4%) if cases are isolated within
2 days of symptom onset, whereas this propor-
tion decreases to 47.5% (95% CI: 41.4 to 53.3%)
if cases are isolated >6 days after symptom
onset (P < 0.001; Mann-Whitney U test).
Next, we adjust for censoring caused by case
isolation and reconstruct the infectiousness
profile of a typical SARS-CoV-2 patient in the
absence of interventions. To do so, we charac-
terize changes in the timeliness of case isola-
tion over time in Hunan. Figure S8 shows the
distributions of time from symptom onset to
isolation during three different phases of epi-
demic control, coinciding with major changes
in COVID-19 case definition (phase I: before
27 January; phase II: 27 January to 4 February;
and phase III: after 4 February; fig. S3) (33). In
phase I, 78% of cases were detected through
passive surveillance; as a result, most cases
Sun et al., Science 371, eabe2424 (2021)
15 January 2021
3 of 8
RES EARCH | R E S E A R C H A R T I C L E
Community
Social
Extended family
Household
A
B
C
D
E
Fig. 2. Heterogeneity in contact rates of SARS-CoV-2 cases and impact
of interventions, separated by contact type. Columns from left to right
represent community contacts (e.g., public transportation, food, and
entertainment), social contacts, extended family contacts, and household
contacts. (A) Violin plots representing the distribution of per-contact
transmission probability by contact type, adjusted for all other covariates in
fig. S3 (probability expressed in percentage; x axis). (B) Complementary
cumulative distribution function (CCDF) (y axis) for duration of exposure
(i.e., the probability that exposure is longer or equal to a certain value).
Dashed vertical lines indicate average values. Household contacts last the
longest, and as expected, contact duration decreases as social ties loosen.
(C) The distribution of the number of distinct contacts (degree distribution)
of the primary cases for each contact type. The y axis indicates probability
mass function (PMF). The dashed vertical lines indicate average values. The
dispersion parameter k is calculated on the basis of the relationship
s2 ¼ m
1þm=k, where m and s2 are the mean and variance of the number of
distinct contacts. Values of k < 1 indicate overdispersion. (D) Age
distribution of SARS-CoV-2 case–contact pairs (contact matrices).
(E) Rate ratios of negative binomial regression of the CCRs against
predictors including the infector’s age, sex, presence of fever or cough,
Wuhan travel history, whether symptom onset occurred before social
distancing was in place (before or after 25 January 2020), and time from
isolation to symptom onset. CCRs represent the sum of relevant contacts
over a 1-week window centered at the date of the primary case’s symptom
onset. Dots and lines indicate point estimates and 95% CIs of the rate
ratios, respectively, and numbers below the dots indicate the numerical
value of the point estimates. Ref., reference category. *P < 0.05;
**P < 0.01; ***P < 0.001.
Sun et al., Science 371, eabe2424 (2021)
15 January 2021
4 of 8
RES EARCH | R E S E A R C H A R T I C L E
A
D
F
B
G
E
C
H
Fig. 3. The impact of interventions on SARS-CoV-2 transmission dynamics.
(A) Violin plot of the generation interval distributions stratified by time from
symptom onset to isolation or presymptomatic quarantine, based on an
ensemble of 100 realizations of the sampled transmission chains. (B) Same as
(A) but for the serial interval distributions. (C) Same as (A) but for the fraction of
presymptomatic transmission, among all transmission events, with vertical line
indicating 50% of presymptomatic transmissions. In (A) to (C), dots represent
the mean, and whiskers represent minimum and maximum. (D) Estimated
average (across 100 realizations of sampled transmission chains) transmission
risk of a SARS-CoV-2–infected individual since time of infection under four
intervention scenarios: the red solid line represents an uncontrolled epidemic
scenario modeled after the early epidemic dynamics in Wuhan before lockdown,
and the dashed lines represent scenarios where quarantine and case isolation
are in place and mimic phases I, II, and III of epidemic control in Hunan. The
shapes of these curves match those of the generation interval distributions in
each scenario, and the areas under the curve are equal to the ratios of the
baseline/effective basic reproduction numbers ðR0=RE
0
with time since symptom onset on the x axis [colors are the same as in (D)].
Þ. (E) Same as in (D) but
The vertical line represents symptom onset. (F) Reduction (percentage) in the
basic reproduction number as a function of mean time from symptom onset (or
from peak infectiousness for asymptomatic cases) to isolation tiso (x axis) and
fraction of SARS-CoV-2 infections being isolated (y axis). The distribution of
onset to isolation follows a normal distribution with mean tiso and standard
deviation of 2 days. The dashed lines indicate 30, 40, and 50% reductions in R0
under interventions. (G) Effective basic reproduction number as a function of
population-level reduction in contact rates (i.e., through physical distancing;
expressed as a percentage, x axis) and isolation rate (fraction of total infections
detected and further isolated). We assume baseline basic reproduction number
R0 = 2.19 and a normal distribution for the distribution from onset to isolation
with a mean of 0 days and a standard deviation of 2 days. The dashed line
represents the epidemic threshold, RE = 1. The blue area indicates the region
below the epidemic threshold (namely, controlled epidemic), and the red area
indicates the region above the epidemic threshold. (H) Same as in (G) but
assuming R0 = 1.57 (a more optimistic estimate of R0 in Wuhan, adjusted for
reporting changes) and a normal distribution for the distribution from onset to
isolation with a mean of 2 days and a standard deviation of 2 days.
were isolated after symptom onset [median
time from onset to isolation, 5.4 days (IQR,
2.7 to 8.2 days); fig. S8A]. By contrast, in phase
III, 66% of cases were detected through active
contact tracing, which shortened the median
time from onset to isolation to −0.1 days (IQR,
−2.9 to 1.8 days; fig. S8C). Timeliness of iso-
lation is intermediate in phase II. We use
mathematical models (detailed in the mate-
rials and methods) to dynamically adjust the
serial interval distribution for censoring, and
we apply the same approach to the time in-
terval between symptom onset of a primary
case and onward transmission (fig. S10). These
censoring-adjusted distributions can be rescaled
by the basic reproduction number R0 to reflect
the risk of transmission of a typical SARS-CoV-
2 case since the time of infection or since
symptom onset (Fig. 3, D and E). We find that in
the absence of interventions, infectiousness peaks
near the time of symptom onset (fig. S10D).
Sun et al., Science 371, eabe2424 (2021)
15 January 2021
5 of 8
RES EARCH | R E S E A R C H A R T I C L E
This is consistent with our regression anal-
ysis, where the higher risk of transmission is
near symptom onset (Table 1).
Evaluating the impact of individual- and
population-based interventions on
SARS-CoV-2 transmission
Next, we use the estimated infectiousness
profile of a typical SARS-CoV-2 infection (Fig.
3, D and E) to evaluate the impact of case iso-
lation on transmission. We first set a baseline
reproduction number R0 for SARS-CoV-2 in
the absence of control. Results from a recent
study (33) suggest that the initial growth
rate in Wuhan was 0.15 day−1 in raw case
data (95% CI: 0.14 to 0.17), although the
growth rate could be substantially lower
(0.08 day−1) if changes in case definition are
considered. Conservatively, we consider the
upper value of the growth rate at 0.15 day−1
together with our generation interval dis-
tribution adjusted for censoring (fig. S10C)
to estimate R0. We obtain a baseline repro-
duction number R0 = 2.19 (95% CI: 2.08 to
2.36) using the renewal equation framework
(34). This represents a typical scenario of
unmitigated SARS-CoV-2 transmissibility in
an urban setting. The reconstructed infec-
tiousness profile in the absence of control is
shown in solid red lines in Fig. 3, D and E,
with respect to time of infection and symptom
onset, respectively. Notably, we find that SARS-
CoV-2 infectiousness peaks slightly before
symptom onset (−0.1 days on average), with
87% of the overall infectiousness concentrated
within ±5 days of symptom onset and 53% of
the overall infectiousness in the presympto-
matic phase (Fig. 3E).
Next, we evaluate the impact of case isola-
tion on transmission by considering three dif-
ferent intervention scenarios mimicking the
timeliness of isolation in the three phases of
the Hunan epidemic control. We further as-
sume that 100% of infections are detected and
isolated and that isolation is fully protective
(i.e., there is no onward transmission after the
patient has been isolated or quarantined). The
infectiousness profiles of the three interven-
tion scenarios are shown in dashed lines in
Fig. 3, D and E. We find that the basic re-
production number decreases in all interven-
tion scenarios, but the projected decrease is
not sufficient to interrupt transmission (Fig.
¼ 1:46 for phase
¼ 1:75 for phase I, RE
3D; RE
0
0
II, and RE
¼ 1:01 for phase III, where RE
0 is
0
the effective basic reproduction number).
We further relax the assumption of 100%
case detection and isolation and relate changes
in the basic reproduction number to two inde-
pendent parameters measuring the strength of
interventions: the effectiveness of case isolation
and contact quarantine (measured as the frac-
tion of total infections isolated) and the time-
liness of isolation (measured as the delay from
symptom onset to isolation; phase diagram in
Fig. 3F). Dashed lines in Fig. 3F illustrate 30,
40, and 50% of reduction in R0. To reduce R0
by half (the minimum amount of transmission
reduction required to achieve control for a
baseline R0 ~ 2), 100% of infections would
need to be isolated even if individuals were
isolated as early as the day of symptom onset.
In practice, epidemic control is unrealistic if
case isolation and quarantine of close contacts
are the only measures in place.
Our data support the idea that case isolation
and quarantine of close contacts are effective
in reducing SARS-CoV-2 transmission, espe-
cially if these interventions occur early in the
infection. To achieve epidemic control, how-
ever, these interventions need to be layered
with additional population-level measures,
including increased teleworking, reduced
operation in the service industry, or broader
adoption of face masks. The synergistic effects
of these interventions are illustrated in Fig.
3G. We find that a 30% reduction in transmis-
sion from population-level measures would
require a 70% case detection rate to achieve
epidemic control, assuming that cases can be
promptly isolated on average upon symptom
presentation. Notably, a 30% reduction in
transmission could also encompass the benefits
of residual population-level immunity from the
first wave of COVID-19, especially in hard-hit
regions (35, 36). As a sensitivity analysis, we
further consider a more optimistic scenario
with a lower baseline R0 = 1.56, correspond-
ing to an epidemic growth rate of 0.08 day−1
(95% CI: 0.06 to 0.10) in Wuhan (33), which is
adjusted for reporting changes. As expected,
control is much easier to achieve in this
scenario: If detected SARS-CoV-2 infections
are effectively isolated on average 2 days after
symptom onset, a 25% population-level re-
duction in transmission coupled with a 42%
infection isolation rate is sufficient to achieve
control (Fig. 3H).
Discussion
Detailed information on 1178 SARS-CoV-2–
infected individuals along with their 15,648
contacts allowed us to dissect the behavioral
and clinical drivers of SARS-CoV-2 transmis-
sion, to evaluate how transmission oppor-
tunities are modulated by individual- and
population-level interventions, and to char-
acterize the typical infectiousness profile of a
case. Informed by this understanding, par-
ticularly the importance of presymptomatic
transmission, we have evaluated the plausibil-
ity of SARS-CoV-2 control through individual-
and population-based interventions.
Health care contacts posed the lowest risk of
transmission in Hunan, which suggests that
adequate protective measures against SARS-
CoV-2 were taken in hospitals and medical
observation centers (Table 1). The average risk
of transmission scales positively with the
closeness of social interactions: The average
per-contact risk is lowest for community ex-
posures (including contacts in the public trans-
portation system and at food and entertainment
venues), intermediate for social and extended
family contacts, and highest in the household.
The average transmission risk in the house-
hold is further elevated when intense physical
distancing is enforced, and the risk is also
elevated for contacts that last longer. These
lines of evidence support the idea that SARS-
CoV-2 transmission is facilitated by close
proximity, confined environment, and high
frequency of contacts.
Regression analysis indicates a higher risk
of transmission when an individual is exposed
to a SARS-CoV-2 patient around the time of
symptom onset, in line with our reconstructed
infectiousness profile. These epidemiological
findings are in agreement with viral shedding
studies (6, 37–40). We estimate that overall in
Hunan, 63% of all transmission events were
from presymptomatic individuals, in concor-
dance with other modeling studies (6, 7, 10, 12, 41).
However, the estimated presymptomatic pro-
portion is affected by case-based measures,
including case isolation and contact quaran-
tine. We estimate that the relative contribution
of presymptomatic transmission drops to 52%
in an uncontrolled scenario where case-based
interventions are absent.
Case isolation reduces the effective infec-
tious period of SARS-CoV-2–infected individ-
uals by blocking contacts with susceptible
individuals. We observe that faster isolation
significantly reduces CCRs across contact types
(Fig. 2E). We also observe shorter serial and
generation intervals and a larger fraction of
presymptomatic transmission when individ-
uals are isolated faster (Fig. 3, A to C). By
contrast, population-level physical distancing
measures have differential impacts on CCRs—
decreasing CCRs for social and community
contacts, while increasing CCRs in the house-
hold and family contacts. As a result, strict
physical distancing confines the epidemic
mostly to families and households (see also
fig. S7). The precise impact of physical dis-
tancing on transmission is difficult to separate
from that of individual-based interventions.
However, our analysis suggests that physical
distancing changes the topology of the trans-
mission network by affecting the number and
duration of interactions. Notably, the topology
of the household contact network is highly
clustered (42), and theoretical studies have
shown that high clustering hinders epidemic
spread (43, 44). These higher-order topological
changes could contribute to reducing trans-
mission beyond the effects expected from an
overall reduction in CCRs. In parallel, the ef-
fectiveness of physical distancing measures
on reducing COVID-19 transmission has been
Sun et al., Science 371, eabe2424 (2021)
15 January 2021
6 of 8
RES EARCH | R E S E A R C H A R T I C L E
demonstrated in empirical data from China
(24, 45) and elsewhere (46).
We have explored the feasibility of SARS-
CoV-2 epidemic control against two important
metrics related to case isolation and contact
quarantine: the timeliness of isolation and the
infection detection rate (Fig. 3F). For a base-
line transmission scenario compatible with the
initial growth phase of the epidemic in Wuhan,
we find that epidemic control solely relying on
isolation and quarantine is difficult to achieve.
Layering these interventions with moderate
physical distancing makes control more likely
over a range of plausible parameters—a situa-
tion that could be further improved by residual
immunity from the first wave of SARS-CoV-2
activity (35, 36). Successful implementation of
contact tracing requires a low level of active in-
fections in the community, as the number of
contacts to be monitored is several folds the
number of infections (~13 contacts were traced
for each SARS-CoV-2–infected individual in
Hunan). The timing of easing of lockdown
measures should align with the capacities of
testing and contact-tracing efforts relative to
the number of active infections in the commu-
nity. In parallel, technology-based approaches
can also facilitate these efforts (7, 47).
Overall, we find that case isolation and
quarantine successfully blocked transmission
to close contacts in Hunan, with an estimated
4.3% of transmission occurring after SARS-
CoV-2 patients were isolated. In this setting,
all SARS-CoV-2 infections were managed under
medical isolation in dedicated hospitals re-
gardless of clinical severity, and contacts were
quarantined in designated medical observation
centers. Self-regulated isolation and quarantine
at home, however, may not be as effective, and
a higher proportion of onward transmission
should be expected.
Several caveats are worth noting. We could
not evaluate the risk of transmission in schools,
workplaces, conferences, prisons, or factories,
as no contacts in these settings were reported
in the Hunan dataset. Our study is likely un-
derpowered to assess the transmission poten-
tial of asymptomatic individuals given the
relatively small fraction of these infections in
our data (13.5% overall and 22.1% of infections
captured through contact tracing). There is no
statistical support for decreased transmission
from asymptomatic individuals (fig. S3A), al-
though we observe a positive, but nonsignificant,
gradient in average transmission risk with
disease severity. Evidence from viral shedding
studies is conflicted; viral load appears to be
independent of clinical severity in some studies
(6, 22, 38, 48), whereas others suggest that
there is faster viral clearance in asymptomatic
individuals (49).
Another limitation relates to changes in
testing practices for contacts of primary cases.
Testing was initially limited to contacts exhibit-
ing symptoms, and this condition was relaxed
after 7 February. The early testing scheme may
lead to underestimation of susceptibility in
children, as younger individuals are less likely
to develop SARS-CoV-2 symptoms (50). How-
ever, sensitivity analyses indicate that the age
gradient of susceptibility is preserved even after
stratification for changes in testing protocol.
Further, our finding of lower susceptibility to
infection among children under 12 years of age
relative to adults remains stable in the period
with comprehensive testing (fig. S4). Overall,
the contribution of asymptomatic infections
to transmission remains debated but has pro-
found implications on the feasibility of control
through individual-based interventions. Careful
serological studies combined with virologic test-
ing in households and other controlled environ-
ments are needed to fully resolve the role of
asymptomatic infections and viral shedding
on transmission.
Detailed contact-tracing data illuminate
heterogeneities in SARS-CoV-2 transmission
driven by biology and behavior and modu-
lated by the impact of interventions. Notably,
and in contrast to SARS-CoV-1, the ability of
SARS-CoV-2 to transmit during the host’s
presymptomatic phase makes it particularly
difficult to achieve epidemic control (51). Our
risk factor estimates can provide evidence to
guide the design of more-targeted and sus-
tainable mitigation strategies, and our recon-
structed transmission kinetics will help calibrate
further modeling efforts.
Materials and methods summary
We combined individual-level data on 1178
SARS-CoV-2 infections with detailed diaries
of exposures collected through contact-tracing
efforts in Hunan, China, to stochastically re-
construct transmission chains and infer infec-
tion times. Reconstructed transmission chains
had to be compatible with highly resolved
individual-level data on symptom onset dates,
daily records of exposure to infected contacts,
and travel history to high-risk regions. On
the basis of the reconstructed transmission
chains, we characterize the distribution of
key SARS-CoV-2 transmission parameters—
including the number of secondary cases,
the generation and serial intervals, and the
interval from infection or symptom onset to
isolation—at different stages of the epidemic.
To further understand the drivers of transmis-
sion heterogeneity and the dispersion in the
number of secondary cases, we study the
degree distribution of SARS-CoV-2–infected
individuals, the duration of exposures, and the
age-specific contact patterns between infec-
tors and infectees, separately by contact type
(household, family, community, transportation,
and health care). We also use logistic regres-
sion analysis to model the per-contact risk of
transmission, with contact type and duration,
symptoms, demographic factors, and different
periods of the outbreak as covariates. Missing
data are addressed through multivariate im-
putation algorithms. We conduct sensitivity
analyses to test the robustness of regression
results.
We next use our data to model the synergistic
effects of case-based and population-level inter-
ventions on transmission. We reconstruct the
average infectiousness profile of a SARS-CoV-2
infection, after adjusting for the truncation
effects of case isolation. On the basis of the
estimated infectiousness profile, we use math-
ematical models to estimate the effect of layered
interventions on transmission (measured as
changes in the effective reproduction number).
We consider different intensities of population-
level physical distancing, case detection, and
timeliness of isolation or quarantine. A full
description of the materials and methods is
provided in the supplementary materials.
REFERENCES AND NOTES
2.
1. W. J. Guan et al., 2019 in China. N. Engl. J. Med. 382,
1708–1720 (2020). doi: 10.1056/NEJMoa2002032;
pmid: 32109013
J. T. Wu et al., Estimating clinical severity of COVID-19 from
the transmission dynamics in Wuhan, China. Nat. Med. 26,
506–510 (2020). doi: 10.1038/s41591-020-0822-7;
pmid: 32284616
3. G. Grasselli et al., Baseline Characteristics and Outcomes of
1591 Patients Infected With SARS-CoV-2 Admitted to ICUs of
the Lombardy Region, Italy. JAMA 323, 1574–1581 (2020).
doi: 10.1001/jama.2020.5394; pmid: 32250385
4. R. Verity et al., Estimates of the severity of coronavirus disease
2019: A model-based analysis. Lancet Infect. Dis. 20, 669–677
(2020). doi: 10.1016/S1473-3099(20)30243-7;
pmid: 32240634
5. K. Sun, J. Chen, C. Viboud, Early epidemiological analysis of
the coronavirus disease 2019 outbreak based on crowdsourced
data: A population-level observational study. Lancet Digit. Heal.
2, e201–e208 (2020). doi: 10.1016/S2589-7500(20)30026-1;
pmid: 32309796
6. X. He et al., Temporal dynamics in viral shedding and
transmissibility of COVID-19. Nat. Med. 26, 672–675 (2020).
doi: 10.1038/s41591-020-0869-5; pmid: 32296168
7. L. Ferretti et al., Quantifying SARS-CoV-2 transmission
suggests epidemic control with digital contact tracing.
Science 368, eabb6936 (2020). doi: 10.1126/science.abb6936;
pmid: 32234805
8. A. Kimball et al., Asymptomatic and presymptomatic
SARS-COV-2 infections in residents of a long-term care
skilled nursing facility - King County, Washington, March 2020.
MMWR Morb. Mortal. Wkly. Rep. 69, 377–381 (2020).
doi: 10.15585/mmwr.mm6913e1; pmid: 32240128
9. W. E. Wei et al., Presymptomatic Transmission of SARS-CoV-2 -
Singapore, January 23-March 16, 2020. MMWR Morb. Mortal.
Wkly. Rep. 69, 411–415 (2020). doi: 10.15585/mmwr.mm6914e1;
pmid: 32271722
10. Z. Du et al., Serial Interval of COVID-19 among Publicly
Reported Confirmed Cases. Emerg. Infect. Dis. 26, 1341–1343
(2020). doi: 10.3201/eid2606.200357; pmid: 32191173
11. E. Lavezzo et al., Suppression of a SARS-CoV-2 outbreak in the
Italian municipality of Vo’. Nature 584, 425–429 (2020).
doi: 10.1038/s41586-020-2488-1; pmid: 32604404
12. S. Hu, W. Wang, Y. Wang, M. Litvinova, K. Luo, L. Ren, Q. Sun,
X. Chen, G. Zeng, J. Li, L. Liang, Z. Deng, W. Zheng, M. Li,
H. Yang, J. Guo, K. Wang, X. Chen, Z. Liu, H. Yan, H. Shi,
Z. Chen, Y. Zhou, K. Sun, A. Vespignani, C. Viboud, L. Gao,
M. Ajelli, H. Yu, Infectivity, susceptibility, and risk factors
associated with SARS-CoV-2 transmission under intensive
contact tracing in Hunan, China. medRxiv
2020.07.23.20160317 [Preprint]. 3 November 2020. .
doi: 10.1101/2020.07.23.20160317
13. Q. Bi et al., Epidemiology and transmission of COVID-19 in 391
cases and 1286 of their close contacts in Shenzhen, China:
A retrospective cohort study. Lancet Infect. Dis. 20, 911–919
Sun et al., Science 371, eabe2424 (2021)
15 January 2021
7 of 8
RES EARCH | R E S E A R C H A R T I C L E
(2020). doi: 10.1016/S1473-3099(20)30287-5;
pmid: 32353347
14. A. Pan et al., Association of Public Health Interventions With
the Epidemiology of the COVID-19 Outbreak in Wuhan, China.
JAMA 323, 1915–1923 (2020). doi: 10.1001/jama.2020.6130;
pmid: 32275295
15. B. J. Cowling et al., Impact assessment of non-pharmaceutical
interventions against coronavirus disease 2019 and influenza in
Hong Kong: An observational study. Lancet Public Health 5,
e279–e288 (2020). doi: 10.1016/S2468-2667(20)30090-6;
pmid: 32311320
16. S. W. Park, K. Sun, C. Viboud, B. T. Grenfell, J. Dushoff,
Potential role of social distancing in mitigating spread of
coronavirus disease, South Korea. Emerg. Infect. Dis. 26,
2697–2700 (2020). doi: 10.3201/eid2611.201099;
pmid: 32795385
17. A. Aleta et al., Modeling the impact of social distancing testing
contact tracing and household quarantine on second-wave
scenarios of the COVID-19 epidemic. Nat. Hum. Behav. 4,
964–971 (2020). doi: 10.1038/s41562-020-0931-9;
pmid: 32759985
18. J. Zhang et al., Evolving epidemiology and transmission
dynamics of coronavirus disease 2019 outside Hubei province,
China: A descriptive and modelling study. Lancet Infect. Dis.
20, 793–802 (2020). doi: 10.1016/S1473-3099(20)30230-9;
pmid: 32247326
19. K. Leung, J. T. Wu, D. Liu, G. M. Leung, First-wave COVID-19
transmissibility and severity in China outside Hubei after
control measures, and second-wave scenario planning: A
modelling impact assessment. Lancet 395, 1382–1393 (2020).
doi: 10.1016/S0140-6736(20)30746-7; pmid: 32277878
20. M. Chinazzi et al., The effect of travel restrictions on the
spread of the 2019 novel coronavirus (COVID-19) outbreak.
Science 368, 395–400 (2020). doi: 10.1126/science.aba9757;
pmid: 32144116
21. H. Tian et al., An investigation of transmission control
measures during the first 50 days of the COVID-19 epidemic in
China. Science 368, 638–642 (2020). doi: 10.1126/science.
abb6105; pmid: 32234804
22. D. Cereda, M. Tirani, F. Rovida, V. Demicheli, M. Ajelli,
P. Poletti, F. Trentini, G. Guzzetta, V. Marziano, A. Barone,
M. Magoni, S. Deandrea, G. Diurno, M. Lombardo, M. Faccini,
A. Pan, R. Bruno, E. Pariani, G. Grasselli, A. Piatti, M. Gramegna,
F. Baldanti, A. Melegaro, S. Merler, The early phase of the
COVID-19 outbreak in Lombardy, Italy. arXiv:2003.09320
[q-bio.PE] (2020).
33. T. K. Tsang et al., Effect of changing case definitions for
COVID-19 on the epidemic curve and transmission parameters
in mainland China: A modelling study. Lancet Public Health 5,
e289–e296 (2020). doi: 10.1016/S2468-2667(20)30089-X;
pmid: 32330458
34. J. Wallinga, M. Lipsitch, How generation intervals shape the
relationship between growth rates and reproductive numbers.
Proc. Biol. Sci. 274, 599–604 (2007). doi: 10.1098/
rspb.2006.3754; pmid: 17476782
35. E. S. Rosenberg et al., Cumulative incidence and diagnosis of
SARS-CoV-2 infection in New York. Ann. Epidemiol. 48, 23–29.
e4 (2020). doi: 10.1016/j.annepidem.2020.06.004;
pmid: 32648546
36. S. Stringhini et al., Seroprevalence of anti-SARS-CoV-2 IgG
antibodies in Geneva, Switzerland (SEROCoV-POP): A
population-based study. Lancet 396, 313–319 (2020).
doi: 10.1016/S0140-6736(20)31304-0; pmid: 32534626
37. R. Wölfel et al., Virological assessment of hospitalized patients
with COVID-2019. Nature 581, 465–469 (2020). doi: 10.1038/
s41586-020-2196-x; pmid: 32235945
38. K. K. W. To et al., Temporal profiles of viral load in posterior
oropharyngeal saliva samples and serum antibody responses
during infection by SARS-CoV-2: An observational cohort
study. Lancet Infect. Dis. 20, 565–574 (2020). doi: 10.1016/
S1473-3099(20)30196-1; pmid: 32213337
39. J. J. A. Van Kampen, D. A. M. C. Van De Vijver, P. L. A. Fraaij,
B. L. Haagmans, M. M. Lamers, M. Sc, N. Okba,
J. P. C. Van Den Akker, H. Endeman, D. A. M. P. J. Gommers,
J. J. Cornelissen, R. A. S. Hoek, M. M. Van Der Eerden,
D. A. Hesselink, H. J. Metselaar, A. Verbon, J. E. M. De Steenwinkel,
G. I. Aron, E. C. M. Van Gorp, S. Van Boheemen, J. C. Voermans,
C. A. B. Boucher, R. Molenkamp, M. P. G. Koopmans,
C. Geurtsvankessel, A. A. Van Der Eijk, Shedding of infectious virus
in hospitalized patients with coronavirus disease-2019 (COVID-19):
duration and key determinants. medRxiv 2020.06.08.20125310
[Preprint]. 9 June 2020. doi: 10.1101/2020.06.08.20125310
40. A. Singanayagam et al., Duration of infectiousness and
correlation with RT-PCR cycle threshold values in cases of
COVID-19, England, January to May 2020. Euro Surveill. 25,
2001483 (2020). doi: 10.2807/1560-7917.
ES.2020.25.32.2001483; pmid: 32794447
41. Y. Liu, S. Funk, S. Flasche; Centre for Mathematical Modelling
of Infectious Diseases nCoV Working Group, The contribution
of pre-symptomatic infection to the transmission dynamics of
COVID-2019. Wellcome Open Res. 5, 58 (2020). doi: 10.12688/
wellcomeopenres.15788.1; pmid: 32685697
23. A. J. Kucharski et al., Early dynamics of transmission and
42. Q. H. Liu et al., Measurability of the epidemic reproduction
control of COVID-19: A mathematical modelling study. Lancet
Infect. Dis. 20, 553–558 (2020). doi: 10.1016/S1473-3099(20)
30144-4; pmid: 32171059
number in data-driven contact networks. Proc. Natl. Acad. Sci.
U.S.A. 115, 12680–12685 (2018). doi: 10.1073/pnas.1811115115;
pmid: 30463945
24. J. Zhang et al., Changes in contact patterns shape the
43. R. Pastor-Satorras, C. Castellano, P. Van Mieghem,
dynamics of the COVID-19 outbreak in China. Science 368,
1481–1486 (2020). doi: 10.1126/science.abb8001;
pmid: 32350060
25. Baidu.com, Baidu Qianxi: Within city mobility index. https://
qianxi.baidu.com/2020/ (2020).
26. E. Goldstein, M. Lipsitch, M. Cevik, On the effect of age on the
transmission of SARS-CoV-2 in households, schools and the
community. J. Infect. Dis. jiaa691 (2020). doi: 10.1093/infdis/
jiaa691; pmid: 33119738
27. R. M. Viner et al., Susceptibility to SARS-CoV-2 Infection
Among Children and Adolescents Compared With Adults: A
Systematic Review and Meta-analysis. JAMA Pediatr. 10.1001/
jamapediatrics.2020.4573 (2020). doi: 10.1001/
jamapediatrics.2020.4573; pmid: 32975552
28. J. Mossong et al., Social contacts and mixing patterns relevant
to the spread of infectious diseases. PLOS Med. 5, e74 (2008).
doi: 10.1371/journal.pmed.0050074; pmid: 18366252
29. M. Litvinova, Q. H. Liu, E. S. Kulikov, M. Ajelli, Reactive school
closure weakens the network of social interactions and reduces
the spread of influenza. Proc. Natl. Acad. Sci. U.S.A. 116,
13174–13181 (2019). doi: 10.1073/pnas.1821298116;
pmid: 31209042
30. Q. Li et al., Early Transmission Dynamics in Wuhan, China, of Novel
Coronavirus-Infected Pneumonia. N. Engl. J. Med. 382, 1199–1207
(2020). doi: 10.1056/NEJMoa2001316; pmid: 31995857
31. T. Ganyani et al., Estimating the generation interval for
coronavirus disease (COVID-19) based on symptom onset
data, March 2020. Euro Surveill. 25, 2000257 (2020).
doi: 10.2807/1560-7917.ES.2020.25.17.2000257; pmid: 32372755
32. S. T. Ali et al., Serial interval of SARS-CoV-2 was shortened over
time by nonpharmaceutical interventions. Science 369, 1106–1109
(2020). doi: 10.1126/science.abc9004; pmid: 32694200
A. Vespignani, Epidemic processes in complex networks.
Rev. Mod. Phys. 87, 925–979 (2015). doi: 10.1103/
RevModPhys.87.925
44. E. M. Volz, J. C. Miller, A. Galvani, L. Ancel Meyers, Effects of
heterogeneous and clustered contact patterns on infectious
disease dynamics. PLOS Comput. Biol. 7, e1002042 (2011).
doi: 10.1371/journal.pcbi.1002042; pmid: 21673864
45. K. E. C. Ainslie et al., Evidence of initial success for China
exiting COVID-19 social distancing policy after achieving
containment. Wellcome Open Res. 5, 81 (2020). doi: 10.12688/
wellcomeopenres.15843.2; pmid: 32500100
46. S. Flaxman et al., Estimating the effects of non-
pharmaceutical interventions on COVID-19 in Europe.
Nature 584, 257–261 (2020). doi: 10.1038/s41586-020-
2405-7; pmid: 32512579
47. J. Hellewell et al., Feasibility of controlling COVID-19 outbreaks
by isolation of cases and contacts. Lancet Glob. Health 8,
e488–e496 (2020). doi: 10.1016/S2214-109X(20)30074-7;
pmid: 32119825
48. Q.-X. Long et al., Clinical and immunological assessment of
asymptomatic SARS-CoV-2 infections. Nat. Med. 26, 1200–1204
(2020). doi: 10.1038/s41591-020-0965-6; pmid: 32555424
49. N. V. V. Chau et al., The natural history and transmission
potential of asymptomatic severe acute respiratory syndrome
coronavirus 2 infection. Clin. Infect. Dis. 71, 2679–2687 (2020).
doi: 10.1093/cid/ciaa711; pmid: 32497212
50. P. Poletti, M. Tirani, D. Cereda, F. Trentini, G. Guzzetta,
G. Sabatino, V. Marziano, A. Castrofino, F. Grosso, G. Del
Castillo, R. Piccarreta, ATS Lombardy COVID-19 Task Force,
A. Andreassi, A. Melegaro, M. Gramegna, M. Ajelli, S. Merler,
Probability of symptoms and critical disease after SARS-CoV-2
infection. arXiv:2006.08471 [q-bio.PE] (2020).
51. C. Fraser, S. Riley, R. M. Anderson, N. M. Ferguson, Factors
that make an infectious disease outbreak controllable. Proc.
Natl. Acad. Sci. U.S.A. 101, 6146–6151 (2004). doi: 10.1073/
pnas.0307506101; pmid: 15071187
52. Protocol on Prevention and Control of COVID-19 (Edition 6).
Natl. Heal. Comm. People’s Repub. China (2020);
www.nhc.gov.cn/jkj/s3577/202003/
4856d5b0458141fa9f376853224d41d7.shtml.
53. K. Sun, W. Wang, L. Gao, Y. Wang, K. Luo, L. Ren, Z. Zhang,
X. Chen, S. Zhao, Y. Huang, Q. Sun, Z. Liu, M. Litvinova,
A. Vespignani, M. Ajelli, C. Viboud, H. Yu, Code and data
sharing for manuscript titled “Transmission heterogeneities,
kinetics, and controllability of SARS-CoV-2,” Zenodo (2020).
doi: 10.5281/zenodo.4129863
AC KNOWLED GME NTS
The authors acknowledge C. Fraser from the University of Oxford;
D. Spiro from Fogarty International Center, National Institutes
of Health; and P. Kilmarx from Fogarty International Center, National
Institutes of Health for their helpful comments on the manuscript.
This article does not necessarily represent the views of the
NIH or the U.S. government. Funding: H.Y. acknowledges financial
support from the National Science Fund for Distinguished Young
Scholars (no. 81525023), the Key Emergency Project of Shanghai
Science and Technology Committee (no. 20411950100), and
the National Science and Technology Major Project of China
(nos. 2017ZX10103009-005, 2018ZX10713001–007, and
2018ZX10201001–010). L.G. acknowledges financial support from
Hunan Provincial Innovative Construction Special Fund: Emergency
response to COVID-19 outbreak (no. 2020SK3012). The funder
of the study had no role in study design, data collection, data
analysis, data interpretation, or writing of the paper. Author
contributions: C.V. and H.Y. are joint senior authors. K.S., C.V., and
H.Y. designed the experiments. L.G., W.W., and Y.W. collected
data. K.S., W.W., L.G., and Y.W. analyzed the data. K.S., W.W.,
L.G., Y.W., K.L., L.R., Z.Z., X.C., S.Z., Y.H., Q.S., Z.L., M.L., A.V., M.A.,
C.V., and H.Y. interpreted the results. K.S., W.W., L.G., Y.W.,
M.L., A.V., M.A., C.V., and H.Y. wrote the manuscript. Competing
interests: M.A. has received research funding from Seqirus.
H.Y. has received research funding from Sanofi Pasteur,
GlaxoSmithKline, Yichang HEC Changjiang Pharmaceutical
Company, and Shanghai Roche Pharmaceutical Company. A.V.
reports a past grant from Metabiota Inc. that is not related to this
work. None of these awards are related to COVID-19. All other
authors declare no competing interests. Ethics statement: The
collection of specimens and epidemiological and clinical data
for SARS-CoV-2–infected individuals was part of a continuing public
health investigation of an emerging outbreak, defined by the
“Protocol on the Prevention and Control of COVID-19” established
by the National Health Commission of the People’s Republic of China
(52). Thus, collection of epidemiological and clinical data was
exempt from institutional review board assessment. This study was
approved by the Institutional Review Board of Hunan Provincial
Center for Disease Control and Prevention (IRB no. 2020005). Data
were deidentified, and informed consent was waived with completion
of an appropriate institutional form. Data and materials availability:
The original database containing confidential patient information
cannot be made public; however, we have created a synthetic database
to demonstrate the structure of the underlying data. All code, the
mock database, and deidentified data to reproduce all figures in
the main text and the supplementary materials are publicly available
on Zenodo (53). This work is licensed under a Creative Commons
Attribution 4.0 International (CC BY 4.0) license, which permits
unrestricted use, distribution, and reproduction in any medium,
provided the original work is properly cited. To view a copy of
this license, visit https://creativecommons.org/licenses/by/4.0/. This
license does not apply to figures/photos/artwork or other content
included in the article that is credited to a third party; obtain
authorization from the rights holder before using such material.
SUPPLEMENTARY MATERIALS
science.sciencemag.org/content/371/6526/eabe2424/suppl/DC1
Materials and Methods
Figs. S1 to S11
Tables S1 to S5
References (54–61)
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
9 August 2020; accepted 19 November 2020
Published online 24 November 2020
10.1126/science.abe2424
Sun et al., Science 371, eabe2424 (2021)
15 January 2021
8 of 8
|
10.1126_science.abf7470
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
IMMUNOLOGY
Marginal zone B cells acquire dendritic cell
functions by trogocytosis
Patrick Schriek, Alan C. Ching, Nagaraj S. Moily, Jessica Moffat, Lynette Beattie, Thiago M. Steiner,
Laine M. Hosking, Joshua M. Thurman, V. Michael Holers, Satoshi Ishido, Mireille H. Lahoud,
Irina Caminschi, William R. Heath, Justine D. Mintern*, Jose A. Villadangos*
INTRODUCTION: Effective immunity is multi-
layered, requiring the cooperation of various
types of molecules and cells. Some types are
components of the fast-responding innate arm
of the immune system, like the molecules that
constitute the complement system and marginal
zone (MZ) B cells. Other types of molecules and
cells participate in adaptive immune responses
that provide long-term protection. These include
conventional dendritic cells (cDCs) and major
histocompatibility complex class II (MHC II)
molecules. This study describes molecular link-
ages between complement and MHC II mole-
cules that enable MZ B cells and cDCs to carry
out cooperatively immunological functions that
neither cell type can perform on its own.
RATIONALE: The initiation of adaptive immu-
nity against infections requires cDCs to detect,
capture, degrade, and present pathogen anti-
gens. cDCs use their MHC II molecules to bind
and display peptide fragments derived from
these antigens. Recognition of the resulting
pMHC II complexes by the antigen receptor
of T cells elicits adaptive immune responses
and, eventually, the establishment of protec-
tive immunological memory against the infec-
tious agent. MZ B cells are specialized in the
MHC II–C3dg complex formation and ubiquitination at the surface of cDCs
Dendritic cell
C3
tickover
CHO
MHC II
MHC II–C3
MHC II–C3dg
MARCH1
Ub
MZ B cells trogocytose from cDC pMHC II–C3 complexes for antigen presentation to CD4+ T cells
Dendritic
cell
MZ B cell
CR2
pMHC II–C3dg
+
CD4
T cell
Trogocytosis
Trogocytic
MZ B cell
Dendritic cell
death (?)
TCR
Marginal zone B cells trogocytose dendritic cells, acquiring peptide-loaded MHC II molecules bound to
complement C3 for antigen presentation to CD4+ T cells. Activated complement C3 binds MHC II on
conventional dendritic cells (cDCs). The complexes are processed into MHC II–C3dg and either internalized via
MARCH1-mediated ubiquitination or recognized by the complement receptor 2 (CR2) of marginal zone
(MZ) B cells. The latter enables MZ B cells to trogocytose and display on their own membrane cDC receptors.
Trogocytic MZ B cells expand their capacity to stimulate helper CD4+ T cells using antigen-loaded MHC II
molecules generated by cDCs. Excessive trogocytosis eliminates cDCs, but MARCH1 prevents this by limiting
the number of MHC II–C3dg complexes on cDCs.
production of polyreactive antibodies that
protect newborns and infants from different
types of microorganisms. In some instances,
MZ B cells require “help” from T cells to
perform this function, which they obtain by
displaying pMHC II complexes. This suggests
that MZ B cells may be able to emulate the
antigen-presenting activity of cDCs.
RESULTS: Complement component 3 (C3) is an
abundant serum protein that constitutively
adopts a reactive form in the absence of
pathogens by a mechanism known as tickover.
We determined that C3 binds to pMHC II
exposed on the surface of mouse and human
cDCs, forming a covalent bond with the car-
bohydrate moiety of the MHC II a chain. Be-
cause C3 can damage healthy cells, it is converted
to inactive C3dg while still bound to pMHC II.
These pMHC II–C3dg complexes are recog-
nized by complement receptor 2 (CR2), which
is highly expressed by MZ B cells. Interaction
between CR2 and C3dg triggers the transfer of
pMHC II–C3dg complexes, along with associ-
ated cDC membrane and additional proteins
embedded in the membrane, from cDCs to
MZ B cells—a process termed trogocytosis. The
trogocytic MZ B cells are thus able to present
pMHC II complexes to T cells they do not
generate themselves but acquire from cDCs.
Although trogocytosis is beneficial for MZ
B cell function, it must be limited to prevent ex-
cessive damage and elimination of the trogo-
cytosed cDCs. This takes place through an
evolutionarily conserved mechanism, namely
pMHC II–C3dg ubiquitination by a highly spe-
cialized ubiquitin ligase, MARCH1, embedded in
the cDC plasma membrane. The ubiquitinated
pMHC II–C3dg complexes are endocytosed and
degraded intracellularly, reducing the number
exposed on the cDC surface in the steady state.
CONCLUSION: Our results describe how C3 and
MHC II interact and how this interaction
enables MZ B cells and cDCs to cooperatively
carry out functions they cannot perform
individually. We demonstrate how an evolu-
tionarily conserved mechanism for the consti-
tutive elimination of potentially damaging C3
has been co-opted by cDCs to tag pMHC II
complexes for capture by MZ B cells via
trogocytosis. This mechanism expands the
range of antigens that MZ B cells can present to
T lymphocytes. The beneficial and deleterious
consequences of trogocytosis are balanced by
MARCH1 ubiquitination.▪
The list of author affiliations is available in the full article online.
*Corresponding author. Email: jmintern@unimelb.edu.au
(J.D.M.); j.villadangos@unimelb.edu.au (J.A.V.)
Cite this article as P. Schriek et al., Science 375, eabf7470
(2022). DOI: 10.1126/science.abf7470
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.abf7470
Schriek et al., Science 375, 630 (2022)
11 February 2022
1 of 1
RES EARCH
R E S E A R C H A R T I C L E
◥
IMMUNOLOGY
Marginal zone B cells acquire dendritic cell
functions by trogocytosis
Patrick Schriek1, Alan C. Ching1, Nagaraj S. Moily1, Jessica Moffat1, Lynette Beattie2,
Thiago M. Steiner2, Laine M. Hosking3, Joshua M. Thurman4, V. Michael Holers4,
Satoshi Ishido5, Mireille H. Lahoud6, Irina Caminschi6, William R. Heath2,
Justine D. Mintern1*, Jose A. Villadangos1,2*
Marginal zone (MZ) B cells produce broad-spectrum antibodies that protect against infection early
in life. In some instances, antibody production requires MZ B cells to display pathogen antigens
bound to major histocompatibility complex class II (MHC II) molecules to T cells. We describe the
trogocytic acquisition of these molecules from conventional dendritic cells (cDCs). Complement
component 3 (C3) binds to murine and human MHC II on cDCs. MZ B cells recognize C3 with
complement receptor 2 (CR2) and trogocytose the MHC II–C3 complexes, which become exposed on
their cell surface. The ubiquitin ligase MARCH1 limits the number of MHC II–C3 complexes displayed
on cDCs to prevent their elimination through excessive trogocytosis. Capture of C3 by MHC II thus
enables the transfer of cDC-like properties to MZ B cells.
E ffective immunity requires orchestrated
cooperation of multiple molecular and
cellular components to maintain homeosta-
sis and respond to infections. Although
major histocompatibility complex class II
(MHC II) molecules and complement compo-
nent 3 (C3) are ancient centerpieces of adapt-
ive and innate immunity, respectively, no
interaction between these two components
has previously been described.
The primary role of MHC II is to bind
peptides derived from protein antigens (Ag)
encountered by antigen-presenting cells (APCs)
(1). The resulting peptide-loaded MHC II
(pMHC II) complexes are displayed on the
APC plasma membrane and detected by CD4+
T cells, initiating adaptive immune responses.
All APCs ubiquitinate the cytosolic tail of MHC
II using the membrane ubiquitin ligase MARCH1
(membrane-associated RING-CH–type finger 1,
encoded by Marchf1) (2). Ubiquitination re-
duces the surface expression and half-life of
pMHC II complexes by promoting their deliv-
ery to lysosomes, where they are degraded
(2). Both MARCH1 and the single MHC II
1Department of Biochemistry and Pharmacology, Bio21
Molecular Science and Biotechnology Institute, University
of Melbourne, Parkville, VIC 3010, Australia. 2Department
of Microbiology and Immunology, Peter Doherty Institute
for Infection and Immunity, University of Melbourne,
Parkville, VIC 3010, Australia. 3Department of Allergy and
Immunology, Royal Children’s Hospital, Parkville, VIC
3052, Australia. 4Division of Rheumatology, University of
Colorado School of Medicine, Aurora, CO 80045, USA.
5Department of Microbiology, Hyogo College of Medicine,
1-1 Mukogawa-cho, Nishinomiya 663-8501, Japan.
6Monash Biomedicine Discovery Institute and Department
of Biochemistry and Molecular Biology, Monash University,
Clayton, VIC 3800, Australia.
*Corresponding author. Email: jmintern@unimelb.edu.au
(J.D.M.); j.villadangos@unimelb.edu.au (J.A.V.)
b-chain residue ubiquitinated by MARCH1,
Lys225, have been conserved through evolu-
tion, but their role in Ag presentation remains
elusive (2–4), raising questions as to whether
MHC II ubiquitination by MARCH1 plays
other functions.
The complement system comprises >30
soluble and membrane proteins that undergo
a cascade of activation upon pathogen en-
counter (5). Its pivotal component is C3, which
can be activated by the classical, lectin, or
alternative pathways (fig. S1A). This third
pathway occurs at low levels in the absence
of pathogens in what is known as tickover.
Activated C3 binds covalently to carbohydrates
on bacterial cell walls (6). This is followed by
recruitment of other complement components
that mediate lysis or phagocytosis of bacteria
(fig. S1A) (5). C3 can also bind to the plasma
membrane of normal host cells during and,
via tickover, in the absence of infection (6).
Deposited C3 is recognized by surface recep-
tors and serum proteases that cleave it into the
inactive forms C3dg and C3d, which remain
attached to the cell membrane (fig. S1, B and C,
and table S1), thereby preventing cell damage.
Activation of C3 by tickover primes the com-
plement system to respond rapidly to infection,
but whether this pathway plays other immuno-
regulatory roles in the steady state is unclear (7).
Here, we show that C3 activated by tickover
specifically binds the carbohydrate of murine
and human MHC II glycoproteins. The result-
ing complexes are recognized by complement
receptor 2 (CR2), expressed by marginal zone
(MZ) B cells, triggering the trogocytic transfer
of pMHC II and other membrane proteins
from conventional dendritic cells (cDCs) to MZ
B cells. This mechanism enables MZ B cells to
present pMHC II complexes generated by
cDCs. Excessive trogocytosis causes cDC elimi-
nation, but MARCH1 ubiquitination prevents
this outcome by limiting the accumulation of
MHC II–C3 complexes.
March1–/– mice have reduced numbers of
splenic cDCs
Relative to wild-type controls, the spleens of
March1–/– mice exhibited reduced numbers of
the two major cDC subsets, cDC1s and cDC2s,
with no alteration in the number of plasma-
cytoid DCs (pDCs), B cells, or T cells (Fig. 1, A
and B, and fig. S2, A to C). By contrast, cDC
numbers in lymph nodes and the thymus were
not altered (fig. S2D). The expression of char-
acteristic cDC markers was comparable between
wild-type and March1–/– cDCs with the excep-
tion of the MARCH1 substrates MHC II and
CD86 (fig. S2E).
March1–/– mice have enriched numbers of MZ
B cells displaying cDC proteins
A splenic CD11cintCD24+CD8int population
that was present in low numbers in wild-type
mice comprised >20% of CD11c+ cells in their
March1–/– counterparts (Fig. 2A). These cells
displayed several surface markers characteristic
of cDC1s, cDC2s, or both, although mostly at
lower levels than cDCs (Fig. 2Β). They also
expressed B cell molecules at levels similar to
B cells but did not express markers charac-
teristic of other cell populations (Fig. 2Β). No
equivalent cell type was found in lymph nodes
or thymus (fig. S3A). The transcriptome of
CD11cintCD24+CD8int cells was similar to those
of wild-type or March1–/– B cells (Fig. 2C), with
high expression of B cell receptor (BCR)–
signaling and B cell–activation genes (Fig. 2D)
and no expression of cDC genes (Fig. 2E). This
suggested that CD11cintCD24+CD8int cells were
B cells that displayed cDC surface proteins but
did not transcribe the corresponding genes.
Additional immunophenotyping revealed that
the majority of these cells were MZ B cells (8)
(Fig. 2F).
MZ B cells trogocytose plasma membrane
from cDCs in a MARCH1-dependent manner
We tested the hypothesis that MZ B cells dis-
played cDC membrane proteins as a result of
trogocytosis (9–11) and the absence of MARCH1-
promoted plasma membrane transfer between
the two cell types (fig. S3B). Indeed, B cells
incubated with cDCs trogocytosed fluorescently
labeled plasma membrane (Fig. 3A) and surface
receptors (Fig. 3B) from cDCs. Trogocytosis was
more prominent if cDCs were March1–/– than if
they were wild-type, even though the two cDC
groups were similarly labeled with fluorescent
membrane dye (fig. S3C) and expressed similar
levels of the surface receptors acquired by the
B cells (fig. S2E). Membrane transfer between
cDCs and B cells was monodirectional, as little
Schriek et al., Science 375, eabf7470 (2022)
11 February 2022
1 of 12
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Mice deficient in
MARCH1 E3 ubiquitin ligase
have reduced numbers of
splenic cDCs. (A and B) Numbers
of the indicated wild-type (WT)
and March1–/– cell types in whole
splenocytes (A) or low-density
splenocyte preparations (B)
before (left) and after (right)
depletion of non-cDCs. Graphs
display data pooled from three
independent experiments, with
each symbol representing an
individual mouse (n = 2 or 3 per
experiment); bars denote mean ±
SD. ***P < 0.0002, ****P <
0.0001 [independent-samples
t test with Welch’s correction
(no assumption of equal varian-
ces), two-tailed P value (95% CI)];
ns, not significant.
cDC1
***
WT
March1−/−
cDC1
****
60
40
20
0
A
)
4
0
1
×
(
r
e
b
m
u
n
B
)
4
0
1
×
(
r
e
b
m
u
n
50
40
30
20
10
0
whole splenocytes
)
4
0
1
×
(
r
e
b
m
u
n
180
150
120
90
60
30
0
cDC2
***
WT
March1−/−
)
4
0
1
×
(
r
e
b
m
u
n
35
28
21
14
7
0
pDC
ns
WT
March1−/−
120
)
6
0
1
×
(
r
e
b
m
u
n
90
60
30
0
B cell
ns
WT
March1−/−
)
6
0
1
×
(
r
e
b
m
u
n
20
15
10
5
0
CD4+ T cell
ns
WT
March1−/−
)
6
0
1
×
(
r
e
b
m
u
n
15
10
5
0
CD8+ T cell
ns
WT
March1−/−
low-density splenocytes
pDC
ns
cDC2
****
)
4
0
1
×
(
r
e
b
m
u
n
35
28
21
14
7
0
)
4
0
1
×
(
r
e
b
m
u
n
150
120
90
60
30
0
low-density splenocytes
+
neg. selection
cDC2
***
cDC1
****
100
)
4
0
1
×
(
r
e
b
m
u
n
80
60
40
20
0
)
4
0
1
×
(
r
e
b
m
u
n
30
20
10
0
WT
March1−/−
WT
March1−/−
WT
March1−/−
WT
March1−/−
WT
March1−/−
transfer was observed from B cells to cDCs (fig.
S3D). MZ B cells were more trogocytic than
their follicular (FO) counterparts (Fig. 3C). Fur-
thermore, B cells did not trogocytose March1–/–
macrophage, neutrophil, or T cell membranes
(fig. S3E). Thus, B cells—particularly MZ B
cells—specifically acquire cDC plasma mem-
brane and surface proteins via trogocytosis, and
MARCH1 deficiency makes cDC membranes
more susceptible to trogocytic transfer.
MARCH1 controls the amount of C3 that
accumulates on the surface of cDCs
Trogocytosis is mediated by surface receptors
(9), so we hypothesized that MARCH1 regu-
lates the expression of a receptor that medi-
ates trogocytosis. The only receptors known to
increase in expression in March1–/– cDCs are
MHC II and CD86 (fig. S2E) (2), but a recent
plasma membrane proteome analysis showed
that the surface of March1–/– cDCs is also highly
enriched in C3 (12). Because C3 is inaccessible to
cytosolic ubiquitination by MARCH1, we sought
to characterize the mechanism that caused its
accumulation on the cell surface before inves-
tigating its potential role in trogocytosis.
First, we confirmed that C3 is present on
wild-type cDC1s and cDC2s and is overex-
pressed on their March1–/– counterparts (Fig. 4A
and fig. S4A). Analysis of C3-deficient mice
demonstrated the specificity of this detection
(Fig. 4A). MARCH1 is active in all hemato-
poietic APCs, where it keeps surface MHC II
expression at intermediate to low levels (12).
We observed an increase in C3 deposition on
both professional and “atypical” APCs from
the spleen, lymph nodes, and thymus but not
on T cells (Fig. 4B). C3 binding to cDCs was
not caused by enzymatic tissue digestion be-
cause it was also observed in cell suspensions
prepared by mechanical disruption (fig. S4B).
Furthermore, when wild-type and March1–/– ×
C3–/– spleens were pooled before cell purifica-
tion, the mutant cDCs remained negative for
C3 expression (fig. S4C).
In mixed–bone marrow (BM) chimeric mice
where 1:1 wild-type and March1–/– BM was
used to reconstitute C3–/– recipients, neither
wild-type nor March1–/– cDCs displayed C3
(Fig. 4C). This indicated that C3 was captured
from the extracellular environment, as expected,
because it is produced mainly by liver cells (13).
If the recipient chimeric mice were wild-type,
the cDCs generated from the March1–/– BM
displayed higher surface expression of C3 than
their wild-type counterparts (Fig. 4C), which
implies that the effect of the mutation was cell-
intrinsic. Thus, C3 is secreted into circulation by
nonhematopoietic cells and is deposited on all
APCs. If the cells do not express MARCH1, C3
deposition increases by as much as a factor of
20 in the case of cDC1s.
C3 activated by tickover binds to MHC II
Proteomic analysis of the March1–/– cDC plasma
membrane (12) (table S2) indicated that the
C3 species found on the cDC surface is one
or both of its inactivated forms (i.e., C3dg or
C3d) (fig. S1B and table S1). Analysis of wild-
type, March1–/–, and C3–/– cDC lysates by im-
munoblot identified several molecular species
of C3 that were only detected in March1–/–
cDCs (Fig. 5A, fig. S1B, and table S1). The most
prominent species had a molecular weight of
~70 kDa and was absent from the serum. We
hypothesized that this species corresponded to
C3dg or C3d covalently bound to MHC IIa or b.
The cDCs of mice deficient in both MARCH1
and MHC II (March1–/– × H2-Aa–/–) showed
elevated CD86 expression (fig. S4D), indicat-
ing that the absence of MHC II did not prevent
surface accumulation of MARCH1 substrates.
However, C3 was barely detectable on the sur-
face of these cells (Fig. 5B) or on cDCs that only
lacked MHC II expression (fig. S4E). Further-
more, the cDCs of knock-in mice—in which
the only MHC II ubiquitination site, Lys225
of the b chain, has been replaced with Arg
(MHC IIKRKI/KI mice) (3, 14)—expressed sim-
ilarly high levels of C3 relative to March1–/–
Schriek et al., Science 375, eabf7470 (2022)
11 February 2022
2 of 12
RES EARCH | R E S E A R C H A R T I C L E
/
i
i
l
-
-
0
0
-2
0
6
-6
0.0
0.5
1.7
1.9
2.1
10
30
40
20
30
20
10
-0.5
-1.0
22.3
IgD
A
B
C
IgM
D
CD4
CD8
CD3
8
D
C
8
D
C
cDC1
T
W
FLT3
B220
Ly6G
Sirpα
B cell
B cell
cDC1
CD24
CD24
CD86
CD19
CD54
F4/80
WT
XCR1
logFC
BST-2
CD11c
CD11c
CD11c
A
C
S
F
A
C
S
F
CD11b
MHC II
CD62L
Clec9A
cDC
f
o
%
g
n
d
a
e
L
Siglec H
DEC205
CD11c+
24.6
CD11c+
43.4
B cell
cDC1
cDC2
)
5
0
1
×
(
2
m
d
C
F
g
o
****
****
6510 genes
r
e
b
m
u
n
+
c
1
1
D
C
−
−
1
h
c
r
a
M
March1−/−
CD11cintCD24+CD8int
B cell activation
March1+/+
March1−/−
March1−/−
March1+/+
BCR signaling pathway
pDC,mac,
T cell,
neutrophil
CD11cintCD24+CD8int
FMO March1−/− B cells March1−/− CD11cintCD24+CD8int March1−/− cDC(1/2)
March1−/− pDC (BST2-, Siglec H), mac (F4/80), T cell (CD3), neutrophlil (Ly6G)
WT cDC1
March1−/− cDC1
WT B cells
March1−/− B cells
March1−/− CD11cintCD24+CD8int
Fig. 2. MARCH1-
deficient mice
harbor MZ B cells
expressing cDC
surface proteins.
(A) Left: Representa-
tive flow cytometry
plots of CD24 versus
CD8 expression in
CD11c+ wild-type and
March1–/– low-density
splenocytes. Right:
Frequencies and
total numbers of
CD11cintCD24+CD8int
cells. Graphs display
data pooled from
three independent
experiments, with each
symbol representing
an individual mouse
(n = 2 or 3 per experi-
ment); bars denote
mean ± SD. ****P <
0.0001 [independent-
samples t test with
Welch’s correction
(no assumption
of equal variances),
two-tailed P value
(95% CI)]. (B) Flow
cytometry analysis
of the indicated surface
molecules on splenic
CD11cintCD24+CD8int
cells (blue), cDCs
(green), or B cells (red)
of March1–/– mice.
Histograms are repre-
sentative of at least
two independent
experiments with
two or three individual
mice per experiment.
(C to E) RNA-seq
of sort-purified
CD11cintCD24+CD8int
cells from March1–/–
mice and B cells
and cDC1s from wild-
type and March1–/–
mice. (C) Top: Heat-
map showing 6510 dif-
ferentially expressed
genes across the
five groups. Bottom:
Two-dimensional scaling
plot of the top 500 differentially expressed genes. (D) Barcode plots showing enrichment of genes in the B cell receptor signaling pathway (BioCarta) (left) and B cell
activation (GO:0042113) (right) based on gene set analysis comparing March1–/– CD11cintCD24+CD8int cells and cDC1s. (E) Reads per kilobase million (RPKM) for
genes encoding characteristic B cell (top) or cDC1 (bottom) surface markers in March1–/– cDC1s, B cells, and CD11cintCD24+CD8int cells. For RNA-seq analysis in (C) to
(E), mRNAs from sort-purified cells of four biological replicates with pooled spleens from five mice were sequenced in technical replicates. (F) Representative gating
strategy for the identification of splenic follicular (FO), MZ, and progenitor MZ (MZP) B cells, based on the surface expression of B220, CD93, IgM, CD21/35, and
CD23 (all gated on CD19+ alive cells), in whole splenocyte preparations of wild-type and March1–/– mice (top and middle rows) or among the CD11cintCD24+CD8int cells
from low-density splenocytes of March1–/– mice (bottom). Data are from at least two independent experiments with two or three individual mice per experiment.
CD11cintCD24+CD8int
CD11cintCD24+CD8int
CD11cintCD24+CD8int
CD11cintCD24+CD8int
CD11cintCD24+CD8int
CD11cintCD24+CD8int
CD11cintCD24+CD8int
CD11cintCD24+CD8int
March1−/− CD11cintCD24+CD8int
WT B cell
MZ
9.28
-1
Leading logFC dim1
March1−/− B cell
t
n
e
m
h
c
i
r
n
E
0
1
C
D
c
−
−
1
h
c
r
a
M
t
n
e
m
h
c
i
r
n
E
0
1
C
D
c
−
−
1
h
c
r
a
M
cDC1
cDC1
cDC1
cDC1
cDC1
cDC1
cDC1
cDC1
B cell
B cell
B cell
B cell
B cell
B cell
B cell
B cell
M
K
P
R
M
K
P
R
CD93/AA4
CD93/AA4
CD93/AA4
CD93−
16.5
CD93−
5.47
CD93−
26.5
CD93+
94.3
CD93+
73.2
CD93+
83.2
Cd19
Cd8a
5
3
/
1
2
D
C
5
3
/
1
2
D
C
5
3
/
1
2
D
C
Ighm
Itgax
t
n
8
D
C
+
4
2
D
C
t
n
8
D
C
+
4
2
D
C
Xcr1
Ighd
MZP
14.3
MZP
34.1
MZP
34.4
FO
83.8
FO
74.1
MZ
18.0
MZ
85.3
FO
22.1
MZ
63.5
MZ
63.8
MZ
72.7
Flt3
Cr2
−
−
1
h
c
r
a
M
−
−
1
h
c
r
a
M
CD23
CD23
CD23
2000
3000
4000
1000
5000
1000
0
2
2
B
0
2
2
B
0
2
2
B
0
2
2
B
0
2
2
B
0
2
2
B
t
n
c
1
1
D
C
t
n
c
1
1
D
C
200
300
100
400
100
150
500
200
100
150
200
600
400
600
400
200
200
100
800
100
200
300
300
IgM
IgM
IgM
E
F
50
50
2
7
.
0
1
-
0
0
.
9
-
8
8
.
0
-
8
1
.
2
-
7
1
.
0
-
7
1
.
0
-
3
4
.
0
-
9
1
.
2
-
9
8
.
0
-
3
4
.
0
-
3
0
.
0
0
3
.
1
3
0
.
0
9
2
.
7
7
6
.
0
1
2
.
0
6
4
.
7
1
3
.
1
7
6
.
0
1
4
.
0
0
2
.
0
0
4
.
0
0
0
0
0
0
0
0
0
2
1
0
3
4
/
/
/
/
i
i
i
i
Schriek et al., Science 375, eabf7470 (2022)
11 February 2022
3 of 12
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. MZ B cells trogocytose
cDC plasma membrane in vitro.
(A) Left: Trogocytic acquisition of
cDC membrane, fluorescently
labeled with PKH26, by wild-type
and March1–/– B cells after in vitro
incubation with PKH26-stained
wild-type or March1–/– cDCs.
Right: Frequency of PKH26+
B cells and the mean fluorescence
intensity (MFI) value of their
PKH26 fluorescence after cocul-
turing. (B) As in (A), but
measuring the indicated cDC
proteins. (C) As in (B), but dis-
playing separately FO and MZ
B cells (as identified in Fig. 2F).
Graphs in (A) and (B) display data
pooled from two independent
experiments, with each data point
(n = 2 or 3 per experiment)
representing a measurement of a
technical replicate; bars denote
mean ± SD. ***P < 0.0002,
****P < 0.0001 [Welch’s analysis
of variance (ANOVA) test (no
assumption of equal variances)
followed by pairwise comparison
(A) or by Games-Howell multiple-
comparisons test (B), adjusted
P value (95% CI)]. Plots in (C) are
representative of at least two
independent experiments with two
or three individual mice per
experiment.
****
****
****
ns
80
60
40
20
0
WT
March1−/−
WT
March1−/−
WT
March1−/−
A
WT
B cell
no
cDC
WT
cDC
March1−/−
cDC
0.22
37.2
69.8
s
l
l
e
c
B
0.33
37.1
60.2
f
o
%
March1−/−
B cell
6
2
H
K
P
B220
B
8
D
C
c
1
1
D
C
1
R
C
X
b
1
1
D
C
no
cDC
WT
cDC
March1−/−
cDC
0.56
6.02
24.9
B220
0.35
2.91
11.9
f
o
%
s
l
l
e
c
B
B220
0.45
4.16
9.05
B220
1.62
11.5
11.9
B220
s
l
l
e
c
B
f
o
%
B cell
CD11c+
B cell
***
CD8+
B cell
****
15
12
9
6
3
30
24
18
12
6
0
March1−/− cDC
WT cDC
no cDC
0
March1−/− cDC
WT cDC
no cDC
CD11b+
B cell
ns
XCR1+
B cell
***
12
9
6
3
15
12
9
6
3
0
March1−/− cDC
WT cDC
no cDC
0
March1−/− cDC
WT cDC
no cDC
10000
)
I
F
M
(
6
2
H
K
P
8000
6000
4000
2000
0
****
ns
****
ns
WT
March1−/−
WT
March1−/−
B cell
co-cultured with
PKH26+:
no cDC
WT cDC
March1−/− cDC
C
FO B cell
MZ B cell
33.5
97.9
8
D
C
c
1
1
D
C
1
R
C
X
B220
8.66
66.5
B220
0.70
23.4
B220
cDCs (Fig. 5B), even though MARCH1 is
functional in these cells as shown by their
wild-type expression of CD86 (fig. S4F).
To confirm the formation of MHC II–C3
complexes in cDCs, we immunoprecipitated
MHC II or C3 from wild-type, March1–/–,
March1–/– × C3–/–, MHC IIKRKI/KI, C3–/–, and
H2-Aa–/– cDCs and detected MHC IIa, MHC
IIb, and C3 by immunoblot. Two ~70-kDa
proteins recognized by anti–I-Aa and anti-C3
but not by anti–I-Ab antibodies were identi-
fied in immunoprecipitates from March1–/–
and MHC IIKRKI/KI cells, but not from C3–/– cells
(Fig. 5, C and D). Mass spectrometry analysis
of this protein confirmed that it contained
C3d/C3dg peptides (table S3). Accumulation
of C3 on cDC1s was accompanied by increased
expression of complement regulators involved
in conversion of C3b to C3dg, namely CR1/
CR2, complement decay-accelerating factor
(DAF), and factor H (fig. S4G). Thus, activated
C3 binds to I-Aa; it is then processed, generat-
ing I-Aa–C3dg and I-Aa–C3d complexes,
which accumulate on the surface of APCs
(fig. S1D). For simplicity, we henceforth use
the term “C3dg” to refer to both C3d and C3dg
isoforms.
A comparison of C3 levels on the surface of
cDCs expressing wild-type, MHC II K225R
mutant molecule, or no MHC II at all (H2-Aa–/–)
indicated that virtually all C3 was associated
with MHC II (Fig. 5B and fig. S4E). This
suggested that some feature in the MHC IIa
glycoprotein made it a target for activated C3.
Activated C3 displays little protein conforma-
tion specificity but reacts preferentially with
mannose (6), so we explored the possibility
that C3dg might be bound to the MHC IIa
carbohydrate. Indeed, when MHC II immu-
noprecipitated from March1–/– cDCs was
deglycosylated with PNGase F, the MHC IIa–
C3dg complex was not detectable (Fig. 5E).
This was accompanied by a change in free I-Aa
mobility in SDS–polyacrylamide gel electropho-
resis due to the loss of the carbohydrate group
(Fig. 5E).
Binding of C3 to MHC II is conserved in mice
and humans
To test whether C3 binding to MHC II was a
peculiarity of C57BL/6 mice, we measured
C3 on cDCs of March1–/– mice backcrossed to
BALB/c (H-2d haplotype) or C3H (H-2k haplotype)
mice. High levels of C3 were present on cDCs
of all three strains (fig. S5A). We also detected
C3 on human blood DCs (fig. S5, B and C), with
highest expression on cDC2s followed by pDCs
and cDC1s (Fig. 6). To determine whether C3
binding required MHC II expression, we as-
sessed DCs in parallel from two human donors
with a 362A>T mutation in RFXANK, which
causes impaired MHC II transcription (15)
and a lack of MHC II at the cell surface (fig. S5B).
All MHC II–deficient DCs showed reduced
C3 expression (Fig. 6). Thus, like MHC II
ubiquitination by MARCH1, constitutive C3
Schriek et al., Science 375, eabf7470 (2022)
11 February 2022
4 of 12
RES EARCH | R E S E A R C H A R T I C L E
A
Surface C3
cDC1
cDC1
cDC1
cDC2
.
x
a
m
f
o
%
C3
FMO
WT
March1−/−
C3−/−
March1−/−×C3−/−
)
O
M
F
r
e
v
o
I
F
M
g
(
3
C
cDC1
**
**
****
25000
20000
15000
10000
5000
0
cDC2
**
**
****
10000
8000
6000
4000
2000
0
WT
C3−/−
March1−/−×C3−/−
March1−/−
WT
C3−/−
March1−/−×C3−/−
March1−/−
B
spleen
total LN
.
x
a
m
thymus
f
o
%
C3
C
cDC1
cDC2
pDC
mac
B cell
T cell
neutrophil eosinophil
FMO
WT March1−/−
cDC1
cDC2
WT+March1−/−BM
C3−/−
(recipient)
WT+March1−/−BM
WT
(recipient)
.
x
a
m
f
o
%
C3
FMO
WT (Ly5.1+Ly5.2−) March1−/−(Ly5.1−Ly5.2+)
cDC2
****
****
ns
ns
cDC1
****
****
*
ns
1500
1200
900
600
300
4000
3200
2400
1600
800
)
O
M
F
r
e
v
o
I
F
M
g
(
3
C
0
WT BM
WT BM
March1−/− BM
March1−/− BM
0
WT BM
WT BM
March1−/− BM
March1−/− BM
C3−/−
(recipient)
WT
(recipient)
C3−/−
(recipient)
WT
(recipient)
Fig. 4. Complement C3 deposition on the surface of splenic cDCs. (A) Representative flow cytometry
histograms (left) and bar graphs with MFI values (right) of C3 surface expression on splenic cDC1s and
cDC2s purified from the indicated mice. (B) Representative histograms of flow cytometry analysis of C3
surface levels on the indicated cell types in spleen, lymph nodes (LN), and thymus of wild-type or March1–/–
mice. Histograms are representative of at least two independent experiments with two or three individual
mice per experiment. (C) Representative flow cytometry histograms (left) and bar graphs with MFI values
(right) of C3 surface expression on wild-type and March1–/– cDC1s and cDC2s from mixed-BM chimeras
where wild-type or C3–/– recipient mice were reconstituted with a 1:1 mix of wild-type and March1–/– BM.
Graphs in (A) and (C) display data pooled from a minimum of two independent experiments, with each
symbol representing an individual mouse (n = 3 to 5 per experiment); bars denote mean ± SD. *P < 0.0332,
**P < 0.002, ****P < 0.0001 [Welch’s ANOVA test (no assumption of equal variances) followed by Games-
Howell multiple-comparisons test (A) or by pairwise comparison (C), adjusted P value (95% CI)].
binding to MHC II has been conserved through
evolution.
CR2 drives B cell trogocytosis of cDC
membranes containing MHC II–C3 complexes
Analysis in vitro confirmed that high surface
MHC II–C3 expression on MHC IIKRKI/KI cDCs
was necessary and sufficient to induce B cell
trogocytosis (Fig. 7A and fig. S6A). C3dg is the
ligand for complement receptor 2 (CR2, also
known as CD21), which is expressed only by
B cells (5). CR2-deficient B cells lacked the
capacity to trogocytose more membrane from
C3-decorated cDCs than from wild-type cDCs
(Fig. 7A and fig. S6A). Furthermore, MZ B cells
express higher levels of CR2 than FO B cells
(Fig. 2F), and virtually all MZ B cells trogocy-
tosed cDC membrane containing MHC II–C3dg
complexes (fig. S6B).
We next tested the validity of the con-
clusions from our in vitro analyses in vivo.
March1–/– and MHC IIKRKI/KI mice accumu-
lated trogocytic (CD8+CD11c+) B cells in their
spleens in a MHC II– and C3-dependent
manner (Fig. 7B). Wild-type spleens contained
a small number of CD8+CD11c+ MZ B cells, but
these were absent from C3–/– spleens (Fig. 7C
and fig. S6C); hence, trogocytosis occurred
constitutively in wild-type mice. Although
lymph node cDCs also displayed MHC II–C3
complexes (Fig. 4B), few trogocytic B cells
were detected in lymph nodes (fig. S6D), as
expected because mouse lymph nodes lack
MZ B cells (8). The number of splenic cDCs
decreased in all mice where cDCs had enriched
C3 surface expression (Fig. 7D) and B cells
expressed CR2 (Fig. 7E), in lockstep with the
increase in trogocytic MZ B cells. The spleens
of mice deficient in C3 did not contain more
cDCs than those of wild-type mice (fig. S6E);
this finding suggests that the limited amount
of trogocytosis occurring in wild-type mice is
insufficient to affect cDC homeostasis.
To further assess whether cDC number re-
duction and trogocytosis were directly correlated,
we produced mixed-BM chimeras where wild-
type or C3–/– recipient mice were reconstituted
with 1:1 wild-type and March1–/– BM. In
wild-type recipient mice, both wild-type and
March1–/– B cells displayed higher trogocytic
activity, but only March1–/– cDCs were re-
duced in numbers. Neither of these events were
observed in C3-deficient recipients (Fig. 7F).
Thus, the loss of cDCs required the accumula-
tion of MHC II–C3 complexes but was caused by
CR2-dependent MZ B cell trogocytosis (fig. S7).
Trogocytic MZ B cells present pMHC II
generated by cDCs
Because the primary mediator of trogocytic
cDC membrane transfer to B cells was CR2
recognition of MHC II–C3dg complexes, we
wondered whether the complexes found on
splenic B cells (Fig. 4B) were in fact acquired
from cDCs. In mixed-BM chimeras where
wild-type recipient mice were reconstituted
with a 1:1 mix of wild-type and Cr2–/– BM, all
B cells, and in particular MZ B cells, displayed
more C3 if they expressed CR2 than if they did
not (Fig. 8A). This suggested that although
some of the MHC II–C3 complexes displayed
by B cells probably formed on the B cells
themselves, most were acquired from cDCs.
This is supported by the absence of C3 de-
tection on B cells of mice where MHC II–C3
complexes could not be generated (i.e., H2-Aa–/–)
or lacked CR2 expression (Fig. 8B), which indi-
cates that virtually all C3 (bound to MHC II)
on the B cell surface was acquired by trogocy-
tosis of cDCs. These results also confirmed that
Schriek et al., Science 375, eabf7470 (2022)
11 February 2022
5 of 12
RES EARCH | R E S E A R C H A R T I C L E
A
cDC lysate
March1−/−
C3−/−
WT
serum
March1−/−
C3−/−
WT
B
Surface C3
cDC1
cDC2
FMO
WT
)
T
W
o
t
kDa
188
98
62
49
38
28
IB: C3
IB: C3
IB: Actin
.
x
a
m
f
o
%
C3
March1−/−
3
C
March1−/−×H2-Aa−/−
ΜΗC IIKRKI/KI
ΜΗC IIKRKI/KI×C3−/−
H2-Aa−/−
d
e
z
i
l
a
m
r
o
n
I
F
M
g
(
5
0
kDa
188
98
62
49
38
28
49
38
cDC1
**
***
****
**
****
25
20
15
10
cDC2
ns
***
****
***
****
25
20
15
10
5
0
H2-Aa−/−
WT
ΜΗC IIKRKI/KI
March1−/−
ΜΗC IIKRKI/KI×C3−/−
March1−/−×H2-Aa−/−
March1−/−
C3−/−
E
WT
H2-Aa−/−
WT
ΜΗC IIKRKI/KI
March1−/−
ΜΗC IIKRKI/KI×C3−/−
March1−/−×H2-Aa−/−
ΜΗC IIKRKI/KI
ΜΗC IIKRKI/KI
March1−/−
March1−/−
−
PNGase F
WT
WT
−
−
C
WT March1−/−×C3−/−
ΜΗC IIKRKI/KI
March1−/−
H2-Aa−/−
C3−/−
WT March1−/−×C3−/−
ΜΗC IIKRKI/KI
March1−/−
H2-Aa−/−
C3−/−
WT March1−/−×C3−/−
ΜΗC IIKRKI/KI
March1−/−
H2-Aa−/−
C3−/−
D
188
98
62
49
38
28
188
98
62
49
38
28
188
98
62
49
38
28
IP: MHC II
IB: MHC IIα
IP: MHC II
IB: MHC IIβ
IP: MHC II
IB: C3
188
98
62
49
188
98
62
49
IP: C3
IB: C3
WT
March1−/−
C3−/−
188
98
62
49
38
28
14
IP: C3
IB: MHC IIα
IP: MHC II
IB: MHC IIα
Fig. 5. C3 binds covalently to the MHC IIa carbohydrate on the surface of
cDCs. (A) Immunoblot (IB) of C3 from lysates of splenic cDCs and serum of wild-
type, March1–/–, or C3–/– mice. (B) Representative flow cytometry histograms
(left) and bar graphs with MFI values (right) of C3 surface expression on
splenic cDC1s and cDC2s of the indicated wild-type and mutant mice. Graphs
display normalized data pooled from two independent experiments, with
each symbol representing an individual mouse (n = 3 per experiment); bars
denote mean ± SD. **P < 0.002, ***P < 0.0002, ****P < 0.0001 [Welch’s
ANOVA test (no assumption of equal variances) followed by Games-Howell
multiple-comparisons test, adjusted P value (95% CI)]. (C) IB analysis of
MHC II immunoprecipitates (IPs) obtained from lysates of cDC1s of the
indicated wild-type or mutant mice, using Abs against MHC IIa, MHC IIb, or C3.
(D) IB detection of MHC IIa in IPs of C3 obtained from cDC1 lysates of the
indicated wild-type or mutant mice. (E) IB analysis of MHC II IPs obtained
from cDC1 lysates of the indicated wild-type or mutant mice after treatment
with (+) or without (–) PNGase F and subsequent detection of MHC IIa. All
immunoblots in (C) to (E) derive from separate gels and membranes for each
immunoblot (instead of sequential detection) and are representative of at
least three independent experiments, each lane loaded with IPs from cell
lysate of 2.5 × 105 purified splenic cDCs.
MZ B cell trogocytosis occurs constitutively in
wild-type mice.
Finally, we tested whether trogocytosis
enabled MZ B cells to “hijack” cDC Ag-
presenting functions by acquiring pMHC II
complexes generated by cDCs. First, we as-
sessed MZ B cell presentation of the model Ag,
I-Ea46-72 (IEpep) (16), after cDC1-targeted im-
munization. Wild-type, MHC IIKRKI/KI, and
MHC IIKRKI/KI × C3–/– mice were immunized
with an Ag consisting of IEpep fused to either
an isotype control monoclonal antibody (mAb)
or a mAb that recognizes Clec9A, a cDC1 re-
ceptor (17). Presentation of IEpep bound to
MHC II (I-Ab) was measured with a mAb (YAe)
that specifically recognizes this complex (16) (Fig.
8C). In wild-type mice immunized with the
mAb that binds Clec9A, only cDC1s, not
cDC2s or B cells, presented IEpep (Fig. 8D).
Additionally, no presentation occurred after
immunization with the isotype control mAb
(Fig. 8D). Immunization of MHC IIKRKI/KI
and MHC IIKRKI/KI × C3–/– mice with the
isotype control mAb led to low presentation of
IEpep by all four cell types (Fig. 8D), which is
likely due to their increased MHC II surface
expression. However, immunization with Clec9A-
targeted mAb led to a much higher IEpep
presentation by cDC1s. This was the case in
both MHC IIKRKI/KI and MHC IIKRKI/KI × C3–/–
mice, implying that presentation of IEpep by cDC1
was C3-independent (Fig. 8D). By contrast, pre-
sentation of the epitope by MZ B cells was
C3-dependent (Fig. 8D). This indicated that
C3-mediated trogocytosis enabled MZ B cells
to display in vivo pMHC II complexes they
cannot generate on their own but can acquire
from cDC1s. To assess whether acquired pMHC
II complexes could be recognized by T cells, we
immunized wild-type, MHC IIKRKI/KI, and MHC
IIKRKI/KI × C3–/– mice with Clec9A mAb con-
jugated to the model Ag ovalbumin (OVA). We
purified splenic FO and MZ B cells and incubated
them in vitro with I-Ab-OVA323-339–specific trans-
genic T cells (OT-II). Only MHC IIKRKI/KI MZ and,
to a lesser extent, FO B cells presented the antigen
and hence stimulated OT-II cells (Fig. 8E).
Discussion
We have described two intersections between
the innate and adaptive immune systems. The
Schriek et al., Science 375, eabf7470 (2022)
11 February 2022
6 of 12
RES EARCH | R E S E A R C H A R T I C L E
Experiment 1
Experiment 2
cDC1
cDC2
pDC
cDC1
cDC2
pDC
.
x
a
m
f
o
%
3
C
)
I
F
M
g
o
t
d
e
z
i
l
a
m
r
o
n
(
1.2
0.9
0.6
0.3
0.0
1.2
0.9
0.6
0.3
0.0
1.2
0.9
0.6
0.3
0.0
C3
FMO
RFXANK
362A>T
healthy blood donor
Healthy
RFXANK362A>T
Healthy
RFXANK362A>T
Healthy
RFXANK362A>T
Fig. 6. MHC II–dependent C3 deposition on human blood dendritic cells. Representative flow
cytometry histograms (left) and bar graphs with MFI values (right) of C3 surface expression on human
blood DCs of healthy donors and two MHC II–deficient (RFXANK362A>T) patients. Graphs display pooled
data (normalized to highest geometric MFI values) from the two experiments, with each symbol representing
an individual blood sample; bars denote mean ± SD.
first is cellular and features cDCs and MZ B cells.
The main role of cDCs is to present Ag to T cells
and initiate adaptive immunity (18). Interac-
tions between cDCs and B cells are less well
characterized (19). MZ B cells produce multi-
specific Abs that protect infants whose adapt-
ive immune systems have not yet generated
the full spectrum of memory B cells that dif-
ferentiate from the FO B cell repertoire (20). This
activity is considered to be T cell–independent.
MZ B cells also engage in T cell–dependent
immunity (8), but this requires Ag presentation,
and it remains unclear whether MZ B cells are
efficient APCs. Here, we demonstrated that MZ
B cells are constitutively in contact with cDCs
and acquire pMHC II complexes from them
using C3- and CR2-dependent trogocytosis.
The second interaction between innate and
adaptive immunity we have described occurs
at the molecular level. Activation of C3 by tick-
over (in the absence of pathogen) “pre-charges”
the complement system to respond to infection
(7). However, activated C3 can bind to healthy
cells and cause autoimmunity, and several
mechanisms are in place to inactivate it (6). We
showed that C3 activated by tickover binds to
the MHC IIa carbohydrate. C3 is then converted
to C3dg, and the resulting MHC II–C3dg com-
plexes are ubiquitinated, internalized, and
degraded (fig. S1D). Formation and ubiquiti-
nation of these complexes are independent
processes conserved in mice and humans.
Their role may be to prevent C3-driven host cell
damage. We have not observed overt inflam-
mation or autoimmunity in mice deficient in
MHC II ubiquitination (21), but such disorders
often do not manifest spontaneously in labo-
ratory mice.
The MHC II carbohydrate appears to have a
property that makes it prone to C3 binding and
is not found in other glycoproteins; possibly this
consists of a high mannose content (6). Because
the proportion of MHC II molecules bound
to C3 at any given time is small, only some
molecules may carry the required carbohy-
drate. It is plausible that mannose removal is
incomplete in a fraction of MHC II molecules
passing through the Golgi complex, causing
microheterogeneity as previously observed (22).
Moreover, the size of this fraction may vary
among cells as a result of differential expres-
sion of glycosidases (23), which would explain
why cells with similar levels of surface MHC II
(e.g., cDC1s, cDC2s, and B cells) displayed dif-
ferent amounts of C3.
Recognition of C3dg by CR2 was sufficient
to trigger the trogocytic acquisition of cDC
membrane by MZ B cells. The mechanism of
trogocytosis is poorly understood (11). It can
be driven by a single receptor-ligand inter-
action, as demonstrated here between C3dg
on cDCs and CR2 on B cells, but it is unclear
whether this interaction simply increases cell-
cell adhesion or triggers active membrane
transfer. Regardless, we showed that B cell
trogocytosis of cDC1s occurs in wild-type mice
and enables MZ B cells to present pMHC II
complexes generated by cDC1s. These results
help to explain how MZ B cells may modulate
T cell–dependent responses. In addition, trogo-
cytic B cells acquired other cDC receptors that
may expand the range of MZ B cell functions—
for instance, capture of Ag recognized by SIGN-R1,
Clec9A, and other receptors (17, 19, 24). MZ B
cells transport Ag to B cell follicles to increase
the efficiency of recognition by Ag-specific
FO B cells (25). Trogocytic acquisition of DC
receptors may expand their capacity for Ag
capture and dissemination.
Notwithstanding the benefits trogocytosis
may confer on MZ B cells, we have shown that
this process must be limited to prevent cDC
elimination. This notion is supported by the
following observations: (i) Reductions in
cDCs were only observed in mice where all
three molecular components required for
trogocytosis—MHC II, C3, and CR2—were
present; (ii) cDCs were lost from the spleen,
where MZ B cells are present, but not from
lymph nodes, where MZ B cells are absent,
even though cDCs displayed similar amounts
of MHC II–C3 complexes in both locations;
(iii) in mice that contained both March1–/–
cDCs, which could act as a source of trogocy-
tosed membrane, and wild-type cDCs, which
could not, only the mutant cDCs were lost.
We propose that wild-type cDCs can tolerate
trogocytic sequestration of a small amount of
plasma membrane but their mutant counter-
parts cannot repair the damage caused by
enhanced trogocytosis and die by “trogoptosis”
(26) (fig. S7). Reductions in cDCs may contribute
to the described defects in T cell priming in
March1–/– or MHC IIKRKI/KI mice (3, 14).
MHC II ubiquitination by MARCH1 plays
two important roles: to enhance the removal of
surface MHC II–C3 complexes on all APCs, and,
as described here, to limit MZ B cell trogocytosis
and elimination of cDCs. It is conceivable that
these two functions, more than the regulation
of MHC II antigen presentation, have been the
major drivers for the conservation of MHC II
ubiquitination through evolution.
Materials and methods
Mice
Experimental wild-type C57BL/6, BALB/c, or
C3H, and mutant March–/– (27), C3–/– (28)
(The Jackson Laboratory 129S4-C3tm1Crr/J;
#0036410), MHC IIKRKI/KI (14), H2-Aa–/–
(29), and Cr2–/– (30) [The Jackson Laboratory
129S7(NOD)-Cr2tm1Hmo/J; #008225] mice
were bred and maintained in specific pathogen-
free conditions in the Melbourne Bioresources
Platform at the Bio21 Molecular Science and
Biotechnology Institute, Victoria, Australia.
Analyses were undertaken with male and female
mice aged 7 to 14 weeks and performed in
accordance with the Institutional Animal Care
and Use Committee guidelines of the University
of Melbourne and the National Health and
Medical Research Council of Australia, approved
by the Animal Ethics Committee at the Univer-
sity of Melbourne (#1714238 and #1513472).
Isolation of primary murine cells and analysis by
flow cytometry
Whole splenocyte suspensions were generated
through digestion of finely chopped spleens
with 0.1% DNase I (Roche) and collagenase
type III (1 mg/ml; Worthington) followed by
lysis of red blood cells through incubation
with 168 mM ammonium chloride (5 min at
room temperature). cDCs were purified from
whole splenocyte suspensions as described
(31). In brief, low-density splenocytes were
Schriek et al., Science 375, eabf7470 (2022)
11 February 2022
7 of 12
A
WT
B cell
Cr2−/−
B cell
B
8
D
C
B220
WT
March1−/−
ΜΗC IIKRKI/KI
ΜΗC IIKRKI/KI
×C3−/−
March1−/−
×C3−/−
March1−/−
×H2-Aa−/−
8
D
C
RES EARCH | R E S E A R C H A R T I C L E
Fig. 7. B cell trogocytosis of cDC
membrane is MHC II–, C3-, and CR2-
dependent. (A) Trogocytic acquisition of CD8
from cDCs by wild-type or Cr2–/– B cells after
their incubation with wild-type or mutant
cDCs as indicated, with bar graphs (right)
representing the percentage of CD8+ B cells,
determined as shown in the flow cytometry
plots (left). Each data point represents
the measurement of a technical replicate (n =
3), displayed as mean ± SD. *P < 0.0332,
***P < 0.0002, ****P < 0.0001 [Welch’s
ANOVA test (no assumption of equal varian-
ces) followed by Games-Howell multiple-
comparisons test, adjusted P value (95%
CI)]. (B) Representative B220 versus CD8
and CD11c plots of flow cytometry analysis of
low-density splenocytes from the indicated
wild-type or mutant mice, showing the
proportion of trogocytic (CD8+ and CD11c+)
B cells. Plots are representative of at least
two independent experiments with two
or three individual mice per experiment.
(C) Frequency of trogocytic MZ B cells
among low-density splenocytes of wild-type
or C3–/– mice. (D) Number of splenic cDC1s
and cDC2s in the indicated wild-type or
mutant mice. (E) Number of cDC1s and
cDC2s (left) and frequency of trogocytic
B cells (right) identified as shown in (B) in
the indicated wild-type or mutant mice.
(F) Number of wild-type or March1–/– trogocytic
B cells, cDC1s, and cDC2s present in the
spleens of mixed-BM chimeras (wild-type or
C3–/– recipient mice), reconstituted with a
1:1 mix of wild-type and March1–/– BM.
Graphs in (C) to (F) display data pooled from
six (C) or two [(D) to (F)] independent
experiments, with each symbol representing
an individual mouse (n = 2 to 5 per experi-
ment); bars denote mean ± SD. *P < 0.0332,
**P < 0.002, ***P < 0.0002, ****P < 0.0001
[independent-samples t test with Welch’s
correction (no assumption of equal variances)
in (C), Welch’s ANOVA test (no assumption of
equal variances) followed by Games-Howell
multiple-comparisons test in (D) and (E),
or followed by pairwise comparison in (F),
two-tailed (C), or adjusted [(D) to (F)]
P value (all 95% CI)].
no
cDC
CD8
MHC IIKRKI/KI
cDC
WT
cDC
MHC IIKRKI/KI×C3−/−
cDC
0.58
12.8
38.9
7.84
0.34
13.6
11.1
7.33
20
f
o
%
10
CD8+ B cell
WT B Cell
***
****
50
40
30
s
l
l
e
c
B
Cr2−/−B Cell
***
*
50
40
30
20
10
0
WT cDC
no cDC
ΜΗC IIKRKI/KI cDC
ΜΗC IIKRKI/KI×C3−/− cDC
0
WT cDC
no cDC
ΜΗC IIKRKI/KI cDC
ΜΗC IIKRKI/KI×C3−/− cDC
CD8+ B cell
CD11c+ B cell
D
1.74
1.79
20.1
11.7
)
4
0
1
×
(
r
e
b
m
u
n
54
45
36
27
18
9
0
16.5
12.2
E
)
4
0
1
×
(
r
e
b
m
u
n
60
45
30
15
0
0.91
1.41
1.53
2.30
cDC1
ns
ns
****
****
March1−/−
ΜΗC IIKRKI/KI×C3−/−
March1−/−×C3−/−
ΜΗC IIKRKI/KI
WT
cDC2
ns
**
cDC1
ns
**
150
120
90
60
30
WT
ΜΗC IIKRKI/KI×Cr2−/−
ΜΗC IIKRKI/KI
CD8+
B cell
****
ns
****ns
0
WT
ΜΗC IIKRKI/KI×Cr2−/−
ΜΗC IIKRKI/KI
CD11c+
B cell
****
ns
****ns
21
14
7
F
2.08
20
15
10
5
)
4
0
1
×
(
r
e
b
m
u
n
)
4
0
1
×
(
r
e
b
m
u
n
200
160
120
80
40
0
cDC2
ns
ns
****
****
March1−/−
ΜΗC IIKRKI/KI×C3−/−
March1−/−×C3−/−
ΜΗC IIKRKI/KI
WT
CD8+ B cell CD11c+ B cell
ns
***
12
e
v
i
l
f
o
%
9
6
3
0
50
40
30
20
10
WT
ΜΗC IIKRKI/KI×Cr2−/−
ΜΗC IIKRKI/KI
cDC1
ns
**
*
***
10
8
6
4
2
0
ns
***
WT
ΜΗC IIKRKI/KI×Cr2−/−
ΜΗC IIKRKI/KI
cDC2
ns
**
**
***
180
150
120
90
60
30
0
WT BM
WT BM
March1−/− BM
March1−/− BM
0
WT BM
WT BM
March1−/− BM
March1−/− BM
0
WT BM
WT BM
March1−/− BM
March1−/− BM
0
WT BM
WT BM
March1−/− BM
March1−/− BM
WT
(recipient)
C3−/−
(recipient)
WT
(recipient)
C3−/−
(recipient)
WT
(recipient)
C3−/−
(recipient)
WT
(recipient)
C3−/−
(recipient)
3.35
c
1
1
D
C
C
B220
B220
low-density splenocytes
8
6
4
2
0
CD8+
MZ B cell
****
WT
C3−/−
20
15
10
5
0
CD11c+
MZ B cell
****
WT
C3−/−
s
l
l
e
c
B
Z
M
f
o
%
enriched by density gradient centrifugation
(at 2237g) with 1.077 g/cm3 Nycodenz (Axis
shield) and subsequent negative depletion
using rat mAbs against CD3 (KT3-1.1), Thy1
(T24/31.7), Ly-76 (Ter119), B220 (RA3-6B2),
and Ly-6C/G (RB6-8C5) and BioMag anti-rat
IgG-coupled magnetic beads (20 ml/106 cells,
Qiagen), resulting in 75 to 90% purity. Similarly,
cDCs from total lymph nodes were enriched
through enzymatic digestion [0.1% DNase I
and collagenase type III (1 mg/ml)] and Nyco-
denz (1.077 g/cm3, Axis shield) gradient cen-
trifugation (at 2237g). Splenic B cells were
isolated from whole-splenocyte suspensions
through gradient centrifugation (at 2237g)
with Ficoll-Paque Plus (GE Healthcare) and
subsequent negative depletion using FITC-
conjugated mAb against CD4 (GK1.5, 1.6 mg/
ml), Ly76 (TER119, 1.6 mg/ml), and CD43 (S7,
1.6 mg/ml) and magnetic anti-FITC MicroBeads
(2 ml/106 cells, Miltenyi Biotec), resulting in 95
to 98% purity.
For flow cytometry analysis, purified murine
cells were washed in EDTA-BSS with 2% (v/v)
FCS, Fc receptor–blocked (1:50, Miltenyi
Biotec), and incubated with mAb for the
detection of surface markers (table S4). All
Schriek et al., Science 375, eabf7470 (2022)
11 February 2022
8 of 12
RES EARCH | R E S E A R C H A R T I C L E
Fig. 8. MZ B cells present pMHC II complexes
trogocytosed from cDC1s in vivo. (A and
B) Representative flow cytometry histograms
(left) and bar graphs with MFI values (right)
of C3 surface expression on wild-type or
Cr2–/– total, FO, or MZ B cells of mixed-BM
chimera mice (wild-type or C3–/– recipients),
reconstituted with a 1:1 mix of wild-type
and Cr2–/– BM (A) or on FO and MZ B cells
of the indicated wild-type or mutant mice
(B). Bar graphs display data pooled from
two independent experiments, with each
symbol representing an individual mouse
(n = 4 or 5 per experiment); bars denote
mean ± SD. **P < 0.002, ***P < 0.0002,
****P < 0.0001 [Welch’s ANOVA test
(no assumption of equal variances) followed
by pairwise comparison (A) or by Games-Howell
multiple-comparisons test (B), adjusted
P value (95% CI)]. (C) Cartoon representing
the capture of anti-Clec9A mAb fused to
the I-E peptide (I-Ea46-72) by cDC1s, followed
by intracellular processing and presentation of
I-E peptide by MHC II (I-Ab) and detection of
the complex on the cell surface with YAe mAb.
(D) Left: Surface expression of I-Ab + IEpep
(YAe) on splenic cDC1s, cDC2s, total B cells,
and MZ B cells of the indicated wild-type
or mutant mice immunized with anti–Clec9A-
IEpep or control isotype-IEpep mAb. Right:
Bar graphs with MFI values of YAe surface
expression; data pooled from two independent
experiments, with each symbol representing an
individual mouse (n = 2 or 3 per experiment),
displayed as means ± SD. *P < 0.0332, **P <
0.002, ***P < 0.0002, ****P < 0.0001 [Welch’s
ANOVA test (no assumption of equal variances)
followed by pairwise comparison within each
group (cell type), P value (all 95% CI)].
(E) Proliferation of OT-II cells incubated with
increasing numbers of sort-purified FO or
MZ B cells from the indicated wild-type or
mutant mice immunized with anti-Clec9A mAb
conjugated with OVA. Left: Dividing OT-II
cells, gated as shown in the representative flow
cytometry plots. Right: Numbers of proliferated
OT-II cells pooled from two independent
experiments, with each symbol representing
a technical replicate (n = 3 per experiment),
displayed as means ± SD. *P < 0.0332,
****P < 0.0001 (one-way ANOVA followed by
Bonferroni’s multiple-comparisons test).
B
.
x
a
m
f
o
%
D
.
x
a
m
f
o
%
all B cell
FO B cell MZ B cell
A
WT
recipient
C3−/−
recipient
.
x
a
m
f
o
%
C3
)
O
M
F
r
e
v
o
I
F
M
g
(
3
C
all B cell
****
****
ns
***
4000
3000
2000
1000
0
WT BM
Cr2−/− BM
WT BM
Cr2−/− BM
MZ B cell
****
ns
****
ns
FO B cell
****
**
ns
***
4000
3000
2000
1000
10000
7500
5000
2500
0
WT BM
Cr2−/− BM
WT BM
Cr2−/− BM
0
WT BM
Cr2−/− BM
WT BM
Cr2−/− BM
FMO
WT (Ly5.1+Ly5.2−)
Cr2−/−(Ly5.1−Ly5.2+)
WT
(recipient)
C3−/−
(recipient)
WT
(recipient)
C3−/−
(recipient)
WT
(recipient)
C3−/−
(recipient)
FO B cell
MZ B cell
FMO
WT
MHC IIKRKI/KI
ΜΗC IIKRKI/KI
×Cr2−/−
Cr2−/−
H2-Aa−/−
C3
)
O
M
F
r
e
v
o
I
F
M
g
(
3
C
C
MZ B cell
***
**
****
****
FO B cell
**
ns
****
****
3000
2500
2000
1500
1000
500
0
20000
15000
10000
5000
4000
3000
2000
1000
0
Cr2−/−
H2-Aa−/−
WT
ΜΗC IIKRKI/KI×Cr2−/−
ΜΗC IIKRKI/KI
Cr2−/−
H2-Aa−/−
WT
ΜΗC IIKRKI/KI×Cr2−/−
ΜΗC IIKRKI/KI
cDC1
cDC2
all B cell MZ B cell
WT
MHC IIKRKI/KI
MHC IIKRKI/KI
×C3−/−
.
l
r
t
c
e
p
y
t
o
s
i
r
e
v
o
)
I
I
C
H
M
n
i
p
e
p
E
I
(
e
A
Y
6000
5000
****
**
****
4000
3000
2000
1000
0
*
***
****
ns
ns
ns
*
*
ns
cDC1
cDC2
all
B cell
MZ
B cell
WT
MHC IIKRKI/KI
MHC IIKRKI/KI×C3−/−
FO B cell
MZ B cell
YAe (IEpep in MHC II)
FMO isotype ctr. WT
MHC IIKRKI/KI
MHC IIKRKI/KI×C3−/−
1800
1500
1200
900
600
E
WT
MHC IIKRKI/KI
OT-II
FO B cell
MZ B cell
0.56
0.69
1.12
4.62
1.12
0.47
I
I
-
T
O
d
e
t
a
r
e
f
i
l
o
r
p
f
o
r
e
b
m
u
n
MHC IIKRKI/KI
×C3−/−
4
D
C
CTV
300
ns
ns
0
****
1800
1500
1200
900
600
ns
ns
*
300
0
0
100 200 300 400
0
100 200 300 400
Number (×103) of FO/MZ B cells
WT
MHC IIKRKI/KI
MHC IIKRKI/KI MHC IIKRKI/KI×C3−/−
fluorescence-minus-one (FMO) controls cor-
respond to cells that were stained with mAbs
for the detection of lineage markers but not
for the indicated molecule. Cell surface C3 was
detected using biotinylated anti-C3 mAbs clone
11H9 (Novus Biological), clone 3d29 [kindly
provided by V. M. Holers and J. M. Thurman,
School of Medicine, University of Colorado,
Denver, Anschutz Medical Campus (32)], or
polyclonal Abs (pAbs) HP8012 (Hycult) and
55500 (MP Biomedical) (table S4) with sub-
sequent incubation with BV605-, APC-, or
APCCy7-conjugated streptavidin. All FMO
controls for C3 surface staining correspond
to cells that were incubated with mAbs for
the detection of lineage markers, including
streptavidin-BV605/APC-Cy7/APC but no bio-
tinylated anti-C3 Abs. All staining steps were
performed on ice (30 min) and isotype-matched
control Abs were used where required. Cells were
analyzed using a LSRFortessa flow cytometer (BD
Biosciences) in the Melbourne Cytometry
Platform (University of Melbourne) and FlowJo
software (Tree Star) with exclusion of cell doublets
and dead cells, in all cases identified on the basis
of forward and side scatter (FSC and SSC)
and propidium iodide (PI) staining (fig. S2).
Schriek et al., Science 375, eabf7470 (2022)
11 February 2022
9 of 12
RES EARCH | R E S E A R C H A R T I C L E
Enumeration of cells via flow cytometry was
performed using a defined number of Sphero
blank calibration beads (BD BioSciences).
B cells were identified as CD19+B220+, CD4+
T cells as CD3+CD4+, CD8+ T cells as CD3+CD8+,
pDCs as MHC II+CD11c+BST-2+Siglec H+, and
cDCs as CD11c+MHC II+, with further discrim-
ination of cDC1s as CD11b–CD8+ and cDC2s
as CD11b+CD8– (fig. S2). Macrophages were
identified as F4/80+CD64+, neutrophils as
B220–CD3–CD4–CD8–CD11clo-intCD11bhiLy6G+,
and eosinophils as B220–CD3–CD4–CD8–
CD11clo-intCD11bhiLy6G–SSC-HhiLy6Clo-int (33).
Isolation of human blood DCs and analysis by
flow cytometry
Blood samples from healthy donors (n = 10,
male and female, 45.8 ± 15.9 years of age) were
obtained as buffy coats from the Australian
Red Cross Lifeblood, Victoria, Australia, with
written and informed consent from the donors
and ethics approval from the University of
Melbourne Human Research and Ethics Com-
mittee (#1035100). Whole blood samples from
MHC II–deficient patients (n = 2, male, 11 and
15 years of age) were collected by staff of the
Royal Children’s Hospital, Victoria, Australia,
with written and informed consent from the
donors and ethics approval from the Royal
Children’s Hospital Research Ethics Committee
(#33146) after receiving regular intravenous
immunoglobulin (IVIg) infusions. Peripheral
blood mononuclear cells (PBMCs) from both
healthy donors as well as from MHC II–deficient
patients were enriched using centrifugation
(at 2237g) with Ficoll-Paque Plus (GE Healthcare).
DCs were purified from PBMCs with the
Pan-DC Enrichment Kit (Miltenyi Biotec) ac-
cording to the manufacturer’s instructions. In
brief, 108 cells were stained with mAb to block
FcR and incubated with a depletion Ab cock-
tail and magnetic microbeads to deplete Ab-
labeled cells using LS magnetic columns and
MidiMACS magnet (both Miltenyi Biotec). This
resulted in 5 to 10% purity of cDCs (CD11c+HLA-
DR+). For flow cytometry analysis, purified
human DCs were washed in EDTA-BSS with
2% (v/v) FCS and incubated with mAb against
CD1c (L161), CD11c (3.9), CD141 (M80), HLA-
DR (LN3), CD123 (6H6), or lineage (lin) mAb
cocktail against CD3, CD14, CD16, CD19,
CD20, and CD56 (OKT3, M5E2, 3G8, HIB19,
2H7, and HCD56). cDC1s were identified as
lin–CD11cint-hiCD123–CD141+CD1c–, cDC2s as
lin–CD11cint-hiCD123–CD141–CD1c+, and pDCs
as lin–CD11cint-hiCD11b–CD123+ (fig. S5B). Cells
were incubated with whole rabbit serum and
cell surface C3 detected using biotinylated
anti-C3–specific rabbit pAb (Abcam, ab48342)
with subsequent incubation with APC-conjugated
streptavidin, including appropriate controls
to ensure specificity (fig. S5C). Cells were analyzed
using an LSRFortessa (BD Biosciences) and
FlowJo software (Tree Star) with exclusion of
cell doublets and dead cells in all cases iden-
tified on the basis of FSC and SSC and staining
with PI.
RNA sequencing and transcriptomic analysis
Splenic cDC1s, B cells, and CD11cintCD24+CD8int
cells from four biological replicates with pooled
spleens from five wild-type or March1–/– mice
per replicate were sorted (95 to 99% purity)
with the Influx Cell Sorter (BD Biosciences)
at Murdoch Children’s Research Institute,
Royal Children's Hospital, Victoria, Australia.
Genomic DNA was removed and total RNA
extracted using the RNeasy Mini Kit (Qiagen).
RNA quality was assessed via Agilent Bioana-
lyzer 2100 using the Agilent RNA 6000 Nano
Kit (Agilent Technologies) and rRNA deple-
tion and library preparation were performed
according to manufacturer protocols (TruSeq,
Illumina) at the Australian Genome Research
Facility (AGRF), Victoria, Australia. Whole-
transcriptome sequencing was undertaken using
Illumina Hi Seq 2500 (Illumina, San Diego,
CA) at the AGRF in replicates on two lanes. All
100–base pair single-end reads were mapped
to the reference mouse genome (GRCm38/
mm10) using STAR aligner (version 2.7) (34),
and gene-wise counts were obtained using
Subread package (v1.6.2) (35). Differential ex-
pression analysis of RNA sequencing data was
carried out in Galaxy/Australia (usegalaxy.org.
au, Melbourne Bioinformatics) (36) using the
differential expression tool (Trinity assembly)
(37, 38) and in RStudio (version 1.1.447)/R (ver-
sion 3.6.1) using the edgeR (version 3.26.6) (39)
and limma (version 3.40.2) (40) packages.
Gene counts were converted to log2 counts per
million and normalized using the trimmed
mean of M-values normalization method in
edgeR (41). Precision weights were applied
with the “voom” function (42) of the limma
package, and a linear model was fitted to each
gene correcting for inter-replicate variation.
Empirical Bayes-moderated t statistics were
used to assess differences in gene expression
and determine P values. Differences in expres-
sion levels were evaluated using a linear model
on replicates, with samples compared across
the different groups. Genes were corrected for
multiple testing, and genes having a false dis-
covery rate (FDR) of <0.05 using the decideTest
function (43) in limma were considered sig-
nificant. Gene set analysis was performed using
the fast implementation of rotation gene set
testing (44) (FRY in limma) for B cell signatures—
GO:0042113 for B cell activation and BCR sig-
naling pathway (Biocarta) from the Molecular
Signatures Database (org.Mm.eg.db version
3.12.0). RNA-seq data from this study were de-
posited in GEO under accession number 185597.
In vitro trogocytosis assay
In vitro trogocytosis assays were carried out as
described (9), excluding Ag priming. In brief,
2 × 105 B cells and 3 × 105 cDCs purified from
spleens were cocultured in RPMI media sup-
plemented with 10% (v/v) FCS (Sigma), 1×
GlutaMAX (Gibco), penicillin (100 U/ml), strep-
tomycin (100 mg/ml; Media Preparation Unit,
Peter Doherty Institute, Victoria, Australia),
and 50 mM b-mercaptoethanol (Life Technol-
ogies) in 96-well U-bottom cell culture plates
for 2 hours at 37°C after quick centrifugation
at 150g to promote cell contact. In some ex-
periments, either cDCs or B cells were stained
with PKH26 Red Fluorescent Cell Linker ac-
cording to the manufacturer’s instructions
(Sigma Aldrich). After incubation, cells were
washed in EDTA-BSS with 2% (v/v) FCS to
disrupt intercellular clusters and analyzed
by flow cytometry as described above.
Mixed–bone marrow chimeric mice
Bone marrow (BM) was harvested from tibia
and femur of donor mice (wild-type, Marchf1–/–,
and Cr2–/–), and recipient mice (wild-type and
C3–/–) were irradiated twice at 550 cGy (rad),
3 hours apart, before intravenously injected
with 50:50 mixed BM cells. Recipient mice
were intraperitoneally injected with anti-Thy1
(clone T24) to eliminate radio-resistant host
T cells the day after irradiation. Mice were
reconstituted for at least 8 weeks before use.
Immunoblotting, immunoprecipitation, and
PNGase F treatment of MHC II and C3
cDCs purified from Flt3L-expanded mice (45)
were lysed on ice in 1% (v/v) IGEPAL CA-630
(Sigma-Aldrich), 50 mM Tris (pH 7.5; Astral
Scientific), 5 mM magnesium chloride (Chem-
Supply), and cOmplete protease inhibitor cocktail
(Roche) at a concentration of 107 cells/ml, and
nuclei were removed by centrifugation at
14,000g at 4°C. For immunoprecipitation of
MHC II and C3, lysates were precleared twice
by incubation with uncoupled protein G–sephar-
ose beads (Walter and Eliza Hall Institute, WEHI)
in the presence of normal rabbit serum. To im-
munoprecipitate MHC II or C3, protein G–
sepharose beads were precoupled with anti–I-
A/I-E (clone M5/115) mAb (10 mg per 107 cells)
or anti-C3dg (clone 3d29) mAb (5 mg per 107
cells) and added to the lysate. To eluate MHC
II/C3 from protein G–sepharose beads for im-
munoblot analysis, beads were incubated in
3× SDS (reducing) sample buffer at 95°C. For
deglycosylation of MHC II molecules, treatment
with PNGase F was carried out according to
manufacturer’s instructions (New England
BioLabs). In brief, washed protein G–sepharose
beads with mAb-bound MHC II molecules were
incubated with denaturing buffer followed by
incubation with 500 U of PNgase F in 1×
GlycoBuffer containing 1% NP-40.
For analysis of MHC II and C3 by immuno-
blotting, serum samples, whole-cell lysates or
immunoprecipitates from equal cell numbers
were separated on a precast NuPAGE gel (4 to
Schriek et al., Science 375, eabf7470 (2022)
11 February 2022
10 of 12
RES EARCH | R E S E A R C H A R T I C L E
12% Bis-Tris Plus) and transferred to iBlot 2
PVDF membranes with an iBlot 2 system
according to manufacturer’s instructions (all
Invitrogen). Samples were probed for C3 with
anti-C3 pAb from rabbit serum (HP8012,
Hycult) or MHC II with anti–I-Aa– or anti–
I-Ab–specific pAb from rabbit serum JV1 and
JV2 (WEHI antibody facility) followed by
HRP-coupled secondary antibodies with
the appropriate species reactivities. Chemi-
luminescence was measured using the ECL
Select Western Blotting reagent (Amersham
GE Healthcare), acquired on a ChemiDoc MP
imaging system (Bio-Rad) and ImageJ.
Mass spectrometry
Immunoprecipitated MHC II was analyzed
using a precast NuPAGE gel (4 to 12% Bis-Tris
Plus, Invitrogen) and stained using Coomassie
Brilliant Blue R-250 (Bio-Rad). Bands of interest
were excised and subjected to reduction, alkyl-
ation, and trypsin digestion before mass spec-
trometry (MS) as described (46). In brief, excised
gel samples were destained in 50 mM ammo-
nium bicarbonate dissolved in 50% (v/v) aceto-
nitrile, reduced with 10 mM dithiothreitol (DTT),
and then alkalized with 55 mM iodoaceta-
mide. Air-dried gel pieces were incubated with
trypsin (15 ng/ml, Promega) in 25 mM ammo-
nium bicarbonate for ~16 hours at 37°C and
peptides were extracted through incubation
with 0.1% (v/v) formic acid in 60% acetonitrile.
Mass spectrometry and data analysis were
carried out at the WEHI Proteomic Laboratory
(Webb laboratory), Victoria, Australia. Extracted
peptides were separated by reverse-phase chro-
matography on a 1.9-mm C18 fused silica column
(I.D. 75 mm, O.D. 360 mm × 25 cm length) packed
into an emitter tip (Ion Opticks, Australia), using
a nano-flow HPLC (M-class, Waters). The HPLC
was coupled to an Impact II UHR-QqTOF mass
spectrometer (Bruker) using a CaptiveSpray
source and nanoBooster at 0.20 bar using
acetonitrile. Peptides were loaded directly onto
the column at a constant flow rate of 400 nl/min
with buffer A (99.9% Milli-Q water, 0.1% formic
acid) and eluted with a 90-min linear gradient
from 2 to 34% buffer B (99.9% acetonitrile, 0.1%
formic acid). Mass spectra were acquired in a
data-dependent manner including an automatic
switch between MS and MS/MS scans using a
1.5-s duty cycle and 4-Hz MS1 spectra rate
followed by MS/MS scans at 8 to 20 Hz de-
pendent on precursor intensity for the re-
mainder of the cycle. MS spectra were acquired
within a mass range of 200 to 2000 m/z. Peptide
fragmentation was performed using collision-
induced dissociation (CID).
All raw files were analyzed by MaxQuant
(1.5.6.5) software using the integrated Androm-
eda search engine. Experiment type was set
as TIMS-DDA with no modification to default
settings. Data were searched against the mouse
Uniprot Reference Proteome with isoforms and
a separate reverse decoy database using a strict
trypsin specificity allowing up to two missed
cleavages. The minimum required peptide
length was seven amino acids.
Immunization with IEpep/OVA-conjugated
Clec9A mAb for YAe detection and ex vivo
antigen presentation assay
IEpep (Ea52-68)–loaded I-Ab surface expres-
sion, as detected by biotinylated YAe mAb (16)
(Thermo Fisher), was analyzed at the surface
of cDCs and B cells after intravenous injection
of mice with 0.5 mg of anti–Clec9A-IEpep
(clone 10B4) (47) or isotype-IEpep mAb. The
IEa epitope (I-Ea46-72) was cloned in-frame with
the heavy chain C terminal region of the Clec9A
mAb (clone 10B4) or isotype mAb via alanine
linkers. After 22 to 24 hours, splenic cDCs and
B cells were examined for IEpep (Ea52-68)–loaded
I-Ab surface expression by flow cytometry
using biotinylated YAe mAb and streptavidin-
PE. FMO controls for YAe surface staining cor-
respond to cells that were incubated with mAb
for the detection of lineage markers, including
streptavidin-PE but not biotinylated YAe mAb.
For ex vivo antigen presentation assays, mice
were intravenously injected with 1 mg of Clec9A-
OVA mAb. After 22 to 24 hours, spleen MZ
and follicular (FO) B cells were sorted to purity.
OT-II cells were purified from lymph nodes and
labeled with 2 mM CellTrace Violet (CTV),
and 5 × 104 cells per well were cultured with
isolated FO or MZ B cells in U-bottom 96-well
plates for 90 hours. Proliferated OT-II cells were
determined by flow cytometry as the number
of CD4+TCRVa2+ cells that had undergone
CTV dilution.
RE FERENCES AND NOTES
1.
J. A. Villadangos, Presentation of antigens by MHC class II
molecules: Getting the most out of them. Mol. Immunol. 38,
329–346 (2001). doi: 10.1016/S0161-5890(01)00069-4;
pmid: 11684289
2. H. Liu, J. D. Mintern, J. A. Villadangos, MARCH ligases in
immunity. Curr. Opin. Immunol. 58, 38–43 (2019).
doi: 10.1016/j.coi.2019.03.001; pmid: 31063934
3. K. R. Wilson et al., MARCH1-mediated ubiquitination of MHC II
impacts the MHC I antigen presentation pathway. PLOS ONE
13, e0200540 (2018). doi: 10.1371/journal.pone.0200540;
pmid: 30001419
4. H. J. Kim et al., Ubiquitination of MHC Class II by March-I
Regulates Dendritic Cell Fitness. J. Immunol. 206, 494–504
(2021). doi: 10.4049/jimmunol.2000975; pmid: 33318291
5. V. M. Holers, Complement and its receptors: New insights into
human disease. Annu. Rev. Immunol. 32, 433–459 (2014).
doi: 10.1146/annurev-immunol-032713-120154;
pmid: 24499275
6. M. K. Pangburn, V. P. Ferreira, C. Cortes, Discrimination
between host and pathogens by the complement system.
Vaccine 26 (Suppl 8), I15–I21 (2008). doi: 10.1016/
j.vaccine.2008.11.023; pmid: 19388159
7. F. Bexborn, P. O. Andersson, H. Chen, B. Nilsson, K. N. Ekdahl,
The tick-over theory revisited: Formation and regulation of the
soluble alternative complement C3 convertase (C3(H2O)Bb).
Mol. Immunol. 45, 2370–2379 (2008). doi: 10.1016/
j.molimm.2007.11.003; pmid: 18096230
8. A. Cerutti, M. Cols, I. Puga, Marginal zone B cells: Virtues of
innate-like antibody-producing lymphocytes. Nat. Rev.
Immunol. 13, 118–132 (2013). doi: 10.1038/nri3383;
pmid: 23348416
9. S. Daubeuf, A.-L. Puaux, E. Joly, D. Hudrisier, A simple
trogocytosis-based method to detect, quantify, characterize
and purify antigen-specific live lymphocytes by flow cytometry,
via their capture of membrane fragments from antigen-
presenting cells. Nat. Protoc. 1, 2536–2542 (2006).
doi: 10.1038/nprot.2006.400; pmid: 17406507
10. E. Joly, D. Hudrisier, What is trogocytosis and what is its
purpose? Nat. Immunol. 4, 815–815 (2003). doi: 10.1038/
ni0903-815; pmid: 12942076
11. D. M. Davis, Intercellular transfer of cell-surface proteins is
common and can affect many stages of an immune response.
Nat. Rev. Immunol. 7, 238–243 (2007). doi: 10.1038/nri2020;
pmid: 17290299
12. P. Schriek et al., Physiological substrates and ontogeny-
specific expression of the ubiquitin ligases MARCH1 and
MARCH8. Curr. Res. Immunol. 2, 218–228 (2021). doi: 10.1016/
j.crimmu.2021.10.004
13. Z. Zhou, M.-J. Xu, B. Gao, Hepatocytes: A key cell type for
innate immunity. Cell. Mol. Immunol. 13, 301–315 (2016).
doi: 10.1038/cmi.2015.97; pmid: 26685902
14. M. Ohmura-Hoshino et al., Cutting edge: Requirement of
MARCH-I-mediated MHC II ubiquitination for the maintenance
of conventional dendritic cells. J. Immunol. 183, 6893–6897
(2009). doi: 10.4049/jimmunol.0902178; pmid: 19917682
15. W. Wiszniewski et al., Novel mutations in the RFXANK gene:
RFX complex containing in-vitro-generated RFXANK mutant
binds the promoter without transactivating MHC II.
Immunogenetics 54, 747–755 (2003). doi: 10.1007/s00251-
002-0521-1; pmid: 12618906
16. Rudensky AYu, S. Rath, P. Preston-Hurlburt, D. B. Murphy,
C. A. Janeway Jr., On the complexity of self. Nature 353,
660–662 (1991). doi: 10.1038/353660a0; pmid: 1656278
I. Caminschi et al., The dendritic cell subtype-restricted C-type
lectin Clec9A is a target for vaccine enhancement. Blood 112,
3264–3273 (2008). doi: 10.1182/blood-2008-05-155176;
pmid: 18669894
17.
18. J. Banchereau, R. M. Steinman, Dendritic cells and the control
of immunity. Nature 392, 245–252 (1998). doi: 10.1038/32588;
pmid: 9521319
19. W. R. Heath, Y. Kato, T. M. Steiner, I. Caminschi, Antigen
presentation by dendritic cells for B cell activation. Curr. Opin.
Immunol. 58, 44–52 (2019). doi: 10.1016/j.coi.2019.04.003;
pmid: 31071588
20. J. F. Kearney, P. Patel, E. K. Stefanov, R. G. King, Natural
antibody repertoires: Development and functional role in
inhibiting allergic airway disease. Annu. Rev. Immunol. 33,
475–504 (2015). doi: 10.1146/annurev-immunol-032713-
120140; pmid: 25622195
21. H. Liu et al., Ubiquitination of MHC Class II Is Required for
Development of Regulatory but Not Conventional CD4+ T Cells.
J. Immunol. 205, 1207–1216 (2020). doi: 10.4049/
jimmunol.1901328; pmid: 32747505
22. S. J. Swiedler, G. W. Hart, A. L. Tarentino, T. H. Plummer Jr.,
J. H. Freed, Stable oligosaccharide microheterogeneity
at individual glycosylation sites of a murine major
histocompatibility antigen derived from a B-cell lymphoma.
J. Biol. Chem. 258, 11515–11523 (1983). doi: 10.1016/
S0021-9258(17)44258-X; pmid: 6604728
23. S. E. Cullen, C. S. Kindle, D. C. Shreffler, C. Cowing, Differential
glycosylation of murine B cell and spleen adherent cell Ia
antigens. J. Immunol. 127, 1478–1484 (1981). pmid: 6792276
24. S. F. Gonzalez et al., Capture of influenza by medullary
dendritic cells via SIGN-R1 is essential for humoral immunity in
draining lymph nodes. Nat. Immunol. 11, 427–434 (2010).
doi: 10.1038/ni.1856; pmid: 20305659
25. G. Cinamon, M. A. Zachariah, O. M. Lam, F. W. Foss Jr.,
J. G. Cyster, Follicular shuttling of marginal zone B cells
facilitates antigen transport. Nat. Immunol. 9, 54–62 (2008).
doi: 10.1038/ni1542; pmid: 18037889
26. H. L. Matlung et al., Neutrophils Kill Antibody-Opsonized
Cancer Cells by Trogoptosis. Cell Rep. 23, 3946–3959.e6
(2018). doi: 10.1016/j.celrep.2018.05.082; pmid: 29949776
27. Y. Matsuki et al., Novel regulation of MHC class II function in B
cells. EMBO J. 26, 846–854 (2007). doi: 10.1038/sj.
emboj.7601556; pmid: 17255932
28. Q. Shi et al., Complement C3-Deficient Mice Fail to Display Age-
Related Hippocampal Decline. J. Neurosci. 35, 13029–13042
(2015). doi: 10.1523/JNEUROSCI.1698-15.2015; pmid: 26400934
29. F. Köntgen, G. Süss, C. Stewart, M. Steinmetz, H. Bluethmann,
Targeted disruption of the MHC class II Aa gene in C57BL/6
mice. Int. Immunol. 5, 957–964 (1993). doi: 10.1093/intimm/
5.8.957; pmid: 8398989
30. H. Molina et al., Markedly impaired humoral immune response
in mice deficient in complement receptors 1 and 2. Proc. Natl.
Schriek et al., Science 375, eabf7470 (2022)
11 February 2022
11 of 12
RES EARCH | R E S E A R C H A R T I C L E
Acad. Sci. U.S.A. 93, 3357–3361 (1996). doi: 10.1073/
pnas.93.8.3357; pmid: 8622941
31. D. Vremec, K. Shortman, The isolation and identification of
murine dendritic cell populations from lymphoid tissues and
their production in culture. Methods Mol. Biol. 415, 163–178
(2008). doi: 10.1007/978-1-59745-570-1_10; pmid: 18370154
32. J. M. Thurman et al., Detection of complement activation using
monoclonal antibodies against C3d. J. Clin. Invest. 123,
2218–2230 (2013). doi: 10.1172/JCI65861; pmid: 23619360
33. S. E. Liyanage et al., Flow cytometric analysis of inflammatory
and resident myeloid populations in mouse ocular
inflammatory models. Exp. Eye Res. 151, 160–170 (2016).
doi: 10.1016/j.exer.2016.08.007; pmid: 27544307
34. A. Dobin et al., STAR: Ultrafast universal RNA-seq aligner.
Bioinformatics 29, 15–21 (2013). doi: 10.1093/bioinformatics/
bts635; pmid: 23104886
35. Y. Liao, G. K. Smyth, W. Shi, The Subread aligner: Fast,
accurate and scalable read mapping by seed-and-vote. Nucleic
Acids Res. 41, e108–e108 (2013). doi: 10.1093/nar/gkt214;
pmid: 23558742
36. E. Afgan et al., The Galaxy platform for accessible, reproducible
and collaborative biomedical analyses: 2018 update. Nucleic
Acids Res. 46, W537–W544 (2018). doi: 10.1093/nar/gky379;
pmid: 29790989
37. D. Blankenberg, N. Coraor, G. Von Kuster, J. Taylor, A. Nekrutenko,
Galaxy Team, Integrating diverse databases into an unified analysis
framework: A Galaxy approach. Database 2011, bar011–bar011
(2011). doi: 10.1093/database/bar011; pmid: 21531983
38. C. Sloggett, N. Goonasekera, E. Afgan, BioBlend: Automating
pipeline analyses within Galaxy and CloudMan. Bioinformatics
29, 1685–1686 (2013). doi: 10.1093/bioinformatics/btt199;
pmid: 23630176
39. M. D. Robinson, D. J. McCarthy, G. K. Smyth, edgeR: A
Bioconductor package for differential expression analysis of
digital gene expression data. Bioinformatics 26, 139–140
(2010). doi: 10.1093/bioinformatics/btp616; pmid: 19910308
40. M. E. Ritchie et al., limma powers differential expression
analyses for RNA-sequencing and microarray studies.
Nucleic Acids Res. 43, e47 (2015). doi: 10.1093/nar/gkv007;
pmid: 25605792
41. M. D. Robinson, A. Oshlack, A scaling normalization method for
differential expression analysis of RNA-seq data. Genome Biol.
11, R25 (2010). doi: 10.1186/gb-2010-11-3-r25; pmid: 20196867
42. C. W. Law, Y. Chen, W. Shi, G. K. Smyth, voom: Precision
weights unlock linear model analysis tools for RNA-seq read
counts. Genome Biol. 15, R29 (2014). doi: 10.1186/gb-2014-15-
2-r29; pmid: 24485249
43. C. W. Law et al., RNA-seq analysis is easy as 1-2-3 with limma,
Glimma and edgeR. F1000 Res. 5, 1408 (2016). doi: 10.12688/
f1000research.9005.3; pmid: 27441086
44. D. Wu et al., ROAST: Rotation gene set tests for complex
microarray experiments. Bioinformatics 26, 2176–2182 (2010).
doi: 10.1093/bioinformatics/btq401; pmid: 20610611
45. N. Mach et al., Differences in dendritic cells stimulated in vivo
by tumors engineered to secrete granulocyte-macrophage
colony-stimulating factor or Flt3-ligand. Cancer Res. 60,
3239–3246 (2000). pmid: 10866317
46. J. R. Wiśniewski, A. Zougman, N. Nagaraj, M. Mann, Universal
sample preparation method for proteome analysis. Nat. Methods
6, 359–362 (2009). doi: 10.1038/nmeth.1322; pmid: 19377485
47. Y. Kato et al., Display of Native Antigen on cDC1 That Have
Spatial Access to Both T and B Cells Underlies Efficient
Humoral Vaccination. J. Immunol. 205, 1842–1856 (2020).
doi: 10.4049/jimmunol.2000549; pmid: 32839238
ACKN OWLED GMEN TS
We thank S. Choo (Royal Children’s Hospital, Melbourne, Australia)
for human blood samples; A. I. Webb, L. F. Dagley, and G. Infusini
(Advanced Technology and Biology Division, Walter and Eliza Hall
Institute of Medical Research, Melbourne, Australia) for the mass
spectrometry analyses; and S. Finch for reviewing the statistical
analyses. Schematics were created with BioRender.com. Funding:
National Health and Medical Research Council of Australia 1058193
(J.A.V.), National Health and Medical Research Council of Australia
1016629 (W.R.H., J.A.V.), National Health and Medical Research
Council of Australia 1113293 (W.R.H.), Australian Research Council
DP110101383 (J.A.V.), Australian Research Council DP160103134
(J.A.V.), Human Frontier Science Program 0064/2011 (S.I., J.A.V.),
National Health Institute R01 DK125823 (J.M.T., V.M.H.), NIH
R01DK076690 (J.M.T.), and Australian Research Training Program
Scholarship (P.S.). Author contributions: Conceptualization:
J.A.V. Methodology: J.M.T., P.S., A.C.C., N.S.M., J.M., L.B., T.M.S.,
L.M.H., V.M.H., S.I., M.H.L., I.C., and W.R.H. Investigation: P.S.,
A.C.C., N.S.M., J.M., L.B., T.M.S., L.M.H., V.M.H., S.I., M.H.L., I.C.,
and W.R.H. Visualization: P.S., N.S.M., J.D.M., and J.A.V. Funding
acquisition: W.R.H., S.I., and J.A.V. Project administration: J.A.V.
Supervision: J.D.M. and J.A.V. Writing–original draft: P.S., J.D.M., and
J.A.V. Writing–review and editing: P.S., J.D.M., and J.A.V. Competing
interests: J.M.T. receives royalties from Alexion Pharmaceuticals
Inc. and is a consultant for Q32 Bio Inc., a company developing
complement inhibitors. He holds stock and will receive royalty
income from Q32 Bio Inc. The authors declare no other competing
interests. Data and materials availability: RNA-seq data from this
study are deposited in GEO under accession number 185597. All
other data needed to evaluate the conclusions in this paper are
present in the manuscript or the supplementary materials. The
March1–/– and MHC IIKRKI/KI mice used in this study were obtained
from RIKEN Yokohama Institute under the terms of a materials
transfer agreement (MTA) with the University of Melbourne. The
anti-C3 antibodies were used under the terms of an MTA between
the University of Colorado and the University of Melbourne. Patient
samples were obtained under the terms of a research collaboration
agreement between Melbourne Health and the University of
Melbourne. Anti-Clec9A-IEpep and anti-Clec9A-OVA mAbs are
available upon establishment of an MTA with Monash University.
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abf7470
Figs. S1 to S7
Tables S1 to S4
17 November 2020; resubmitted 3 July 2021
Accepted 6 January 2022
10.1126/science.abf7470
Schriek et al., Science 375, eabf7470 (2022)
11 February 2022
12 of 12
|
10.1126_science.abg4020
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
NEUROSCIENCE
Multiscale representation of very large environments
in the hippocampus of flying bats
Tamir Eliav†, Shir R. Maimon†, Johnatan Aljadeff, Misha Tsodyks, Gily Ginosar,
Liora Las, Nachum Ulanovsky*
INTRODUCTION: Place cells are neurons in the
hippocampus that represent the animal’s posi-
tion in space and are important for support-
ing navigation behaviors. These cells increase
their spiking activity when the animal passes
through a specific region of space, called the
neuron’s “place field.” Since the discovery of
place cells half a century ago, nearly all the
research on spatial representations in the mam-
malian brain has focused on rats and mice as
animal models and used small laboratory en-
vironments as experimental setups—usually
small boxes or short linear tracks ~1 to 2 m in
size. In such small environments, individual
place cells typically have one place field, with
a small field size. However, outdoor naviga-
tion of all mammals occurs in natural environ-
ments that span much larger spatial scales,
of hundreds of meters or kilometers, and
nothing is known about the neural codes for
such large spatial scales.
RATIONALE: We reasoned that in very large
environments, the hippocampus must exhibit
a different coding scheme than seen in small
environments because large environments
Question: What is the neural code for very large spaces?
Methods
Bat flying in 200-m-long tunnel with wireless
electrophysiology system
Findings
Individual place-cells in dorsal hippocampus CA1
showed multiple fields with highly variable sizes,
from day 1 in the tunnel
Wireless
neural logger
Cell 1
Cell 2
Cell 3
e
t
a
r
g
n
i
r
i
F
200 m
Localization
antennas
Largest fields ≈ 32m
Smallest fields ≈ 0.6m
0
Position in tunnel (m)
200
Function
Decoding analysis showed that the multifield
multiscale code outperforms classical place-codes
Modeling
Multifield multiscale coding can be explained with 1D
interacting attractor networks and feedforward models
20-fold better
100-fold better
s
n
o
r
u
e
n
d
e
r
i
u
q
e
r
f
o
r
e
b
m
u
N
r
o
r
r
e
g
n
d
o
c
e
D
i
Single
field
Multifield
Multiscale
Single
field
Multifield
Multiscale
g
n
i
t
c
a
r
e
t
n
i
l
e
p
i
t
l
u
M
s
r
o
t
c
a
r
t
t
a
s
u
o
u
n
i
t
n
o
c
g
n
i
r
i
F
e
t
a
r
0
Position in tunnel (m)
200
Multiscale hippocampal spatial code for very large environments. (Methods) We wirelessly recorded
neural activity from hippocampal neurons of bats flying in a 200-m tunnel. (Findings) Single neurons
exhibited multiple place fields with highly heterogeneous field sizes for the same neuron. (Function) This
multiscale neural code for space strongly outperforms classical single-field place codes. (Modeling) Modeling
by using interacting attractor networks and feedforward models recapitulated the multiscale coding.
cannot be tiled fully by the limited number
of hippocampal neurons. We set out to dis-
cover this alternative coding scheme and
thus to close the longstanding gap between
the neurobiology of navigation as studied
in the laboratory and natural large-scale
navigation. To this end, we studied bats
flying in a 200-m-long tunnel while we
recorded the activity of hippocampal dor-
sal CA1 neurons using a custom wireless-
electrophysiology system.
RESULTS: We found that place cells recorded in
the large environment exhibited a multifield,
multiscale representation of space: Individ-
ual neurons exhibited multiple place fields of
diverse sizes, ranging from <1 m to more than
30 m, and the fields of the same neuron could
differ up to 20-fold in size. This multifield,
multiscale code was observed already from
the first day in the environment and was sim-
ilar between wild-born and laboratory-born
bats that were never exposed to large environ-
ments. By contrast, recordings from a small-
scale 6-m environment did not reveal such a
multiscale code but rather classical single
fields. Theoretical decoding analysis showed
major advantages of the multiscale code over
classical single-field codes, both in the num-
ber of required neurons and in the decoding
errors. Thus, the multiscale code provides an
efficient population code with a high capacity
for representing very large environments. We
conducted neural-network modeling, which
suggested that the multiscale code may arise
from interacting attractor networks with mul-
tiple scales or from feedforward networks,
which yielded experimentally testable predic-
tions for the inputs into CA1.
CONCLUSION: Using this experimental setup,
our study uncovered a new coding scheme
for large spaces, which was never observed
before in small spaces: a multiscale code for
space. This coding scheme existed from day 1 in
the environment and was observed in both
wild-born and laboratory-born bats, suggest-
ing that it does not require previous experi-
ence. These findings provide a new notion
for how the hippocampus represents space.
The large naturalistic scale of our experimen-
tal environment was crucial for revealing this
type of code. More generally, this study dem-
onstrates the power of studying brain circuits
under naturalistic conditions.▪
The list of author affiliations is available in the full article online.
*Corresponding author. Email: nachum.ulanovsky@weizmann.
ac.il
†These authors contributed equally to this work.
Cite this article as T. Eliav et al., Science 372, eabg4020
(2021). DOI: 10.1126/science.abg4020
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.abg4020
Eliav et al., Science 372, 933 (2021)
28 May 2021
1 of 1
RES EARCH
R E S E A R C H A R T I C L E
◥
NEUROSCIENCE
Multiscale representation of very large environments
in the hippocampus of flying bats
Tamir Eliav1†, Shir R. Maimon1†, Johnatan Aljadeff1,2, Misha Tsodyks1,3, Gily Ginosar1,
Liora Las1, Nachum Ulanovsky1*
Hippocampal place cells encode the animal’s location. Place cells were traditionally studied in small
environments, and nothing is known about large ethologically relevant spatial scales. We wirelessly
recorded from hippocampal dorsal CA1 neurons of wild-born bats flying in a long tunnel (200 meters).
The size of place fields ranged from 0.6 to 32 meters. Individual place cells exhibited multiple fields
and a multiscale representation: Place fields of the same neuron differed up to 20-fold in size.
This multiscale coding was observed from the first day of exposure to the environment, and also in
laboratory-born bats that never experienced large environments. Theoretical decoding analysis showed
that the multiscale code allows representation of very large environments with much higher precision
than that of other codes. Together, by increasing the spatial scale, we discovered a neural code that is
radically different from classical place codes.
N avigation and spatial memory are cru-
cial for the survival of animals in the
wild. The hippocampal formation con-
tains several types of spatial neurons
whose activity represents the animal’s
position and direction in space (1–10). One of
these spatial cell types is the “place cell,” hip-
pocampal neurons that increase their spiking
activity when the animal passes through a
specific region of space, in turn called the
neuron’s “place field” (1, 2, 11–15). Individual
place cells typically have only one (or two)
place fields in a small environment (2, 11, 16),
whereas multiple place fields are found in
dentate-gyrus neurons upstream (16). Nearly
all of the research on spatial representations
in the mammalian brain has focused on rats
and mice as animal models and used small lab-
oratory environments as experimental setups—
usually small boxes or short linear tracks ~1 to
2 m in size. Consequently, almost all current
knowledge on spatial neurons in the hippocam-
pal formation is based on data from animals
moving in small laboratory environments.
Two studies of place cells examined larger
spatial scales (17, 18). However, these studies
used either a zig-zagging track composed of
~1-m segments or a track that passed through
several small rooms; thus, the largest single-
compartment environment in which place cells
were recorded to date was <10 m in size.
By contrast, outdoor navigation of all mam-
mals occurs in natural environments that span
1Department of Neurobiology, Weizmann Institute of Science,
Rehovot 76100, Israel. 2Section of Neurobiology, Division of
Biological Sciences, University of California, San Diego, CA
92093, USA. 3The Simons Center for Systems Biology,
Institute for Advanced Study, Princeton, NJ 08540, USA.
*Corresponding author. Email: nachum.ulanovsky@weizmann.
ac.il
†These authors contributed equally to this work.
spatial scales much larger than 10 m. For exam-
ple, wild rats were shown to navigate outdoors
>1 km per night (19, 20). Navigation over such
distances requires spatial representation of
very large environments, on the scale of hun-
dreds of meters or kilometers (21). Egyptian
fruit bats fly every night distances of up to
~30 km to their favorite fruit trees, with fly-
ways spanning ~2 km in width and 0.5 km in
height (21, 22). A simple calculation shows that
tiling this space with typical place fields as
measured in the laboratory (~10 to 20 cm di-
ameter, single field per neuron) would require
~1013 neurons. This is ~108 times more neu-
rons than the number of cells in the entire
dorsal hippocampal area CA1 (3), suggesting
that it is simply not feasible to represent such
large spatial scales with laboratory-sized place
fields. Thus, there is a fundamental gap be-
tween the neurobiology of navigation as studied
in the laboratory and kilometer-scale natural
navigation outdoors.
Neural recordings in bats flying in a
200-m environment
We studied wild-born Egyptian fruit bats, a
mammal that has rodent-like hippocampal
spatial representations in small laboratory
environments (23–26). We developed a min-
iaturized wireless neural-logging system that
stores all the data on board (Fig. 1A). This sys-
tem enabled neural recordings to be conducted
over great distances in freely behaving animals,
with uninterrupted experiments lasting up to
~3 hours (27). Using this system, we conducted
tetrode recordings from dorsal CA1, in flight
(Fig. 1, B to D, and fig. S1). We built a 200-m-
long flight tunnel (Fig. 1E), composed of a
long arm and a shorter arm, with landmarks
dispersed along it (fig. S2). We used a medium
light level (5 lux), allowing these bats—which
have excellent vision (21)—to see several distal
landmarks from each location in the tunnel
(fig. S2B). We used a radio frequency–based
localization system, with a small mobile tag
placed on the bat that measured the bat’s dis-
tances to a ground-based antenna array (Fig.
1F). This system yielded a high spatial local-
ization accuracy of ~9 cm (Fig. 1G) along with
a high temporal resolution (27). We harnessed
the natural behavioral tendency of bats to fly
long distances in straight trajectories (22)
and trained them to fly in the tunnel be-
tween two landing balls that were placed at
the two ends of the tunnel, on which food was
given. The bats flew continuously back and
forth between the landing balls (fig. S3A).
Flight trajectories were rather stereotyped,
with bats flying at the center-top portion of
the tunnel, with only very small deviations
perpendicular to the flight direction (Fig. 1H
and fig. S3, B and C). Thus, the bats exhibited
nearly perfect one-dimensional (1D) back-and-
forth trajectories. Hence, in all subsequent
analyses, we projected the behavioral data
onto the main axis of the tunnel and included
only long unidirectional flights that were
>100 m in length (27). This 1D tunnel bears sim-
ilarities to bats’ natural behaviors because
these bats navigate underground in 1D cave
tunnels, and also their flight trajectories out-
doors are largely 1D (22). Flight speed was high
and showed very little variation across differ-
ent locations (Fig. 1, I and J). Bats flew dozens
of flights per direction in each recording ses-
sion (Fig. 1K), covering on average 14.1 km per
session and up to 22.5 km in a single session
(Fig. 1L).
Hippocampal place cells exhibit a multifield,
multiscale spatial code
We recorded 235 well-isolated putative pyram-
idal cells from the dorsal CA1 of five bats; all
235 neurons were active in flight, and 83.4%
of them (n = 196) were place cells, showing
significant spatial tuning with distinct and
stable place fields (Fig. 2A and figs. S4 and S5;
the numbers of place cells in individual bats
are provided in table S1) (27). By contrast, in
both rodents and bats, the reported percent-
age of place cells in small environments is
typically 30 to 40% of all the recorded cells,
whereas the remaining cells are virtually
silent during behavior (11, 23, 24, 28). Place
cells in the 200-m tunnel exhibited strong
spatial tuning (Fig. 2, B to D), and the spatial
tuning was stable across flights (Fig. 2E). The
place cells fired differently in different flight
directions (Fig. 2, A, red and blue raster plots,
and F, map correlations between directions),
similar to the directionality shown previously
for place cells in rats and bats in small 1D en-
vironments (29, 30). However, we found seve-
ral surprising characteristics of place cell firing
Eliav et al., Science 372, eabg4020 (2021)
28 May 2021
1 of 12
RES EARCH | R E S E A R C H A R T I C L E
A
D
E
F
10mm
CA1
1mm
20m
B
Spikes recorded in-flight
C
900
)
V
(
4
h
C
V
0
0
5
20ms
G
0.06
y
t
i
l
i
b
a
b
o
r
P
0
0
Ch3 ( V)
400
1300
Ch2 ( V)
0
50
100
Position (m)
150
200
I
)
s
/
m
(
d
e
e
p
s
t
h
g
i
l
F
8
0
0
0
-40 -20 0 20 40
Localization error (cm)
20m
Antenna
Bat
North
H
Z
J
i
s
n
o
s
s
e
s
f
o
.
o
N
20
10
0
K
i
s
n
o
s
s
e
s
f
o
.
o
N
20
10
0
L
i
s
n
o
s
s
e
s
f
o
.
o
N
20
10
0
Y
50cm
0
0.04 0.08
CV of speed
0
30
60
No. of flights
0 5 10 15 20 25
Distance flown (km)
Fig. 1. Neuronal and behavioral recordings from bats flying over large
spatial scales. (A) Sixteen-channel wireless neural logger. (B) Neural traces
from one tetrode, recorded in bat dorsal hippocampal area CA1, showing spikes
in-flight. (C) Spike-sorting of one tetrode (data from full session, 108 min).
Shown are spike clusters from different neurons, with spike amplitudes plotted
for three of the tetrode’s channels; well-isolated units are shown in different
colors. Same session and tetrode as in (B). (D) Histology of one recording site in
dorsal CA1. Red arrowhead, electrolytic lesion; Black lines, proximal and distal
borders of CA1. (E) Aerial photograph showing top-view of the large-scale
environment. The flight-tunnel was composed of long and short arms (27), which
the bat traversed without slowing down (I). Vertical lines indicate location
where neural data in (B) were recorded. (F) Localization system, showing
positions of ground-based antennas (red dots), the tunnel (dark gray thick line),
snapshot of measured distances from each antenna to the localization tag
on the bat’s head (large black circles; cropped for visualization purposes), and
the bat’s estimated location (blue dot; computed as the intersection of the
black circles). (G) Precision of the localization system (27), showing localization
error of s = 8.9 cm along the tunnel’s major axis. (H) Example session,
showing the y-z positions of the bat’s passages (blue dots) through a cross-
section in the tunnel’s midpoint (black outline). There are relatively small
deviations of the blue dots in the y and z axes, indicating the bat flew essentially
in 1D trajectories (fig. S3, B and C). (I) Example session showing speed
profiles along the tunnel, pooled over both flight directions. Gray areas indicate
locations of low flight speeds, owing to takeoff and landing, which were
removed from further analysis of place fields (27). (J) Distribution of the
coefficient of variation (CV) of the flight speed per session (n = 60 sessions;
five bats). The CV was computed over the tunnel’s high-speed portion [excluding
the gray areas from (I)]; mean CV = 0.042. (K) Distribution of number of
flights (laps) per direction per session. Shown are only valid unidirectional long
flights, longer than 100 m (27). Red and blue colors in (K) and (J) indicate
the two flight directions (arrows). (L) Distribution of total distance flown per
session, based on valid long flights only (n = 60 sessions; five bats).
in our 200-m environment. First, unlike the
typical single place field reported for CA1
neurons in small environments (11), we found
that many cells exhibited multiple place fields
(Fig. 2A and fig. S5, examples; Fig. 2G, popu-
lation). The mean number of fields per direc-
tion was 4.9, and some neurons had more than
10 fields in each flight direction (Fig. 2G). This
result extends similar findings in enlarged en-
vironments in rodents, which showed several
fields per neuron (18, 31, 32). The fields were
strongly tuned and contained the large ma-
jority of the neuron’s spikes; the background
firing was relatively low (fig. S6, A and B).
Second, many cells had very large place fields,
often >10 m in size, and up to 32 m (Fig. 2, A,
cells 1 and 5, examples, and H, population; and
figs. S5 and S7A). On the other hand, some
cells had very small place fields of <1 m in size
and down to 0.6 m (Fig. 2, A, cells 3 and 7,
zoom-in, and H, leftmost bar). The distribu-
tion of field sizes was skewed (Fig. 2H) and
was well-fitted by a log-normal distribution
(fig. S8) (33). Third, and most surprisingly,
many place cells showed highly variable field
sizes, with up to 20-fold ratio between the size
of the largest and smallest field for the same
neuron (Fig. 2A, cells 1 to 7, examples; and
Fig. 2, I and J, population; mean ratio, 4.4).
This multifield, multiscale code was found
Eliav et al., Science 372, eabg4020 (2021)
28 May 2021
2 of 12
A
e
t
a
r
g
n
i
r
i
F
22
)
z
H
(
0
57
.
o
n
t
h
g
i
l
F
58
1
0
e
t
a
r
g
n
i
r
i
F
27
)
z
H
(
0
41
.
o
n
t
h
g
i
l
F
40
1
0
e
t
a
r
g
n
i
r
i
F
9
)
z
H
(
0
38
.
o
n
t
h
g
i
l
F
38
1
0
B
s
l
l
e
c
f
o
.
o
N
10 2
10 1
10 0
RES EARCH | R E S E A R C H A R T I C L E
Cell 1
max=20.0m
min=3.9m
ratio=5.1
35
10
7
0
38
Cell 2
16
max=9.5m
min=1.7m
ratio=5.6
0
3
1 38
Cell 3
max=11.9m
min=1.0m
ratio=11.5
15
15
1.0m
50
100
150
200
Cell 4
max=9.9m
min=2.3m
ratio=4.3
38
1
0
30
38
1
0
24
50
100
150
200
Cell 5
max=31.3m
min=4.9m
ratio=6.4
50
100
150
200
Cell 6
max=17.1m
min=1.7m
ratio=10.2
2
0
0
45
7
8
0
23
5
3
50
100
150
200
0.9m
Cell 7
max=9.0m
min=0.9m
ratio=9.6
44
1
0
31
23
1
0
38
50
100
150
200
Cell 8
max=5.1m
min=3.3m
ratio=1.6
50
100
150
200
Cell 9
single field=4.1m
9
7
0
32
2
4
0
22
1
0
50
100
Position (m)
150
200
C
s
l
l
e
c
f
o
.
o
N
10 2
10 1
10 0
H
s
d
e
i
f
l
f
o
.
o
N
10 2
10 1
10 0
5
0
Spatial information
(bits/spike)
G
s
l
l
e
c
f
o
.
o
N
10 2
10 1
10 0
0
10
No. of fields per direction
20
0
10
20
30
Field size (m)
32
1
0
D
s
l
l
e
c
f
o
.
o
N
10 2
10 1
10 0
21
1
0
50
100
Position (m)
150
200
50
100
Position (m)
150
200
Coverage (%)
0
30
60
90
E
s
l
l
e
c
f
o
.
o
N
150
100
50
0
Stability
-1 -0.5
0
Map correlation
0.5
F
y
t
i
l
i
b
a
b
o
r
P
n
o
i
t
c
n
u
f
y
t
i
s
n
e
d
4
2
0
1
Directionality
P KS = 0.12
-1 -0.5
0
Map correlation
0.5
1
I
)
m
(
e
z
s
i
l
d
e
F
i
30
20
10
0
J
s
l
l
e
c
f
o
.
o
N
10 1
10 0
Smallest
field
Largest
field
1
2
5 10 20
Field size ratio
largest/smallest
K
o
i
t
a
r
e
z
s
i
l
d
e
F
i
20
15
10
5
3
2
1
= 0.03
P = 0.67
0.8
1
Speed ratio
1.2
0
0.5
Sparsity
1
0
50 100 150
Coverage (m)
Fig. 2. Dorsal CA1 hippocampal neurons represented very large space
using many fields with multiscale coding. (A) Examples of firing-rate maps
and raster plots for nine cells. For each neuron, (Top) firing-rate maps calculated
separately for each flight direction (red and blue; arrows above cell 1); (Bottom),
raster plots of spike positions (x axis) for different flights, or laps (y axis); the
detected place-fields are indicated with red and blue thick horizontal lines above
the raster plots [fields inside the low-flight-speed zones (gray) were excluded
(27)]. In each example, the smallest and largest field sizes are indicated (min,
max), together with the ratio between them; the numbers of fields in each direction
are indicated in blue and red on the right. For cells 3 and 7, shown also are
zoom-ins on their smallest field (cell 3, field size 1.0 m; cell 7, field size 0.9 m).
(B to D) Distribution of (B) spatial information, (C) sparsity, and (D) the total
coverage of the environment by place fields, calculated for the firing-rate map
in each flight direction separately (“No. of cells” here refers to significant
Eliav et al., Science 372, eabg4020 (2021)
28 May 2021
3 of 12
RES EARCH | R E S E A R C H A R T I C L E
cells × directions; n = 331). (D) Bottom x axis, total coverage in meters; top x axis,
total coverage in percent of tunnel length. (E) Distribution of firing-map correlations
between odd and even flights (n = 331 cells × directions), showing high correlation
values [median correlation coefficient (r) = 0.87]. (F) Distribution of firing-map
correlations between the two flight directions (gray; n = 135 cells, including only cells
where both directions were significantly tuned) was similar to cell-shuffled distribution
(black) [Kolmogorov-Smirnov test: P = 0.12 (DKS 135,13566 = 0.10)]. (G) Distribution
of number of place fields per neuron per flight direction (n = 331 cells × directions).
Rightmost bar, cases with ≥20 fields per direction. Mean number of fields per direction
was 4.9. (H) Distribution of place field size (n = 1629 fields). The field size ranged from
<1 m (leftmost bar of histogram) up to 32 m. (I and J) Single cells exhibited multiscale
field sizes (plotted are n = 172 cells with ≥2 fields). (I) Distributions of smallest
and largest field sizes per neuron (shown cells with ≥2 fields). (J) Distribution of the
ratio between largest and smallest field sizes for each neuron. Both axes here are in
log scale. (K) Lack of correlation between largest-to-smallest field size ratio and
the speed ratio at the locations of those fields (plotted are n = 172 cells with ≥2 fields).
For all histograms except (F), the red vertical line indicates mean of distribution, and the
red dot and red horizontal line indicate median and interquartile range, respectively.
in all the five individual animals (fig. S9).
Although most cells showed heterogeneous
field sizes, some neurons also exhibited a more
uniform scale across their place fields (Fig. 2A,
cell 8), and a small minority of neurons had a
single place field (Fig. 2A, cell 9) (only 12.2%
of the neurons had one field overall, with an
average field size of 5.9 ± 3.5 m; mean ± SD).
Individual neurons exhibited similar multi-
scale firing properties in both flight directions:
a similar number of fields per direction, sim-
ilar median field-size, and similar field size
ratios (fig. S10). This suggests a characteristic
firing propensity per neuron (34) while still
exhibiting widely varying field sizes. Taken
together, most neurons exhibited these two
key properties: many fields per neuron (Fig.
2G) and a multiscale mixture of small fields
and large fields for the same neuron (Fig. 2J).
We next examined several possible alterna-
tive explanations for the multiscale code that
we observed. First, the multiscale property
could not be explained as arising from varia-
tions in flight-speed—for example, larger fields
at high flight speeds—because the flight speed
was in fact highly consistent along the entire
tunnel (Fig. 1, I and J). Further, the field-size
ratio (largest/smallest fields per neuron) was
not correlated with the speed ratio at the
locations of the largest and smallest fields
(Spearman r = 0.03, df = 170, P = 0.67) (Fig.
2K). Moreover, the speed ratio was narrowly
distributed around 1, indicating that the speed
was similar at large and small fields (Student’s
t test of speed ratio versus 1: t = 0.83, df = 171,
P = 0.41; SD of the speed ratio was 0.10) (Fig.
2K). Second, the multiscale property also could
not be explained by systematic differences in
field sizes in the long versus short arms of the
tunnel because we found no significant dif-
ference in field sizes between the two arms
[Kolmogorov-Smirnov test comparing field
sizes in the long versus short arm: P = 0.60 and
0.12 for the two flight-directions (DKS 586,180 =
0.06 and DKS 628,235 = 0.09)], and there was no
significant difference in field sizes between
the long arm and the full tunnel [Kolmogorov-
Smirnov test: P = 0.96 (DKS 1214,1629 = 0.02)]
(fig. S7A). We also found multiscale coding
when restricting the analysis only to the long
arm (fig. S7B). Third, the multiscale property
did not stem from an unusual recording lo-
cation within CA1. All the recordings were
done in the dorsal part of the hippocampus and
spanned rather central proximo-distal locations
in CA1 (fig. S1, A and B); these are the classical
recording-locations used in rodents and bats in
small laboratory setups. Fourth, the multiscale
property of CA1 neurons could not be explained
by spike-sorting quality (fig. S11). Fifth, the re-
sults were robust to the detailed criteria of field
detection (fig. S12).
We then looked for possible contributions
of landmarks to the multiscale code. First,
we considered several landmark-based com-
partmentalization models of the environment,
in which the tunnel is assumed to be seg-
mented into smaller portions at the landmark
locations, allowing fields to merge at the seg-
ment borders (27). These models could not
explain the wide distribution of place field
sizes observed in the data (fig. S13). Second,
we examined the possibility that the multi-
scale code could be explained by an over-
representation (concentration) of place fields
near the landmarks, and in particular small
place fields. However, the cumulative distri-
bution of field locations was linear as a func-
tion of position along the tunnel (Fig. 3A),
with no apparent overrepresentation near
landmarks [but with an overrepresentation
of fields at the two ends of the tunnel, in the
reward areas (fig. S14)]. We computed the dis-
tance of each field’s peak to its nearest land-
mark and compared the distribution of these
distances to the distribution of distances for
shuffled place field locations (27); we found
no significant difference between the two
(Kolmogorov-Smirnov test, P ≥ 0.18 for both
directions) (Fig. 3B), indicating that place fields
did not concentrate near landmarks but were
distributed rather uniformly along the tun-
nel. This uniform distribution was supported
also by an analysis of the gaps between fields,
which showed an exponential distribution
(Fig. 3C), indicating lack of structure in the
spatial arrangement of place fields. Addition-
ally, the entire range of field sizes was repre-
sented rather uniformly along the tunnel, with
no prominent concentration of small (or large)
fields near landmarks (Fig. 3, D and E, and fig.
S15)—likely because of the low saliency of
these landmarks for the bats—except for a few
landmarks that possibly showed slight con-
centration of fields (Fig. 3D). Further, there
was no strong relation between the interland-
mark distance and the field size (fig. S15B)
(however, this does not rule out that very large
fields would be found in extremely impover-
ished large regions of space, where absolute
spatial information is not available over long
distances). Together, these analyses suggest
that the multiscale statistics were not driven
by landmarks.
Comparison between large and
small environments
To examine whether multiscale coding may
be found also in small environments, we re-
corded from the dorsal CA1 of an additional
three bats flying in a short 6-m segment of
the tunnel, which we blocked off (table S1,
dataset 2) (Fig. 4A). This allowed testing di-
rectly the effect of environment size on the
spatial coding of neurons in the dorsal CA1
of bats, using the same experimental design.
The percentage of neurons that were active
during flight in the short 6-m tunnel (36 of
67 cells, 53.7%) was much smaller than in the
full 200-m tunnel (235 of 235 cells, 100%)
(table S1) (27). The majority of the active cells
were significant place cells (30 of 36, 83.3%);
thus, almost half of the neurons recorded in
the 6-m tunnel were significant place cells
(30 of 67 cells, 44.8%). Next, we systematically
compared the spatial tuning properties of cells
in the large versus small environments (Fig. 4,
B to G). In the 6-m small environment, dorsal
CA1 place cells showed only one or two place
fields (Fig. 4, A and B, bottom), in contrast to
the high number of place fields observed in
the large 200-m environment (Fig. 4, B, top,
and E). Across cells, the place field sizes in the
small environment were much smaller than in
the large environment (Fig. 4, C and F). At the
single-cell level, neurons in the small environ-
ment had a significantly lower ratio between
their largest and smallest fields as compared
with that in the large environment (Fig. 4, D
and G). Thus, neurons in the small environ-
ment showed virtually no multiscale coding.
Multiscale coding of space is independent of
both early and recent experience
Does multiscale coding of large environments
emerge over time, as a function of experience?
First, we asked whether prior experience in
the long tunnel is needed for the multiscale
code. We conducted recordings of place cells
Eliav et al., Science 372, eabg4020 (2021)
28 May 2021
4 of 12
RES EARCH | R E S E A R C H A R T I C L E
Landmarks
A
1
l
e
v
i
t
a
u
m
u
C
n
o
i
t
c
a
r
f
0
0
1
l
e
v
i
t
a
u
m
u
C
n
o
i
t
c
a
r
f
0
0
D
20
)
m
(
e
z
s
i
l
d
e
F
i
50
100
Position (m)
150
200
50
100
Position (m)
150
200
B
y
t
i
l
i
b
a
b
o
r
P
y
t
i
l
i
b
a
b
o
r
P
0.1
n
o
i
t
c
n
u
f
y
t
i
s
n
e
d
0
0.1
n
o
i
t
c
n
u
f
y
t
i
s
n
e
d
0
E
0.2
y
t
i
l
i
b
a
b
o
r
P
n
o
i
t
c
n
u
f
y
t
i
s
n
e
d
C
y
t
i
l
i
b
a
b
o
r
P
n
o
i
t
c
n
u
f
y
t
i
s
n
e
d
10 -1
10 -2
10 -3
10 -4
Data
Exponential fit
50
0
Gap between fields (m)
100
P
KS
= 0.82
Data
Shuffle
-10
0
Distance of fields
to nearest landmark (m)
10
P
KS
= 0.18
Data
Shuffle
-10
0
Distance of fields
to nearest landmark (m)
10
P
KS
= 0.80
Distance > 5m
Distance < 5m
0
0
20
)
m
(
e
z
s
i
l
d
e
F
i
50
100
Position (m)
150
200
0
0
10
Field size (m)
20
P
KS
= 0.25
Distance > 5m
Distance < 5m
0.2
n
o
i
t
c
n
u
f
y
t
i
s
n
e
d
y
t
i
l
i
b
a
b
o
r
P
0
0
50
100
Position (m)
150
200
0
0
10
Field size (m)
20
Fig. 3. Place fields were distributed uniformly along the tunnel.
(A) Cumulative fraction of peak firing-rate locations for all the place fields
along the tunnel, pooled across all the five bats and 196 place cells;
plotted for each flight-direction separately (East direction, blue, n = 863 fields;
West direction, red, n = 766 fields). Gray vertical lines, locations of landmarks
(we did not treat the landing balls as “landmarks”). (B) Distributions of the
distances of each field’s peak to its nearest landmark (blue and red, flight
directions) were similar to shuffle distributions (black) [Kolmogorov Smirnov
test, P = 0.82 (DKS 782,7820000 = 0.02) and P = 0.18 (DKS 661,6610000 = 0.04)
for the two flight directions]. (C) Distribution of gaps between fields
(gray bars), overlaid with exponential fit (black line), plotted on a logarithmic
y scale. The good fit to the exponential distribution indicates lack of
spatial structure in the field locations. (D) Field size versus the location
of field peak, pooled across all bats and neurons. Gray vertical lines, locations
of landmarks; open circles, fields larger than 20 m. The entire range of
field sizes was represented along the entire tunnel. (E) Distribution of field
size for the two directions (blue and red), plotted separately for fields
located close to landmarks (thin line, fields <5 m from nearest landmark)
or far from landmarks (thick line, fields ≥ 5 m from nearest landmark).
No significant differences in field-size were found between fields located close
or far from a landmark [Kolmogorov Smirnov test, P = 0.80 (DKS 577,205 =
0.05) and P = 0.25 (DKS 469,192 = 0.09) for the two flight directions]. In
(B) and (E) we excluded fields whose peak occurred before the first landmark
or after the last landmark in the tunnel, where the assignment of “nearest
landmark” is one-sided and hence biased [the same was done for the
shuffles in (B)].
Eliav et al., Science 372, eabg4020 (2021)
28 May 2021
5 of 12
RES EARCH | R E S E A R C H A R T I C L E
Small environment (6 m)
13
1
0
0
36
37
5
6
1
0
1
n = 331
0
5
10 15 20
n = 40
A
e
t
a
r
g
n
i
r
i
F
)
z
H
(
.
o
n
t
h
g
i
l
F
24
0
36
37
1
0
1
2
3
Position (m)
4
Large
environment
(200 m)
Small
environment
(6 m)
B
s
l
l
e
c
f
o
.
o
N
10 2
10 1
10 0
s
l
l
e
c
f
o
.
o
N
10 1
10 0
6
0
1
0
36
37
5
6
1
0
1
2
3
Position (m)
4
2
3
Position (m)
4
3
1
1
0
41
41
5
6
1
0
1
11
0
1
0
27
26
5
6
1
0
1
2
0
5
6
2
3
Position (m)
4
2
3
Position (m)
4
C
s
d
e
i
f
l
f
o
.
o
N
s
d
e
i
f
l
f
o
.
o
N
10 2
10 1
10 0
10 1
10 0
E
n = 196
n = 1629
D
s
l
l
e
c
f
o
.
o
N
10 1
10 0
0
10
20
30
1
2
5 10 20
n = 45
101
100
0 1 2 3
s
l
l
e
c
f
o
.
o
N
10 1
10 0
n = 30
F
n
o
i
t
c
e
r
i
d
r
e
p
s
d
e
i
f
l
f
o
.
o
N
)
m
i
(
e
z
s
d
e
F
i
l
*****
*****
6
4
2
0
6
4
2
0
0
5
10 15 20
No. of fields per direction
0
10
20
30
Field size (m)
1
2
5 10 20
Field size ratio
largest/smallest
Fig. 4. No multiscale coding was found in a small-scale environment. Experiments were done in a small 6-m segment of
the long 200-m tunnel, which we blocked with two curtains (table S1, dataset 2). (A) Examples of firing-rate maps and raster
plots for dorsal CA1 place cells recorded in this small-scale environment. Same graphical conventions as in Fig. 2A. Most
cells had a single field per direction, or two fields with a similar scale. (B to D) Distributions of (B) number of fields, (C) field
sizes, and (D) field size ratio for the two different dorsal CA1 datasets: large-scale 200 m environment (top row) and small-scale
6 m environment (bottom row). Red vertical lines, mean of distribution; red horizontal lines and red dot, interquartile range
and median, respectively. (C) (Inset) Zoom-in. Black bars in (D) show neurons with one field. (E to G) Comparison between
the two datasets (large versus small environment), for the three quantities: (E) number of fields, (F) field sizes, and (G) field size
ratio (here, we included only neurons with ≥2 fields). Error bars, mean ± SEM. There was a highly significant difference in all of
these three quantities between large-scale and small-scale environments, indicating that the multifield, multiscale coding is
expressed most prominently in large-scale environments. (E) Student’s t test with unequal variances, P = 6.7 × 10−44, t = 15.94;
Wilcoxon rank-sum test, P = 1.5 × 10−15, z = 7.89. (F) Student’s t test with unequal variances, P = 4.5 × 10−58, t = 29.56; Wilcoxon rank-sum test, P = 3.9 × 10−26, z = 10.51.
(G) Student’s t test with unequal variances, P = 1.8 × 10−14, t = 8.91; Wilcoxon rank-sum test, P = 4.6 × 10−5, z = 3.91; ***** P < 10−5 for the t tests in (E) to (G).
Large env.
Small env.
o
i
t
a
r
e
z
s
d
e
F
a
m
s
/
t
s
e
g
r
a
t
s
e
3
4
1
2
l
l
l
i
l
i
G
5
*****
from the first exposure to the novel large
environment. We recorded 125 place cells from
two bats flying in a 130-m portion that was
blocked out of the 200-m tunnel, with neural
recordings commencing from the very first
day in the tunnel (day 1) and continuing over
several weeks (with new cells being recorded
every day) (table S1, dataset 3). Cells were
spatially tuned already in the first sessions
and exhibited many place fields with differ-
ent sizes (Fig. 5A). The multifield, multiscale
properties were seen from day 1 and were
stable across several weeks of recordings,
showing no significant trend in the number
of fields (Fig. 5B), field sizes (Fig. 5C), or field-
size ratio (Fig. 5D) [overall, place fields in the
130-m tunnel exhibited somewhat smaller
numbers of fields, field sizes, and field-size
ratios as compared with those of the 200-m
tunnel (Fig. 5, B to D, bars on the right)].
This suggests that the multiscale coding does
not require substantial recent experience with
the long tunnel. Although the general multi-
scale properties were stable over days (Fig. 5,
B to D), the cells occasionally exhibited within-
day dynamics in the form of fields appearing
and disappearing (Fig. 5E). The rate of within-
day changes was larger during the first 2 days
of the bat in the tunnel (two-proportion z test,
P < 0.001) (Fig. 5F) but also occurred many
days after the first exposure (Fig. 5, E, cells 7
and 8, and F), which is consistent with pre-
vious findings in mice of ongoing changes in
the tuning of place cells (35, 36).
Second, we asked whether laboratory-born
bats that were never exposed to large environ-
ments would lack a multiscale code. We re-
corded from an additional three adult bats
that were born in the laboratory and grew up
in an enriched environment but had never ex-
perienced during development any large-scale
environments bigger than a few meters (table
S1, dataset 4, and fig. S16) (27)—in contrast to
the wild-caught bats that navigated long dis-
tances outdoors during development (37). The
laboratory-born bats were trained to fly in
the 200-m tunnel for several weeks and were
thus familiar with the environment before the
Eliav et al., Science 372, eabg4020 (2021)
28 May 2021
6 of 12
RES EARCH | R E S E A R C H A R T I C L E
A
e
t
a
r
g
n
i
r
i
F
.
o
n
t
h
g
i
l
F
0
11
13
1
0
Examples of multiscale coding in the first days of exposure to the tunnel
Cell 1 - Day 1
Cell 2 - Day 2
21
)
z
H
(
max=11.8m
min=2.2m
ratio=5.5
2
2
22
0
23
22
max=8.1m
min=2.5m
ratio=3.2
3
6
43
0
24
23
Cell 3 - Day 5
max=7.2m
min=1.8m
ratio=3.9
2
2
38
0
34
34
Cell 4 - Day 6
max=8.5m
min=1.1m
ratio=7.7
3
6
Position (m)
130
1
0
Position (m)
130
1
0
Position (m)
130
1
0
Position (m)
130
Population: Stability of multiscale coding across weeks, starting from day 1 in the tunnel
B
s
d
e
l
i
f
f
o
.
o
N
n
o
i
t
c
e
r
i
d
r
e
p
20
15
10
5
n = 188
= 0.02
P = 0.81
C
)
m
(
e
z
s
i
l
d
e
F
i
n = 611
= 0.06
P = 0.13
25
20
15
10
5
D
o
i
t
a
r
e
z
s
i
l
d
e
F
i
= -0.04
P = 0.75
10
n = 83
t
s
e
l
l
a
m
s
/
t
s
e
g
r
a
l
8
6
4
2
0
10
20
Day
30
m
6
m
0
3
1
m
0
0
2
0
10
20
Day
30
m
6
m
0
3
1
m
0
0
2
0
10
20
Day
30
m
6
m
0
3
1
m
0
0
2
E
Examples of within-day dynamics in spatial tuning
Cell 5 - Day 1
Cell 6 - Day 3
Cell 7 - Day 31
Cell 8 - Day 31
24
25
.
o
n
t
h
g
i
l
F
38
40
30
31
30
31
1
0
Position (m)
130
1
0
Position (m)
130
1
0
Position (m)
130
1
0
Position (m)
130
F
0.2
****
*****
Population: Within-day dynamics is most prominent
during the first 2 days, but occurs also later
Fig. 5. Multiscale coding exists already from the first days of exposure to
the tunnel. Experiments from day 1 were conducted in a 130-m portion of
the 200-m tunnel (table S1, dataset 3). (A) Examples of firing-rate maps and
raster plots for four cells recorded in a large-scale environment during the
first days of exposure. Same graphical conventions as in Fig. 2A. The days since
first exposure are indicated for each cell (day 1 is the very first day of exposure;
day count represents experimental days). (B to D) Population scatter plots of
(B) the number of fields per direction, (C) field sizes, and (D) field size ratio as a
function of days since first exposure. There is a lack of trend across days
(Spearman r, P > 0.13 for all three scatters), suggesting that the multiscale coding
exists already from day 1. For display purposes only, dots were jittered along the x
axis (uniform jitter of ±0.5 days); in (B), dots were jittered also along the y axis
(uniform jitter of ±0.3 fields); all correlations were computed without the jitter.
Error bars in main plots, mean ± SD (using 5-day bins with no overlap). (Insets) Gray bars are
mean ± SEM for the three tunnel lengths used in this study: 6, 130, and 200 m. (E) Four
examples of within-day dynamics in spatial tuning. Raster plots show spike positions in each flight
(blue and red dots indicate two flight directions), with the behavioral coverage shown with light gray lines. Arrowheads
indicate field appearance (filled arrowheads) or disappearance (empty arrowheads). These dynamics occurred in both small and large fields and happened both on the first
days of exposure (cells 5 and 6) and after ≥1 month (cells 7 and 8). (F) Probability of appearance and disappearance of fields (per-flight probability of change in any of the fields), grouped
by the day from first exposure: days 1 and 2, days 3 and 4, days 5 and 6, and ≥7 days. Error bars, mean ± standard error of the proportion (27). In the first 2 days after
exposure, the cells exhibited a higher probability of appearance or disappearance of fields than on later days (two-proportion z test, P < 0.001 for all six tests comparing days 1 and
2 versus the other days). The probabilities for appearance and disappearance were similar over the entire course of exposure (compare black versus white bars; two-proportion z test:
P = 0.64, pooled over all days), which is consistent with the overall stability over weeks in the number of fields per neuron (B). *****P < 10−5, ****P < 10−4, ***P < 10−3.
Appearance
Disappearance
7
Days
Days 3-4
Days 5-6
Day 1-2
b
a
b
o
r
P
***
y
t
i
l
i
0
Eliav et al., Science 372, eabg4020 (2021)
28 May 2021
7 of 12
RES EARCH | R E S E A R C H A R T I C L E
neural recordings, similar to the wild-born
bats (Fig. 6A). The laboratory-born bats were
in good flight shape and flew similar distances
in the tunnel as those of the wild-born bats
(fig. S16B). Thus, the main difference between
the laboratory-born and wild-born bats was
their experience during development, with all
other experimental conditions being kept iden-
tical (Fig. 6A) (27). We recorded 113 cells in the
dorsal CA1 of the laboratory-born bats, out of
which 95 were place cells (84.1%), which is
very similar to the percentage of place cells
in wild-born bats (83.4%). The place cells of
laboratory-born bats showed a multifield, mul-
tiscale code, with individual neurons exhibit-
ing many fields with varying sizes per-neuron
(Fig. 6B, examples; Fig. 6, C to E, population),
which is similar to place cells recorded from
the wild-born bats. We then compared the
multiscale properties between the two groups
(Fig. 6, C to H): (i) The number of fields per
direction was not significantly different (Fig.
6, C and F). (ii) Both groups exhibited wide
distributions of place field sizes, with wild-born
bats having slightly larger fields (Fig. 6, D and
G) [this difference was not due to differences
in dorso-ventral recording positions along the
longitudinal axis of CA1, which were very sim-
ilar in both groups (Fig. 6I), but could be due
to the slightly different recording positions
along the proximo-distal axis of CA1 (Fig. 6I)].
(iii) The field-size ratio was not significantly
different between the groups (Fig. 6, E and H),
despite the difference in field sizes—indicating
a similar multiscale code between laboratory-
born and wild-born bats.
The multiscale code yields substantial
advantages for large environments
We next turned to a theoretical analysis to
understand the possible functional advan-
tage of the multiscale representation of large
environments. We compared the perform-
ance of six spatial encoding schemes (Fig. 7A)
(27): (scheme 1) a single small place field per
neuron; (scheme 2) a single large place field
per neuron; (scheme 3) a single place field
with a gradual increase in field size across the
population, mimicking the dorso-ventral ana-
tomical gradient of field sizes in the hippo-
campus (17); (scheme 4) multiple small fields
per neuron, identical in size for all the neurons
(18); (scheme 5) multiple fields per neuron, all
with the same size for each neuron but with
different scales across different neurons; and
(scheme 6) multiple fields with multiscale
coding per neuron, as in our data. The distri-
bution of field sizes for schemes 5 and 6 was
matched to our data [field sizes were drawn
from a g-distribution fitted to the data (fig.
S8) (27); the field-size ratio for scheme 6 also
closely matched the data (fig. S17G); variants
of schemes 5 and 6 in which we matched also
the total coverage of fields to the data are
shown in fig. S17]. We used two types of
decoders, a Bayesian maximum-likelihood
decoder (Fig. 7) and a population-vector de-
coder (fig. S18), and two integration time win-
dows, Dt = 500 ms (Fig. 7) and Dt = 200 ms (fig.
S19) (27). We compared the decoding error
of simulated data for each of these six encod-
ing schemes for progressively larger environ-
ments. For small environments, all six encoding
schemes performed qualitatively equally well,
but for very large environments (hundreds of
meters), the experimentally observed encoding
scheme with multiscale place fields substan-
tially outperformed the other schemes (Fig. 7,
B to E, and fig. S18, B to E). Specifically, for
encoding schemes with either a single field
(schemes 1, 2, or 3) or multiple fields of small
size (scheme 4), the number of neurons re-
quired to accurately decode the animal’s posi-
tion was extremely large for large environments
(Fig. 7B, left; the red, green, pink, and yel-
low lines go out of bounds). By contrast, the
two schemes with multiple fields of vary-
ing sizes (schemes 5 and 6) required only
~50 neurons for accurately decoding the bat’s
position, even in a very large environment of
1000 m in size (Fig. 7B, left; a 2-m decoding
accuracy). Furthermore, the mean decoding
error for schemes 1 and 4 increased dramatically
for large environments (Fig. 7C, red and green),
but for schemes 5 and 6, the mean decoding
error barely increased as a function of the envi-
ronment size (Fig. 7C, inset, blue and purple),
maintaining a small decoding error of 5 to 10 m
for a 1000-m environment, even for a very small
ensemble of 50 neurons (Fig. 7C, inset). We thus
conclude that encoding schemes 1 to 4 are less
suitable for very large environments.
Next, we asked whether scheme 6, which
closely matches our experimental results, offers
any functional advantage over scheme 5. We
reasoned that scheme 5, in which all the fields
of the same neuron have the same field size, is
problematic because when a neuron emits a
spike, it could mean that the animal is located
in any of the neuron’s fields; this creates large
positional ambiguity. By contrast, scheme 6, in
which each neuron has multiscale fields, alle-
viates this problem because the neuron’s spike
count during an integration time Dt differs
between different fields—the neuron produces
many spikes in large fields but only a few
spikes in small fields—and this variability
serves to disambiguate which field the ani-
mal passed through; this in turn improves the
decoding accuracy. For large, 1000-m environ-
ments, the mean decoding error was substan-
tially smaller for scheme 6 than for scheme 5
(Fig. 7C, inset, compare purple and blue lines).
Moreover, scheme 6 led to much smaller and
fewer catastrophic decoding errors (Fig. 7, D
and E, compare purple and blue lines) [There is
a ~10-fold difference in the size of catastrophic
decoding errors, defined as the 99th percentile
of the decoding errors (Fig. 7D, inset) and an
approximately two- or threefold difference in
the probability of catastrophic errors, which is
defined as the probability of decoding error
larger than 5% of the environment size (Fig.
7E)]. All of these theoretical results were robust
to the choice of decoder type (fig. S18), the choice
of integration time window of the decoder (fig.
S19), and choice of the parameter that controls
the scaling of encoding schemes with environ-
ment size (fig. S17H) (27). Together, this theoretical
analysis suggested that for small environments,
all the encoding schemes perform equally well
(Fig. 7, B to E; all six lines meet at the environ-
ment size of 20 m); by contrast, for very large
environments, of hundreds of meters or more,
scheme 6—which matches the multiscale coding
that we found in bat CA1—outperforms all the
other coding schemes.
Last, we suggest that the absence of a mul-
tiscale code in small environments might stem
from energy considerations. We used published
experimental estimates of the energy [adeno-
sine triphosphate (ATP) molecules] required
to generate one action potential (27, 38) to
approximate the energy required to represent
environments of different sizes for the various
coding schemes (Fig. 7F). In small environ-
ments, classical single-field codes (schemes 1
to 3) were more energetically efficient than
our multiscale code (scheme 6). Because all of
the codes exhibit a similar localization per-
formance in small environments, the energetic
consideration becomes more important, and
therefore the single-field codes are preferable
for small environments. By contrast, in large
environments our multiscale code becomes
energetically closer to the single-field codes
and even surpasses some of them in terms of
energy consumption (Fig. 7F, compare scheme 6
with the other schemes). Further, the localization
accuracy of classical single-field codes deterio-
rates so greatly in large environments (Fig. 7, C
to E) that the energetic consideration becomes
largely irrelevant, and the superior localiza-
tion accuracy of the multiscale code becomes
the central consideration. Thus, we propose
that this energetic consideration—and in par-
ticular, the tradeoff between energy expendi-
ture and coding performance—may explain
why in small environments there is no multi-
scale code. Taken together, the theoretical de-
coding analyses suggest that the multiscale
code is better suited than classical place codes
for representing very large spaces, such as real-
world natural environments.
Neural network modeling of multiscale codes:
Attractor networks and feedforward models
Classical models of hippocampal place cells
are characterized by a single spatial scale per
neuron in a given environment (39–47). We
investigated two types of models that might
support multiscale representations (figs. S20
Eliav et al., Science 372, eabg4020 (2021)
28 May 2021
8 of 12
RES EARCH | R E S E A R C H A R T I C L E
A
B
e
t
a
r
g
n
i
r
i
F
17
)
z
H
(
0
25
.
o
n
t
h
g
i
l
F
26
1
0
e
t
a
r
g
n
i
r
i
F
30
)
z
H
(
0
15
.
o
n
t
h
g
i
l
F
28
1
0
Cell 1
max=9.7m
min=1.2m
ratio=7.9
39
8
9
0
25
Cell 2
Cell 3
32
max=6.5m
min=1.4m
ratio=4.7
7
5
0
31
max=9.6m
min=2.1m
ratio=4.6
7
9
50
100
150
200
29
1
0
50
100
150
200
31
1
0
Cell 4
max=7.9m
min=2.1m
ratio=3.8
35
5
5
0
35
Cell 5
40
max=10.1m
min=2.4m
ratio=4.3
2
8
0
29
50
100
150
200
Cell 6
max=10.8m
min=2.3m
ratio=4.7
8
5
40
1
0
50
100
Position (m)
150
200
29
1
0
50
100
Position (m)
150
200
50
100
Position (m)
150
200
Lab
Wild
0.2
0.1
0
0
10 20
C
10 -1
10 -2
s
l
l
e
c
f
o
n
o
i
t
c
a
r
F
0
10
No. of fields per direction
20
D
s
d
e
i
f
l
10 -1
f
o
n
o
i
t
c
a
r
F
10 -2
10 -3
0.3
0.15
0
0
10
20
0
10
20
30
Field size (m)
F
s
d
e
l
i
f
f
o
.
o
N
n
o
i
t
c
e
r
i
d
r
e
p
10
5
0
G
)
m
10
P = 0.18
P = 5e-19
i
(
e
z
s
d
e
F
i
l
5
0
E
s
l
l
e
c
f
o
n
o
i
t
c
a
r
F
H
o
i
t
i
a
r
e
z
s
d
e
F
i
l
0.3
0.15
0
1 2 5
20
1 2
5 10 20
Field size ratio
largest/smallest
I
)
%
(
i
s
x
a
l
i
a
n
d
u
t
i
g
n
o
L
22
20
18
16
0
Lab
Wild
50
100
P = 0.64
Proximo-distal axis (%)
10 -1
10 -2
t
s
e
l
l
a
m
s
/
t
s
e
g
r
a
l
6
4
2
0
Lab
Wild
Lab
Wild
Lab
Wild
Fig. 6. Multiscale coding does not require early exposure to large-scale
environments during development. Comparison of multiscale properties
between laboratory-born bats that were raised in a 5-m-sized room (27)
and have never experienced large-scale environments during development
(green) (table S1, dataset 4) versus wild-born bats that were caught as adults
outdoors (gray). Both groups of bats were tested under identical conditions
in the 200-m tunnel. (A) Schematic of experimental design. The only difference
between laboratory-born and wild-born bats occurred during early life;
subsequent stages were identical: Both groups spent several months in the
same colony-room before surgery, and then the training and recording
procedures were identical for both groups. (B) Examples of firing-rate maps
and raster plots for six cells recorded from laboratory-born bats flying in
the large-scale environment (200-m tunnel). Same graphical conventions as in
Fig. 2A. (C to E) Distributions of (C) number of fields per direction, (D) field
sizes and (E) field-size ratio for laboratory-born bats (green) and wild-born
bats (gray), recorded in the same large-scale environment. (C) nlab = 161 cells ×
directions, nwild = 331. (D) nlab = 649 fields, nwild = 1629. (E) nlab = 82 cells,
nwild = 172 [only cells with ≥2 fields shown in (E)]. y axes are in log scale.
(Insets) Same histograms with y axis in linear scale. (F to H) Population
comparisons between laboratory-born and wild-born bats: (F) number of fields
per direction, (G) field sizes, and (H) field-size ratio. Boxplots denote the
median (horizontal line), 25 to 75% (box), and 10 to 90% (whiskers); P values
of Wilcoxon rank-sum tests are indicated. (F) df = 490, z = –1.33. (G) df =
2276, z = –8.92. (H) df = 252, z = –0.47. Despite significant difference in the
Eliav et al., Science 372, eabg4020 (2021)
28 May 2021
9 of 12
RES EARCH | R E S E A R C H A R T I C L E
field-sizes distribution [(G) P = 5 × 10−19], the field-size ratio distribution
did not differ significantly [(H) P = 0.64], indicating that the multiscale code
exists also in neurons recorded from laboratory-born bats. (I) Anatomical
positions of tetrodes along the CA1 longitudinal (dorso-ventral) axis and
proximo-distal axis [0% longitudinal: dorsal (septal) pole of CA1; 0% proximo-
distal, proximal border with CA2]. Tetrodes from both groups had similar
longitudinal coordinates in dorsal CA1, but laboratory-born bats’ tetrodes
concentrated more proximally along the proximo-distal axis of CA1.
to S23 and supplementary text). First, we used
a continuous attractor neural-network frame-
work (fig. S20, A to C) (40, 42–44, 47, 48). We
generated a network with multiple interact-
ing attractors at various scales, in which each
neuron could participate in any of the attrac-
tors at a random location (fig. S20A) (27). Net-
work simulations showed coherent bumps of
activity at each attractor, with different bump
widths (fig. S21, A and B), and single neurons
exhibited multifield, multiscale coding (fig.
S20B) that was consistent with our experimen-
tal data. Second, we explored a set of feedfor-
ward models, in which CA1 neurons received
inputs from CA3 and medial entorhinal cortex
(MEC) with diverse synaptic strengths (fig.
S20, D to J) (27). The modeling suggested that
the experimental data were inconsistent with
a strong periodic grid input and were most
consistent with a model in which the major
input into CA1 comes from CA3, in which in-
dividual CA3 neurons exhibit a single place-field
(supplementary text and fig. S20J). We thus pre-
dict that in very large environments, (i) MEC
neurons should not exhibit strong periodic-
ity, and (ii) place cells in CA3 (unlike those in
CA1) should exhibit single place fields.
Discussion
We found a multiscale neural code for large
environments: Single hippocampal neurons
in the dorsal CA1 area of bats exhibited many
fields, and the different fields of the same
neuron varied dramatically in size, with an
up to 20-fold ratio in the size of different place
fields for the same neuron. This unknown
coding scheme was revealed through the use
of an extremely large environment. This find-
ing constitutes a fundamentally different phe-
nomenon from the well-known gradient of
place field sizes along the longitudinal ana-
tomical axis of the hippocampus (14, 17, 49)—
where each neuron has one characteristic
spatial scale, and this scale changes between
neurons according to anatomical position.
In this study, by contrast, all the recordings were
conducted in the same anatomical position,
the dorsal CA1 (fig. S1), and we found that in-
dividual neurons did not have a single scale
but rather that the spatial scale of the same
neuron varied dramatically across the envi-
ronment. Further, this neural code was observed
from the first day of exposure to the environ-
ment and was similar between laboratory-
born and wild-born bats, suggesting that the
multiscale code is a very robust phenomenon
that does not require substantial recent ex-
perience with the test environment nor
early experience with large environments
in general.
Previous studies in rodents have reported
multiple place fields for individual CA1 neurons
in (relatively) large environments (18, 31, 32)—
although the number of fields per neuron was
much smaller than we found here—but no
study to date has found the multiscale prop-
erty that we discovered here for individual
neurons. Our theoretical decoding analysis
provides a simple functional explanation for
this multiscale code: For very large environ-
ments, multiscale coding outperforms all the
other codes that we considered, in terms of
reducing the number of required neurons and
minimizing the decoding errors. We hypoth-
esize that the reason why previous studies
(18, 31, 32) did not find a multiscale code was
that they used much smaller environments, or
concatenated small compartments, where such
a code does not provide a functional advantage.
Recordings from bats flying in a small environ-
ment did not show a multiscale code (Fig. 4).
The absence of a multiscale code in the small
environment can be interpreted in two ways:
(i) Neurons in small environments exhibit the
classical place code and switch to a multi-
scale code in large environments. (ii) Multi-
scale coding is the underlying representation
in all environmental scales, but the multiscale
nature of the code cannot be revealed in small
environments, where the firing reflects a small
“pinhole view” of the larger multiscale map,
and therefore the largest fields are too big to
be seen because they cover the entire space.
However, option (ii) seems unlikely because we
would then expect to see in the 6-m setup many
neurons that fire over the entire environment,
thus reducing substantially the percentage of
place cells out of the neurons active in flight—
but in fact, these percentages were remarkably
similar between the 6-m and 200-m environ-
ments (83.3 and 83.4%, respectively).
Our multiscale findings open the way for
numerous future questions on the neuro-
biology of large-scale navigation. For exam-
ple: What are the mechanisms that underlie
this multiscale coding that we discovered? Our
network modeling suggested that one possi-
bility is a feedforward convergence of inputs
from CA3, where each CA3 neuron has a single
field (fig. S20, D, left, and J), and also predicted
that MEC neurons should not exhibit spatial
periodicity in large environments (fig. S20, G
to I). Further, what is the biological decoder
that may read this code downstream? How are
such large spaces learned by the hippocampal
system? Are there ultralong compressed firing
sequences during rest and sleep, similar to
sequences observed in laboratory environments
(50–52), but extending over hundreds of meters
or more? If so, what are the mechanisms that
could create these sequences under this multi-
scale code, in which each neuron would par-
ticipate multiple times in each sequence, each
time with a different resolution? More broadly,
these findings call for performing neuro-
physiological research in very-large-scale envi-
ronments on all types of hippocampal and
entorhinal spatial neurons. We posit that such
research is crucial for understanding the brain’s
“navigation circuit” for two reasons: First, most
animals and humans evolved to navigate in
multicompartment environments with differ-
ent spatial scales, including very large scales,
so it is important to conduct neurobiological
research on large scales. Second, studies in
humans have emphasized that spatial scale
is important for navigation; people navigate
differently in large versus small environments,
which calls for conducting navigation experi-
ments in very large environments (53). Our
study provides direct single-neuron evidence
that the use of a real-world spatial scale can re-
veal a fundamentally new kind of spatial coding
in the hippocampus. This work thus makes a
step toward bridging the major gap between
the neurobiological tradition of studying the
brain’s navigation circuit in small-scale labora-
tory setups and the ecological tradition of study-
ing large-scale animal navigation outdoors.
Materials and methods summary
We conducted tetrode-based recordings of
single neurons in the dorsal hippocampus
area CA1 of Egyptian fruit bats (Rousettus
aegyptiacus), in both wild-born and laboratory-
born bats, using a wireless electrophysiology
system, while the bats were flying in a very
large environment (200-m-long tunnel), in
either familiar or novel conditions. For com-
parison, we also recorded from bats flying in a
6-m segment of the tunnel. The experimental
datasets are summarized in table S1. We local-
ized the bat’s position in the tunnel using a
radio frequency–based system yielding ~9-cm
precision. We computed firing-rate maps sepa-
rately for each flight direction and used spatial
information and a shuffling procedure to iden-
tify significant place cells. Individual place fields
were detected as prominent, stable, and sig-
nificantly tuned peaks in the firing-rate maps.
To theoretically compare the observed spatial
Eliav et al., Science 372, eabg4020 (2021)
28 May 2021
10 of 12
RES EARCH | R E S E A R C H A R T I C L E
A
.
o
n
n
o
r
u
e
n
l
e
p
m
a
x
E
10
5
1
0
Scheme 1
Scheme 2
Scheme 3
Scheme 4
Scheme 5
Scheme 6
10
5
1
10
5
1
10
5
1
10
5
1
10
5
1
100
Position (m)
200
0
100
Position (m)
200
0
100
Position (m)
200
0
100
Position (m)
200
0
100
Position (m)
200
0
100
Position (m)
200
B
m
2
<
r
o
r
r
e
r
o
f
d
e
r
i
u
q
e
r
N
l
i
a
m
n
M
i
Minimal no. of neurons (N)
required for decoding
250
200
150
100
50
20
200
400
600
800
1000
Environment size (m)
1.7
)
r
e
t
e
m
r
e
p
N
(
e
p
o
S
l
0
***
***
***
***
*
1: Single small field
2: Single large field
3: Single field, dorso-ventral gradient
4: Multiple small fields (Rich et al. 2014)
5: Multiscale (population)
6: Multiscale (single-cell), matching bat data in 200 meters
C
)
m
(
r
o
r
r
e
i
g
n
d
o
c
e
d
n
a
e
M
500
20
10
0
20
400
300
200
100
0
20
Mean decoding errors
D
1000
Catastrophic errors: size
)
m
(
r
o
r
r
e
i
g
n
d
o
c
e
d
%
9
9
100
800
10
1
100
600
400
200
0
20
1000
100
1000
Environment size (m)
100
1000
100
1000
Environment size (m)
10 0
10 -1
10 -2
10 -3
E
)
e
z
s
i
t
n
e
m
n
o
r
i
v
n
e
f
o
%
5
>
r
o
r
r
e
(
.
b
o
r
P
Catastrophic errors: probability
20
100
1000
Environment size (m)
F
10 11
Energy considerations
Fig. 7. Theoretical analysis showed that multiscale coding decreases the decoding error for large
environments. Decoding accuracy analysis for simulations of six different models (encoding schemes), using
maximum likelihood decoder and integration time window of 500 ms (27). (A) We examined six different
encoding schemes for spatial representations; shown here are 10 simulated example neurons for each scheme:
(scheme 1) single field with small size; (scheme 2) single field with large size; (scheme 3) single field with
gradually increasing field-size across neurons – mimicking the dorso-ventral anatomical gradient of field
sizes; (scheme 4) multiple small fields [the distribution of field-propensity was taken from (18)]; (scheme 5)
multiple fields with fixed size per neuron, but with variable sizes across the population; and (scheme 6) multiple
fields with multiscale per neuron (as in the bat data). In schemes 5 and 6, we matched the distribution of
field sizes to our data (fig. S8). The mean coverage in schemes 2, 5, and 6 was identical (27). (B) (Left) Minimal
number of neurons required for reaching mean decoding error <2 m, plotted as a function of different
environment sizes (from 20 to 1000 m). (Right) Slopes of the curves on the left, representing how many
additional neurons are required on average when increasing the environment size by 1 m. Colors represent
the six encoding schemes. (C to E) Decoding errors when using n = 50 neurons. (C) Mean decoding error
versus environment size, showing that schemes 1, 3, and 4 exhibit huge decoding errors for large environments.
(Inset) Zoom-in on errors smaller than 20 m (y axis), showing that per-neuron multiscale encoding (scheme 6, purple) outperforms fixed scale per-neuron
(scheme 5, blue) in terms of mean decoding error. [(D) and (E)] Catastrophic errors. (D) Rare large errors (99th percentile of decoding error), plotted versus
environment size. (Inset) Same plot in log-scale for the y axis. (E) Probability of decoding error larger than 5% of the environment size, plotted as a function of
environment size. (F) Theoretical estimate of energy expenditure under the various coding schemes: Shown is the number of ATP molecules per second required to
represent the environment with mean decoding error < 2 m, plotted against the environment size (27).
Environment size (m)
s
e
u
c
e
o
m
P
T
A
d
n
o
c
e
s
10 10
1000
10 9
r
e
p
100
20
l
l
Eliav et al., Science 372, eabg4020 (2021)
28 May 2021
11 of 12
RES EARCH | R E S E A R C H A R T I C L E
coding scheme with a set of five other coding
schemes, we generated synthetic data for each
coding scheme and then used maximum-
likelihood and population-vector decoders
to test their decoding performance. To theo-
retically explore the possible neural-network
mechanisms underlying the observed coding,
we considered both an attractor network
model, based on multiple interacting attrac-
tors that randomly share neurons between
them, as well as four feedforward models,
based on inputs from MEC and CA3. Further
details can be found in the supplementary
materials, materials and methods.
RE FE RENCES AND N OT ES
1.
2.
J. O’Keefe, J. Dostrovsky, The hippocampus as a spatial map.
Preliminary evidence from unit activity in the freely-moving rat.
Brain Res. 34, 171–175 (1971). doi: 10.1016/0006-8993(71)
90358-1; pmid: 5124915
J. O’Keefe, L. Nadel, The Hippocampus as a Cognitive Map
(Oxford Univ. Press, 1978).
3. P. Andersen, R. G. M. Morris, D. G. Amaral, T. V. Bliss, J. O’Keefe,
4.
5.
Eds., The Hippocampus Book (Oxford Univ. Press, 2007).
J. B. Ranck Jr., Head direction cells in the deep layer of dorsal
presubiculum in freely moving rats. Soc. Neurosci. Abstr. 10,
599 (1984).
J. S. Taube, R. U. Muller, J. B. Ranck Jr., Head-direction cells
recorded from the postsubiculum in freely moving rats. I.
Description and quantitative analysis. J. Neurosci. 10, 420–435
(1990). doi: 10.1523/JNEUROSCI.10-02-00420.1990;
pmid: 2303851
6. T. Hafting, M. Fyhn, S. Molden, M.-B. Moser, E. I. Moser,
Microstructure of a spatial map in the entorhinal cortex.
Nature 436, 801–806 (2005). doi: 10.1038/nature03721;
pmid: 15965463
7. C. Barry, R. Hayman, N. Burgess, K. J. Jeffery, Experience-
dependent rescaling of entorhinal grids. Nat. Neurosci. 10,
682–684 (2007). doi: 10.1038/nn1905; pmid: 17486102
8. T. Solstad, C. N. Boccara, E. Kropff, M.-B. Moser, E. I. Moser,
Representation of geometric borders in the entorhinal cortex.
Science 322, 1865–1868 (2008). doi: 10.1126/
science.1166466; pmid: 19095945
9. C. Lever, S. Burton, A. Jeewajee, J. O’Keefe, N. Burgess,
Boundary vector cells in the subiculum of the hippocampal
formation. J. Neurosci. 29, 9771–9777 (2009). doi: 10.1523/
JNEUROSCI.1319-09.2009; pmid: 19657030
10. E. I. Moser, M.-B. Moser, B. L. McNaughton, Spatial representation
in the hippocampal formation: A history. Nat. Neurosci. 20,
1448–1464 (2017). doi: 10.1038/nn.4653; pmid: 29073644
11. M. A. Wilson, B. L. McNaughton, Dynamics of the hippocampal
ensemble code for space. Science 261, 1055–1058 (1993).
doi: 10.1126/science.8351520; pmid: 8351520
12. S. Leutgeb et al., Independent codes for spatial and episodic
memory in hippocampal neuronal ensembles. Science 309,
619–623 (2005). doi: 10.1126/science.1114037; pmid: 16040709
13. C. D. Harvey, F. Collman, D. A. Dombeck, D. W. Tank,
Intracellular dynamics of hippocampal place cells during virtual
navigation. Nature 461, 941–946 (2009). doi: 10.1038/
nature08499; pmid: 19829374
14. S. Royer, A. Sirota, J. Patel, G. Buzsáki, Distinct
15.
representations and theta dynamics in dorsal and ventral
hippocampus. J. Neurosci. 30, 1777–1787 (2010). doi: 10.1523/
JNEUROSCI.4681-09.2010; pmid: 20130187
I. Lee, D. Yoganarasimha, G. Rao, J. J. Knierim, Comparison of
population coherence of place cells in hippocampal subfields
CA1 and CA3. Nature 430, 456–459 (2004). doi: 10.1038/
nature02739; pmid: 15229614
16. J. K. Leutgeb, S. Leutgeb, M.-B. Moser, E. I. Moser, Pattern
separation in the dentate gyrus and CA3 of the hippocampus.
Science 315, 961–966 (2007). doi: 10.1126/science.1135801;
pmid: 17303747
17. K. B. Kjelstrup et al., Finite scale of spatial representation in
the hippocampus. Science 321, 140–143 (2008). doi: 10.1126/
science.1157086; pmid: 18599792
18. P. D. Rich, H.-P. Liaw, A. K. Lee, Large environments reveal the
statistical structure governing hippocampal representations.
Science 345, 814–817 (2014). doi: 10.1126/science.1255635;
pmid: 25124440
19. K. D. Taylor, Range of movement and activity of common rats
(Rattus norvegicus) on agricultural land. J. Appl. Ecol. 15,
663–677 (1978). doi: 10.2307/2402767
20. J. C. Russell, A. J. C. McMorland, J. W. B. MacKay, Exploratory
behaviour of colonizing rats in novel environments. Anim. Behav.
79, 159–164 (2010). doi: 10.1016/j.anbehav.2009.10.020
21. M. Geva-Sagiv, L. Las, Y. Yovel, N. Ulanovsky, Spatial cognition
in bats and rats: From sensory acquisition to multiscale maps
and navigation. Nat. Rev. Neurosci. 16, 94–108 (2015).
doi: 10.1038/nrn3888; pmid: 25601780
22. A. Tsoar et al., Large-scale navigational map in a mammal.
Proc. Natl. Acad. Sci. U.S.A. 108, E718–E724 (2011).
doi: 10.1073/pnas.1107365108; pmid: 21844350
23. N. Ulanovsky, C. F. Moss, Hippocampal cellular and network
activity in freely moving echolocating bats. Nat. Neurosci. 10,
224–233 (2007). doi: 10.1038/nn1829; pmid: 17220886
24. M. M. Yartsev, M. P. Witter, N. Ulanovsky, Grid cells without
theta oscillations in the entorhinal cortex of bats. Nature 479,
103–107 (2011). doi: 10.1038/nature10583; pmid: 22051680
25. M. M. Yartsev, N. Ulanovsky, Representation of three-dimensional
space in the hippocampus of flying bats. Science 340, 367–372
(2013). doi: 10.1126/science.1235338; pmid: 23599496
26. A. Finkelstein et al., Three-dimensional head-direction coding in
the bat brain. Nature 517, 159–164 (2015). doi: 10.1038/
nature14031; pmid: 25470055
27. Materials and methods are available as supplementary materials.
28. L. T. Thompson, P. J. Best, Place cells and silent cells in the
hippocampus of freely-behaving rats. J. Neurosci. 9, 2382–2390
(1989). doi: 10.1523/JNEUROSCI.09-07-02382.1989;
pmid: 2746333
29. B. L. McNaughton, C. A. Barnes, J. O’Keefe, The contributions
of position, direction, and velocity to single unit activity in the
hippocampus of freely-moving rats. Exp. Brain Res. 52, 41–49
(1983). doi: 10.1007/BF00237147; pmid: 6628596
30. M. Geva-Sagiv, S. Romani, L. Las, N. Ulanovsky, Hippocampal
global remapping for different sensory modalities in flying bats.
Nat. Neurosci. 19, 952–958 (2016). doi: 10.1038/nn.4310;
pmid: 27239936
31. A. A. Fenton et al., Unmasking the CA1 ensemble place code by
exposures to small and large environments: More place cells
and multiple, irregularly arranged, and expanded place fields in
the larger space. J. Neurosci. 28, 11250–11262 (2008).
doi: 10.1523/JNEUROSCI.2862-08.2008; pmid: 18971467
32. E. Park, D. Dvorak, A. A. Fenton, Ensemble place codes in
hippocampus: CA1, CA3, and dentate gyrus place cells have
multiple place fields in large environments. PLOS ONE 6, e22349
(2011). doi: 10.1371/journal.pone.0022349; pmid: 21789250
33. G. Buzsáki, K. Mizuseki, The log-dynamic brain: How skewed
distributions affect network operations. Nat. Rev. Neurosci. 15,
264–278 (2014). doi: 10.1038/nrn3687; pmid: 24569488
34. J. S. Lee, J. J. Briguglio, J. D. Cohen, S. Romani, A. K. Lee, The
statistical structure of the hippocampal code for space as
a function of time, context, and value. Cell 183, 620–635
(2020). doi: 10.1016/j.cell.2020.09.024; pmid: 33035454
35. Y. Ziv et al., Long-term dynamics of CA1 hippocampal place
codes. Nat. Neurosci. 16, 264–266 (2013). doi: 10.1038/
nn.3329; pmid: 23396101
36. A. Rubin, N. Geva, L. Sheintuch, Y. Ziv, Hippocampal ensemble
dynamics timestamp events in long-term memory. eLife 4,
e12247 (2015). doi: 10.7554/eLife.12247; pmid: 26682652
37. L. Harten, A. Katz, A. Goldshtein, M. Handel, Y. Yovel, The ontogeny
of a mammalian cognitive map in the real world. Science 369,
194–197 (2020). doi: 10.1126/science.aay3354; pmid: 32647001
38. S. Hallermann, C. P. J. de Kock, G. J. Stuart, M. H. P. Kole,
State and location dependence of action potential metabolic
cost in cortical pyramidal neurons. Nat. Neurosci. 15,
1007–1014 (2012). doi: 10.1038/nn.3132; pmid: 22660478
39. D. Zipser, A computational model of hippocampal place fields.
Behav. Neurosci. 99, 1006–1018 (1985). doi: 10.1037/
0735-7044.99.5.1006; pmid: 3843299
40. M. Tsodyks, T. Sejnowski, Associative memory and
44. M. Tsodyks, Attractor neural network models of spatial maps
in hippocampus. Hippocampus 9, 481–489 (1999).
doi: 10.1002/(SICI)1098-1063(1999)9:4<481::AID-HIPO14>3.0.
CO;2-S; pmid: 10495029
45. T. Hartley, N. Burgess, C. Lever, F. Cacucci, J. O’Keefe,
Modeling place fields in terms of the cortical inputs to the
hippocampus. Hippocampus 10, 369–379 (2000). doi: 10.1002/
1098-1063(2000)10:4<369::AID-HIPO3>3.0.CO;2-0; pmid: 10985276
46. T. Strösslin, D. Sheynikhovich, R. Chavarriaga, W. Gerstner,
Robust self-localisation and navigation based on hippocampal
place cells. Neural Netw. 18, 1125–1140 (2005). doi: 10.1016/
j.neunet.2005.08.012; pmid: 16263241
47. J. J. Knierim, K. Zhang, Attractor dynamics of spatially
correlated neural activity in the limbic system. Annu. Rev.
Neurosci. 35, 267–285 (2012). doi: 10.1146/annurev-neuro-
062111-150351; pmid: 22462545
48. R. Ben-Yishai, R. L. Bar-Or, H. Sompolinsky, Theory of orientation
tuning in visual cortex. Proc. Natl. Acad. Sci. U.S.A. 92,
3844–3848 (1995). doi: 10.1073/pnas.92.9.3844; pmid: 7731993
49. M. W. Jung, S. I. Wiener, B. L. McNaughton, Comparison of
spatial firing characteristics of units in dorsal and ventral
hippocampus of the rat. J. Neurosci. 14, 7347–7356 (1994).
doi: 10.1523/JNEUROSCI.14-12-07347.1994; pmid: 7996180
50. D. J. Foster, M. A. Wilson, Reverse replay of behavioural sequences
in hippocampal place cells during the awake state. Nature 440,
680–683 (2006). doi: 10.1038/nature04587; pmid: 16474382
51. K. Diba, G. Buzsáki, Forward and reverse hippocampal place-
cell sequences during ripples. Nat. Neurosci. 10, 1241–1242
(2007). doi: 10.1038/nn1961; pmid: 17828259
52. T. J. Davidson, F. Kloosterman, M. A. Wilson, Hippocampal
replay of extended experience. Neuron 63, 497–507 (2009).
doi: 10.1016/j.neuron.2009.07.027; pmid: 19709631
53. T. Wolbers, J. M. Wiener, Challenges for identifying the neural
mechanisms that support spatial navigation: The impact of
spatial scale. Front. Hum. Neurosci. 8, 571 (2014).
doi: 10.3389/fnhum.2014.00571; pmid: 25140139
54. T. Eliav et al., Data for “Multiscale representation of very
large environments in the hippocampus of flying bats”.
Zenodo (2021). doi: 10.5281/zenodo.4646728
AC KNOWLED GME NTS
We thank K. D. Harris and A. Treves for suggestions; D. Omer, A. Rubin,
T. Stolero, M. Naim, A. Sarel, S. Palgi, and S. Ray for comments
on the manuscript; S. Futerman, I. Shulman, B. Pevzner, K. Dor,
S. Kodenzik, E. Solomon, C. Cohen, A. Shalev, N. Raish, and L. Hartman
for help with bat training; G. Ankaoua and B. Pasmantirer for
mechanical designs; A. Tuval for veterinary support; C. Ra’anan and
R. Eilam for histology; and G. Brodsky for graphics. Funding: This
study was supported by research grants to N.U. from the European
Research Council (ERC-CoG– NATURAL_BAT_NAV), Deutsche
Forschungsgemeinschaft (DFG-SFB 1372), and Yehuda and
Judith Bronicki, and by the André Deloro Prize for Scientific Research
and the Kimmel Award for Innovative Investigation to N.U.; N.U. is the
incumbent of the Barbara and Morris Levinson Professorial Chair in
Brain Research. T.E. was supported by the Otto Schwarz Scholarship,
the Horowitz KKL-JNF foundation, and by the Maccabim Foundation
Excellence Fellowship for PhD students. Author contributions:
T.E., L.L., and N.U. conceived the initial experiments. T.E., S.R.M., L.L.,
and N.U. set up experimental systems. T.E., S.R.M., L.L., and N.U.
designed experiments. T.E. conducted experiments for dataset 1,
and S.R.M. conducted experiments for datasets 2 to 4. T.E.
and S.R.M. analyzed the experimental data. L.L. and N.U. guided
the data analysis. J.A. and M.T. conducted the theoretical
decoding analysis and neural network modeling. G.G. analyzed the
energy-decoding tradeoff. T.E. and N.U. wrote the first draft of the
manuscript, with major contribution from L.L.; all authors participated
in writing and editing of the manuscript. N.U. supervised the project.
Competing interests: The authors declare no competing interests.
Data and materials availability: The data and code that support the
conclusions of this study are freely accessible online at Zenodo (54).
hippocampal place cells. Int. J. Neural Syst. 6, 81–86 (1995).
SUPPLEMENTARY MATERIALS
41. J. O’Keefe, N. Burgess, Geometric determinants of the
place fields of hippocampal neurons. Nature 381, 425–428
(1996). doi: 10.1038/381425a0; pmid: 8632799
42. A. Samsonovich, B. L. McNaughton, Path integration and
cognitive mapping in a continuous attractor neural network
model. J. Neurosci. 17, 5900–5920 (1997). doi: 10.1523/
JNEUROSCI.17-15-05900.1997; pmid: 9221787
43. F. P. Battaglia, A. Treves, Attractor neural networks storing
multiple space representations: A model for hippocampal place
fields. Phys. Rev. E 58, 7738–7753 (1998). doi: 10.1103/
PhysRevE.58.7738
science.sciencemag.org/content/372/6545/eabg4020/suppl/DC1
Materials and Methods
Supplementary Text
Figs. S1 to S23
Table S1
References (55–75)
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
4 January 2021; accepted 6 April 2021
10.1126/science.abg4020
Eliav et al., Science 372, eabg4020 (2021)
28 May 2021
12 of 12
|
10.1126_science.abg7071
|
RES EARCH
R E S E A R C H A R T I C L E
◥
MOLECULAR BIOLOGY
Regulation of biomolecular condensates by
interfacial protein clusters
Andrew W. Folkmann1,2†, Andrea Putnam1,2†, Chiu Fan Lee3, Geraldine Seydoux1,2*
Biomolecular condensates are cellular compartments that can form by phase separation in the absence
of limiting membranes. Studying the P granules of Caenorhabditis elegans, we find that condensate
dynamics are regulated by protein clusters that adsorb to the condensate interface. Using in vitro
reconstitution, live observations, and theory, we demonstrate that localized assembly of P granules is
controlled by MEG-3, an intrinsically disordered protein that forms low dynamic assemblies on P
granules. Following classic Pickering emulsion theory, MEG-3 clusters lower surface tension and slow
down coarsening. During zygote polarization, MEG-3 recruits the DYRK family kinase MBK-2 to accelerate
spatially regulated growth of the P granule emulsion. By tuning condensate-cytoplasm exchange,
interfacial clusters regulate the structural integrity of biomolecular condensates, reminiscent of the role
of lipid bilayers in membrane-bound organelles.
L iquid-liquid phase separation has emerged
as a new principle for cellular organi-
zation (1). Phase separation of proteins,
frequently with RNA, can create dense
condensates that are visible by micros-
copy as micron-scale assemblies containing
dozens or more protein and RNA species. Con-
densates have been reconstituted in vitro using
purified proteins that interact by using multi-
valent, low-affinity binding sites to generate
large, interconnected networks. Some proteins
form condensates that exhibit fast internal
dynamics and rapid exchange with the dilute
phase on a time scale of seconds or minutes
(1). Other proteins form more viscous conden-
sates with slower internal and exchange dy-
namics, in some cases becoming glass-like with
time (2, 3). All condensate systems (emulsions)
are predicted to evolve (coarsen) toward a
single-condensate equilibrium at a rate that
correlates positively with their internal and
exchange dynamics (4–8) (see supplemen-
tary text). Like condensates assembled in vitro,
condensates in cells exhibit a range of dynam-
ic behaviors, but these do not always fit the-
oretical predictions (9, 10). For example, some
condensates exhibit little coarsening over hour
time scales, yet maintain low viscosity, main-
tain fast exchange dynamics, and dissolve
within minutes in response to changes in the
cellular environment (11). Several hypothe-
ses have been put forward to explain the lack
of coarsening of cellular condensates with fast
1Department of Molecular Biology and Genetics, Johns
Hopkins University, Baltimore, MD 21205, USA. 2Howard
Hughes Medical Institute, Johns Hopkins University,
Baltimore, MD 21205, USA. 3Department of Bioengineering,
Imperial College London, London SW7 2AZ, UK.
*Corresponding author. Email: gseydoux@jhmi.edu
†These authors contributed equally to this work.
internal dynamics, including physical barriers
that keep condensates away from each other
(5, 12, 13), active mechanisms that continu-
ously regenerate small condensates (14), and
chemical reactions and protein gradients that
suppress Ostwald ripening (15–18) (see sup-
plementary text). In this study, we investigate
the mechanisms that control the coarsening
and dynamics of P granules, condensates in
the Caenorhabditis elegans germline. We find
that P granule coarsening is controlled by
nanoscale protein clusters that adsorb to the
condensate interface, a phenomenon first de-
scribed by Ramsden (in 1904) and Pickering
(in 1907) for inorganic emulsions (19, 20).
P granules were the first cellular conden-
sates proposed to form by liquid-liquid phase
separation (21). At the core of P granules is a
liquid-like phase assembled by PGL proteins,
paralogs PGL-1 and PGL-3 (21–23). During
most of the C. elegans life cycle, PGL conden-
sates associate stably with the cytoplasmic
face of nuclei (24). During the oocyte-to-zygote
transition, PGL condensates redistribute to
the cytoplasm and undergo two rapid cycles
of dissolution and condensation. The first
cycle occurs during oocyte maturation when
most PGL condensates dissolve before reas-
sembling after fertilization in the zygote. The
second cycle occurs during zygote polarization
when PGL condensates dissolve in anterior
cytoplasm and assemble in posterior cytoplasm.
Both cycles are completed within minutes with-
out substantial coarsening. Factors that regulate
P granule dynamics during oocyte maturation
have not yet been identified. Factors that reg-
ulate dynamics during polarization include
MEX-5, an RNA-binding protein; MEG pro-
teins MEG-3 and MEG-4, two paralogous in-
trinsically disordered proteins; and MBK-2, a
DYRK family kinase that interacts physical-
ly and genetically with MEG-3 (22, 25–28).
During polarization, MEX-5 becomes enriched
in the anterior cytoplasm, where it promotes
P granule dissolution, possibly by competing
with P granule proteins for RNA (22, 27). In
zygotes lacking mbk-2 or meg-3 and meg-4
activity, P granules do not dissolve in the an-
terior cytoplasm, despite a normal MEX-5
gradient (28). In this study, we investigate
how the MEGs and MBK-2 collaborate with
MEX-5 to regulate P granule dynamics.
MEG-3 forms low-dynamic clusters that
adsorb to the PGL-3 interface
MEG-3 has been reported to form assemblies
on the surface of PGL-3 condensates in newly
fertilized zygotes (28, 29) and in PGL-3 and
MEG-3 co-condensates reconstituted in vitro
(29). Superresolution three-dimensional con-
focal microscopy (see methods) confirmed
that MEG-3 forms diffraction-limited clusters
(<160 nm) at the PGL-3 interface in vivo and
in vitro (Fig. 1A and fig. S1). Consistent with
MEG-3 clusters adsorbing to the interface,
MEG-3 modifies the wetting behavior of PGL-3
condensates assembled in vitro, reducing the
extent to which PGL-3 droplets wet the surface
of untreated glass slides (Fig. 1, B and C).
MEG-3 clusters are resistant to dilution, high
temperature, and salt treatment; by contrast,
PGL-3 condensates readily dissolve in dilute
conditions and at increased temperatures
(29). MEG-3 condensates exchange more slow-
ly than PGL-3 condensates, as measured by
fluorescence recovery after photobleaching
(FRAP) in vitro and in vivo (29). Using a single-
molecule method adapted from Wu et al. (30),
we measured the dynamics of MEG-3 and
PGL-3 molecules in P granules in vivo (Fig. 1,
D to F, and movies S1 and S2). Most PGL-3
molecules exhibited short-lived trajectories
in P granules with an average apparent diffu-
sion coefficient of D = 0.056 mm2/s (Fig. 1, E
and F). By contrast, all MEG-3 molecules
exhibited restricted long-lived trajectories
with an average apparent diffusion coefficient
of D = 0.0018 mm2/s (Fig. 1, E and F). In three
of three cases where we captured the trajec-
tories of three labeled MEG-3 molecules in
the same P granule, their relative position re-
mained fixed over time (Fig. 1D and movie
S2). Together these observations confirm that
PGL-3 molecules exist primarily in a dynamic
liquid-like phase (albeit highly viscous), where-
as MEG-3 molecules experience much slower
dynamics, resembling solid clusters within our
experimental time scales.
MEG-3 reduces the surface tension of PGL-3
condensates without changing viscosity
Solid particulates that adsorb to liquid sur-
faces reduce surface tension without affect-
ing internal dynamics (viscosity) (31). Previous
studies have shown that PGL-3 condensates
Folkmann et al., Science 373, 1218–1224 (2021)
10 September 2021
1 of 7
RES EARCH | R E S E A R C H A R T I C L E
B
A
l
150
0 s
97 5
C
57 2
64 s
PGL-3
PGL-3
PGL-3
MERGE
MERGE
Time:
8 s
in vitro
MEG-3
in vivo
MEG-3
PGL-3
+ MEG-3
Fig. 1. MEG-3 forms low-dynamic clusters that
adsorb to the surface of PGL-3 condensates.
(A) Photomicrographs of a P granule in vivo labeled
with PGL-3::mCherry and MEG-3::meGFP (GFP,
green fluorescent protein) and of a P granule
reconstituted in vitro with purified PGL-3 and MEG-3
trace-labeled with Dylight 488 and Alexa 647,
respectively. Scale bars are 500 nm (in vivo, each
scale bar applies to all images in the same row)
and 3 mm (in vitro, scale bar applies to all images in
the set). The top panels are a maximum projection
of a z-stack through the granule. The middle panels
are a single x-y plane through the middle of the
same granule. The lower panels are a single z-x plane
through the middle of the same granule. See fig. S1A
for additional examples of P granules captured
in vivo. (B) In vitro time-lapse series showing PGL-3
droplets trace-labeled with Dylight 488 wetting the
surface of a glass slide with 3 mM PGL-3, with
80 ng/ml nos-2 RNA, and with or without 0.5 mM
MEG-3. Average contact angles at 64 s are indicated.
Scale bars are 1 mm, and each scale bar applies to
all images in the same row. (C) Contact angles
measured as in (B). Each dot represents a droplet,
and red lines represent the mean. (D) In vivo time-lapse series showing single molecules (green) of PGL-3::Halo and MEG-3::Halo in P granules (magenta). Scale bars
are 1 mm, and each scale bar applies to all images in the respective set. (E) Graph depicting the apparent diffusion coefficients of PGL-3::Halo and MEG-3::Halo
molecules in P granules. Each dot represents one trajectory, and the red line represents the mean. (F) Graph depicting the dwell time of PGL-3::Halo and MEG-3::Halo
molecules in P granules. Each dot represents one trajectory, and the red line represents the mean.
PGL-3 PGL-3 + MEG-3
MEG-3
MEG-3
PGL-3
PGL-3
e
g
n
A
a
c
a
f
r
e
t
n
I
:
:
3
-
G
E
M
)
s
(
e
u
n
a
r
G
n
t
n
e
c
i
f
f
e
o
C
3
-
L
G
P
t
n
e
r
a
p
p
A
n
o
s
u
f
f
i
o
a
H
o
a
H
-0.2 s
0.6 s
1.2 s
2.9 s
3.5 s
3.8 s
4.1 s
D
e
m
T
e
w
D
12 s
10 s
100
10-1
10-4
10-3
10-2
10-1
m
µ
(
E
100
)
s
/
2
100
102
101
F
1 s
2 s
4 s
8 s
6 s
0 s
0 s
50
:
:
D
0
l
l
l
l
l
i
i
i
i
l
i
Fig. 2. MEG-3 reduces the surface tension of
PGL-3 condensates and prevents coarsening.
(A) Photomicrographs of PGL-3 droplets (3 mM PGL-3
and 80 ng/ml nos-2 RNA) coalescing with or without
0.5 mM MEG-3. (B) Relaxation time (t) of fusing
PGL-3 droplets (as above) is plotted versus length
scale (‘) with varying concentrations of MEG-3 as
indicated. Each dot represents a single fusion event.
The linear slope represents the inverse capillary
velocity (h/g). (C) Photomicrographs of a PGL-3
emulsion (max projections) at the indicated
time points after assembly. Three micromolar PGL-3
and 80 ng/ml nos-2 RNA were incubated in
condensation buffer in the presence or absence of
0.5 mM MEG-3. Scale bar is 5 mm and applies to
all images in the set. (D and E) Histograms plotting
the size distribution of PGL condensates assembled
as in (C). Each data point indicates the fraction
of total PGL-3 condensate volume represented by
condensates binned by radius from 80 images [as in
(C)] collected in four replicates. Lines were fit to
a log normal distribution.
are “aging Maxwell fluids” whose viscosity
increases over days, eventually adopting glass-
like properties [see supplementary text and
(2, 32)]. To focus on the time scales expe-
rienced by PGL-3 condensates during the
oocyte-to-embryo transition, we examined
PGL-3 condensates within the first 3 hours
of assembly for all in vitro studies. To probe
internal dynamics, we measured the move-
ment of fluorescent microspheres embedded
in PGL-3 condensates with and without MEG-3
(fig. S2, A and B, and movie S3). Fitting the mean
squared displacement (MSD) of microspheres
to the equation MSD = 4Dta, we calculated a
diffusion coefficient (D) and anomalous diffu-
sion exponent (a). We observed an a ≈ 1 for
PGL-3, consistent with newly formed PGL-3
condensates behaving as viscous liquids (fig.
S2C). Using the Stokes-Einstein relation, we
calculated a viscosity of h = 5.4 Pa⋅s in the range
estimated previously for PGL-3 condensates
Folkmann et al., Science 373, 1218–1224 (2021)
10 September 2021
2 of 7
RES EARCH | R E S E A R C H A R T I C L E
2.5 µM PGL-3
A
P
T
G
/
3
K
R
Y
D
P
T
A
/
3
K
R
Y
D
B
r
e
b
m
u
N
e
t
a
s
n
e
d
n
o
C
r
e
b
m
u
N
e
t
a
s
n
e
d
n
o
C
2.5 µM PGL-3 + GTP + DYRK3
200
0 min
10 min
25 min
45 min
60 min
150
100
50
0
4
6
8
10
Log(Intensity)
2.5 µM PGL-3 + ATP + DYRK3
200
150
100
50
0
0 min
10 min
25 min
45 min
60 min
4
6
8
10
Log(Intensity)
C
l
e
n
o
A
P
T
A
P
T
A
/
3
K
R
Y
D
5.0 µM PGL-3
5.0 µM PGL-3 + MEG-3
D
10-1
)
s
/
2
m
µ
(
10-2
10-3
n
o
s
u
i
f
f
i
D
t
n
e
r
a
p
p
A
t
i
n
e
c
i
f
f
e
o
C
ATP alone
DYRK3/ATP
No MEG-3
MEG-3
10-4
10-5
10-6
300
200
100
0
0
2
4
Log(Intensity)
6
8
E
r
e
b
m
u
N
e
t
a
s
n
e
d
n
o
C
Fig. 3. MEG-3 stabilizes PGL-3 condensates against kinase accelerated
coarsening. (A) Photomicrographs of a 2.5 mM PGL-3488 emulsion after a 60-min
treatment with 100 nM DYRK3 kinase in the presence of GTP (top) or ATP
(bottom). Scale bar is 50 mm and applies to both images. (B) Histograms of PGL
condensates assembled as in (A) with GTP and DYRK3 or ATP and DYRK3 at
indicated time points. Circles indicate the number of PGL-3 condensates binned by
the log(intensity) of each condensate. Colors indicate the time after addition of
DYRK3. (C) Photomicrographs of PGL-3 condensates (max projections) assembled
with and without 70 nM MEG-3 and captured 60 min after addition of 100 mM
ATP with and without 100 nM DYRK3. Arrows point to small PGL-3 and MEG-3
co-condensates. The inset is a high-resolution image with PGL-3 in magenta and MEG-3
in green (see fig. S4, F and G, for additional examples). Scale bars are 50 mm
(applies to all images in the set) and 5 mm (inset). (D) Diffusion coefficients of
200-nm microspheres in PGL-3 condensates (5 mM PGL-3 and 100 mM ATP) with
and without 100 nM DYRK3. Each dot represents a single microsphere trajectory.
The red line represents the mean. (E) Histograms of PGL-3 condensates assembled
with and without 70 nM MEG-3 and incubated for 60 min in 100 mM ATP with
100 nM DYRK3. Circles indicate the number of PGL-3 condensates binned by the log
(intensity) of each condensate captured from 17 images. See fig. S4, D and E, for
additional MEG-3 concentrations and an additional time point.
in vivo and in vitro (fig. S2D) (2, 21). Addition of
MEG-3 did not have a notable effect on PGL-3
viscosity (h = 5.6 Pa⋅s) (fig. S2, B to D, and movie
S3), as expected for a surface agent that, on its
own, does not affect internal dynamics.
To measure PGL-3 surface tension, we
examined the coalescence behavior of PGL-3
droplets (Fig. 2A). Relaxation time of coalesc-
ing condensates can be expressed by t ≈ ‘(h/g),
where ‘ is the geometric mean of the diam-
eters of the droplets at the onset of fusion, h is
the viscosity of the droplets, g is the surface
tension, and h/g is the inverse capillary ve-
locity. PGL-3 condensates coalesce with a
linear relationship over a range of condensate
sizes, yielding an inverse capillary velocity of
0.25 s/mm and surface tension of 19.4 mN/m
(Fig. 2B and movie S4), comparable to pre-
vious estimates for P granules in vivo (21).
Addition of MEG-3 slowed coalescence of
PGL-3 droplets, yielding an inverse capil-
lary velocity of 1.2 s/mm and a surface ten-
sion of 4.7 mN/m (500 nM MEG-3; Fig. 2B and
movie S5).
In addition to modulating surface tension,
Pickering agents can also inhibit coalescence
by steric hindrance and slow rearrangement
at the condensate interface (33). We observed
examples of MEG-3–coated PGL-3 droplets
that did not relax to a sphere (movie S6) or
remained in contact without fusing during
imaging (movie S7). Higher concentrations
of MEG-3 (1 mm) increased surface cover-
age and caused PGL-3 droplets to flocculate
without fusing (fig. S2E). We conclude that,
as expected for a Pickering agent, MEG-3
clusters lower the surface tension of PGL-3
condensates and also form a physical barrier
to coalescence.
MEG-3 prevents coarsening of the PGL-3
emulsion without eliminating
surface exchange
To determine whether MEG-3 stabilizes the
PGL-3 emulsion against coarsening, we exam-
ined the evolution of a newly assembled PGL-3
emulsion over time. Over the course of 180 min,
the PGL-3 emulsion coarsened substantially:
Droplets increased in size on average and
decreased in number without a change in the
total volume of PGL-3 in droplets (Fig. 2, C and
D, and fig. S2, F to H). Addition of MEG-3
reduced coarsening, stabilizing droplet size
and number over the 180 min of the experi-
ment (Fig. 2, C and E, and fig. S2, F to H).
MEG-3 concentrations constrained the size of
PGL-3 droplets in a dose-dependent manner
(fig. S2, I to M).
MEG-3 does not affect the internal dynam-
ics of PGL-3 condensates as measured by
FRAP (29). To examine whether MEG-3 pre-
vents exchange of PGL-3 molecules at the
interface, we used PGL-3 preparations trace-
labeled with different fluorophores and exam-
ined mixing kinetics. Unlike MEG-3 clusters,
which do not mix after assembly, PGL-3 con-
densates continue to exchange after assembly
(fig. S3, A to C). Addition of soluble PGL-3 to a
preformed emulsion leads to the formation of
new condensates that, over the course of an
hour, completely mix with the old condensates
(fig. S3, D to F). Addition of MEG-3 prevented
coarsening but did not affect the rate of new
and old PGL-3 mixing (fig. S3, D to F). We con-
clude that MEG-3 clusters stabilize the PGL-3
emulsion by lowering surface tension with-
out completely blocking surface exchange, as
described for other Pickering agents (34).
MEG-3 stabilizes PGL-3 condensates against
DYRK3-accelerated coarsening
The relative high viscosity of the PGL-3 emul-
sion is consistent with the apparent stability of
P granules during most of germline develop-
ment and suggests that active processes must
operate to dissolve P granules during oocyte
maturation and polarization. MBK-2 kinase is
Folkmann et al., Science 373, 1218–1224 (2021)
10 September 2021
3 of 7
RES EARCH | R E S E A R C H A R T I C L E
A
P granule
Dissolution
-3
-2
Oocytes
-1
Wild-type
meg-3 meg-4
n
o
i
t
a
z
i
l
i
t
r
e
F
Zygote
B
e
p
y
t
-
d
W
l
i
4
-
g
e
m
3
-
g
e
m
C
)
3
m
µ
(
e
m
u
o
V
l
e
t
a
s
n
e
d
n
o
C
G
-1 Oocyte
in utero
Wild-type
meg-3 meg-4
8
6
4
2
0
0.0
1.0
0.5
Radius (µm)
Pickering Agent
in silico
)
3
m
µ
(
e
m
u
o
V
l
2.5
2.0
1.5
1.0
0.5
0.0
Oocytes
in utero
Zygote
in utero
Zygote
ex utero
D
)
3
m
µ
(
e
m
u
o
V
l
e
t
a
s
n
e
d
n
o
C
H
-1 Oocyte Dissolved
in utero
Wild-type
meg-3 meg-4
8
6
4
2
0
0.0
1.0
0.5
Radius (µm)
No Pickering Agent
in silico
)
3
m
µ
(
e
m
u
o
V
l
2.5
2.0
1.5
1.0
0.5
0.0
E
)
3
m
µ
(
e
m
u
o
V
l
e
t
a
s
n
e
d
n
o
C
I
Zygote
in utero
Wild-type
meg-3 meg-4
8
6
4
2
0
0.0
1.0
0.5
Radius (µm)
Experimental
F
)
3
m
µ
(
e
m
u
o
V
l
e
t
a
s
n
e
d
n
o
C
J
Zygote
ex utero
Wild-type
meg-3 meg-4
8
6
4
2
0
0.0
1.0
0.5
Radius (µm)
in silico
)
3
l
m
µ
(
e
m
u
o
V
e
g
a
r
e
v
A
Wild-type
meg-3 meg-4
0.6
0.4
0.2
0.0
)
3
l
m
µ
(
e
m
u
o
V
e
g
a
r
e
v
A
Pickering Agent
No Pickering Agent
0.6
0.4
0.2
0.0
0
200
Time (s)
400
600
0
200
400
Time (s)
600
Oocyte
Zygote
Oocyte
Zygote
Fig. 4. MEG-3 and MEG-4 stabilize P granules
against coarsening during the oocyte-to-zygote
transition. (A) Schematics depicting the dissolution
of P granules (magenta) during the transition
from oocyte to fertilized zygote. The numbers
indicate the relative position of each oocyte in the
germline, and blue represents DNA. (B) Photo-
micrographs of wild-type and meg-3 meg-4 oocytes
and zygotes expressing PGL-3::mCherry (white).
Photomicrographs were captured in live adult her-
maphrodites (in utero) or after dissection out of the
uterus (ex utero). Representative photomicrographs
are max projections corresponding to ~20% of oocyte
volume and ~80% of zygote volume. The anterior
(left) bias for PGL condensates in meg-3 meg-4
zygotes correlates with anterior displacement of the
oocyte nucleus (and associated P granules) that
occurs immediately before fertilization. The white
dashed lines indicate the boundary of each oocyte or
zygote. Scale bars are 10 mm. (C to F) Histograms of
PGL condensate volumes measured from images
captured as in (B) representing 100% of oocyte and
zygote volumes. Circles indicate the volume of
individual PGL-3 in condensates binned by condensate
radius in wild-type (green) and meg-3 meg-4 (black)
oocytes and zygotes. Volumes are higher in (F) than in
(E) owing to higher detection sensitivity ex utero.
(G and H) Graphs showing the evolution of individual
PGL condensates in a 10-min period starting after
dissolution in wild-type and meg-3 meg-4 oocytes and
zygotes under simulated conditions. Each line repre-
sents the evolution of a single condensate over time.
(I and J) Graphs showing the average volume of
individual PGL condensates in oocytes and zygotes
under experimental and simulated conditions. For (I),
each dot corresponds to an oocyte [same dataset
as shown in (D)] or zygote [same data set as shown in
(F)] of the indicated genotypes. For (J), each dot
corresponds to one simulation. Simulations were run in
the presence or absence of the Pickering agent.
Horizontal and vertical lines represent the mean SD.
required for P granule dissolution during po-
larization, and its mammalian homolog DYRK3
has been implicated in the dissolution of con-
densates in mammalian cells (28, 35, 36). We
found that recombinant DYRK3 phosphoryl-
ates PGL-3 efficiently in vitro (fig. S4A). Addi-
tion of DYRK3 and adenosine triphosphate
(ATP) to preassembled PGL-3 condensates
(2.5 mM) led to their complete dissolution over
60 min (Fig. 3, A and B). The effect did not
depend on the hydrotrope properties of ATP
(37) because dissolution was not observed in
the presence of guanosine triphosphate (GTP)
(Fig. 3, A and B). Sedimentation experiments
confirmed that phosphorylation by DYRK3
increases the fraction of PGL-3 in the soluble
pool (fig. S4B). At a higher concentration of
PGL-3 (5 mM), condensates could be main-
tained in the presence of DYRK3 (Fig. 3C).
We used these supersaturated conditions to
conduct microrheology experiments on PGL-3
droplets treated with DYRK3 (Fig. 3D and
movie S8). We found that DYRK3 decreases
the viscosity of PGL-3 droplets (Fig. 3D). Con-
sistent with accelerated internal dynamics,
DYRK3-treated PGL-3 condensates also coars-
ened rapidly (Fig. 3, C and E). Addition of
MEG-3 did not interfere with PGL-3 phos-
phorylation but led to a ~3-fold increase in
the frequency of small PGL-3 droplets (Fig. 3, C
and E, and fig. S4D). The small droplets were
covered with MEG-3 and remained stable for
120 min, even in the presence of much larger
condensates (fig. S4, E to G). These observa-
tions indicate that MEG-3 stabilizes the PGL-3
emulsion against coarsening, even under con-
ditions where PGL-3 dynamics have been
accelerated by phosphorylation.
MEG-3 and MEG-4 stabilize P
granules against coarsening during
the oocyte-to-zygote transition
To examine the impact of MEG-3 on P granule
dynamics in vivo, we used quantitative live-
cell imaging to measure the number and size
of PGL-3 condensates in wild type and meg-3
meg-4 mutants (where meg-3 and meg-4
are deleted). We began by imaging eggs in
utero as they progress through oocyte matu-
ration and fertilization (Fig. 4, A and B). We
found that the volume of PGL-3 in conden-
sates decreased 10-fold during oocyte matu-
ration and remained low in newly fertilized
zygotes as they completed the meiotic divi-
sions (Fig. 4, C to F, and fig. S5A). Total PGL-3
levels did not change during this period (fig.
S5B), consistent with a transient increase
in PGL-3 solubility. We observed the same
Folkmann et al., Science 373, 1218–1224 (2021)
10 September 2021
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
A
e
p
y
t
-
d
W
l
i
4
-
g
e
m
3
-
g
e
m
B
Meiosis
Pronuclear
Formation
Pronuclear
Meeting
Mitosis
Two-cell
Polarization (~15 min)
i
n
o
s
v
D
i
i
l
l
e
C
Wild-type
C
meg-3 meg-4
D
Wild-type
E
meg-3 meg-4
150
Posterior
Anterior
Posterior
Anterior
40
)
3
m
µ
(
Posterior
Anterior
40
)
3
m
µ
(
Posterior
Anterior
150
100
R
E
B
M
U
N
e
t
a
s
n
e
d
n
o
C
50
0
Meiosis PF PM Mitosis Two-
cell
Meiosis PF PM Mitosis Two-
cell
30
20
10
E
M
U
L
O
V
e
t
a
s
n
e
d
n
o
C
0
Meiosis PF PM Mitosis Two-
cell
30
20
10
E
M
U
L
O
V
e
t
a
s
n
e
d
n
o
C
Wild-type
Posterior
G
)
s
/
m
n
(
e
a
R
t
Anterior
0
0.2
0.6
0.4
Radius (µm)
0.8
1.0
1
0
-1
-2
-3
-4
-5
0
0.2
Wild-type
meg-3 meg-4
K
80
n
o
i
t
a
z
i
l
a
c
o
o
C
l
t
n
e
c
r
e
P
60
40
20
0
Meiosis PF PM Mitosis Two-
cell
meg-3 meg-4
Posterior
Anterior
0
3
2
1
0
meg-3 meg-4
Posterior
Anterior
H
Wild-type
1.0
Posterior
Anterior
0.5
)
3
m
µ
(
e
m
u
o
V
l
I
)
3
m
µ
(
e
m
u
o
V
l
0.6
0.4
Radius (µm)
0.8
1.0
0.0
0
Meiosis
P. Formation
P. Meeting
Mitosis
L
)
3
m
µ
(
e
m
u
o
V
l
1.0
0.5
0.0
0
Wild-type meg-3 meg-4
100
200
Time (s)
300
0
100
200
Time (s)
300
in silico Wild-type
Posterior
Anterior
in silico meg-3 meg-4
Posterior
Anterior
M
)
3
m
µ
(
e
m
u
o
V
l
3
2
1
0
100
200
Time (s)
300
0
100
200
Time (s)
300
R
E
B
M
U
N
e
100
t
a
s
n
e
d
n
o
C
50
0
1
0
-1
-2
-3
-4
-5
F
)
s
/
m
n
(
e
a
R
t
J
3
-
L
G
P
2
-
K
B
M
Fig. 5. MEG-3 and MEG-4 drive asymmetric growth of the P granule emulsion during polarization.
(A) Top row: Cartoons depicting PGL-3 condensates (magenta) and MEX-5 (gray) at different
stages during the transition from unpolarized to polarized zygote. Blue represents DNA. Bottom
two rows: Photomicrographs of wild-type and meg-3 meg-4 zygotes (ex utero) expressing
PGL-3::mCherry (white) and matching the stages shown in the cartoons above. Scale bar is 10 mm
and applies to all images. (B to E) Graphs showing the total number [(B) and (C)] and total volume
[(D) and (E)] of PGL-3::mCherry condensates in the posterior (light color) or anterior (dark color) half of
wild-type and meg-3 meg-4 zygotes calculated from the photomicrographs shown in (A). Circles
represent the average from five zygotes, and error bars represent the SD. PF, pronuclear formation;
PM, pronuclear meeting. (F and G) Graphs showing the rate of change in radius of PGL-3::mCherry
condensates in wild-type and meg-3 meg-4 zygotes calculated from traces shown in (H) and (I).
(H and I) Graphs showing the evolution of individual PGL-3::mCherry condensates in anterior and
posterior regions during polarization in wild-type and meg-3 meg-4 zygotes. Traces begin at pronuclear
formation and end just before pronuclear meeting. (J) Photomicrographs of fixed zygotes (mitosis)
of indicated genotypes showing the distribution of MBK-2::OLLAS (green) and PGL-3::mCherry
(magenta). Note that, in addition to P granules, MBK-2 localizes to centrosomes. Scale bar is 10 mm and
applies to all images in the set. (K) Graph showing the percentage of PGL-3:mCherry condensates
colocalized with MBK-2::OLLAS puncta in wild-type or meg-3 meg-4 zygotes at the indicated
developmental stages. Each circle represents one zygote (>50 puncta), and each line represents
the mean. (L and M) Graphs showing the evolution of individual anterior (dark color) and posterior
(light color) PGL-3 condensates under conditions simulating “wild-type” (starting condensate sizes as in
wild-type zygotes, high conversion rates, and Pickering agent) and “meg-3 meg-4” (starting condensate
sizes as in meg-3 meg-4 zygotes, low conversion rates, and no Pickering agent). Compare with
experimental data in (H) and (I).
decrease in PGL-3 condensate volume in wild-
type and meg-3 meg-4 oocytes, indicating that
the increase in PGL-3 solubility during the
oocyte-to-zygote transition is not dependent
on meg-3 and meg-4 (Fig. 4, C to F, and fig.
S5, A and B). The size distribution of PGL-3
condensates after dissolution, however, was
different in the two genotypes. The PGL-3
emulsion coarsened rapidly in meg-3 meg-4
zygotes with fewer larger condensates domi-
nating the emulsion (Fig. 4, B and E). By con-
trast, wild-type zygotes maintained many
small PGL-3 condensates, consistent with
MEG-3 stabilizing the PGL-3 emulsion against
coarsening (Fig. 4, B and E). Zygotes in late
meiosis can survive outside of the uterus,
allowing for the acquisition of high-resolution
images ex utero. These images confirmed that
wild-type zygotes contain dozens of <1-mm
condensates not observed in meg-3 meg-4 zy-
gotes (Fig. 4, B and F).
These observations suggest that MEG-3
functions as a Pickering agent for the PGL-3
emulsion. To examine the physical plausibility
of this hypothesis, we modeled in silico the
kinetics of an idealized PGL-3 emulsion in
the presence or absence of a Pickering agent
that lowers surface tension. To account for
the intrinsically slow dynamics of PGL-3
condensates, we modeled PGL-3 dynamics
under a “conversion-limited” scheme, where
the soluble-to-condensate conversion rate of
PGL-3 molecules is much slower than their
diffusion-limited adsorption-desorption rate
and is therefore rate limiting for condensate
growth and degrowth (see supplementary text).
To model dissolution of PGL-3 condensates
during oocyte maturation, we assigned a rela-
tively high conversion rate to PGL-3 conden-
sates and a high critical concentration for
PGL-3 phase separation because most PGL-3
condensates dissolve during this period. To
model MEG-3 as a Pickering agent, we re-
duced the Gibbs-Thomson length (propor-
tional to surface tension) of MEG-3–coated
PGL-3 condensates by 100-fold compared with
PGL-3–only condensates. Parameter sweeps
revealed that reductions in the Gibbs-Thomson
length as low as twofold, within the range
observed in vitro (Fig. 2B), were sufficient to
show the same quantitative behavior (see sup-
plementary text). We used the model to run
simulations tracking the dynamics of conden-
sates with starting sizes matching the distribu-
tion of PGL-3 condensates in oocytes after
dissolution (Fig. 4, G and H). Coarsening was
notably different in the presence or absence of
the Pickering agent, with the stabilization of
smaller condensates requiring the Pickering
agent. The simulations reproduced the increase
in average condensate volume observed in
meg-3 meg-4 zygotes in comparison to wild-
type zygotes (Fig. 4, I and J). These results
support the hypothesis that MEG-3 stabilizes
Folkmann et al., Science 373, 1218–1224 (2021)
10 September 2021
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
the PGL-3 emulsion against an increase in
PGL-3 dynamics in newly fertilized zygotes
by lowering the surface tension of PGL-3
condensates.
MEG-3 and MEG-4 drive asymmetric growth of
the P granule emulsion during polarization
Next, we examined PGL-3 dynamics as zy-
gotes transition from unpolarized to polar-
ized. During this period, zygotes exit meiosis,
assemble pronuclei that migrate to the center
of the zygote, and fuse and initiate the first
mitotic division, and MEX-5 redistributes
in an anterior-rich gradient (Fig. 5A). During
pronuclear migration, we observed new PGL-3
condensates appearing throughout the cyto-
plasm in both wild-type and meg-3 meg-4 zy-
gotes (Fig. 5, B and C, and fig. S6A). The total
volume of PGL-3 in condensates also increased
(fig. S6B), without a change in total PGL-3
(fig. S6C), consistent with a return to low
PGL-3 solubility in both genotypes. During
this period, total condensate volume in the
anterior half of the zygote decreased steadily,
eventually reaching zero, whereas total con-
densate volume in the posterior increased
(Fig. 5D). Notably, in meg-3 meg-4 zygotes,
we also observed a decrease and increase in
total condensate volume in the anterior and
posterior, respectively, but the amplitude of
the change was greatly reduced (Fig. 5E). By
mitosis, in wild-type zygotes, all PGL-3 con-
densates were restricted to the posterior cyto-
plasm; by contrast, in meg-3 meg-4 zygotes,
PGL-3 condensates remained stable through-
out the cytoplasm through the first cell divi-
sion (Fig. 5, A to E), suggesting that in meg-3
meg-4 mutants, the PGL-3 emulsion does not
respond efficiently to the MEX-5–driven sol-
ubility gradient.
To examine condensate dynamics directly,
we tracked individual PGL-3 condensates in
live zygotes from pronuclear formation to
pronuclear meeting (Fig. 5, F to I, and movie
S9). These observations confirmed that in
meg-3 meg-4 zygotes, condensates experience
very slow growth and dissolution rates during
polarization, with a slight bias for decay in the
anterior [anterior rate (kA,avg) = −0.19 nm/s;
posterior rate (kP,avg) = 0.06 nm/s; Fig. 5, G
and I]. Growth and decay rates were more
than fivefold faster in wild-type zygotes, with
a clear bias for dissolution in the anterior
and condensation in the posterior (kA,avg =
−1.18 nm/s; kP,avg = 0.50 nm/s; Fig. 5, F and H).
These findings suggest that unlike in oocytes
where PGL-3 dissolution occurs independent-
ly of meg-3 and meg-4, during polarization,
meg-3 and meg-4 are required to accelerate
PGL-3 condensate dynamics.
In addition to meg-3 and meg-4, P granule
dissolution during polarization requires the
DYRK family kinase MBK-2 (38, 39). Genetic
epistasis experiments have shown that dissolu-
A
Slow dynamics - Slow coarsening
+ Kinase
B
Undersaturated Conditions
C
Supersaturated Conditions
+ Pickering Agent
- Pickering Agent
Fast dynamics - Dissolution
Fast dynamics - Slow coarsening
Fast dynamics - Fast coarsening
Fig. 6. Pickering agents stabilize dynamic emulsions against kinase-accelerated coarsening.
Schematics showing an idealized emulsion of a self-interacting polymer (blue). (A) In untreated condensates,
polymer-polymer binding and unbinding events are slow (red dots), allowing the emulsion to persist with
minimal coarsening over short time scales. (B and C) Phosphorylation by kinase increases solubility
and accelerates internal dynamics [green dots in (C)]. In undersaturated conditions (B), kinase-accelerated
dynamics cause the condensates to dissolve, as is observed in the anterior of wild-type polarized zygotes.
In saturated conditions (C), kinase-accelerated dynamics allow the condensates to rapidly grow by allowing
new molecules from the dilute phase to enter the condensates. In the presence of the Pickering agent
(yellow), the condensates are stabilized against coarsening, as observed in the posterior of wild-type
polarized zygotes. In the absence of the Pickering agent, the condensates coarsen, as observed in meg-3
meg-4 mutants during the oocyte-to-embryo transition. During polarization, MEG-3 and MEG-4 function both
as Pickering agents and recruiters of the kinase MBK-2. In meg-3 meg-4 zygotes undergoing polarization,
PGL condensates are maintained throughout the cytoplasm with minimal coarsening, because MBK-2 kinase
is not recruited to the condensates and PGL dynamics remain slow, as in (A).
tion activity of MBK-2 is dependent on meg-3
and meg-4 (28) (see supplementary text). Con-
sistent with these observations, using an epitope-
tagged allele of endogenous MBK-2, we found
that MBK-2 is recruited to PGL-3 conden-
sates during polarization and that this recruit-
ment is diminished in meg-3 meg-4 mutant
embryos (Fig. 5, J and K). MEG-3 and MBK-2
levels varied more than fivefold among PGL-3
condensates (fig. S6, D to G). Condensate growth
rates during polarization were also heteroge-
neous in a manner that did not correlate with
initial condensate size (Fig. 5, F and G). To-
gether, these observations suggest that heter-
ogeneous recruitment of MEG-3 and MBK-2
variably accelerates PGL-3 condensate dynam-
ics during polarization.
Based on these findings, we built on our
theoretical model of the PGL-3 emulsion,
adding three new features specific to polar-
ization: (i) higher variable conversion rates for
wild-type versus meg-3 meg-4 PGL-3 conden-
sates to reflect heterogeneous fluidization by
MEG proteins and MBK-2, (ii) higher solubil-
ity for PGL-3 in anterior cytoplasm to reflect
the influence of the MEX-5 gradient, and (iii)
lower Gibbs-Thomson lengths for posterior
condensates to reflect sustained coverage of
posterior condensates by MEG proteins (“asym-
metric Pickering agent”). Parameters were
benchmarked to allow for complete dissolution
of anterior granules in ~6 min, as observed
in vivo. We used the model to run simulations
tracking the growth and decay of condensates
matching the size distribution of PGL-3 conden-
sates in vivo before polarization. The simulations
faithfully recapitulated the coarsening-free dis-
solution of anterior condensates and growth
of posterior condensates that were observed
in wild-type zygotes (Fig. 5L and movie S10).
Simulations mimicking conditions in meg-3
meg-4 mutants reproduced the slow dynamics
Folkmann et al., Science 373, 1218–1224 (2021)
10 September 2021
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
observed in those mutants (Fig. 5M and movie
S10). The simulations also reproduced the
rapid increase in PGL-3 condensate volume
(fig. S6H) that is observed during mitosis in
wild-type but not meg-3 meg-4 zygotes (figs.
S5A and S6B). The rapid rise is a consequence
of the faster approach to equilibrium in the
supersaturated environment of the poste-
rior cytoplasm by PGL-3 condensates with
accelerated dynamics driven by MEG pro-
teins and MBK-2.
To test the importance of each feature in the
model, we reran the wild-type simulations,
changing one model feature at a time. Simula-
tions where condensates were assigned uni-
formly low conversion rates yielded decay
rates that were too slow to clear anterior
granules under the in vivo time constraints
(fig. S6I and movie S10). Simulations with
low PGL-3 solubility across the cytoplasm led
to the coarsening of anterior condensates
(fig. S6J and movie S10). Simulations lacking
the Pickering agent led to the coarsening of
posterior condensates (fig. S6K and movie
S10). Together with the in vitro and in vivo
findings, the theory supports a model where
MEG proteins facilitate P granule polarization
in response to the MEX-5–induced saturation
gradient by accelerating PGL-3 conversion dy-
namics (through recruitment of MBK-2) and
by functioning as Pickering agents to lower
surface tension on posterior condensates,
thus preventing coarsening during periods
of fast PGL dynamics (Fig. 6; see supple-
mentary text for a full discussion of the-
oretical considerations).
Discussion
Since their description by Ramsden and Picker-
ing in the early 1900s (19, 20), Pickering agents
have been used widely to stabilize emulsions in
the pharmaceutical, energy, and food indus-
tries (40–42). Unlike surfactants (amphiphilic
molecules that insert at interfaces), Pickering
agents are nanoscale solid particulates that
adsorb to interfaces upon partial wetting by
both phases. Adsorption is energetically fa-
vored and balances the drive to reduce inter-
facial area, stabilizing the emulsion against
coarsening (43). Many types of solid particu-
lates have been shown to function as Picker-
ing agents, from silica to denatured proteins
(44). The first described intrinsically dis-
ordered protein, casein, functions as a natural
Pickering agent in homogenized milk (45). We
characterized intrinsically disordered protein
MEG-3 as an intracellular Pickering agent, and
we speculate that other self-assembling bio-
polymers will exhibit similar properties. In
somatic cells, PGL droplets are covered by
EPG-2 clusters that may function like MEG-3
to regulate the size and dynamics of PGL
droplets in preparation for autophagy (46).
Artificial protein-RNA assemblies that adsorb
to the surface of stress granules have been re-
ported to influence their size and coalescence
(47). mRNAs that accumulate on the surface of
protein condensates could also serve as stabi-
lizing agents (48, 49). Given the rich diversity
of biopolymers in cells, it is tempting to spec-
ulate that biopolymers acting as Pickering
agents will prove a general organizing prin-
ciple for biomolecular condensates. By inter-
acting with enzymes like the DYRK family
kinase MBK-2, biological Pickering agents
also regulate interfacial exchange to control
the flux of molecules in and out of conden-
sates in response to environmental changes.
In this respect, surface-adsorbed biopolymers
fulfill a boundary function for biomolecular
condensates, reminiscent of the role of lipid
bilayers and associated machineries (e.g.,
channels) in membrane-bound organelles.
Indeed, it has been speculated that lipid mem-
branes arose from liposomes that functioned
as Pickering stabilizers for aqueous emul-
sions (50).
RE FERENCES AND NOTES
1. A. A. Hyman, C. A. Weber, F. Jülicher, Annu. Rev. Cell Dev. Biol.
30, 39–58 (2014).
2. L. Jawerth et al., Science 370, 1317–1323 (2020).
3. S. Boeynaems et al., Trends Cell Biol. 28, 420–435
(2018).
4. E. S. Freeman Rosenzweig et al., Cell 171, 148–162.e19
(2017).
5. M. Feric, C. P. Brangwynne, Nat. Cell Biol. 15, 1253–1259
6.
(2013).
J. Berry, S. C. Weber, N. Vaidya, M. Haataja, C. P. Brangwynne,
Proc. Natl. Acad. Sci. U.S.A. 112, E5237–E5245 (2015).
7. S. Elbaum-Garfinkle et al., Proc. Natl. Acad. Sci. U.S.A. 112,
7189–7194 (2015).
8. S. Ranganathan, E. I. Shakhnovich, eLife 9, e56159
9.
(2020).
J. Berry, C. P. Brangwynne, M. Haataja, Rep. Prog. Phys. 81,
046601 (2018).
10. D. T. McSwiggen, M. Mir, X. Darzacq, R. Tjian, Genes Dev. 33,
1619–1634 (2019).
11. J. R. Wheeler, T. Matheny, S. Jain, R. Abrisch, R. Parker, eLife 5,
e18413 (2016).
12. J. E. Lee, P. I. Cathey, H. Wu, R. Parker, G. K. Voeltz, Science
367, eaay7108 (2020).
13. D. S. W. Lee, N. S. Wingreen, C. P. Brangwynne, Nat. Phys. 17,
531–538 (2021).
14. C. F. Lee, C. P. Brangwynne, J. Gharakhani, A. A. Hyman,
F. Jülicher, Phys. Rev. Lett. 111, 088101 (2013).
15. C. A. Weber, C. F. Lee, F. Jülicher, New J. Phys. 19, 053021
(2017).
16. D. Zwicker, A. A. Hyman, F. Jülicher, Phys. Rev. E 92, 012317
(2015).
17. P. C. Bressloff, Phys. Rev. E 101, 042804 (2020).
18. J. D. Wurtz, C. F. Lee, Phys. Rev. Lett. 120, 078102
(2018).
19. W. Ramsden, F. Gotch, Proc. R. Soc. London 72, 156–164
(1904).
20. S. U. Pickering, J. Chem. Soc. Trans. 91, 2001–2021
(1907).
21. C. P. Brangwynne et al., Science 324, 1729–1732
(2009).
22. S. Saha et al., Cell 166, 1572–1584.e16 (2016).
23. M. Hanazawa, M. Yonetani, A. Sugimoto, J. Cell Biol. 192,
929–937 (2011).
24. U. Sheth, J. Pitt, S. Dennis, J. R. Priess, Development 137,
1305–1314 (2010).
27. J. Smith et al., eLife 5, e21337 (2016).
28. J. T. Wang et al., eLife 3, e04591 (2014).
29. A. Putnam, M. Cassani, J. Smith, G. Seydoux, Nat. Struct. Mol.
Biol. 26, 220–226 (2019).
30. Y. Wu et al., Mol. Biol. Cell 30, 333–345 (2019).
31. T. Ngai, S. Bon, Eds., Particle-Stabilized Emulsions and Colloids
(Royal Society of Chemistry, 2014); https://pubs.rsc.org/en/
content/ebook/978-1-84973-881-1.
32. L. M. Jawerth et al., Phys. Rev. Lett. 121, 258101 (2018).
33. A. B. Pawar, M. Caggioni, R. Ergun, R. W. Hartel, P. T. Spicer,
Soft Matter 7, 7710–7716 (2011).
34. C. D. Crowe, C. D. Keating, Interface Focus 8, 20180032
(2018).
35. K. M. Pang et al., Dev. Biol. 265, 127–139 (2004).
36. A. K. Rai, J.-X. Chen, M. Selbach, L. Pelkmans, Nature 559,
211–216 (2018).
37. A. Patel et al., Science 356, 753–756 (2017).
38. S. Quintin, P. E. Mains, A. Zinke, A. A. Hyman, EMBO Rep. 4,
1175–1181 (2003).
39. J. Pellettieri, V. Reinke, S. K. Kim, G. Seydoux, Dev. Cell 5,
451–462 (2003).
40. C. C. Berton-Carabin, K. Schroën, Annu. Rev. Food Sci. Technol.
6, 263–297 (2015).
41. F. Goodarzi, S. Zendehboudi, Can. J. Chem. Eng. 97, 281–309
(2019).
42. C. Albert et al., J. Control. Release 309, 302–332
(2019).
43. D. Gonzalez Ortiz, C. Pochat-Bohatier, J. Cambedouzou,
M. Bechelany, P. Miele, Engineering 6, 468–482 (2020).
44. Y. Yang et al., Front. Pharmacol. 8, 287 (2017).
45. E. M. Redwan, B. Xue, H. A. Almehdar, V. N. Uversky,
Curr. Protein Pept. Sci. 16, 228–242 (2015).
46. G. Zhang, Z. Wang, Z. Du, H. Zhang, Cell 174, 1492–1506.e22
(2018).
47. M. Garcia-Jove Navarro et al., Nat. Commun. 10, 3230
(2019).
48. B. Hampoelz et al., Cell 179, 671–686.e17 (2019).
49. D. Tauber et al., Cell 180, 411–426.e16 (2020).
50. D. C. Dewey, C. A. Strulson, D. N. Cacace, P. C. Bevilacqua,
C. D. Keating, Nat. Commun. 5, 4670 (2014).
51. C. F. Lee, chiufanlee/Pickering-stabilization, Version v1.0,
Zenodo (2021); https://doi.org/10.5281/zenodo.5083065.
AC KNOWLED GME NTS
We thank the Johns Hopkins Integrated Imaging Center
(S10OD023548) for microscopy support. We thank the Lavis lab
for HaloTag ligands JF549 and JF646, the Griffin lab for the MEG-3::
Halo strain, the Waugh lab for TEV protease (pRK793, Addgene),
T. Hyman, the Baltimore Worm Club, and the Seydoux lab for many
helpful discussions. Funding: This work was supported by the
National Institutes of Health (grant numbers R37HD037047
to G.S. and F32GM134630 to A.P.). G.S. is an investigator of the
Howard Hughes Medical Institute. Author contributions: A.W.F.,
A.P., and G.S. designed the research. A.W.F. and A.P. performed all
experiments and collected and analyzed data. C.F.L. performed
the theoretical analysis. A.W.F., A.P., C.F.L., and G.S. prepared the
manuscript with contributions from all authors. Competing
interests: G.S. serves on the scientific advisory board of Dewpoint
Therapeutics, Inc. A.W.F., A.P., and G.S. are inventors on
provisional application #63/094,987 filed on 10/22/2020 held
by Johns Hopkins University that covers the use of intrinsically
disordered proteins as Pickering agents. Data and materials
availability: All data are available in the manuscript or the
supplementary materials. Code for simulations is deposited and
accessible at Zenodo (51).
SUPPLEMENTARY MATERIALS
https://science.org/doi/10.1126/science.abg7071
Materials and Methods
Supplementary Text
Figs. S1 to S9
Tables S1 and S2
References (52–69)
MDAR Reproducibility Checklist
Movies S1 to S10
25. C. M. Schubert, R. Lin, C. J. de Vries, R. H. Plasterk, J. R. Priess,
View/request a protocol for this paper from Bio-protocol.
Mol. Cell 5, 671–682 (2000).
26. J.-X. Chen et al., Mol. Cell. Proteomics 15, 1642–1657
(2016).
22 January 2021; accepted 21 July 2021
10.1126/science.abg7071
Folkmann et al., Science 373, 1218–1224 (2021)
10 September 2021
7 of 7
|
10.1126_science.abc6663
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
CHROMATIN
Regulation of the Dot1 histone H3K79
methyltransferase by histone H4K16 acetylation
Marco Igor Valencia-Sánchez*, Pablo De Ioannes*, Miao Wang*, David M. Truong, Rachel Lee,
Jean-Paul Armache, Jef D. Boeke, Karim-Jean Armache†
ubiquitinated nucleosomes, and established
in vitro and in vivo assays allowed us to
determine the detailed mechanisms of yeast
Dot1 regulation through histone acetylation
and ubiquitination.
RESULTS: We tested nucleosomes with differ-
ent acetylation states of histone H4 in vitro,
and we show that Dot1 is allosterically stimu-
lated by acetylation of H4 and that this effect
INTRODUCTION: Nucleosomes—the primary,
repeating unit of chromatin—package and
protect the genome while also transmitting
various regulatory signals through, in part,
posttranslational modifications of histones.
These histone modifications (for example,
acetylation, ubiquitination, or methylation)
affect critical processes such as transcription,
replication, recombination, and repair, often
forming complicated networks to ensure finely
tuned signaling for chromatin enzymes.
One such enzyme—evolutionarily con-
served disruptor of telomeric silenc-
ing (Dot1)—catalyzes mono-, di-, and
trimethylation of histone H3 lysine 79
(H3K79). H3K79 methylation by Dot1
is a prominent example of trans-histone
cross-talk, a process in which one his-
tone and its modification affects the
modification of another histone. In
mammals, the human homolog Dot1L
plays critical roles in embryogenesis
and hematopoiesis. Although recent
advances have provided insights into
Dot1L stimulation through histone
H2B ubiquitination, how other mod-
ifications mechanistically regulate
Dot1 activity is not known. We were
particularly interested in how histone
lysine acetylation contributes to the
regulation of chromatin enzymes.
Lysine acetylation plays pivotal roles
in chromatin decondensation, tran-
scriptional activation, and mainte-
nance of euchromatin, serving as a
general antisilencing mark in eukar-
yotes. Here, we present mechanistic
studies that show how histone acety-
lation regulates the activity of Dot1.
cryo-EM structures: one of Dot1 in complex,
with nucleosomes bearing both H4K16ac
and H2Bub, and the second of Dot1, with a
nucleosome bearing only H2BUb. Upon
examining our cryo-EM dataset of Dot1
bound to the nucleosome containing un-
acetylated H4, the particles classified into
two main three-dimensional (3D) classes. In
the first class, Dot1 is bound to the nucle-
osome in a catalytic conformation. In the
second class, it is bound in a noncatalytic
conformation. This is different from the
dataset in which Dot1 is bound to an H4K16
acetylated nucleosome. When H4K16ac is
present, the cryo-EM data is more homo-
geneous, and Dot1 is bound to the nucleosome
predominantly in a catalytic conformation.
This suggests a model in which acetylation
of the H4 tail restricts the sampling space
of Dot1, resulting in an active conforma-
tion leading to increased activity. We there-
fore propose that H2BUb partially restricts
the conformation of yeast Dot1 on the nu-
cleosome and that H4K16ac further
restricts and stabilizes the active con-
formation. Comparing both of these
cryo-EM structures allowed us to
identify residues that are critical
for Dot1 stimulation by H4K16ac and
H2BUb. Site-directed mutagenesis of
Dot1 coupled with enzymatic assays
on nucleosomes revealed the details
of these interfaces. These results
show that the allosteric stimulation
of Dot1 by H4K16ac and H2BUb
plays a crucial role in H3K79 di-
and trimethylation.
CONCLUSION: This work demonstrates
how Dot1 is regulated by histone ace-
tylation and how H4K16ac coordi-
nates with H2BUb to regulate Dot1.
H4K16ac plays a critical role in open-
ing chromatin structure by counteract-
ing the binding of silencing proteins,
while simultaneously stimulating an
enzyme that is important for transcrip-
tion. We provide an example in which
the activity of the fundamental methyl-
transferase Dot1 is modulated through
cross-talk between distinct histone
modifications to ensure optimal main-
tenance and propagation of an epigenetic
state. Cross-talk such as this may represent a
general property of chromatin enzymes.▪
The list of author affiliations is available in the full article online.
*These authors contributed equally to this work.
†Corresponding author. Email: karim-jean.armache@
nyulangone.org
Cite this article as M. I. Valencia-Sánchez et al., Science
371, eabc6663 (2021). DOI: 10.1126/science.abc6663
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.abc6663
H3K79 methylation by Dot1 is allosterically stimulated by H4K16
acetylation and by H2BK123 ubiquitination. Cross-talk between
different histone posttranslational modifications can orchestrate distinct
chromatin states to regulate transcription and gene silencing.
RATIONALE: Histone H4 acetylation, histone
H2B ubiquitination, and H3 methylation are
conserved cotranscriptional histone mod-
ifications that work together to ensure the
appropriate regulation of chromatin struc-
ture during transcription. The mechanisms
of cross-talk between these modifications
and enzymes that deposit them are crucial
for understanding transcription. Advances
in cryo–electron microscopy (cryo-EM), the
ability to make specifically acetylated and
is specific to lysine 16 (H4K16ac). The other
known acetylation targets on histone H4
(H4K5ac, H4K8ac, and H4K12ac) do not
stimulate the activity of Dot1, which high-
lights the distinctive role of H4K16 acetyla-
tion in regulating chromatin structure. We
also show that the effect of H4K16 acetylation
is direct and further enhanced by H2B ubiq-
uitination (H2BUb), resulting in an optimal
catalytic rate for Dot1. To gain mechanistic
insights into stimulation by H4K16ac and its
coordination with H2BUb, we determined two
Valencia-Sánchez et al., Science 371, 363 (2021)
22 January 2021
1 of 1
Corrected 28 January 2021. See full text. RES EARCH
R E S E A R C H A R T I C L E
◥
CHROMATIN
Regulation of the Dot1 histone H3K79
methyltransferase by histone H4K16 acetylation
Marco Igor Valencia-Sánchez1*, Pablo De Ioannes1*, Miao Wang1*, David M. Truong2, Rachel Lee1,
Jean-Paul Armache3, Jef D. Boeke2, Karim-Jean Armache1†
Dot1 (disruptor of telomeric silencing-1), the histone H3 lysine 79 (H3K79) methyltransferase, is
conserved throughout evolution, and its deregulation is found in human leukemias. Here, we provide
evidence that acetylation of histone H4 allosterically stimulates yeast Dot1 in a manner distinct from but
coordinating with histone H2B ubiquitination (H2BUb). We further demonstrate that this stimulatory
effect is specific to acetylation of lysine 16 (H4K16ac), a modification central to chromatin structure. We
provide a mechanism of this histone cross-talk and show that H4K16ac and H2BUb play crucial roles in
H3K79 di- and trimethylation in vitro and in vivo. These data reveal mechanisms that control H3K79
methylation and demonstrate how H4K16ac, H3K79me, and H2BUb function together to regulate gene
transcription and gene silencing to ensure optimal maintenance and propagation of an epigenetic state.
H istones are subject to a vast array of
posttranslational modifications (PTMs)
that influence chromatin structure and
function by altering interactions be-
tween nucleosomes or acting as docking
sites for recruitment of effector proteins (1).
Modifications such as acetylation of histones
H3 and H4 (H3ac and H4ac), methylation of
H3K79 (H3K79me), or monoubiquitination
of histone H2B (H2BUb) are associated with
active transcription (2–4). Regulation of these
modifications and the enzymes that deposit
them is critical to transcription, and disrup-
tion of their deposition leads to aberrant gene
expression and disease (5).
Histone acetylation is perhaps the most ex-
tensively studied PTM (6–8). There are many
lysines that can be acetylated, and they play
critical roles in gene regulation (8–10). Acety-
lation of lysines in H3 and H4 N-terminal tails
correlates positively with gene transcription,
peaking sharply at active promoters and pre-
sent in gene bodies, with the levels of acety-
lation being proportional to the transcription
rates (7, 11, 12). Acetylation neutralizes positive
charges of lysines, which results in the open-
ing of the chromatin, allowing greater access
to transcription factors, and facilitating passage
of the RNA polymerase II (13, 14). Acetylated
histones are easier to displace from DNA both
1Skirball Institute of Biomolecular Medicine, Department of
Biochemistry and Molecular Pharmacology, New York
University Grossman School of Medicine, New York, NY
10016, USA. 2Institute for Systems Genetics, Department of
Biochemistry and Molecular Pharmacology, New York
University Langone Health, New York, NY 10016, USA.
3Department of Biochemistry and Molecular Biology, The
Huck Institutes of the Life Sciences, Pennsylvania State
University, University Park, PA 16802, USA.
*These authors contributed equally to this work.
†Corresponding author. Email: karim-jean.armache@
nyulangone.org
in vivo (15) and in vitro (16). In particular,
acetylation of histone H4 lysine 16 (H4K16ac)
inhibits the formation of compact 30-nm
fibers (17).
The conserved factor Dot1 (disruptor of
telomeric silencing-1) is the only known methyl-
transferase that catalyzes mono-, di-, and
trimethylation of H3K79 (H3K79me1, -me2,
and -me3) (18–20). Genome-wide analyses
of H3K79me in yeast, fly, mouse, and human
have demonstrated a high correlation between
this modification and transcriptional activity
(21–25). In Saccharomyces cerevisiae, for ex-
ample, ~90% of the genome is methylated at
H3K79 (20). In mammals, the homologous
Dot1L (Dot1-like) is essential for embryogenesis,
hematopoiesis, and cardiac function (26, 27).
Aberrant transcriptional activation through
Dot1L is found in leukemias that result from
oncogenic chromosomal translocations involv-
ing the MLL gene (28–31). A prerequisite for
efficient H3K79me2 and -me3 by Dot1 and
Dot1L is monoubiquitination of histone H2B
lysine K123 in yeast and K120 in humans
(hereafter H2BUb) (32–36). Recent struc-
tural studies by us and other groups found
that interaction of Dot1L with a hydrophobic
patch on ubiquitin (Ub) reduces sampling
space of Dot1L on the nucleosome, resulting
in higher methylation states (37–41).
The activity of Dot1 has also been linked to
histone acetylation. For example, the activity
of the histone deacetylase Rpd3L restricts
H3K79me3 at its euchromatic targets in bud-
ding yeast, and inactivation of the Rpd3 homo-
log HDAC1 in mouse thymocytes leads to an
increase in H3K79me (42). H4K16ac regulates
the Dot1-mediated distribution of H3K79me
on euchromatin (43). H4K16A and H4K16R mu-
tations decrease the global level of H3K79me3
(43). Genome-wide H3K79me2 and -me3 were
decreased in a null mutant of the H4K16-specific
acetyltransferase Sas2p (sas2D) (43). These ob-
servations suggest a possible existence of
evolutionarily conserved cross-talk between
histone acetylation and H3K79me by Dot1 and
Dot1L. It is unknown whether histone acety-
lation stimulates Dot1 and Dot1L directly, and
the mechanism of such putative cross-talk is
unknown.
Dot1 was originally implicated in the silenc-
ing of genes in yeast telomeres (44). Telomeric
silencing is established through the recruitment
and binding of SIR (silent information regula-
tor) complex (Sir2, -3, and -4) to chromatin
(45, 46). Activity of Sir2, nicotinamide adenine
dinucleotide (NAD+)–dependent H4K16 de-
acetylase, is essential to create a nucleosomal
binding site for Sir3 (47–49). Nucleosome-Sir3
binding is maximally perturbed when H4K16
is acetylated and H3K79 is methylated (50, 51).
Overexpression of Dot1 spreads H3K79me
into silent chromatin, displacing Sir proteins,
and mutation of H3K79 or deletion of Dot1
compromises telomeric silencing by mislocal-
izing the Sir complex (20). Furthermore, it has
been shown that a basic patch on the histone
H4 tail is critical for Dot1 binding and H3K79
methylation (52, 53) and that Sir3 competes
with Dot1 for this site (53, 54).
Given the strong correlation between H4
acetylation, H2B ubiquitination, and H3K79
methylation and the critical role of Dot1, we
sought to determine the nature of this puta-
tive cross-talk. The cross-talk between H4K16ac
and Dot1 could involve (i) an indirect effect,
through structural changes to the nucleosome
caused by acetylation; (ii) an interaction be-
tween Dot1 and H4K16ac mediated by another
accessory protein; or (iii) direct interaction
and stimulation of Dot1 by H4K16ac. To dis-
tinguish among these different mechanisms,
we used a combination of biochemical, struc-
tural, and functional approaches. We demon-
strate that direct interaction of H4K16ac with
Dot1 stimulates its catalytic activity. We also
show that Dot1 H3K79me activity on the doubly
modified substrate (H4K16ac-H2Bub nucleo-
some) is higher when compared with that on
the singly modified nucleosome. We define
the structural basis of H4K16ac recognition
by Dot1 that results in stimulation of its
H3K79me activity and propose a molecular
mechanism for this trans-histone cross-talk.
Last, we assess the biological impact of our
proposed biochemical mechanism for Dot1
stimulation by H4K16ac and H2BUb. On the
basis of our structural, biochemical, and in vivo
functional data, we propose a direct interac-
tion model that describes the rules of Dot1
stimulation by the combined action of histone
acetylation and ubiquitination. An extended
view of the implications of our data suggests a
scheme for how H3K79me by Dot1 regulates
the kinetics of gene silencing in yeast.
Valencia-Sánchez et al., Science 371, eabc6663 (2021)
22 January 2021
1 of 9
Corrected 28 January 2021. See full text. RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Acetylation of H4K16 directly stimulates
the catalytic activity of Dot1 on nucleosome
substrates. (A) (Top) Multiple sequence alignment
of the H4 tail showing the conservation of the
N-terminal region. Residues acetylated for biochemistry
experiments are indicated with numbers. (Bottom)
Endpoint methyltransferase assay for yeast Dot1 using
Unmod nuc, tetraacetylated (K5acK8acK12acK16ac), or
monoacetylated (H4Kxac) nucleosomes. Numbers
above the columns are ratios between Dot1 activity
on acetylated H4 and unmodified nucleosome sub-
strates. (B) Representative endpoint methyltransferase
assay for Dot1 by using Unmod nuc, H4K16ac nuc, Ub
nuc, or ac/Ub nuc substrates. Each data point and error
bar indicate the mean ± SD from three independent
experiments. (C) Representative HMT (histone
methyltransferase) assay measuring activity of Dot1 on
different nucleosome substrates. HMT assays were
performed with an increasing amount of Dot1 in the
presence of Unmod nuc, H4K16ac nuc, Ub nuc, or ac/Ub
nuc substrates, and reaction products were identified
by using Western blot. (D) Binding curves of Dot1 to
Unmod nuc (Kd = 70.8 nM) or H4K16ac nuc (Kd =
83.3 nM) measured with EMSA. Each data point and
error bar indicate the mean ± SD from three independent
experiments. The standard errors of dissociation
constants (Kd) are indicated. (E) Michaelis–Menten
saturation curves for Dot1 on Unmod nuc or modified
nucleosomes. The Km and kcat values of the fitted
data are reported in the graph. Each data point and error
bar represent the mean ± SD from three independent
experiments. The reported errors of the fitted kcat
and Km correspond to the standard error. (F) 3.1-Å
cryo-EM reconstruction of Dot1-H4K16ac structure
displayed in two separate views related by 90°.
(G) Structural model of Dot1-H4K16ac complex. The
Dot1 catalytic domain is depicted in purple, ubiquitin in
cyan, DNA in gray, histone H2A in pale yellow, histone
H2B in red salmon, histone H3 in marine blue, and
histone H4 in lime green.
A
Sc H4
Hs H4
Xl H4
1
1
1
S G R G K G G K G L G K G G A K R H R K
S G R G K G G K G L G K G G A K R H R K
S G R G K G G K G L G K G G A K R H R K
20
20
20
B
)
U
L
R
(
e
c
n
e
c
s
e
n
m
u
L
i
7.2
250000
200000
150000
100000
2.0
50000
0
0.4
0.1
0.5
Dot1
Unmod nuc
Tetra ac
H4K5ac
H4K8ac
H4K12ac
H4K16ac
E
Unmod nuc
app (nM)
K
d
70.8 ± 6.5
H4K16ac nuc
83.3 ± 15.0
100
Dot1 [nM]
1000
d
n
u
o
b
n
u
n
o
i
t
c
a
r
F
-
1
1.0
0.5
0.0
10
D
F
G
Dot1
400
300
200
100
0
1
i
1
-
n
m
d
e
c
u
d
o
r
p
H
A
S
90°
90°
)
U
L
R
(
e
c
n
e
c
s
e
n
m
u
L
i
9.9
6.1
Unmod nuc
H4K16ac nuc
3.3
Ub
nuc
H4K16ac/Ub nuc
1000000
800000
600000
400000
200000
0
C
H3-
H3-
H3-
Dot1-
Ub H2B-
H3-
H2A/H2B
H4-
Unmod
H4K16ac
Ub
ac/Ub
Nuc
Dot1
H3K79me3
H3K79me2
H3K79me1
Dot1
Blue-51
K
m
(nM)
k
cat
(min-1)
Unmod nuc
381 ± 227
0.85 ± 0.21
H4K16ac nuc
264 ± 40.0
17.5 ± 0.88
Ub nuc
387 ± 59.2
9.71 ± 0.54
H4K16ac/Ub nuc
456 ± 46.5
35.5 ± 1.37
10
100
1000
10000
Nucleosome [nM]
DOT1
Ub
H2A
H2B
H3
H4
DNA
Results
Acetylation of H4K16 directly stimulates the
catalytic activity of Dot1 on nucleosomes
Four invariant lysine residues in the histone
H4 N terminus (K5, K8, K12, and K16) can
be acetylated in eukaryotes (Fig. 1A, top). To
test whether global acetylation of histone H4
stimulates Dot1, we measured activity of the
catalytic domain of yeast Dot1 (residues 158
to 582) using semisynthetic tetraacetylated
(H4K5acK8acK12acK16ac) “designer” nucleo-
somes as substrates (Fig. 1A, bottom). We ob-
served a twofold stimulation of Dot1 activity
by tetraacetylated H4 nucleosomes as com-
pared with unmodified nucleosomes (“Unmod
nuc”). To understand whether any specific H4
residue underlies stimulation, we performed
the enzymatic reactions using designer nucleo-
somes acetylated at individual sites (H4K5ac,
H4K8ac, H4K12ac, or H4K16ac) (Fig. 1A, bot-
tom). Of all individually acetylated tested
residues, only acetylated lysine 16 showed a
substantial (sevenfold) stimulation of Dot1 cat-
alytic activity (Fig. 1A, bottom). This obser-
vation established that acetylation of H4K16
directly stimulates catalytic activity of Dot1
and suggested that the modification of either
K5, K8, or K12 may inhibit this stimulatory
effect in the context of nucleosome tetraace-
tylated on H4 tail. This result is in line with the
biological roles of K5, K8, and K12, which are
acetylated during chromatin assembly, and K16
acetylation, which regulates chromatin open-
ing and gene activation (17).
H2BUb stimulates both yeast and human
Dot1 activity (32, 34, 55–57). To understand
how this well-described stimulation compares
with that of H4K16 acetylation, we reconsti-
tuted singly modified H4K16ac and H2BUb
nucleosome substrates (fig. S1) and measured
methyltransferase activity of Dot1 (Fig. 1B).
Dot1 stimulation by singly modified H4K16ac
nucleosomes is higher (roughly twofold) than
that by control H2BUb-only nucleosomes (Fig.
1B). We asked whether the combination of
these two histone modifications on a single
nucleosome results in further stimulation of
Dot1 activity. We reconstituted H4K16ac/
H2BUb (“doubly modified”; ac/Ub) nucleo-
somes and used them as a substrate (Fig. 1B
and fig. S1). We observed that the doubly mod-
ified substrate increased the stimulation effect
on Dot1 compared with single modifications
(Fig. 1B). Then, we performed enzymatic
assays probing levels of H3K79 methylation
(me1, me2, and me3) on Unmod nuc, H4K16ac
(H4K16ac nuc), H2BUb (Ub nuc), and doubly
modified (ac/Ub nuc) nucleosomes by using
immunoblot (Fig. 1C). Although unmodified
nucleosomes can be monomethylated, the
higher-level methylation states (me2 and me3)
can only be achieved when stimulatory histone
modifications, such as H4K16ac or H2BUb or
both, coreside on nucleosomes (Fig. 1C). We
further confirmed these results using full-
length yeast Dot1 (fig. S2).
Valencia-Sánchez et al., Science 371, eabc6663 (2021)
22 January 2021
2 of 9
Corrected 28 January 2021. See full text.
RES EARCH | R E S E A R C H A R T I C L E
Then, we tested whether stimulation by
H4K16ac is due to higher binding affinity of
Dot1 by using electrophoretic mobility shift
assays (EMSAs) with unmodified and H4K16ac
nucleosomes (Fig. 1D and fig. S3). We did not
observe any difference in affinity as a func-
tion of H4K16ac [Dissociation constant (Kd)
on unmod nuc, 70.8 nM; on H4K16ac nuc,
83.3 nM]. To further understand the mech-
anism, we performed enzyme kinetics using
unmodified, H4K16ac, H2BUb, and doubly
modified nucleosomes (Fig. 1E). We observed
that both acetylated and ubiquitinated nu-
cleosomes result in a higher apparent uni-
molecular rate constant (kcat) compared with
that of unmodified nucleosomes and that this
effect is even stronger on doubly modified
nucleosomes, whereas the Michaelis constant
(Km) for methyl transfer is similar for all assayed
substrates (average Km = ~375 nM) (Fig. 1E).
Binding and kinetic data for H4K16ac sug-
gest an allosteric role of this modification in
stimulating Dot1.
Structural basis for Dot1 stimulation by
H4K16 acetylation
To gain mechanistic insight of Dot1 stimula-
tion by the combined action of H4K16ac and
H2BUb modifications, we determined the
cryo–electron microscopy (cryo-EM) struc-
ture of the catalytic domain of yeast Dot1
(residues 158 to 582) bound to the H4K16ac/
H2BUb/H3K79M nucleosome (hereafter “Dot1-
H4K16ac structure”) (Fig. 1F). We used the
H3K79M mutation because it was shown to
increase the affinity of methyltransferases
for nucleosomes (58, 59). We reconstituted the
complex, cross-linked it with glutaraldehyde
using the GraFix method (60), froze grids, and
collected cryo-EM data (fig. S4 and S6) on a
Titan Krios (300 kV). This resulted in a 3.1 Å–
resolution map of the complex, which we used
to unambiguously model the structures of the
nucleosome, catalytic domain of yeast Dot1, and
ubiquitin (Fig. 1, F and G). The structure re-
vealed Dot1 (residues 176 to 580) bound to the
nucleosome in a catalytic conformation, with
clearly resolved nucleosome-Dot1, Ub-Dot1,
and H4-Dot1 interfaces (Fig. 1, F and G, and
fig. S7). The cryo-EM map is of high quality
and allows unambiguous interpretation of
key interactions (Fig. 1F and figs. S7, S8, and
S9). The understanding of the mechanism of
stimulation by H4K16ac and H4K16ac/H2BUb
nucleosomes required a control structure of
Dot1 bound to nucleosomes with unacetylated
histone H4K16 (hereafter “Dot1-unacetylated
H4 structure”) (fig. S5) and followed the same
experimental procedures to determine the cryo-
EM structure of this complex at a comparable
resolution of 3.2 Å (figs. S5, S6, and S9). These
structures allowed us to make detailed compar-
isons to understand the structural changes
underlying Dot1 stimulation by H4K16ac.
Upon careful examination of cryo-EM data,
we found that in the Dot1-unacetylated H4
dataset, most of the particles classified into
two main three-dimensional (3D) classes (fig.
S6). In the first class, Dot1 is bound to the nu-
cleosome in a catalytic conformation, and in
the second class, it is bound in a noncatalytic
conformation, resembling the “poised state”
previously described for Dot1L (figs. S6 and
S10) (38, 40). By contrast, in the Dot1-H4K16ac
dataset, only the catalytic conformation can be
observed, leading us to propose a model in
which acetylation of the H4 tail restricts the
sampling space of Dot1, resulting in an active
conformation and leading to increased activity.
To investigate the mechanisms of this stabili-
zation, we compared the Dot1-unacetylated H4
and Dot1-H4K16ac structures in catalytically
competent conformations. Overall, both struc-
tures are very similar (Fig. 2, A and B, and fig.
S7), but with two major differences in histone
H4 and in the Acetyl-Gate loop (“AcG loop”;
residues 252 to 264) of Dot1. These differences
imply a stabilization of the Dot1-H4 tail in-
terface when K16 is acetylated. In the Dot1-
H4K16ac structure, the H4 tail is ordered
starting from the backbone of residue K12
(Fig. 2A and figs. S11 and S12), whereas in
the Dot1-unacetylated H4 structure, the tail
density is only visible from the backbone of
residue K16 (Fig. 2B and figs. S11 and S12).
Similarly, in the Dot1-H4K16ac structure, resi-
dues 255 to 262 of the AcG loop of Dot1 are
ordered, whereas the density for this region is
spurious in the Dot1-unacetylated H4 struc-
ture (figs. S11 and S12). In both structures, his-
tidine H355 in Dot1 makes potential H-bonds
between main-chain carbonyls of H4R17 and
H4K16 and between H347 in Dot1 and the
main-chain carbonyl of H4K16 (Fig. 2, A and
B, and fig. S11). In the Dot1-H4K16ac structure,
we observed an additional potential H-bond
between the main chain carbonyl of A15 of H4
and Dot1 H347 (Fig. 2A). In the Dot1-H4K16ac
structure, a clear density of the H4K16ac side
chain allowed us to analyze the key interac-
tions (fig. S11). The carbonyl component of the
acetyl group is within H-bonding distance
(~3.5 Å) of the imidazole group of H355 (Fig.
2A). Additionally, the methyl component of
the acetyl group establishes van der Waals
interactions with the side chain of Dot1 I261,
stabilizing its backbone. The I261 main chain
carbonyl forms a H-bond with Dot1 H355, re-
sulting in overall stabilization of the AcG loop
of Dot1 (Fig. 2A). By contrast, the side chain of
H4K16 is not visible in the structure of the Dot1-
unacetylated H4 nucleosome (Fig. 2B and figs.
S11 and S12). On the basis of this, we hypoth-
esize that in the unacetylated nucleosome, the
imidazole of H347 and H355 of Dot1 may repel
the positively charged, protonated amino group
of the lysine H4K16, preventing stabilization of
I216, which in turn results in destabilization of
the AcG loop in Dot1 and H4 N terminus
(residues 12 to 16). Acetylation of H4K16 would
neutralize its charge, thereby favoring stabiliz-
ing interactions.
We examined whether stabilization of H4
and Dot1 as a function of H4K16 acetylation
affects the Dot1 active site (Fig. 2C, blue arrow).
The active site of Dot1 consists of hydropho-
bic and aromatic residues of Dot1, cofactor
S-adenosyl methionine (SAM), and residues of
the H4 amino tail (Fig. 2C). H3K79 is inserted
into the hydrophobic active site and anchored
through van der Waals interactions of residues
including W543, F481, V371, and L482 of Dot1.
Furthermore, W543, F481, and F367 are present
in loops that are stabilized upon binding of H3
substrate and critical for activity (40). In par-
ticular, W543 might be involved in establish-
ing cation-p interactions with K79 and its
different methyl states (61, 62). Dot1 binding
induces a conformational change in loop L1
of histone H3 residues Q76, F78, K79, T80,
and D81, reorienting the side chain of K79
(mutated in this work to methionine) ~95°
and directing it ~10 Å away of the nucleosome,
which moves it closer to the SAM cofactor and
into the Dot1 active site (Fig. 2C and fig. S13).
The position of hydrophobic residues in the
Dot1 active site and the contacts of this re-
gion with nucleosome are similar in the Dot1-
unacetylated H4 and Dot1-H4K16ac structures
and are conserved in human Dot1L (fig. S13).
W543 and F367 of Dot1 and the H4 tail resi-
dues R17, H18, and R19 play critical roles in the
remodeling of H3 (40). H4R19 can make con-
tacts with the backbone of H3 residues T80,
K79, and Q76 and is also in proximity to con-
tact the carbonyl of the main chain of the
W543 that is part of the hydrophobic pocket
(Fig. 2C and fig. S13). H4H18 contacts S542 of
Dot1, probably helping to stabilize the loop of
W543, and is also close in distance to establish
a stacking interaction with Dot1 Y372 that is
part of the SAM binding site. H4R17 is well
defined in the Dot1-H4K16ac structure. It es-
tablishes several contacts with Dot1, including
interactions with E374 and N404. Previous
studies established a critical role of H4 tail
overall and for residues 17 to 19 in particular
in regulating Dot1 in vivo and in vitro (52, 53).
Substitution of R17 or R19 for neutral or nega-
tive charges retained only wild-type levels of
monomethylation, abolishing higher methyla-
tion states (53). Our structures suggest that
because of direct contacts of R17, H18, and R19
of H4 in the active site, any conformational
change in H4 such as that resulting from the
stabilization by H4K16ac would directly affect
the structure at the active site in a manner that
increases catalysis (Fig. 2, A to C).
To further characterize Dot1 stimulation by
H4K16ac, we mutated the residues at the Dot1:
H4K16ac interface and measured Dot1 enzy-
matic activity (Fig. 2D and fig. S14). Double
Valencia-Sánchez et al., Science 371, eabc6663 (2021)
22 January 2021
3 of 9
Corrected 28 January 2021. See full text. RES EARCH | R E S E A R C H A R T I C L E
Dot1-H4K16ac structure
Dot1-unacetylated H4 structure
A
B
K12
H347
A15
K12
H347
R17
H355
I261
K16ac
K16
R17
I261
H355
C
Dot1-H4K16ac structure
N404
R17
H18
Y372
SAM
E374
G373
N479
V371
K16ac
AcG
Loop
S542
W543
D81
R19
Loop 1
K79M
L482
F481
T80
F78
E H347E/H355E (H4 mutant)
Unmod
H4K16ac
Ub
ac/Ub
Nuc
Dot1
H3K79me3
H3K79me2
H3K79me1
Dot1
Blue-51
H3-
H3-
H3-
Dot1-
Ub H2B-
H3-
H2A/H2B
H4-
H3-
H3-
H3-
Dot1-
Ub H2B-
H3-
H2A/H2B
H4-
D
2500000
0.125 pmols
0.250 pmols
)
U
L
R
(
e
c
n
e
c
s
e
n
m
u
L
i
2000000
1500000
1000000
500000
0
WT
H347E/H355E
H355E
I261E
Unmod nuc
H4K16ac nuc
Ub nuc
H4K16ac/Ub nuc
I261E
WT
H347E/H355E
H355E
H355E (H4 mutant)
I261E (H4K16ac mutant)
Unmod
H4K16ac
Ub
ac/Ub
Nuc
Dot1
H3K79me3
H3K79me2
H3K79me1
Dot1
Blue-51
H3-
H3-
H3-
Dot1-
Ub H2B-
H3-
H2A/H2B
H4-
Unmod
H4K16ac
Ub
ac/Ub
Nuc
Dot1
H3K79me3
H3K79me2
H3K79me1
Dot1
Blue-51
Fig. 2. Structural basis for Dot1 stimulation by H4K16 acetylation. (A and B) Side-by-side overview and
close-up of interactions of Dot1 with the H4 tail in the (A) Dot1-H4K16ac structure and (B) Dot1-unacetylated
H4 structure. The structure is color-coded as in Fig. 1. The yellow dashed lines show distances of ≤3.5 Å. (C) Close-up
of interactions of Dot1 with H3K79M at the catalytic site (blue arrow), the position of residues in H4 tail, and SAM.
The red square indicates the region of Dot1 and H4 that is stabilized by H4K16 acetylation as shown in (A).
(D) Representative endpoint methyltransferase assay for Dot1 and mutants (H347E-H355E, H335E, and I261E) in
the presence of Unmod nuc or H4K16ac nuc or Ub nuc or ac/Ub nuc substrates. Reactions were performed with
0.125 and 0.250 pmol of Dot1. Each data point and error bar indicate the mean ± SD from three independent
experiments. (E) Representative HMT assay measuring activity of Dot1 and its mutants on different nucleosome
substrates. HMT assays were performed with an increasing amount of Dot1 mutants (H347E-H355E, H335E,
and I261E) in the presence of Unmod nuc or H4K16ac nuc or Ub nuc or ac/Ub nuc substrates, and reaction products
were identified by using Western blot.
mutation of H347E-H355E and point muta-
tion of H355E resulted in almost complete loss
of Dot1 activity on unmodified and H4K16ac
nucleosomes and a decreased activity on H2BUb
and doubly modified nucleosomes (Fig. 2D and
fig. S14, alanine mutants). Mutation of AcG loop
residue I261 resulted in decreased stimulation
by H4K16ac, whereas stimulation by H2BUb
was not affected (Fig. 2D and fig. S14, alanine
mutant). Immunoblot analysis is consistent
with the essential role of H347 and H355 in
histone H4 binding, in which their mutations
resulted in loss of all methylation states on
unmodified and H4K16ac nucleosomes (Fig.
2E and fig. S14, alanine mutants). The analysis
of methylation patterns in the I261E mutant
confirmed that this mutant is active, whereas
stimulation by H4K16ac is reduced (Fig. 2E and
fig. S14, alanine mutant). These results support
our structural data, showing that interaction
between I261 with acetylated H4K16 leads to
stabilization of the H4-Dot1 interface.
Interactions of Dot1 with H2BUb and the
acidic patch
Yeast Dot1 uses a different general mode of
interaction with ubiquitin, resulting in a dis-
tinct interface from that observed with human
Dot1L (Fig. 3, A and B, and fig. S15). Unlike
human Dot1L, which interacts with a non-
canonical “I36 patch” on ubiquitin, we observed
interactions of the canonical “I44 patch” on
ubiquitin (residues L8, I44, H68, and V70),
as defined in (63) with hydrophobic residues
in yeast Dot1 (Fig. 3C and fig. S15). Specifically,
the conserved F564 of Dot1 interacts with
ubiquitin residues H68, L8, and V70, and Dot1
V559 and V522 interact with ubiquitin residue
I44 (Fig. 3C). Dot1 I517 establishes potential
interactions with ubiquitin residues V70 and
L8 (Fig. 3C). In addition to the canonical I44
patch, Dot1 F519 establishes a cation-p inter-
action with ubiquitin residue R72. We tested
the relevance of the Dot1-Ub interface using
site-directed mutagenesis coupled with enzy-
matic assays. We showed that this surface is
indeed critical for the observed stimulatory
role of H2BUb, and its mutation mainly dis-
rupts effects by Ub (with only a marginal effect
on H4K16ac stimulation) (Fig. 3, D and E). The
Dot1 point mutant V522D is active and shows
a decrease in stimulation on H2BUb nucleo-
somes, with only small effects on stimulation
by H4K16ac (Fig. 3, D and E, and fig. S14,
alanine mutant). The analysis of methylation
by using immunoblot confirmed that V522
mutations reduce the stimulation by H2BUb,
as evidenced by loss of di- and trimethyla-
tion of H3K79, whereas monomethylation
was not affected (Fig. 3E and fig. S14, alanine
mutant).
There is an unconventional, extended, bi-
partite interface between Dot1 and H2A-H2B
dimer in the nucleosome (fig. S15). One part
of this interface is formed by interactions
between Dot1 residues E485 and T514 with
residues in the aC helix of H2B (fig. S15). Other
canonical interactions are established through
salt bridges between positively charged resi-
dues in Dot1 and the nucleosome acidic patch
(Fig. 3F). Interactions with the acidic patch
involve two arginines: a stable one between
R571 and H2A residues E92, D90, and E61 and
a more flexible one between R572 and H2A
E61 and possibly E64 (Fig. 3F and fig. S15). The
relevance of these residues is highlighted by
a loss of activity of the R571E-R572E double
mutant in enzymatic assays on unmodified
and singly modified nucleosomes (Fig. 3G).
This mutant is active only on doubly modified
(H4K16ac/H2BUb) nucleosomes (Fig. 3G),
where it is able to perform mono-, di-, and
trimethylation of H3K79 (Fig. 3H). Both in
yeast and in human Dot1, regions that interact
with ubiquitin and with the acidic patch are in
close spatial proximity (Fig. 3, A and B). In the
main interface with Ub, Dot1 residues V559
Valencia-Sánchez et al., Science 371, eabc6663 (2021)
22 January 2021
4 of 9
Corrected 28 January 2021. See full text.
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Interactions of Dot1 with H2BUb and the
acidic patch. (A) Sequence alignment of budding
yeast Dot1 (purple) and human Dot1L (red) showing
primary and secondary structure of the catalytic
domain and residues interacting with the ubiquitin
(cyan) and the acidic patch (red). (B) Superposition
of yeast Dot1 (purple) and Dot1L (red) [PDB ID
6NJ9 (40)] nucleosome structures showing the
different surface of interaction with ubiquitin (shown
in cyan in the yeast Dot1 structure and in green
in the Dot1L structure). (C) Detailed view of inter-
actions between Dot1 (purple) and ubiquitin (cyan).
(D) Representative endpoint methyltransferase assay
for Dot1 and V522D mutant in the presence of
Unmod nuc or H4K16ac nuc or Ub nuc or ac/Ub nuc
substrates. Reactions were performed with 0.125
and 0.250 pmol of Dot1. Each data point and error bar
indicate the mean ± SD from three independent
experiments. (E) Representative HMT assay measuring
activity of Dot1 and V522D mutant on different
nucleosome substrates. HMT were performed with an
increasing amount of Dot1 mutant (V522D) in the
presence of Unmod nuc or H4K16ac nuc or Ub nuc or
ac/Ub nuc substrates, and reaction products were
identified by using Western blot. (F) Detailed
view of interactions between Dot1 (purple) and
nucleosome acidic patch (H2A, yellow; H2B, red).
(G) Representative endpoint methyltransferase assay
for Dot1 and (R571E-R572E) mutant in the presence of
Unmod nuc or H4K16ac nuc or Ub nuc or ac/Ub
nuc substrates. Reactions were performed with
0.125 pmol and 0.250 pmol of Dot1. Each data
point and error bar indicate the mean ± SD from
three independent experiments. (H) Representative
HMT assay measuring activity of Dot1 and mutant on
different nucleosome substrates. HMT assays were
performed with an increasing amount of Dot1 mutant
(R571E-R572E) in the presence of Unmod nuc,
H4K16ac nuc, Ub nuc, or ac/Ub nuc substrates, and
reaction products were identified by using Western blot.
A
B
Dot1L
9
10
11
K
R I V S S K P F A P L N F R I N S R N L S D I G T I M R V V E L S P L K G S V SWT G K P V S Y Y L H T I D R T I L E N Y F S S L K N P K L R E E
I S L K S L R S L T Y Q I N F Y N V E N I F N R L K V Q R Y D L K E D S V SWT H S G G E Y Y I S T V M E D V D E S L F S P A A R G R R N R G
K I
9
10
11
C
Ub-Dot1L
Ub-Dot1
V70
F564
L8
I44
H68
V522
V559
I517
E
V522D (Ub mutant)
Unmod
H4K16ac
Ub
ac/Ub
Dot1
Ub-Dot1
)
U
L
R
(
e
c
n
e
c
s
e
n
m
u
L
i
2500000
2000000
1500000
1000000
500000
0
H4
H2B
0.125 pmols
0.250 pmols
WT
V522D
WT
V522D
G
D
F
R572
E64
R571
E61
E92
D90
)
U
L
R
(
e
c
n
e
c
s
e
n
m
u
L
i
2500000
2000000
1500000
1000000
500000
Unmod nuc
H4K16ac nuc
Ub nuc
H4K16ac/Ub nuc
H3-
H3-
H3-
Dot1-
Ub H2B-
H3-
H2A/H2B
H4-
Nuc
Dot1
H3K79me3
H3K79me2
H3K79me1
Dot1
Blue-51
0.125 pmols
0.250 pmols
H
R571E/R572E (Ap mutant)
Unmod nuc
H4K16ac nuc
Ub nuc
H4K16ac/Ub nuc
0
WT
R571E/R572E
WT
R571E/R572E
Unmod
H4K16ac
Ub
ac/Ub
Nuc
Dot1
H3K79me3
H3K79me2
H3K79me1
Dot1
Blue-51
H3-
H3-
H3-
Dot1-
Ub H2B-
H3-
H2A/H2B
H4-
and F564 are inside and flank the acidic motif
557-561 EDVDE (fig. S15). On the basis of our
structure, we predict that deletion of this
motif would affect the interaction of Dot1
with Ub and/or with the acidic patch, which
explains previous data that show that dele-
tion of these residues limits catalysis by Dot1
to only monomethylation (52).
Regulation of H3K79 methyltransferase activity
of Dot1 by extended cross-talk in vivo.
Having provided the structural and biochemi-
cal mechanism for Dot1 stimulation by H4K16ac
and H2BUb, we decided to test the effects of our
mutants in vivo (Fig. 4, A and B). We used an
established approach for assessing the biologi-
cal impact of Dot1 mutants in S. cerevisiae (64).
We introduced single-copy plasmids contain-
ing Dot1 mutants to dot1D cells containing a
telomeric URA3 reporter gene. Loss of H3K79
methylation affects URA3 silencing and leads to
growth toxicity on 5–fluoroorotic acid (5-FOA)
(64), although the exact mechanism is not well
understood (65, 66). As a control, we included
Dot1 mutants known to reduce H3K79 meth-
ylation (G401A and G401R) (Fig. 4A). The mu-
tants affected URA3 silencing by two to four
orders of magnitude (Fig. 4A). To confirm these
genetic observations by examining in vivo
methylation status, we assessed with immu-
noblotting global H3K79 mono-, di-, and tri-
methylation (Fig. 4B). Consistent with the
growth assays, the mutants disrupted di- and
trimethylation of H3K79 (Fig. 4B). Dot1 mu-
tants H355E and V522D, which retain residual
methylation activity, lead to URA3 silencing
similar to that by the catalytically dead G401R
mutant. This echoes previous work that showed
that catalytically reduced Dot1 G401A also
affects in vivo URA3 silencing similarly to that
by G401R (64). These results suggest that ace-
tylation of lysine 16 and ubiquitination of
histone H2B are both required for efficient
methylation of H3K79 in yeast.
On the basis of our structural, biochemical,
and in vivo functional data, we propose a
model that describes the rules of Dot1 stimula-
tion by histone acetylation and ubiquitination
(Fig. 4C). The outcome of enzymatic reac-
tion by Dot1 is directly linked to its ability to
interact with the nucleosome using several
“anchors” that stabilize its conformation for
optimal activity. The most relevant intrinsic
anchor is established through the interaction
of Dot1 with the acidic patch. This anchor,
together with the Dot1–histone H4 interac-
tion, suffices to allow Dot1-mediated mono-
methylation of H3K79. H4K16ac and H2BUb
individually provide critical regulatory anchors
that allow Dot1 to perform di- and trimethy-
lation (Fig. 4C). Nucleosomes that are both
H4K16 acetylated and H2B ubiquitinated
allow optimal positioning of Dot1 on the nu-
cleosome, restricting its orientation and allow-
ing mostly productive conformations, resulting
in increased trimethylation (Fig. 4C).
Valencia-Sánchez et al., Science 371, eabc6663 (2021)
22 January 2021
5 of 9
Corrected 28 January 2021. See full text.
RES EARCH | R E S E A R C H A R T I C L E
dot1
+plasmid
empty vector
DOT1
dot1(G401A)
dot1(G401R)
dot1(I261E)
dot1(H355E)
dot1(V522D)
A
C
SC–Leu
(2 days)
SC–Leu+5-FOA
(3 days)
B
y
t
i
s
n
e
t
n
I
e
v
i
t
a
e
R
l
1.5
1.0
0.5
0
Dot1
Me
Ac
Dot1
Me
Dot1
Me
Ub
dot1(G401R)
dot1(I261E)
dot1(H355E)
dot1(V522D)
empty
DOT1
-H3K79me3
-H3K79me2
-H3K79me1
-H3K4me3
(control)
-H3
(control)
Dot1
Me
Ac
Ub
Nuc
Unmod
H3K79Me1
H3K79Me2
H3K79Me3
Intrinsic anchors
Regulatory anchors
Combined anchors
+
-
-
2
0
2
Ub
+++
++
+
2
1
3
H4K16ac
ac / Ub
+++
++
++
2
1
3
++
+++
+++
2
2
4
Fig. 4. Regulation of H3K79 methyltransferase activity of Dot1 through extended cross-talk in vivo.
(A) Tenfold serial dilutions of UCC1783 yeast containing Dot1 mutants on media with or without 5-FOA
(n = 3 experiments). Proper silencing of a URA3 gene on telomere VIIIL confers resistance to 5-FOA.
(B) (Top) Representative Western blot showing H3K79 methylation by Dot1 mutants. (Bottom)
Quantitative image analysis of Western blot in (B) based on n = 2 blots. Errors bars show standard deviation.
(C) General model of how Dot1 activity is regulated by histone posttranslational modifications. Intrinsic
anchors represent interactions between Dot1 and histone H4 and acidic patch. These anchors allow Dot1
to monomethylate H3K79. Regulatory anchors are provided when histone H4K16 is acetylated or when
H2B is ubiquitinated. The expression anchor is used here to denominate an interaction that stabilizes a
productive conformation of Dot1 on the nucleosome. The illustration is color-coded as in Fig. 1.
Discussion
Our findings suggest a possible mechanism for
how H3K79 methylation by Dot1 regulates the
kinetics of gene silencing in yeast. In this
process, SIR complex binds to nucleosomes to
form a repressive chromatin structure. Spe-
cifically, the NAD+–dependent deacetylase
Sir2 removes the acetyl group from H4K16ac,
allowing binding of Sir3 and thus promotion
of chromatin silencing (67, 68). There are data
that show that methylation of lysine 79 by
Dot1 slows down kinetics of establishment of
silent chromatin in yeast (69). Competition for
access to H4K16ac between deacetylase Sir2
and Dot1 could be central to this mechanism,
especially at regions distal to SIR recruitment
sites. Acetylated H4K16 stimulates H3K79
methylation, and once this happens, nucleo-
somes that harbor both H4K16ac and H3K79me3
could prevent SIR complex from binding to
chromatin. Because H2BUb has also been
shown to reduce silencing in yeast (24, 70),
nucleosomes modified at multiple sites would
be predicted to have an even stronger effect.
Additionally, Dot1 binds to the acidic patch,
a nucleosome region with key functions in
chromatin structure (71). It has been shown
that the H4 tail interacts with the acidic patch
of another nucleosome through charge com-
plementarity (72). Acetylation of H4K16 neu-
tralizes the charge and prevents the formation
of chromatin fiber (17, 73). Therefore, acetyla-
tion of H4K16 would (i) obstruct contacts of
this tail with the acidic patch, (ii) block con-
tacts of this tail with Sir3, and (iii) allow
Dot1 to bind and (iv) stimulate its catalytic
activity, and (v) Dot1 binding would further
occlude the acidic patch from the binding of
silencing proteins such as Sir3.
What could be the mechanism of Dot1 sti-
mulation by H4K16ac? Our data point to
H4K16ac playing a role in restricting the
conformation of Dot1. This is based on the
comparison of Dot1-unacetylated H4 and Dot1-
H4K16ac structures where in the latter, stabi-
lization of Dot1 and histone H4 is observed.
Conformational restriction has been proposed
as a mechanism for stimulation of human
Dot1L by H2Bub (33, 40). On the basis of our
data, H2BUb acts in the similar way, but per-
haps less efficiently in yeast. The presence of a
3D class in which Dot1 is in the noncatalytic
position exclusively in the Dot1-unacetylated
H4 dataset supports the role of H4K16 ace-
tylation in further restricting Dot1 in the pro-
ductive configuration, which results in the
increased catalytic rate. Given the substantial
structural differences between Dot1-H4K16ac
and Dot1-unacetylated H4 datasets, and the
functional experiments that support our struc-
tures, it is unlikely that the presence of H3K79M
masked the mechanisms of stimulation by
H4K16ac. There is no major change of Kd or
Km for Dot1 on H4K16ac nucleosomes. This is
also in line with the stimulation of human
Dot1L by the H2BK120Ub nucleosome, for
which it was postulated that the energy from
Ub binding is used on conformational changes
to bring the active site into catalytically com-
patible conformation (33, 40). In the structure
of Dot1L bound to the H2BK120Ub nucleo-
some in the catalytically active conformation,
it was speculated in (40) that the energetic
cost of the H3 distortion is paid off by energy
gained by binding to ubiquitin. This is sup-
ported in our structure, in which the distor-
tion around H3K79 was also observed, and
there are additionally changes in the AcG loop
of Dot1 and histone H4.
How can the PTMs change the outcome of
methylation by Dot1? A recent review raised
the possibility that perhaps histone modifica-
tions render human Dot1L processive through
multivalent interactions with chromatin, such
as was shown for other chromatin enzymes
(74). In this Dot1 binding scenario, stimula-
tory histone modifications would facilitate a
search for Dot1’s lysine substrate, and Dot1
would methylate H3K79 in a processive man-
ner, after which it would dissociate. If Dot1 is a
distributive enzyme, as has been shown before
(75), at each round of methylation there would
be costs associated with the conformational
sampling of the different positions on the nu-
cleosome. In such a case, at each round of
methylation Dot1 would bind nucleosome,
and H4K16ac and H2BUb would help Dot1
reach the active conformation. Because there
is no substantial increase in affinity in the pres-
ence of H2BUb and H4K16ac, the reduction
of the conformational sampling space would
still be predicted to be the most relevant con-
tribution of these histone modifications. These
PTMs would together optimally position Dot1
to allow a more robust deposition of higher-
order methyl marks.
As stated, our current structure is different
in terms of ubiquitin recognition from that
by Dot1L but similar to that by MLL1/3 sub-
complexes. Dot1L uses the I36 patch on ubiq-
uitin, and Dot1 and MLL1/3 use the I44 patch
(fig. S15) (40, 76). The recognition of ubiquitin
by the COMPASS subcomplex shares both
patches of recognition I36 and I44 (fig. S15)
(77). These findings highlight the plasticity
of ubiquitin recognition, as summarized be-
fore (74, 77). Why ubiquitin is differently
recognized by these proteins and complexes
Valencia-Sánchez et al., Science 371, eabc6663 (2021)
22 January 2021
6 of 9
Corrected 28 January 2021. See full text.
RES EARCH | R E S E A R C H A R T I C L E
remains an unresolved question that will re-
quire further work.
There are several examples of trans-histone
cross-talk described in chromatin biology
(78, 79). The concept of positive- and negative-
feedback loops is also quite well established.
We hypothesize that during transcription, his-
tone modifications—as well as enzymes that
catalyze their deposition—are tightly regulated
through trans-histone cross-talk to provide
positive feed-forward loops that result in an
optimal chromatin environment. There are
many examples in which this happens. H2BUb
regulates both MLL complex deposition of
H3K4 methylation and Dot1 deposition of
H3K79 methylation. In turn, Dot1 helps to
catalyze histone H2B ubiquitination (80). In
this work, we show that acetylation of lysine
16, perhaps the most critical modification on
histone H4, provides the means to stimulate
K79 methylation. It would be interesting to
see whether H3K79 methylation in turn stim-
ulates histone acetylation, providing a self-
sustaining chromatin state maintenance circuit.
Methods summary
Yeast Dot1 (158 to 582) and its mutants were
expressed in Escherichia coli ONESHOT
BL21(DE3) and purified by using nickel affin-
ity chromatography, ion exchange, and size-
exclusion liquid chromatography. Ubiquitin
(Ub) G76C was expressed as soluble protein
in E. coli SoluBL21 and purified by using
nickel affinity chromatography (81). Wild-
type and H2BK120C Xenopus histones were
expressed in E. coli BL21(DE3) pLysS cells,
extracted from inclusion bodies (82), purified
by size exclusion chromatography, and lyophi-
lized. Ubiquitination of histone H2BK120C
was performed by using a previously published
protocol (81). Ub G76C and histone H2BK120C
were mixed at the ratio of 2:1 and cross-linked
by using 1,3-dichloroacetone. The ubiquitinated
histone H2BK120C (H2BK120Ub) was puri-
fied by using nickel affinity chromatography
and lyophilized. The genetically encoded histone
H4 acetyl K16 (H4K16ac) was produced accord-
ing to Wilkins et al. (83). Plasmid pASB567_9 g7
(amber codon expression vector) for expres-
sion of the histone H3(D93-98)-H4 K16ac fusion
and pAcKRS-3 plasmid containing acetyl-lysyl-
tRNA synthetase/tRNACUA pair were trans-
formed into E. coli C321.D A.exp cells (Addgene),
and protein expression was induced by the
addition of 0.2% arabinose in presence of Ne-
Acetyl-L-lysine. The histone 6xHis-H3(D93-98)-
H4K16ac fusion was purified from inclusion
bodies by nickel affinity purification and H4
peptide was excised with TEV protease and
lyophilized.
Nucleosome substrates were assembled as
described (54, 82). Octamers were reconsti-
tuted by mixing equimolar amounts of each
lyophilized histone and dialysis into refold-
ing buffer. Nucleosomes were assembled by
combining equimolar ratios of purified Widom
601 DNA and histone octamers and dialyzing
the mix overnight with gradient salt dialysis.
Designer, individually acetylated (H4K5ac,
H4K8ac, H4K12ac, or H4K16ac), or tetra-
acetylated nucleosomes were purchased from
EpiCypher.
For the nucleosome binding assay, a twofold
serial dilution of Dot1 (50 to 1000 nM) was
incubated with 12 nM recombinant nucleo-
somes at room temperature for 30 min and
analyzed by using native polyacrylamide gels
[6% polyacrylamide gel electrophoresis (PAGE),
0.2× tris-boric acid–EDTA buffer], stained
with SYBR Gold (Thermo Fisher), imaged in
a Typhoon Trio+ scanner, and quantified with
the program ImageQuant 5.2v (Molecular
Dynamics).
The Dot1 global methyltransferase activity
and Michaelis-Menten kinetic analysis were
determined by monitoring the production
of S-(5′-adenosyl)-L-homocysteine (SAH) with
MTase-Glo methyltransferase kit (Promega)
on nucleosome substrates. The Dot1 global
methyltransferase activity (Endpoint meth-
ylation assay) reactions were performed by
mixing 50 or 100 nM of wild-type or mutant
Dot1 proteins with 1 mM unmodified or mod-
ified nucleosomes [in the presence of 20 mM
S-(5′-adenosyl)-L-methionine (SAM)] at 30°C
for 30 min and stopped with 5 ml of 0.5% TFA
(trifluoroacetic acid). The Michaelis-Menten
kinetic analysis of Dot1 methyltransferase ac-
tivity on nucleosome was performed by mix-
ing 10 nM Dot1 containing 20 mM SAM with a
twofold dilution series of nucleosomes (4 to
4000 nM). The reaction was incubated for 5 min
at 30°C and stopped on ice by the addition of
4 ml of 0.5% TFA. The kinetic parameters Km
and kcat of Dot1 methyltransferase activity
were determined by fitting the initial veloc-
ities into the Michaelis-Menten equation using
the enzymatic kinetics analysis (kcat analysis)
in Prism 8 (GraphPad).
For the histone methyltransferase assay
(HMT) by use of Western blot, reactions of
Dot1 (0.250 pmol at the highest concentration)
and nucleosomes (10 pmol) in presence of
20 mM SAM were incubated at 30°C for 1 hour
and stopped by the addition of 5 ml of 5× SDS
buffer. The products were resolved in a 15%
SDS-PAGE gel and transferred to polyvinylidene
difluoride membrane. The histone methyla-
tion level was determined by incubating the
membranes with anti-H3K79Me3 (Abcam
Ab2621), anti-H3K79Me2 (Abcam Ab3594),
or anti-H3K79Me1 (sera 58) antibody. The
Western blots were developed by using the
ECL reagent (ThermoFisher).
Gradient Fixation (GraFix) of Dot1 with the
nucleosome for structural studies was per-
formed according to (60). Nucleosomes were
mixed with purified Dot1 protein and sup-
plemented with 5× molar ratio of SAM. The
sample was applied to a 10 to 30% glycerol
and 0 to 0.1% glutaraldehyde gradient and
centrifuged for 16 hours at 30,000 revolutions
per minute (rpm) (Beckman Coulter Optima
XL-100K). Selected fractions were quenched,
dialyzed into dialysis buffer (20 mM Tris 7.0,
50 mM KCl, 1 mM MgCl2, 1 mM dithiothreitol),
and concentrated.
Cryo-EM grids of the Dot1-H4K16ac com-
plex and Dot1-unacetylated H4 complex were
prepared following an established protocol
(84): 3.0 ml of the samples at 0.45 mg/ml were
applied to glow-discharged Quantifoil gold
grids (400 mesh, 1.2-mm hole size). The grids
were then blotted for 3 s at 4°C and 100%
humidity and plunge-frozen by using Vitrobot
Mark IV (FEI Company). All sample images
were recorded on FEI Titan Krios operated
at 300 kV by using a Gatan K2 Summit direct
electron detector camera in counting mode at
a nominal magnification of 130,000× (calibrated
pixel size of 1.035 Å per pixel). Total accumulated
electron exposure was 74.5 electrons per Å2 for
the Dot1-H4K16ac complex and 52 electrons per
Å2 for the Dot1-unacetylated H4 complex.
Images were motion corrected by using UCSF
MotionCor2 v1.2.1 (85); GCTF (86) was used to
calculate CTF, Gautomatch (www.mrc-lmb.cam.
ac.uk/kzhang) for particle picking, Relion3 (87)
for particle extraction, cryoSPARC (88) for Ab
Initio reconstruction and 3D refinement, and
cisTEM (89) for the final reconstruction. This
procedure led to obtaining reconstructions of
the Dot1-H4K16ac complex at 3.1 Å and Dot1-
unacetylated H4 complex at 3.2 Å (Fourier
shell correlation = 0.143). To validate that the
cryo-EM reconstructions and our interpreta-
tion were not distorted by an overrepresen-
tation of certain views in the particle datasets,
we conducted additional distribution analysis
and normalization that is described in the sup-
plementary materials.
For the model building of the Dot1-H4K16ac
complex, we used the following available x-ray
crystal structures for rigid body fit into the
3.1 Å reconstruction: for the nucleosome, PDB
IDs 3TU4 (54) and 5AV5 (90); for catalytic
domain of Dot1, 1U2Z (91); and for ubiquitin,
1UBQ (92). We used 1NW3 (93) to inform the
placement of SAM. We then used Coot (94) for
local adjustments of secondary elements and
side chains, and the complete model was refined
by using real-space refinement implemented in
PHENIX (95). For the Dot1-unacetylated H4
complex, we used the cryo-EM reconstruction
at 3.2 Å and the PDB of the Dot1-H4K16ac
structure as a starting point, removing the
missing atoms and applying the same protocol
of refinement.
For in vivo experiments, we used yeast spot
assays performed with Dot1 deletion yeast
strain (UCC7183), as described in (64). Yeast
strains were cultured in synthetic complete
Valencia-Sánchez et al., Science 371, eabc6663 (2021)
22 January 2021
7 of 9
Corrected 28 January 2021. See full text. RES EARCH | R E S E A R C H A R T I C L E
medium (SC). UCC7183 was transformed
with Dot1 wild-type and mutant plasmids
containing pRS315 and selected for 2 days
at 30°C on SC–Leu plates. Spot assays were
performed by preparing 10-fold dilutions of
overnight cultures (A600 of ~10) and then
spotted on both SC–Leu and SC–Leu+5-FOA
plates.
Histone H3K79 modification in yeast was
tested with protein Western blotting. Histone
protein fractions were prepared by growing
Dot1 yeast strains in SC–Leu to an A600 of ~1 to
1.2. Cells were disrupted in presence of 0.25 M
HCl, histones were solubilized in 0.25 M HCL/
97.5% ethanol and precipitated with acidi-
fied acetone. Solubilized protein pellet in 1X
NuPAGE LDS loading buffer (Invitrogen)
was electrophoresed on 12% Bis-Tris NuPAGE
gels. Immunoblots were performed as described
(96) by using the Licor Odyssey system. Com-
mercial primary antibodies used were mouse
anti H3 (Abcam) and mouse anti H3K4me3
(Abcam). H3K79 and yeast Dot1 primary anti-
bodies were a gift from F. Van Leeuwen (64):
rabbit anti H3K79me1 (sera #58), rabbit anti
H3K79me2 (sera #30), rabbit anti H3K79me3
(sera #34), and rabbit anti Dot1 (affinity pu-
rified). Secondary antibodies were IRDye 800
goat anti mouse and IRDye 680 goat anti rab-
bit (LiCor).
RE FE RENCES AND N OT ES
1. A. J. Bannister, T. Kouzarides, Regulation of chromatin by
histone modifications. Cell Res. 21, 381–395 (2011).
doi: 10.1038/cr.2011.22; pmid: 21321607
2. S. L. Berger, The complex language of chromatin regulation
during transcription. Nature 447, 407–412 (2007).
doi: 10.1038/nature05915; pmid: 17522673
3. D. Schübeler et al., The histone modification pattern of active
genes revealed through genome-wide chromatin analysis of a
higher eukaryote. Genes Dev. 18, 1263–1271 (2004).
doi: 10.1101/gad.1198204; pmid: 15175259
4. Z. Wang et al., Combinatorial patterns of histone
acetylations and methylations in the human genome.
Nat. Genet. 40, 897–903 (2008). doi: 10.1038/ng.154;
pmid: 18552846
J. E. Audia, R. M. Campbell, Histone Modifications and Cancer.
Cold Spring Harb. Perspect. Biol. 8, a019521 (2016).
doi: 10.1101/cshperspect.a019521; pmid: 27037415
5.
6. B. M. Turner, L. P. O’Neill, Histone acetylation in chromatin and
chromosomes. Semin. Cell Biol. 6, 229–236 (1995).
doi: 10.1006/scel.1995.0031; pmid: 8562915
7. S. M. Görisch, M. Wachsmuth, K. F. Tóth, P. Lichter, K. Rippe,
Histone acetylation increases chromatin accessibility.
J. Cell Sci. 118, 5825–5834 (2005). doi: 10.1242/jcs.02689;
pmid: 16317046
8. M. H. Kuo, C. D. Allis, Roles of histone acetyltransferases and
deacetylases in gene regulation. BioEssays 20, 615–626
(1998). doi: 10.1002/(SICI)1521-1878(199808)20:8<615::AID-
BIES4>3.0.CO;2-H; pmid: 9780836
9. D. J. Clarke, L. P. O’Neill, B. M. Turner, Selective use of H4
acetylation sites in the yeast Saccharomyces cerevisiae.
Biochem. J. 294, 557–561 (1993). doi: 10.1042/bj2940557;
pmid: 8373369
10. S. K. Kurdistani, S. Tavazoie, M. Grunstein, Mapping global
histone acetylation patterns to gene expression. Cell 117,
721–733 (2004). doi: 10.1016/j.cell.2004.05.023;
pmid: 15186774
11. W. Zhang, J. R. Bone, D. G. Edmondson, B. M. Turner,
S. Y. Roth, Essential and redundant functions of histone
acetylation revealed by mutation of target lysines and loss of
the Gcn5p acetyltransferase. EMBO J. 17, 3155–3167 (1998).
doi: 10.1093/emboj/17.11.3155; pmid: 9606197
12. D. K. Pokholok et al., Genome-wide map of nucleosome
34. H. H. Ng, R. M. Xu, Y. Zhang, K. Struhl, Ubiquitination of
acetylation and methylation in yeast. Cell 122, 517–527 (2005).
doi: 10.1016/j.cell.2005.06.026; pmid: 16122420
13. M. D. Shahbazian, M. Grunstein, Functions of site-specific
histone acetylation and deacetylation. Annu. Rev. Biochem. 76,
75–100 (2007). doi: 10.1146/annurev.
biochem.76.052705.162114; pmid: 17362198
14. B. Li, M. Carey, J. L. Workman, The role of chromatin during
transcription. Cell 128, 707–719 (2007). doi: 10.1016/
j.cell.2007.01.015; pmid: 17320508
15. H. Reinke, W. Hörz, Histones are first hyperacetylated and then
lose contact with the activated PHO5 promoter. Mol. Cell 11,
1599–1607 (2003). doi: 10.1016/S1097-2765(03)00186-2;
pmid: 12820972
16. M. Chandy, J. L. Gutiérrez, P. Prochasson, J. L. Workman,
SWI/SNF displaces SAGA-acetylated nucleosomes. Eukaryot.
Cell 5, 1738–1747 (2006). doi: 10.1128/EC.00165-06;
pmid: 17030999
17. M. Shogren-Knaak et al., Histone H4-K16 acetylation controls
chromatin structure and protein interactions. Science 311,
844–847 (2006). doi: 10.1126/science.1124000;
pmid: 16469925
18. Q. Feng et al., Methylation of H3-lysine 79 is mediated by a
new family of HMTases without a SET domain. Curr. Biol. 12,
1052–1058 (2002). doi: 10.1016/S0960-9822(02)00901-6;
pmid: 12123582
19. N. Lacoste, R. T. Utley, J. M. Hunter, G. G. Poirier, J. Côte,
Disruptor of telomeric silencing-1 is a chromatin-specific
histone H3 methyltransferase. J. Biol. Chem. 277,
30421–30424 (2002). doi: 10.1074/jbc.C200366200;
pmid: 12097318
20. F. van Leeuwen, P. R. Gafken, D. E. Gottschling, Dot1p
modulates silencing in yeast by methylation of the nucleosome
core. Cell 109, 745–756 (2002). doi: 10.1016/S0092-8674(02)
00759-6; pmid: 12086673
21. A. Kouskouti, I. Talianidis, Histone modifications defining active
genes persist after transcriptional and mitotic inactivation.
EMBO J. 24, 347–357 (2005). doi: 10.1038/sj.emboj.7600516;
pmid: 15616580
22. N. J. Krogan et al., The Paf1 complex is required for histone H3
methylation by COMPASS and Dot1p: Linking transcriptional
elongation to histone methylation. Mol. Cell 11, 721–729
(2003). doi: 10.1016/S1097-2765(03)00091-1; pmid: 12667454
histone H2B by Rad6 is required for efficient Dot1-mediated
methylation of histone H3 lysine 79. J. Biol. Chem. 277,
34655–34657 (2002). doi: 10.1074/jbc.C200433200;
pmid: 12167634
35. S. Nakanishi et al., Histone H2BK123 monoubiquitination is the
critical determinant for H3K4 and H3K79 trimethylation by
COMPASS and Dot1. J. Cell Biol. 186, 371–377 (2009).
doi: 10.1083/jcb.200906005; pmid: 19667127
36. B. Zhu et al., Monoubiquitination of human histone H2B: The
factors involved and their roles in HOX gene regulation.
Mol. Cell 20, 601–611 (2005). doi: 10.1016/j.molcel.2005.09.025;
pmid: 16307923
37. M. I. Valencia-Sánchez et al., Structural Basis of Dot1L
Stimulation by Histone H2B Lysine 120 Ubiquitination. Mol. Cell
74, 1010–1019.e6 (2019). doi: 10.1016/j.molcel.2019.03.029;
pmid: 30981630
38. C. J. Anderson et al., Structural Basis for Recognition of
Ubiquitylated Nucleosome by Dot1L Methyltransferase.
Cell Rep. 26, 1681–1690.e5 (2019). doi: 10.1016/
j.celrep.2019.01.058; pmid: 30759380
39. S. Jang et al., Structural basis of recognition and
destabilization of the histone H2B ubiquitinated nucleosome by
the DOT1L histone H3 Lys79 methyltransferase. Genes Dev. 33,
620–625 (2019). doi: 10.1101/gad.323790.118;
pmid: 30923167
40. E. J. Worden, N. A. Hoffmann, C. W. Hicks, C. Wolberger,
Mechanism of Cross-talk between H2B Ubiquitination and H3
Methylation by Dot1L. Cell 176, 1490–1501.e12 (2019).
doi: 10.1016/j.cell.2019.02.002; pmid: 30765112
41. T. Yao et al., Structural basis of the crosstalk between histone
H2B monoubiquitination and H3 lysine 79 methylation on
nucleosome. Cell Res. 29, 330–333 (2019). doi: 10.1038/
s41422-019-0146-7; pmid: 30770869
42. H. Vlaming et al., Conserved crosstalk between histone
deacetylation and H3K79 methylation generates DOT1L-dose
dependency in HDAC1-deficient thymic lymphoma. EMBO J.
38, e101564 (2019). doi: 10.15252/embj.2019101564;
pmid: 31304633
43. S. Lee et al., Dot1 regulates nucleosome dynamics by its
inherent histone chaperone activity in yeast. Nat. Commun. 9,
240 (2018). doi: 10.1038/s41467-017-02759-8;
pmid: 29339748
23. D. J. Steger et al., DOT1L/KMT4 recruitment and H3K79
44. M. S. Singer et al., Identification of high-copy disruptors of
methylation are ubiquitously coupled with gene transcription in
mammalian cells. Mol. Cell. Biol. 28, 2825–2839 (2008).
doi: 10.1128/MCB.02076-07; pmid: 18285465
24. H. H. Ng, D. N. Ciccone, K. B. Morshead, M. A. Oettinger,
K. Struhl, Lysine-79 of histone H3 is hypomethylated
at silenced loci in yeast and mammalian cells: A potential
mechanism for position-effect variegation. Proc. Natl. Acad.
Sci. U.S.A. 100, 1820–1825 (2003). doi: 10.1073/
pnas.0437846100; pmid: 12574507
25. K. Wood, M. Tellier, S. Murphy, DOT1L and H3K79 Methylation
in Transcription and Genomic Stability. Biomolecules 8, 11
(2018). doi: 10.3390/biom8010011; pmid: 29495487
26. B. Jones et al., The histone H3K79 methyltransferase Dot1L is
essential for mammalian development and heterochromatin
structure. PLOS Genet. 4, e1000190 (2008). doi: 10.1371/
journal.pgen.1000190; pmid: 18787701
27. C. M. McLean, I. D. Karemaker, F. van Leeuwen, The emerging
roles of DOT1L in leukemia and normal development.
Leukemia 28, 2131–2138 (2014). doi: 10.1038/leu.2014.169;
pmid: 24854991
28. K. M. Bernt et al., MLL-rearranged leukemia is dependent on
aberrant H3K79 methylation by DOT1L. Cancer Cell 20, 66–78
(2011). doi: 10.1016/j.ccr.2011.06.010; pmid: 21741597
29. C. W. Chen, S. A. Armstrong, Targeting DOT1L and HOX gene
expression in MLL-rearranged leukemia and beyond. Exp. Hematol.
43, 673–684 (2015). doi: 10.1016/j.exphem.2015.05.012;
pmid: 26118503
30. A. V. Krivtsov et al., H3K79 methylation profiles define murine
and human MLL-AF4 leukemias. Cancer Cell 14, 355–368
(2008). doi: 10.1016/j.ccr.2008.10.001; pmid: 18977325
31. Y. Okada et al., hDOT1L links histone methylation to
leukemogenesis. Cell 121, 167–178 (2005). doi: 10.1016/
j.cell.2005.02.020; pmid: 15851025
32. S. D. Briggs et al., Gene silencing: Trans-histone regulatory
pathway in chromatin. Nature 418, 498 (2002).
doi: 10.1038/nature00970; pmid: 12152067
33. R. K. McGinty, J. Kim, C. Chatterjee, R. G. Roeder, T. W. Muir,
Chemically ubiquitylated histone H2B stimulates hDot1L-
mediated intranucleosomal methylation. Nature 453, 812–816
(2008). doi: 10.1038/nature06906; pmid: 18449190
telomeric silencing in Saccharomyces cerevisiae. Genetics 150,
613–632 (1998). pmid: 9755194
45. M. Grunstein, S. M. Gasser, Epigenetics in Saccharomyces
cerevisiae. Cold Spring Harb. Perspect. Biol. 5, a017491 (2013).
doi: 10.1101/cshperspect.a017491; pmid: 23818500
46. A. Hecht, T. Laroche, S. Strahl-Bolsinger, S. M. Gasser,
M. Grunstein, Histone H3 and H4 N-termini interact with SIR3
and SIR4 proteins: A molecular model for the formation of
heterochromatin in yeast. Cell 80, 583–592 (1995).
doi: 10.1016/0092-8674(95)90512-X; pmid: 7867066
47. A. Kimura, T. Umehara, M. Horikoshi, Chromosomal gradient of
histone acetylation established by Sas2p and Sir2p functions
as a shield against gene silencing. Nat. Genet. 32, 370–377
(2002). doi: 10.1038/ng993; pmid: 12410229
48. M. Oppikofer et al., A dual role of H4K16 acetylation in the
establishment of yeast silent chromatin. EMBO J. 30,
2610–2621 (2011). doi: 10.1038/emboj.2011.170;
pmid: 21666601
49. N. Suka, K. Luo, M. Grunstein, Sir2p and Sas2p opposingly
regulate acetylation of yeast histone H4 lysine16 and spreading
of heterochromatin. Nat. Genet. 32, 378–383 (2002).
doi: 10.1038/ng1017; pmid: 12379856
50. R. Behrouzi et al., Heterochromatin assembly by interrupted
Sir3 bridges across neighboring nucleosomes. eLife 5, e17556
(2016). doi: 10.7554/eLife.17556; pmid: 27835568
51. D. Goodnight, J. Rine, S-phase-independent silencing
establishment in Saccharomyces cerevisiae. eLife 9, e58910
(2020). doi: 10.7554/eLife.58910; pmid: 32687055
52. I. M. Fingerman, H. C. Li, S. D. Briggs, A charge-based
interaction between histone H4 and Dot1 is required for H3K79
methylation and telomere silencing: Identification of a new
trans-histone pathway. Genes Dev. 21, 2018–2029 (2007).
doi: 10.1101/gad.1560607; pmid: 17675446
53. M. Altaf et al., Interplay of chromatin modifiers on a short basic
patch of histone H4 tail defines the boundary of telomeric
heterochromatin. Mol. Cell 28, 1002–1014 (2007).
doi: 10.1016/j.molcel.2007.12.002; pmid: 18158898
54. K. J. Armache, J. D. Garlick, D. Canzio, G. J. Narlikar,
R. E. Kingston, Structural basis of silencing: Sir3 BAH domain
in complex with a nucleosome at 3.0 Å resolution. Science
Valencia-Sánchez et al., Science 371, eabc6663 (2021)
22 January 2021
8 of 9
Corrected 28 January 2021. See full text. RES EARCH | R E S E A R C H A R T I C L E
334, 977–982 (2011). doi: 10.1126/science.1210915;
pmid: 22096199
particle at 2.8 A resolution. Nature 389, 251–260 (1997).
doi: 10.1038/38444; pmid: 9305837
J. Biol. Chem. 279, 43296–43306 (2004). doi: 10.1074/
jbc.M405902200; pmid: 15292170
55. A. Wood, J. Schneider, J. Dover, M. Johnston, A. Shilatifard,
73. B. Dorigo, T. Schalch, K. Bystricky, T. J. Richmond,
The Paf1 complex is essential for histone monoubiquitination
by the Rad6-Bre1 complex, which signals for histone
methylation by COMPASS and Dot1p. J. Biol. Chem. 278,
34739–34742 (2003). doi: 10.1074/jbc.C300269200;
pmid: 12876294
56. Z. W. Sun, C. D. Allis, Ubiquitination of histone H2B regulates
H3 methylation and gene silencing in yeast. Nature 418,
104–108 (2002). doi: 10.1038/nature00883; pmid: 12077605
57. A. Wood et al., Bre1, an E3 ubiquitin ligase required for recruitment
and substrate selection of Rad6 at a promoter. Mol. Cell 11,
267–274 (2003). doi: 10.1016/S1097-2765(02)00802-X;
pmid: 12535539
58. H. Jayaram et al., S-adenosyl methionine is necessary for
inhibition of the methyltransferase G9a by the lysine 9
to methionine mutation on histone H3. Proc. Natl. Acad. Sci.
U.S.A. 113, 6182–6187 (2016). doi: 10.1073/pnas.1605523113;
pmid: 27185940
59. P. W. Lewis et al., Inhibition of PRC2 activity by a gain-of-
function H3 mutation found in pediatric glioblastoma. Science
340, 857–861 (2013). doi: 10.1126/science.1232245;
pmid: 23539183
60. H. Stark, GraFix: Stabilization of fragile macromolecular
Chromatin fiber folding: Requirement for the histone H4
N-terminal tail. J. Mol. Biol. 327, 85–96 (2003). doi: 10.1016/
S0022-2836(03)00025-1; pmid:12614610
74. E. J. Worden, C. Wolberger, Activation and regulation of
H2B-Ubiquitin-dependent histone methyltransferases. Curr. Opin.
Struct. Biol. 59, 98–106 (2019). doi: 10.1016/j.sbi.2019.05.009;
pmid: 31229920
75. I. J. Stulemeijer et al., Dot1 histone methyltransferases share a
distributive mechanism but have highly diverged catalytic
properties. Sci. Rep. 5, 9824 (2015). doi: 10.1038/srep09824;
pmid: 25965993
76. H. Xue et al., Structural basis of nucleosome recognition and
modification by MLL methyltransferases. Nature 573,
445–449 (2019). doi: 10.1038/s41586-019-1528-1;
pmid: 31485071
77. E. J. Worden, X. Zhang, C. Wolberger, Structural basis for
COMPASS recognition of an H2B-ubiquitinated nucleosome.
eLife 9, e53199 (2020). doi: 10.7554/eLife.53199;
pmid: 31922488
78. J. S. Lee, E. Smith, A. Shilatifard, The language of histone
crosstalk. Cell 142, 682–685 (2010). doi: 10.1016/
j.cell.2010.08.011; pmid: 20813257
complexes for single particle cryo-EM. Methods Enzymol. 481,
109–126 (2010). doi: 10.1016/S0076-6879(10)81005-5;
pmid: 20887855
79. T. Suganuma, J. L. Workman, Crosstalk among Histone
Modifications. Cell 135, 604–607 (2008). doi: 10.1016/
j.cell.2008.10.036; pmid: 19013272
61. J. E. Beaver, B. C. Peacor, J. V. Bain, L. I. James, M. L. Waters,
Contributions of pocket depth and electrostatic interactions to
affinity and selectivity of receptors for methylated lysine in
water. Org. Biomol. Chem. 13, 3220–3226 (2015).
doi: 10.1039/C4OB02231A; pmid: 25437861
62. J. J. Kamps et al., Chemical basis for the recognition of
trimethyllysine by epigenetic reader proteins. Nat. Commun. 6,
8911 (2015). doi: 10.1038/ncomms9911; pmid: 26578293
63. D. Komander, M. Rape, The ubiquitin code. Annu. Rev.
Biochem. 81, 203–229 (2012). doi: 10.1146/annurev-biochem-
060310-170328; pmid: 22524316
64. F. Frederiks et al., Nonprocessive methylation by Dot1 leads to
functional redundancy of histone H3K79 methylation states.
Nat. Struct. Mol. Biol. 15, 550–557 (2008). doi: 10.1038/
nsmb.1432; pmid: 18511943
65. Y. H. Takahashi et al., Dot1 and histone H3K79 methylation in
natural telomeric and HM silencing. Mol. Cell 42, 118–126
(2011). doi: 10.1016/j.molcel.2011.03.006; pmid: 21474073
66. M. P. Rossmann, W. Luo, O. Tsaponina, A. Chabes, B. Stillman,
A common telomeric gene silencing assay is affected by
nucleotide metabolism. Mol. Cell 42, 127–136 (2011).
doi: 10.1016/j.molcel.2011.03.007; pmid: 21474074
67. S. Imai, C. M. Armstrong, M. Kaeberlein, L. Guarente,
Transcriptional silencing and longevity protein Sir2 is an
NAD-dependent histone deacetylase. Nature 403, 795–800
(2000). doi: 10.1038/35001622; pmid: 10693811
68. F. Martino et al., Reconstitution of yeast silent chromatin:
Multiple contact sites and O-AADPR binding load SIR
complexes onto nucleosomes in vitro. Mol. Cell 33, 323–334
(2009). doi: 10.1016/j.molcel.2009.01.009; pmid: 19217406
69. E. A. Osborne, S. Dudoit, J. Rine, The establishment of gene
silencing at single-cell resolution. Nat. Genet. 41, 800–806
(2009). doi: 10.1038/ng.402; pmid: 19543267
70. N. C. Emre et al., Maintenance of low histone ubiquitylation by
Ubp10 correlates with telomere-proximal Sir2 association
and gene silencing. Mol. Cell 17, 585–594 (2005). doi: 10.1016/
j.molcel.2005.01.007; pmid: 15721261
71. R. K. McGinty, S. Tan, Recognition of the nucleosome by
chromatin factors and enzymes. Curr. Opin. Struct. Biol. 37,
54–61 (2016). doi: 10.1016/j.sbi.2015.11.014; pmid: 26764865
80. T. van Welsem et al., Dot1 promotes H2B ubiquitination by a
methyltransferase-independent mechanism. Nucleic Acids Res.
46, 11251–11261 (2018). doi: 10.1093/nar/gky801;
pmid: 30203048
81. L. Long, M. Furgason, T. Yao, Generation of nonhydrolyzable
ubiquitin-histone mimics. Methods 70, 134–138 (2014).
doi: 10.1016/j.ymeth.2014.07.006; pmid: 25063569
82. P. N. Dyer et al., Reconstitution of nucleosome core particles
from recombinant histones and DNA. Methods Enzymol. 375,
23–44 (2004). doi: 10.1016/S0076-6879(03)75002-2;
pmid: 14870657
83. B. J. Wilkins et al., Genetically encoding lysine modifications
on histone H4. ACS Chem. Biol. 10, 939–944 (2015).
doi: 10.1021/cb501011v; pmid: 25590375
84. X. Li et al., Electron counting and beam-induced motion
correction enable near-atomic-resolution single-particle
cryo-EM. Nat. Methods 10, 584–590 (2013). doi: 10.1038/
nmeth.2472; pmid: 23644547
85. S. Q. Zheng et al., MotionCor2: Anisotropic correction of beam-
induced motion for improved cryo-electron microscopy.
Nat. Methods 14, 331–332 (2017). doi: 10.1038/nmeth.4193;
pmid: 28250466
86. K. Zhang, Gctf: Real-time CTF determination and correction.
J. Struct. Biol. 193, 1–12 (2016). doi: 10.1016/j.jsb.2015.11.003;
pmid: 26592709
87. J. Zivanov et al., New tools for automated high-resolution
cryo-EM structure determination in RELION-3. eLife 7, e42166
(2018). doi: 10.7554/eLife.42166; pmid: 30412051
88. A. Punjani, J. L. Rubinstein, D. J. Fleet, M. A. Brubaker,
cryoSPARC: Algorithms for rapid unsupervised cryo-EM
structure determination. Nat. Methods 14, 290–296 (2017).
doi: 10.1038/nmeth.4169; pmid: 28165473
89. T. Grant, A. Rohou, N. Grigorieff, cisTEM, user-friendly software
for single-particle image processing. eLife 7, e35383 (2018).
doi: 10.7554/eLife.35383; pmid: 29513216
90. M. Wakamori et al., Intra- and inter-nucleosomal interactions of
the histone H4 tail revealed with a human nucleosome core
particle with genetically-incorporated H4 tetra-acetylation.
Sci. Rep. 5, 17204 (2015). doi: 10.1038/srep17204;
pmid: 26607036
92. S. Vijay-Kumar, C. E. Bugg, W. J. Cook, Structure of ubiquitin
refined at 1.8 A resolution. J. Mol. Biol. 194, 531–544 (1987).
doi: 10.1016/0022-2836(87)90679-6; pmid: 3041007
93. J. Min, Q. Feng, Z. Li, Y. Zhang, R. M. Xu, Structure of the
catalytic domain of human DOT1L, a non-SET domain
nucleosomal histone methyltransferase. Cell 112, 711–723
(2003). doi: 10.1016/S0092-8674(03)00114-4; pmid:
12628190
94. P. Emsley, K. Cowtan, Coot: Model-building tools for molecular
graphics. Acta Crystallogr. D Biol. Crystallogr. 60, 2126–2132
(2004). doi: 10.1107/S0907444904019158; pmid:15572765
95. P. D. Adams et al., PHENIX: A comprehensive Python-based
system for macromolecular structure solution. Acta Crystallogr.
D Biol. Crystallogr. 66, 213–221 (2010). doi: 10.1107/
S0907444909052925; pmid: 20124702
96. D. M. Truong, J. D. Boeke, Resetting the yeast epigenome with
human nucleosomes. Cell 171, 1508–1519.e13 (2017).
doi: 10.1016/j.cell.2017.10.043; pmid: 29198523
AC KNOWLED GME NTS
We thank W. Rice and B. Wang for helping with data collection at
the New York University (NYU) cryo-EM Shared Resource. We
thank the staff at the NYU Microscopy Laboratory for helping with
negative stain microscopy. We thank M. Costantino and HPC Core
at NYU Langone Health for computer access and support. We
thank C. Yun for helping with luminescence assays at the NYUMC
High Throughput Biology Laboratory. We thank F. Van Leeuwen for
sharing the rabbit anti-H3K79me1 (sera 58), rabbit anti-H3K79me2
(sera 30), rabbit anti-H3K79me3 (sera 34), and rabbit anti-Dot1.
We thank the members of the Boeke and Armache laboratories for
critical comments and discussion. We thank R. E. Kingston,
J. Cochrane, D. Smith, and A. Armache for comments and critical
review of this manuscript. Funding: The work in the Boeke
laboratory is supported by NSF grant MCB-1921641. The work in
Armache laboratory is supported by grants from the David
and Lucile Packard Foundation and the National Institutes of
Health (5R01GM115882). Author contributions: M.I.V.-S., P.D.,
M.W., and K.-J.A. conceptualized and designed the study. M.I.V.-S.,
P.D., M.W., R.L., and J.-P.A., conducted structural and biochemical
experiments. D.M.T. and J.D.B. performed yeast assays. All
authors contributed to data analysis, interpretation, and writing of
the manuscript. Competing interests: J.D.B. is a founder and
director of CDI Labs, a founder of Neochromosome, a founder of
and consultant to ReOpen Diagnostics, and serves or served on the
Scientific Advisory Board of the following: Sangamo, Modern
Meadow, Sample6, and the Wyss Institute. Data and materials
availability: All data are available in the manuscript or the
supplementary materials. The structure models and the cryo-EM
density maps have been deposited in the Protein Data Bank (PDB)
and Electron Microscopy Data Bank (EMDB) with accession codes
7K6Q, 7K6P, EMD-22691, EMD-22692, EMD-22693, EMD-22694,
and EMD-2265. Further inquiries and requests for resources and
reagents should be directed to and will be fulfilled by the lead
contact, K.-J.A. (karim-jean.armache@nyulangone.org).
SUPPLEMENTARY MATERIALS
science.sciencemag.org/content/371/6527/eabc6663/suppl/DC1
Materials and Methods
Supplementary Text
Figs. S1 to S19
Tables S1 to S3
References (97–107)
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
72. K. Luger, A. W. Mäder, R. K. Richmond, D. F. Sargent,
T. J. Richmond, Crystal structure of the nucleosome core
91. K. Sawada et al., Structure of the conserved core of the yeast
Dot1p, a nucleosomal histone H3 lysine 79 methyltransferase.
11 May 2020; accepted 27 October 2020
10.1126/science.abc6663
Valencia-Sánchez et al., Science 371, eabc6663 (2021)
22 January 2021
9 of 9
Corrected 28 January 2021. See full text.
|
10.1126_science.abh3209
|
RES EARCH
RESPIRATORY ENZYMES
The three-spin intermediate at the O–O cleavage and
proton-pumping junction in heme–Cu oxidases
Anex Jose1, Andrew W. Schaefer1, Antonio C. Roveda Jr.1, Wesley J. Transue1, Sylvia K. Choi2,
Ziqiao Ding2, Robert B. Gennis2, Edward I. Solomon1,3*
Understanding the mechanistic coupling of molecular oxygen reduction and proton pumping for
adenosine triphosphate synthesis during cellular respiration is the primary goal of research on
heme-copper oxidases—the terminal complex in the membrane-bound electron transport chain. Cleavage
of the oxygen-oxygen bond by the heme-copper oxidases forms the key intermediate PM, which
initiates proton pumping. This intermediate is now experimentally defined by variable-temperature,
variable-field magnetic circular dichroism spectroscopy on a previously unobserved excited state feature
associated with its heme iron(IV)-oxo center. These data provide evidence that the iron(IV)-oxo in
PM is magnetically coupled to both a copper(II) and a cross-linked tyrosyl radical in the active site.
These results provide new insight into the oxygen-oxygen bond cleavage and proton-pumping
mechanisms of heme-copper oxidases.
A erobic organisms use O2 to drive the
respiratory electron transport chain that
governs production of adenosine tri-
phosphate (ATP), the universal energy
currency in biology. The four-electron
reduction of O2 to H2O underlying this bio-
logical energy transduction process is cata-
lyzed by the superfamily of metalloenzymes
called heme-copper oxidases (HCOs), most
notably cytochrome c oxidases (CcOs) and
ubiquinol oxidases (UbOs). The free energy
gained in this exergonic reaction is used by
the HCOs to pump protons across the in-
ner mitochondrial or bacterial membrane
to create a proton gradient used by ATP syn-
thase, also bound to the membrane, for ATP
synthesis (1, 2). Oxygen reduction in HCOs
takes place at an active site that consists of
a heme center ligated by an axial histidine
(His) residue and a copper center (CuB) li-
gated by three His residues, one of which is
posttranslationally modified to form a co-
valent cross-link with a nearby tyrosine (Tyr)
residue (fig. S1A). HCOs also have a bis-His
ligated low spin (LS) heme and a binuclear
CuA center (only in CcOs) for electron transfer
(ET) to the active site.
The current consensus on the mechanism
of O2 reduction by HCOs (fig. S1B) (1–7) is that
when O2 binds to the mixed valent enzyme
(MV), the O–O bond is rapidly cleaved to
form intermediate PM. Because MV comprises
a fully reduced heme–Cu active site and fully
oxidized other redox centers, this PM forma-
tion indicates that all four requisite elec-
trons for O–O bond cleavage are supplied by
1Department of Chemistry, Stanford University, Stanford, CA
94305, USA. 2Department of Biochemistry, University of
Illinois, Urbana, IL 61801, USA. 3Stanford Synchrotron
Radiation Lightsource, SLAC National Accelerator Laboratory,
Stanford University, Menlo Park, CA 94025, USA.
*Corresponding author. Email: solomone@stanford.edu
the active site. Only the source of two of the
electrons has been clearly identified as the
heme iron (FeII→FeIV); whereas, one electron
each is commonly thought to be donated by CuB
(CuI→CuII) and the cross-linked Tyr (Tyr→Tyr●;
tyrosyl radical) (8–11).
Babcock and co-workers have used radio-
active iodine to probe the possible formation
of the Tyr● in PM through its labeling (9).
However, only ~2 × 10−3 mole fraction of the
PM formed was found to be labeled. Later
studies have also suggested a Tyr● in PM (12–15),
and electron paramagnetic resonance (EPR)
studies on PM generated by the reaction of the
oxidized enzyme (O) with H2O2 have shown
~2 to 5% of Tyr● (not necessarily the cross-
linked Tyr) but still lacked the Cu(II) signal
(16, 17). Thus, evidence for the presence of Cu
(II) and Tyr● during the oxygen reduction
reaction in HCOs remains elusive. Studies of
intermediates in the catalytic cycle of HCOs
are thus far limited to the heme side of the
heme–Cu active site using resonance Raman
(rR) and ultraviolet-visible (UV-vis) absorp-
tion spectroscopies. The intense features of
the two hemes present in these enzymes have
precluded the study of the oxidation and
protonation states of the CuB and Tyr in the
active site.
PM is the first intermediate in the catalytic
cycle where the O–O bond is fully cleaved,
and HCOs pump the first proton across the
membrane upon its one-electron reduction
(fig. S1B). Thus, an understanding of the geo-
metric and electronic structural contributions
of the heme–Cu active site to O–O bond
cleavage and how this is coupled to proton
pumping are important issues remaining
to be explored in HCO literature that can be
addressed through the experimental defini-
tion of PM. This is accomplished in this study
by variable-temperature, variable-field magnetic
circular dichroism (VTVH-MCD) spectroscopy
on a previously unobserved excited-state fea-
ture associated with the heme FeIV=O in PM to
study its ground state.
Results and analysis
Generating PM in ubiquinol oxidase bo3
We chose to study PM in Escherichia coli
cytochrome bo3 UbO because it lacks strong
absorption features of the CuA center in the
near-infrared (NIR) energy region important
for this study. We performed the photoreaction
of the CO-bound MV with O2 in D2O at pH 10
(CAPS buffer), thereby generating PM (supple-
mentary text and figs. S2 and S3), which is stable
for several minutes (10, 18–22).
MCD spectrum of PM
We characterized the electronic and magnetic
properties of PM and its decayed form F and O
by MCD spectroscopy. At 24 K and 6 T (Fig. 1A),
these spectra are dominated by the intense
features from p→p* intraligand and charge
transfer transitions of hemes: The LS heme-
b shows characteristic features at energies
<11,000 cm−1, and both hemes (heme-b and
heme-o) have overlapping features at ener-
gies >15,000 cm−1. Between these regions, a
relatively weak derivative-shaped feature is
present in PM, centered at ~13,900 cm−1 (Fig. 1B,
red), that is absent in the next intermediate
F (Fig. 1B, orange) and in O (Fig. 1B, blue).
We therefore conclude that this is a clear
characteristic feature of PM. In CcO, this rela-
tively weak feature at ~13,900 cm−1 would be
obscured by overlap with an intense MCD
feature from CuA (the Y→Y* transition) (23).
A similar MCD feature (~13,500 cm−1) is ob-
served for the alkaline form of myoglobin
compound-II (Mb-II) (from horse heart),
characteristic of its isolated FeIV=O heme
(24). On the basis of the similar natures of
these MCD features (band shapes and ener-
gies) and the similarities in their heme FeIV=O
centers (supplementary text and fig. S4), we
use Mb-II as an isolated FeIV=O heme refer-
ence (i.e., lacking the nearby CuB center) for
comparison with this center in PM.
VTVH-MCD of PM
The VTVH interrogation of this derivative-
shaped MCD feature of PM and Mb-II gives
the sets of isofield curves shown in Fig. 2, A
and B, respectively. These show similar
isofield behaviors at higher temperatures
(>15 K; 1/T < 0.07 K−1) but major differences
in their low-temperature behavior (<15 K).
Whereas the isofield MCD intensity plots of
Mb-II decrease to a temperature-independent
value at low temperatures, the isofield MCD
plots of PM decrease below 20 K to a nadir
around 8 K before rising again as the tem-
perature is further decreased to 1.7 K. This
rise in intensity at low temperature indicates
that PM has a paramagnetic ground state. This
Jose et al., Science 373, 1225–1229 (2021)
10 September 2021
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. MCD spectrum of O, PM, and F. (A) MCD spectrum of O, pH 10 (blue); PM (red); and F (orange) at 24 K and 6 T [the enlarged region shown in (B) is
marked with a square]. (B) MCD spectrum in the NIR region of O (blue), PM (red), and F (orange) at 24 K and 6 T (the derivative-shaped MCD feature in PM is
highlighted). mdeg, millidegrees.
can only be true if the FeIV=O [S = 1, axial zero-
field splitting (ZFS), D ~ +20 cm−1] heme in PM
is magnetically coupled to additional para-
magnetic center(s).
Fit to the VTVH-MCD data of PM: Coupling to
additional paramagnetic centers
The analysis and fit of the VTVH-MCD data
of Mb-II, including orientation averaging and
out-of-state magnetic field effects, are presented
in the supplementary text and figs. S5 and S6.
Because the VTVH-MCD data of PM reveal the
presence of additional paramagnetic center(s)
interacting with the FeIV=O, the influence of
isotropic exchange coupling (cid:1)2J
of
this zero-field split S = 1 center with additional
spin center(s) was assessed.
→^S 1 (cid:3) ^S 2
(cid:3)
(cid:1)
First, the exchange coupling of the FeIV=O
with a single S = ½ center was evaluated.
Floating the FeIV=O ZFS (D) and exchange
interaction (J12) gave D = +19.7 cm−1 and J12 =
−4.6 cm−1. This best fit greatly deviates from
the data (Fig. 2C) at low-temperature–low-
field conditions (T < 2.5 K, H = 2 to 4 T), where
the fit intensity continues to increase as the
data rapidly saturate, and the data at low-
temperature–high-field conditions (T < 10 K,
H = 7 T), where the fit greatly underestimates
the observed intensity. Thus, the VTVH-MCD
data of PM require the presence of a third spin
center (S = ½), where all spin centers have the
superexchange pathways required for exchange
coupling. We therefore evaluated three-spin
models, where the FeIV=O center is exchange-
coupled with an S = ½ center (J12), which is in
turn exchange-coupled with the second S = ½
center (J23), and a model with the FeIV=O
center exchange-coupled to both the S = ½
centers (J12 and J13). Both models greatly
improved the fit to the data, demonstrating
the requirement of a three-spin system in PM
(fig. S7). The best fit was obtained when all
spin centers were allowed to have exchange
interactions with each other. The final fit to
the VTVH-MCD data and the spin Hamiltonian
parameters of PM are given in Fig. 2A (con-
tributions to the MCD intensity of PM relative
to Mb-II are compared in fig. S8).
On the basis of the geometric and electronic
structural model of PM obtained through
proton-initiated O–O bond cleavage of the
generally assumed peroxo-bridged FeIII/CuII
precursor (IP in fig. S1B), we assign the two
additional S = ½ spin centers in PM revealed
by the VTVH-MCD data as the Cu(II) and the
cross-linked Tyr● (7). One of these S = ½
centers has a larger exchange coupling with
the FeIV=O than the other, which requires
better orbital overlap with the FeIV=O; we
therefore assign this to the Cu(II) center. The
third spin has superexchange pathways with
both the FeIV=O and the Cu(II) centers, and
these exchange interactions triangulate its
location to the cross-linked Tyr●. On the basis
of these data, we obtain the magnetic coupling
scheme of PM (Fig. 2D). The superexchange
pathways that enable these exchange inter-
actions are evaluated in the final section of the
analysis using the density functional theory
(DFT) model of PM derived from the proton-
initiated O–O bond cleavage in (7).
Spin Hamiltonian analysis
The best three-spin fit of the VTVH-MCD data
provides the energy levels in Fig. 2E, where the
magnetic field is along the z axis, which is the
dominant contributor to MCD intensity of an
xy-polarized transition (25). With D ≫ all Jj
j
values, the magnetic sublevels form a lower-
energy group of four (Fig. 2E, red and black)
and a higher-energy group of eight (Fig. 2E,
orange and blue). VTVH-MCD intensity is
dictated by the spin expectation value and
Boltzmann population of each sublevel, so
the lowest four-level grouping dominates
the observed low-temperature (T < 10 K)
MCD behavior. These levels originate from
coupling the nonmagnetic MS = 0 FeIV=O
sublevel with the two S = ½ spin centers,
providing direct insight into their exchange
couplings. The paramagnetic MS,tot = −1 (Stot =
1, black) is the ground sublevel when H > 2 T,
giving rise to MCD intensity at the lowest
temperature, 1.7 K in Fig. 2A. As the temper-
ature is increased from 1.7 K, population
density is redistributed into two MS,tot = 0
sublevels (from Stot = 1 and 0, black and red)
and an MS,tot = +1 sublevel (Stot = 1, black),
reducing the MCD intensity and producing
the intensity dip at ~8 K in Fig. 2A. Upon
further increase in temperature, the higher
energy MS,tot = −1 and −2 sublevels (from the
FeIV=O S = 1, MS = ±1 coupled to the two S =
½, orange and blue) are populated, and the
MCD intensity increases, leading to a peak at
~20 K. At still higher temperatures, the MCD
intensity decreases according to the Curie
behavior.
The two-spin system provided a much
poorer best fit to the VTVH-MCD data of PM
in Fig. 2C because it would have an Stot = ½
lowest energy spin state (fig. S9). The Zeeman
splitting of the MS,tot = ±½ is half of that of
the MS,tot = ±1, and thus the faster rate of
saturation of the MCD intensity at lower
temperatures observed in the isofield plots is
not feasible in a two-spin system (2 to 4 T;
Fig. 2C). Therefore, it is the low-temperature,
low-magnetic field region of the VTVH-MCD
data in PM that conclusively establishes a
magnetically coupled three-spin system of the
FeIV=O, the Cu(II), and the cross-linked Tyr●
centers.
Jose et al., Science 373, 1225–1229 (2021)
10 September 2021
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. VTVH-MCD data and analysis of PM and Mb-II. (A and B) VTVH-MCD
isofield data at 2 to 7 T and the best fit for PM (three-spin, S1 = 1 and
S2 = S3 = ½) (A) and Mb-II (S = 1) (B). (C) VTVH-MCD data of PM fit using
two-spin model (S1 = 1 and S2 = ½). (D) The best fit [three-spin model, solid lines
in (A)] derived magnetic coupling scheme for PM. (E) Spin state energy-level
diagram of PM [magnetic field along the z-axis (Hz); Stot and corresponding
MS,tot levels are labeled and color coded]. VTVH-MCD intensities are obtained
by measuring peak-to-trough intensity of the derivative-shaped MCD
features of PM and Mb-II (fig. S4A). The data are normalized to the
highest intensity (PM: 1.7 K, 7 T; Mb-II: 24 K, 7 T) and are means ± standard
deviations. All fits with xy-polarization. The 1-T data have been omitted
because of the poor signal-to-noise quality. T, Tesla; T, temperature;
WRSS, weighted residual squared-sum. See supplementary text for details
of spin Hamiltonian parameters used.
Geometric and electronic structure of PM
The geometric and electronic structural model
of PM that was obtained in (7) had a singlet
ground state arising from weak antiferromag-
netic (AF) coupling between the FeIV=O and
CuII–OH and weak ferromagnetic coupling
between the CuII–OH and Tyr●. We obtain a
lower-energy structure [DH = −1.8 kcal/mol
with respect to the structure in (7)] upon
reoptimizing it after including a small
structural perturbation based on features
present in the crystal structures of HCOs
(Fig. 3A; calculated spin densities and J
values are given at the bottom). The N(His)–
C(Tyr) cross-link has rotated by ~9° and tilted
by ~4° with respect to the structure in (7). The
calculated J values of this structure agree with
the results from the VTVH-MCD data and
further validate that the second and third spin
centers are the Cu(II) and the cross-linked Tyr●.
To understand the nature of the magnetic
interactions and their associated superexchange
pathways, we examined the molecular orbitals
(MOs) responsible for the exchange interactions
(schematic representation of the MOs are shown
in Fig. 3, B to E; the corresponding calculated
MO contours are given in fig. S10).
The AF coupling between FeIV=O and
CuII–OH arises from the superexchange pathway
mediated by the hydrogen bonding interac-
tion between the oxo(FeIV) and hydroxo(CuII)
ligands (Fig. 3A, heavy atom distance, 2.85 Å;
and Fig. 3, B and C), whereas the AF cou-
pling between Cu(II) and Tyr● is facilitated
by the superexchange pathway through the
His s-orbital, which results in the overlap be-
tween the Cu(II) and Tyr● magnetic orbitals
(Fig. 3D). The distortion of the N(His)–C(Tyr)
cross-link described above is responsible for
this s/p overlap. The ferromagnetic coupling
between Fe(IV) and Tyr● is promoted by the
hydrogen bonding interaction between the hy-
droxyl of the hydroxyethylfarnesyl of the heme
and the Tyr● (Fig. 3A, heavy atom distance,
2.82 Å; and Fig. 3E).
Jose et al., Science 373, 1225–1229 (2021)
10 September 2021
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. DFT-optimized structure and MOs involved
in the superexchange pathways of PM. (A) DFT-
optimized structure of PM in agreement with the
VTVH-MCD data. Hydrogen bonding interactions
between FeIV=O and CuII–OH and the hydroxyl of
hydroxyethylfarnesyl and Tyr(cid:129) are shown with red
arrows; heavy atom (O–O) distances are given.
Calculated exchange couplings between magnetic
centers and their spin densities are given at the
bottom. H2O hydrogen bonded to Tyr(cid:129) is removed for
clarity. (B to E) Representative schematics of MOs
involved in the superexchange pathway between Fe
(IV), Cu(II), and Tyr(cid:129): a-hole on CuII dx2-y2 (+OCu p)
mixed into the FeIV dp (+OFe p) (B), b-hole on FeIV dp
(+OFe p) mixed into the CuII dx2-y2 (+OCu p) (C),
a-hole on CuII dx2-y2 mixed into the Tyr p-orbital
through the His s-orbital (D), and b-hole on Tyr(cid:129)
p-orbital mixed into FeIV dp [orthogonal to the FeIV dp
in (C)] through the Ofarnesyl p-orbital (E). The valine
residue in (A) is removed for clarity in (B) to (E).
Discussion
The VTVH-MCD on the newly observed excited
state feature associated with the heme FeIV=O
in PM reveals a three-spin exchange-coupled
system involving FeIV=O (S = 1), Cu(II) (S = ½),
and Tyr● (S = ½) in the active site. Definition
of the geometric and electronic structure of
PM gives further insight into the O–O bond
cleavage mechanism in HCOs. Our past DFT
results showed that the Tyr donates the fourth
electron for the O–O bond cleavage through
the His–Tyr cross-link superexchange path-
way, which agrees with the VTVH-MCD data
showing the exchange interaction between
Cu(II) and Tyr●. It has also been concluded
that the Tyr supplies the proton required to
cleave the O–O bond in Ip (7). Because a valine
(Val) residue near the active site prevents the
direct proton transfer (PT) from the Tyr to the
O bound to Cu (OCu) (7), the proton must first
transfer to the OFe (through a water molecule
in the active site) and then rotate to transfer to
the OCu. This leads to the O–O bond cleavage
resulting in the PM structure in Fig. 3A, which
agrees with the VTVH-MCD results from this
study (Fig. 2A).
The ground state exchange coupling con-
stant, J, is related to the electronic coupling
matrix element, HDA, for ET as both involve
the same superexchange pathway (26, 27). By
experimentally determining the ground state
exchange interactions between the spin centers
in PM and by defining the superexchange path-
ways involved using DFT, our results elucidate
an ET pathway from the heme FeIV=O to the
cross-linked Tyr● in HCOs. Proceeding along
the catalytic cycle, when the FeIV=O center
in PM receives an electron from the LS heme,
the hydrogen bond–enabled superexchange
pathway facilitates ET from the FeIV=O to
the CuII–OH and then to the Tyr● through the
pathway enabled by the s/p overlap of the
His and the Tyr. From calculations, the ad-
dition of an electron to the structure of PM in
Fig. 3A results in the reduction of Tyr● to
tyrosinate (Tyr−) (fig. S11A).
Intensive efforts by researchers over several
decades have contributed to a molecular-level
understanding of proton pumping in HCOs
(1, 2). This ET from the LS heme to PM initiates
proton pumping (28). Coupled to this ET is
the arrival of a pump proton (pH+) to a proton-
loading site (PLS) located close to the active
site. The PLS has been generally proposed to
be a propionate of the active site heme or
nearby residues (1, 2, 29); however, because
the calculations show that reduction of PM
involves Tyr●→Tyr–, this Tyr– will have higher
proton affinity than the propionate and other
nearby residues, which indicates that the
tyrosinate can be the initial residue that accepts
pH+ to compensate its additional negative
charge. A plausible PT pathway from the
glutamic acid (Glu) residue to the Tyr− is shown
in fig. S12, A and B [this Glu is at the end of
the D proton channel and is thought to be the
branching point that supplies chemical pro-
tons (cH+) to the active site and pH+ to the PLS
(1, 2)]. Although the DFT calculations of the
one-electron reduced form of PM show that
protonation of the CuII–OH is thermodynami-
cally favored (DG ~9 kcal/mol) relative to the
protonation of the Tyr− (fig. S11, B and C)
(30, 31), because of the higher barrier of this
PT to the CuII–OH (32) (forming the thermo-
dynamic product), initial PT from the Glu to
the Tyr− would be kinetically favored [PT path-
way from the Glu to the CuII–OH proposed
in (2) is shown in fig. S12, C and D]. Once the
Tyr− is protonated, this proton likely will not
transfer to the CuII–OH. This is because the
direct PT is hindered by the Val near the active
site, and the pathway that enabled this PT for
O–O bond cleavage in IP is not available in PM
(as the cleaved O–O bond results in an FeIV=O
with a low proton affinity) (7). The subsequent
cH+ transfer from the Glu to the CuII–OH
repels pH+ from the Tyr to outside the mem-
brane, forming intermediate F. The structure
of F with a cross-linked Tyr− is consistent with
Fourier transform infrared studies on F where
the Tyr was found to be deprotonated at
pH 6.5 to 9.0 (33). This model accounts for con-
sumption of two protons from the N (negative)
side of the membrane and a proton that is
ejected to the P (positive) side during PM→F
(34). Thus, Tyr can be directly involved in
proton pumping either by serving as an initial
residue that accepts pH+ and transfers it to PLS
or by serving as the PLS itself. In this context, it
is also notable that the only proton channel in
the type-B family of HCOs ends near this Tyr.
The combination of our experimentally
calibrated DFT study of O–O cleavage in (7)
and the VTVH-MCD results presented here
for the resultant intermediate PM also provide
initial descriptions of the geometric and electronic
sructures of the one-electron reduced form
of PM, its protonated forms, and its thermo-
dynamic product F. Further experiments are
underway to spectroscopically define these
key intermediates for insight into how HCOs
couple the O–O bond cleavage to the function
of proton pumping for ATP synthesis.
Jose et al., Science 373, 1225–1229 (2021)
10 September 2021
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
RE FE RENCES AND N OT ES
1. M. Wikström, K. Krab, V. Sharma, Chem. Rev. 118, 2469–2490
(2018).
2. S. Yoshikawa, A. Shimada, Chem. Rev. 115, 1936–1989
(2015).
3. S. M. Adam et al., Chem. Rev. 118, 10840–11022 (2018).
4. A. W. Schaefer, M. T. Kieber-Emmons, S. M. Adam, K. D. Karlin,
E. I. Solomon, J. Am. Chem. Soc. 139, 7958–7973 (2017).
5. W.-G. Han Du, L. Noodleman, Inorg. Chem. 54, 7272–7290
(2015).
6. F. Poiana et al., Sci. Adv. 3, e1700279 (2017).
7. A. W. Schaefer, A. C. Roveda Jr., A. Jose, E. I. Solomon, J. Am.
Chem. Soc. 141, 10068–10081 (2019).
8. M. Fabian, G. Palmer, Biochemistry 34, 13802–13810 (1995).
9. D. A. Proshlyakov et al., Science 290, 1588–1591 (2000).
10. D. A. Proshlyakov, M. A. Pressler, G. T. Babcock, Proc. Natl.
Acad. Sci. U.S.A. 95, 8020–8025 (1998).
11. V. R. I. Kaila, M. P. Johansson, D. Sundholm, L. Laakkonen,
M. Wiström, Biochim. Biophys. Acta 1787, 221–233 (2009).
12. T. Uchida, T. Mogi, T. Kitagawa, Biochemistry 39, 6669–6678
(2000).
13. M. Iwaki, A. Puustinen, M. Wikström, P. R. Rich, Biochemistry
45, 10873–10885 (2006).
14. V. S. Oganesyan et al., J. Biol. Inorg. Chem. 15, 1255–1264
(2010).
15. D. A. Proshlyakov, Biochim. Biophys. Acta 1655, 282–289 (2004).
16. M. A. Yu et al., J. Am. Chem. Soc. 134, 4753–4761 (2012).
17. K. Budiman et al., Biochemistry 43, 11709–11716 (2004).
I. Szundi, J. Ray, A. Pawate, R. B. Gennis, O. Einarsdóttir,
18.
Biochemistry 46, 12568–12578 (2007).
19. M. Karpefors, P. Ädelroth, A. Namslauer, Y. Zhen, P. Brzezinski,
Biochemistry 39, 14664–14669 (2000).
20. B. Chance, C. Saronio, J. S. Leigh Jr., J. Biol. Chem. 250,
9226–9237 (1975).
21. K. Oda, T. Ogura, E. H. Appelman, S. Yoshikawa, FEBS Lett.
36. W. J. Transue, wtransue/mcdps: Initial particle swarm MCD
570, 161–165 (2004).
22. E. Pinakoulaki et al., J. Biol. Chem. 288, 20261–20266
fitting routine, version 0.4, Zenodo (2021); https://doi.org/10.
5281/zenodo.4969304.
(2013).
23. D. R. Gamelin et al., J. Am. Chem. Soc. 120, 5246–5263
AC KNOWLED GME NTS
(1998).
24. N. Foote, P. M. A. Gadsby, C. Greenwood, A. J. Thomson,
Biochem. J. 261, 515–522 (1989).
25. F. Neese, E. I. Solomon, Inorg. Chem. 38, 1847–1865
(1999).
26. T. C. Brunold, D. R. Gamelin, E. I. Solomon, J. Am. Chem. Soc.
122, 8511–8523 (2000).
27. R. G. Hadt, S. I. Gorelsky, E. I. Solomon, J. Am. Chem. Soc. 136,
15034–15045 (2014).
28. I. Belevich, M. I. Verkhovsky, M. Wikström, Nature 440,
829–832 (2006).
29. J. Lu, M. R. Gunner, Proc. Natl. Acad. Sci. U.S.A. 111,
12414–12419 (2014).
30. W.-G. Han Du, A. W. Götz, L. Noodleman, Inorg. Chem. 58,
13933–13944 (2019).
31. M. R. A. Blomberg, Biochemistry 55, 489–500 (2016).
32. R. Liang, J. M. J. Swanson, Y. Peng, M. Wikström, G. A. Voth,
Proc. Natl. Acad. Sci. U.S.A. 113, 7420–7425 (2016).
We thank the Stanford Research Computing Center for providing
computational resources. Funding: This study received funding
from the National Institutes of Health grant R01DK031450 (to
E.I.S.); Coordenação de Aperfeiçoamento de Pessoal de Nível
Superior - Brasil (CAPES), Finance Code 001 (to A.C.R.);
and the Ruth L. Kirschstein National Research Service Award
F32GM131602 (to W.J.T.). Author contributions: E.I.S., A.W.S.,
A.C.R., and A.J. designed the experiments. S.K.C. and Z.D.
prepared and purified the protein. A.W.S., A.C.R., and A.J.
performed the experiments. W.J.T. wrote the VTVH-MCD fitting
program. A.J. and A.W.S. performed the DFT calculations. A.J.,
A.W.S., A.C.R., and E.I.S. analyzed the data. A.J. and E.I.S. wrote
the manuscript with assistance from all authors. Competing
interests: The authors declare that they have no competing
interests. Data and materials availability: All spectroscopic data
in the main text and supplementary materials, computer code
for fitting the VTVH-MCD data, and atomic coordinates of DFT
structures are freely available through Zenodo (35, 36).
33. E. A. Gorbikova, M. Wikström, M. I. Verkhovsky, J. Biol. Chem.
SUPPLEMENTARY MATERIALS
283, 34907–34912 (2008).
34. K. Faxén, G. Gilderson, P. Ädelroth, P. Brzezinski, Nature 437,
286–289 (2005).
35. A. Jose, A. W. Schaefer, A. C. Roveda Jr., W. J. Transue,
S. K. Choi, Z. Ding, R. B. Gennis, E. I. Solomon, Spectroscopic
data and DFT coordinates for “The three-spin intermediate at
the O–O cleavage and proton pumping junction in heme–Cu
oxidases,” Zenodo (2021); https://doi.org/10.5281/zenodo.
4961011.
https://science.org/doi/10.1126/science.abh3209
Materials and Methods
Supplementary Text
Figs. S1 to S12
References (37–51)
MDAR Reproducibility Checklist
1 March 2021; accepted 30 July 2021
10.1126/science.abh3209
Jose et al., Science 373, 1225–1229 (2021)
10 September 2021
5 of 5
|
10.1126_science.abh3629
|
RES EARCH
FOREST ECOLOGY
Multidimensional tropical forest recovery
Lourens Poorter1*, Dylan Craven2, Catarina C. Jakovac1,3, Masha T. van der Sande1, Lucy Amissah4,
Frans Bongers1, Robin L. Chazdon5,6, Caroline E. Farrior7, Stephan Kambach8, Jorge A. Meave9,
Rodrigo Muñoz1,9, Natalia Norden10, Nadja Rüger8,11,12, Michiel van Breugel13,14,15,
Angélica María Almeyda Zambrano16, Bienvenu Amani17, José Luis Andrade18, Pedro H. S. Brancalion19,
Eben N. Broadbent20, Hubert de Foresta21, Daisy H. Dent12,22, Géraldine Derroire23, Saara J. DeWalt24,
Juan M. Dupuy18, Sandra M. Durán25,26, Alfredo C. Fantini27, Bryan Finegan28, Alma Hernández-Jaramillo29,
José Luis Hernández-Stefanoni18, Peter Hietz30, André B. Junqueira31, Justin Kassi N’dja32,
Susan G. Letcher33, Madelon Lohbeck1,34, René López-Camacho35, Miguel Martínez-Ramos36,
Felipe P. L. Melo37, Francisco Mora36, Sandra C. Müller38, Anny E. N’Guessan32, Florian Oberleitner39,
Edgar Ortiz-Malavassi40, Eduardo A. Pérez-García9, Bruno X. Pinho37, Daniel Piotto41,
Jennifer S. Powers42,43, Susana Rodríguez-Buriticá10, Danaë M. A. Rozendaal44,45, Jorge Ruíz46,
Marcelo Tabarelli37, Heitor Mancini Teixeira44,47,48, Everardo Valadares de Sá Barretto Sampaio49,
Hans van der Wal50, Pedro M. Villa51,52, Geraldo W. Fernandes53, Braulio A. Santos54, José Aguilar-Cano10,
Jarcilene S. de Almeida-Cortez55, Esteban Alvarez-Davila56, Felipe Arreola-Villa36, Patricia Balvanera36,
Justin M. Becknell57, George A. L. Cabral37, Carolina Castellanos-Castro10, Ben H. J. de Jong58,
Jhon Edison Nieto10, Mário M. Espírito-Santo59, Maria C. Fandino60, Hernando García10,
Daniel García-Villalobos10, Jefferson S. Hall13, Alvaro Idárraga61, Jaider Jiménez-Montoya62,
Deborah Kennard63, Erika Marín-Spiotta64, Rita Mesquita65, Yule R. F. Nunes59, Susana Ochoa-Gaona58,
Marielos Peña-Claros1, Nathalia Pérez-Cárdenas36, Jorge Rodríguez-Velázquez36,
Lucía Sanaphre Villanueva66,18, Naomi B. Schwartz67, Marc K. Steininger68, Maria D. M. Veloso59,
Henricus F. M. Vester69, Ima C. G. Vieira70, G. Bruce Williamson65,71, Kátia Zanini38, Bruno Hérault72,73,74
Tropical forests disappear rapidly because of deforestation, yet they have the potential to regrow
naturally on abandoned lands. We analyze how 12 forest attributes recover during secondary succession
and how their recovery is interrelated using 77 sites across the tropics. Tropical forests are highly
resilient to low-intensity land use; after 20 years, forest attributes attain 78% (33 to 100%) of their
old-growth values. Recovery to 90% of old-growth values is fastest for soil (<1 decade) and plant
functioning (<2.5 decades), intermediate for structure and species diversity (2.5 to 6 decades), and
slowest for biomass and species composition (>12 decades). Network analysis shows three independent
clusters of attribute recovery, related to structure, species diversity, and species composition.
Secondary forests should be embraced as a low-cost, natural solution for ecosystem restoration,
climate change mitigation, and biodiversity conservation.
T ropical forests are converted at alarm-
ing rates to other land uses (1), yet they
also have the potential to regrow natu-
rally on abandoned agricultural fields and
pastures. Widespread land abandonment
because of fertility loss, migration, or alter-
native livelihood options has led to a rapid
increase in the extent of regrowing forests.
Currently, regrowth covers as much as 28%
(2.4 million km2) of the neotropics alone (2).
Regrowing secondary forests (SFs) form a
large and important component of human-
modified tropical landscapes and have the
potential to play a key role in biodiversity
conservation (3), climate change mitigation
(2), and landscape restoration (4). A holistic,
quantitative understanding of the recovery
of multiple SF functions is needed to inform
and design effective policies that benefit na-
ture and people from local to global scales.
In this study, we assess the resilience of 12
forest attributes to recover from agriculture
and pasture use. Resilience is the ability of
a system to absorb disturbances and return
to its previous state (5). Resilience is driven
by two underlying components: the ability
to resist disturbance and the ability to re-
cover after disturbance (6). We defined “re-
sistance” as the difference between the value
of the forest attribute at the start of suc-
cession and the average old-growth forest
(OGF) values [compare (7)], which reflects
the combined legacies of previous forest and
previous land use, and “recovery” as the abil-
ity to return to OGF attribute values after
succession. Succession is defined as a change
in vegetation structure, species composition
(SC), and ecosystem functioning over time
after a disturbance (8). Secondary succession
occurs on previously vegetated lands when
a disturbance removes most of the above-
ground vegetation and can proceed at fast
rates due to legacy effects of previous forest
or previous land use, such as a developed
soil, seed bank, remnant trees, and resprout-
ing stumps. Successional pathways are, to
some extent, predictable but, because of local
stochastic factors, are also, to some extent,
uncertain (9).
Most successional theories have focused on
specific ecosystem attributes, such as SC (10),
species richness (SR) (11), forest structure (12),
or soils (13), but they have rarely been con-
ceptually integrated and assessed together.
Recovery of these different ecosystem attrib-
utes (i.e., dimensions) is likely to depend on
one another. For example, rapid recovery of
biomass may lead to high litter production
and decomposition and, hence, rapid recovery
of soil organic carbon.
Chronosequence studies allow us to infer
long-term trends in forest recovery by com-
paring forests with similar land-use history
that differ in age since agricultural or pas-
ture abandonment. Single-site studies have
assessed the recovery of multiple attributes
[summarized in (14)], and several synthetic
analyses have assessed the recovery of single
attributes. They have found that ecosystem
functioning, such as nitrogen fixation, re-
covers fast [in about three decades (15)],
whereas aboveground biomass (AGB) and
SR recover more slowly [three to seven dec-
ades (16, 17)] and SC recovers slowest [i.e.,
centuries (17)]. To date, we lack a comprehen-
sive understanding on how multiple attributes
differ in recovery rates and how recovery of
these attributes is interrelated.
In this study, we analyze how 12 forest at-
tributes recover during secondary succession
and how their recovery is interrelated. We fo-
cused on four complementary groups of attrib-
utes that capture successional changes in soils
[bulk density (BD), carbon (C), and nitrogen
(N)], ecosystem functioning [community nitro-
gen fixers, wood density (WD), and specific
leaf area (SLA)], forest structure [AGB, maxi-
mum tree diameter, and structural heteroge-
neity (SH)], and diversity and composition
(SR, species diversity, and similarity to OGF).
These four groups are key components of eco-
system functioning (18), and knowledge of
their recovery during succession is a prereq-
uisite for the formulation of global policies
on biodiversity conservation, climate change
mitigation, and forest restoration. We ask (i)
how multiple forest attributes recover during
succession, (ii) how their relative recovery is
interrelated, and (iii) whether one (or several)
attribute(s) can be used as a simple proxy for
multidimensional recovery. We advance pre-
vious analyses by (i) including a wider range
of forest attributes for a larger number of sites
(77) compared with those in previous studies,
(ii) developing and applying an original con-
ceptual framework to model forest recovery,
(iii) examining how recovery among forest at-
tributes is interrelated, and (iv) identifying
simple indicators to monitor the progress of
forest restoration.
We compiled original chronosequence data
from three continents, 77 sites, 2275 plots, and
226,343 stems, spanning the major environmental
and latitudinal gradients in the lowland neo-
tropics and West Africa [(18); Fig. 1D and table
S1]. Chronosequences do not monitor plots
Poorter et al., Science 374, 1370–1376 (2021)
10 December 2021
1 of 7
RES EARCH | R E S E A R C H A R T I C L E
over time but rather substitute space for time
to infer recovery. Plots were, on average, 0.1 ha,
in which all woody plants were identified and
measured for their stem diameter. Forest at-
tributes were measured for 21 sites (for soils)
up to 77 sites for the other variables. To quan-
tify to what extent SF attributes recover toward
OGF values, recovery was modeled for each
chronosequence as a process in which SF values
return exponentially to OGF values (Fig. 1A).
When available, OGF plots were used to esti-
mate OGF chronosequence reference values
(supplementary text, section S1). For each study
site, relative forest recovery was expressed as
the similarity (ranging between 0 and 100%)
between the predicted values for SF plots and
OGF plots, thereby enabling direct compar-
isons of recovery across forests and attributes
(Fig. 1, A and B). To assess how recovery of
different forest attributes was connected during
succession and which attributes can serve as
proxies for multidimensional recovery, we
carried out a network analysis (Fig. 1C).
Pace of recovery
Forest attributes differ in their starting val-
ues after land abandonment (i.e., resistance)
and subsequent recovery (Fig. 2). Starting
values varied from 1 to 90% (Fig. 3A), re-
covery after 20 years (R20y) varied from 33 to
100% (Fig. 3B), and recovery time (RT) to
90% of OGF values varied from 0 to 120 years
(Fig. 3D). The ranking in recovery of the four
different groups is maintained when recov-
ery is evaluated in terms of intrinsic recov-
ery rate (l) instead of percentage of recovery
(fig. S2). In the coming sections, we first
briefly introduce each group of attributes.
See (18) for a detailed explanation of their
importance and how they recover during
succession.
Soil functioning was evaluated in terms of
organic C, N, and BD of the topsoil. Soil C
concentration scales positively with soil organ-
ic matter content and, hence, with nutrients in
organic material and water holding capacity.
Abandoned agricultural fields and pastures
may have low soil C because of combustion
during slash and burn (19). Soil N concen-
tration is an indicator of soil fertility and
may be low in abandoned fields because of
uptake by crops and cattle, volatilization, ero-
sion, and leaching (19). Soil BD is soil dry mass
over soil volume and may be high because of
soil compaction by agricultural practices and
livestock.
We expected soil recovery to occur more
slowly than vegetation recovery because soil
recovery depends on leaf and root litter inputs.
Yet recovery of soil attributes was surprisingly
fast [compare (7)], with an R20y of 98 to 100%
and an RT of 1 to 9 years (Fig. 3, B and D). Just
after land abandonment of agriculture or pas-
ture (t0), the starting values of C, N, and BD
were relatively high (62 to 90%; Fig. 3A),
which indicates that they are less affected by
slashing or burning than aboveground veg-
etation, contain more legacies of previous land
use, and have a high resistance to disturbance.
Most of our data come from regrowth after
light- to mid-intensity land uses during which
soil degradation is not extreme. Soils may also
recover quickly due to rapid recovery of the
soil biotic community, because slash-and-burn
management has transferred nutrients from
the aboveground vegetation to the soil, or
because productive grass roots and nitrogen-
fixing herbs have increased soil C and N (19).
Soil C recovered in ~5 years to 90% of OGF
values, probably because it is weakly affected
by aboveground disturbances associated with
land-use change, such as fire and clearing.
A meta-analysis found that soil C of SF was
similar to that of OGF and did not change
during succession (20). Most soil nutrients
may recover quickly because plants may ac-
quire nutrients from deeper soil layers, because
of high litter production early in succession
due to ample light availability, and because
of the high rates of leaf and root turnover of
pioneer species (21). Litter quality may also be
higher early in succession, because pioneers
tend to have high concentrations of leaf nu-
trients (22) and nitrogen fixers are especially
abundant (15) and active early in succession
(23). Recovery of phosphorus may be slow be-
cause it can only be replenished through atmo-
spheric deposition and mineral weathering (7).
The observed fast soil recovery is important
1Forest Ecology and Forest Management Group, Wageningen University, Wageningen, Netherlands. 2Centro de Modelación y Monitoreo de Ecosistemas, Universidad Mayor, Santiago, Chile.
3Departamento de Fitotecnia, Universidade Federal de Santa Catarina. Rod. Admar Gonzaga, Florianópolis, SC, Brazil. 4CSIR-Forestry Research Institute of Ghana, KNUST, Kumasi, Ghana.
5Department of Ecology and Evolutionary Biology, University of Connecticut, Storrs, CT, USA. 6Tropical Forests and People Research Centre, University of the Sunshine Coast, Maroochydore DC,
QLD, Australia. 7University of Texas at Austin, Austin, TX, USA. 8German Centre for Integrative Biodiversity Research (iDiv) Halle-Jena-Leipzig, Leipzig, Germany. 9Departamento de Ecología y
Recursos Naturales, Facultad de Ciencias, Universidad Nacional Autónoma de México, Coyoacán, Mexico City, Mexico. 10Instituto de Investigación de Recursos Biológicos Alexander von Humboldt,
Bogotá, Colombia. 11Department of Economics, University of Leipzig, Leipzig, Germany. 12Smithsonian Tropical Research Institute, Ancón, Balboa, Panama. 13SI ForestGEO, Smithsonian Tropical
Research Institute, Ancón, Balboa, Panama. 14Yale-NUS College, Singapore, Singapore. 15Department of Biological Sciences, National University of Singapore, Singapore, Singapore. 16Center for
Latin American Studies, University of Florida, Gainesville, FL, USA. 17UFR Agroforesterie, Université Jean Lorougnon Guédé Daloa, Daloa, Côte d’Ivoire. 18Centro de Investigación Científica de
Yucatán A.C. Unidad de Recursos Naturales, Colonia Chuburná de Hidalgo, Mérida, Yucatán, Mexico. 19Department of Forest Sciences, “Luiz de Queiroz” College of Agriculture, University of São
Paulo, Piracicaba, São Paulo, Brazil. 20Spatial Ecology and Conservation Lab, School of Forest Resources and Conservation, University of Florida, Gainesville, FL, USA. 21UMR AMAP, Institut de
Recherche pour le Développement (IRD), Montpellier, France. 22Biological and Environmental Sciences, University of Stirling, Stirling, UK. 23CIRAD, UMR EcoFoG (AgroParistech, CNRS, INRAE,
Université des Antilles, Université de la Guyane), Campus Agronomique, Kourou, French Guiana. 24Department of Biological Sciences, Clemson University, Clemson, SC, USA. 25Earth and
Atmospheric Sciences Department, University of Alberta, Edmonton, AB, Canada. 26Department of Ecology and Evolutionary Biology, University of Minnesota, St. Paul, MN, USA. 27Universidade
Federal de Santa Catarina, Brazil. 28CATIE-Centro Agronómico Tropical de Investigación y Enseñanza, Turrialba, Costa Rica. 29Neotropical Primate Conservation Colombia, Bogotá, Colombia.
30Institute of Botany, University of Natural Resources and Life Sciences, Vienna, Austria. 31Institut de Ciència i Tecnologia Ambientals, Universitat Autònoma de Barcelona, Cerdanyola del
Vallès, Barcelona, Spain. 32Departement of Bioscience, University Felix Houphouet-Boigny, Abidjan, Côte d’Ivoire. 33College of the Atlantic, Bar Harbor, ME, USA. 34World Agroforestry Centre,
ICRAF, United Nations Avenue, Gigiri, Nairobi, Kenya. 35Universidad Distrital Francisco José de Caldas, Facultad de Medio Ambiente y Recursos Naturales, Bogotá, Colombia. 36Instituto de
Investigaciones en Ecosistemas y Sustentabilidad, Universidad Nacional Autónoma de México, Morelia, Michoacán, Mexico. 37Departamento de Botânica, Universidade Federal de Pernambuco,
Recife, Brazil. 38Departamento de Ecologia, Instituto de Biociências, Universidade Federal do Rio Grande do Sul, Porto Alegre, RS, Brazil. 39Department of Ecology, University of Innsbruck,
Innsbruck, Austria. 40Instituto Tecnológico de Costa Rica, Escuela de Ingeniería Forestal, Cartago, Costa Rica. 41Centro de Formação em Ciências Agroflorestais, Universidade Federal do Sul da
Bahia, Itabuna, BA, Brazil. 42Department of Ecology, Evolution and Behavior, University of Minnesota, St. Paul, MN, USA. 43Department of Plant and Microbial Biology, University of Minnesota,
St. Paul, MN, USA. 44Plant Production Systems Group, Wageningen University and Research, Wageningen, Netherlands. 45Centre for Crop Systems Analysis, Wageningen University and Research,
Wageningen, Netherlands. 46Programa de Estudios de Posgrado en Geografia, Convenio Universidad Pedagogica y Tecnológica de Colombia-Instituto Geografico Agustin Codazzi, Bogotá,
Colombia. 47Farming Systems Ecology, Wageningen University, Wageningen, Netherlands. 48Copernicus Institute, Utrecht University, Utrecht, Netherlands. 49Departamento de Energia Nuclear–
CTG, Universidade Federal de Pernambuco, Recife, Pernambuco, Brazil. 50Departamento de Agricultura, Sociedad y Ambiente, El Colegio de la Frontera Sur – Unidad Villahermosa, Centro,
Tabasco, México. 51Program of Botany, Departamento de Biologia Vegetal, Laboratório de Ecologia e Evolução de Plantas, Universidade Federal de Viçosa, Viçosa, Minas Gerais, Brazil. 52Fundación
para la Conservación de la Biodiversidad (PROBIODIVERSA), Mérida, Mérida, Venezuela. 53Ecologia Evolutiva e Biodiversidade/DBG, ICB, Universidade Federal de Minas Gerais, Belo Horizonte,
Brazil. 54Federal University of Paraíba, João Pessoa, Brazil. 55Departamento de Botânica–CCB, Universidade Federal de Pernambuco, Pernambuco, Brazil. 56Escuela ECAPMA – Universidad
Nacional Abierta y a Distancia, Bogotá, Colombia. 57Environmental Studies Program, Colby College, Waterville, ME, USA. 58Department of Sustainability Science, El Colegio de la Frontera Sur,
Lerma, Campeche, Mexico. 59Departamento de Biologia Geral, Universidade Estadual de Montes Claros, Montes Claros, Minas Gerais, Brazil. 60Fondo Patrimonio Natural para la Biodiversidad y
Areas Protegidas, Bogota, Colombia. 61Fundación Jardín Botánico de Medellín, Herbario JAUM, Medellín, Colombia. 62Instituto de Biología, Universidad de Antioquia, Antioquia, Colombia.
63Department of Physical and Environmental Sciences, Colorado Mesa University, Grand Junction, CO, USA. 64Department of Geography, University of Wisconsin–Madison, Madison, WI, USA.
65Biological Dynamics of Forest Fragments Project, Environmental Dynamics Research Coordination, Instituto Nacional de Pesquisas da Amazonia, Manaus, Amazonas, Brazil. 66Consejo Nacional
de Ciencia y Tecnologia, Centro del Cambio Global y la Sustentabilidad, Tabasco, Mexico. 67Department of Geography, University of British Columbia, Vancouver, BC, Canada. 68Department of
Geographical Sciences, University of Maryland, College Park, MD, USA. 69Institute for Biodiversity and Ecosystem Dynamics (IBED), University of Amsterdam, Amsterdam, Netherlands. 70Museu
Paraense Emilio Goeldi, Belém, Pará, Brazil. 71Department of Biological Sciences, Louisiana State University, Baton Rouge, LA, USA. 72CIRAD, UPR Forêts et Sociétés, Yamoussoukro, Côte d’Ivoire.
73Forêts et Sociétés, Université Montpellier, CIRAD, Montpellier, France. 74Institut National Polytechnique Félix Houphouët-Boigny, INP-HB, Yamoussoukro, Côte d’Ivoire.
*Corresponding author. Email: lourens.poorter@wur.nl
Poorter et al., Science 374, 1370–1376 (2021)
10 December 2021
2 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Study approach to analyze recovery of different forest attributes. (A to C) Absolute recovery
of SF attributes toward OGF values (A) can be standardized to relative recovery rates (B), which express how
close each SF attribute is to OGF values, thereby allowing direct comparisons across attributes, such as
in network analyses (C), which show how recovery is coordinated across forest attributes. The widths
of paths among attributes indicate the strength of the coordination. The different colors indicate attribute
category: soil (brown), plant functioning (purple), structure (green), and diversity (turquoise). (D) Map of
the 77 study sites in the neotropics and West Africa (for site numbers, see table S1). Potential forest cover is
shown in green.
for the sustainability of shifting cultivation
agriculture, which coincides with agronomic
studies that indicate that a fallow period of
more than 8 to 10 years allows agricultural
productivity to be maintained (21).
Plant functioning was evaluated in terms of
basal area–weighted community WD and SLA
and the percentage basal area of nitrogen-
fixing trees. WD is the stem-wood dry mass
divided by stem volume, and it increases tissue
longevity and carbon residence time in trees
and forests. SLA is the leaf area divided by the
leaf mass. It reflects leaf display cost and scales
positively with photosynthetic capacity and
forest productivity and negatively with leaf
longevity. WD and SLA change during sec-
ondary succession because pioneer species are
typically replaced by later-successional species
with opposite trait values (24). Nitrogen fixa-
tion indicates the potential for biological nitro-
gen input to trees and forests. Nitrogen fixation
is generally high early in succession when ir-
radiance is high and trees can support their
nitrogen-fixing symbionts with carbohydrates
(23) and declines over time as forests regrow
(15), light availability in the stand drops, and
nitrogen fixation becomes too costly (23).
Recovery of ecosystem processes depends
on the characteristics of species that make up
the community. Although SC may recover slow-
ly, we expected plant functioning to recover at
an intermediate pace because many OGF spe-
cies have similar (i.e., redundant) trait values.
We found that plant functioning recovers sur-
prisingly fast (R20y of 82 to 100% and an ave-
rage RT of 3 to 27 years; Fig. 3, B and D).
During succession, short-lived pioneer spe-
cies (with life spans of 10 to 30 years and
extreme trait values) are rapidly replaced by
later-successional species that are functionally
similar to one another but different from
pioneer species (25), which leads to a fast
functional recovery. Additionally, resprouting
is a common mode of regeneration on aban-
doned fields, which explains why the func-
tional composition rapidly resembles that of
the previous OGF [(26); Fig. 3A]. Finally, fast
recovery also occurs because traits such as
SLA and WD never have values of zero and,
therefore, start closer to OGF values (85 and
76%, respectively) than, for example, the pro-
portion of nitrogen-fixing trees (40%; Fig. 3A).
Forest structure was evaluated in terms of
AGB, maximum tree size (Dmax), and SH. AGB
is a strong driver of ecosystem processes (27)
and important for carbon storage and climate
change mitigation (16). Dmax reflects the pres-
ence of large trees that have a high conserva-
tion value, providing habitat and food for many
organisms. SH refers to the tree size variation
in a plot; it increases light capture and eco-
system productivity (18) and contributes to
biodiversity conservation by providing a hab-
itat for different species.
We expected forest structure to recover fast-
er than soil and trait attributes because all
trees and species contribute to forest struc-
ture, but we found that it occurs at an inter-
mediate pace (R20y of 33 to 83% and an
average RT of 27 to 119 years; Fig. 3, B and
D), probably because it often starts close to
zero. SH recovered at an intermediate pace,
probably because it increases with Dmax and
because it reflects a gradual transition from
even-aged forest that establishes just after
land abandonment toward uneven-aged forest
with continuous regeneration and multiple
cohorts. Recovery of Dmax took more time
because it depends on the identity and growth
of individual trees. AGB had the slowest re-
covery because large trees drive AGB (28) and
because of low productivity in later succes-
sional stages (16). The RT of 12 decades for
AGB is substantially longer than the seven
decades we previously estimated, owing to dif-
ferences in the number of study sites (77 ver-
sus 43) and modeling approach (18).
Diversity was evaluated in terms of SR,
Simpson diversity (SD), and SC. SR is directly
relevant for conservation, because it indicates
the number of locally co-occurring species.
SD indicates the diversity of common species
and reflects successional shifts in community
structure from young forests dominated by
few pioneer species to diverse forests with
many rare species. SC indicates to what ex-
tent the SC in an area (i.e., the identity of
species and their relative abundance) resem-
bles that of an OGF and thus indicates the
quality of diversity and the value of SFs for
the conservation of old-growth species. SR,
species diversity, and SC usually start close to
zero (Fig. 3A) with few or no woody plants,
owing to biomass and species removal for
previous land use, and increase over time as
seeds germinate from the seed bank and new
species arrive and get established.
Recovery of species diversity and SC occurred
at an intermediate to slow pace. SR recovered
fastest (R20y = 78%, RT = 37 years) because early
in succession, biodiversity can be high because
both light-demanding early-successional and
shade-tolerant later-successional species coexist
(11, 29). SD recovered more slowly (R20y = 69%,
Poorter et al., Science 374, 1370–1376 (2021)
10 December 2021
3 of 7
RES EARCH | R E S E A R C H A R T I C L E
RT = 59 years) because it takes time before
competition leads to more equal species abun-
dances. SC recovered slowest (R20y = 33%, RT =
120 years) because it depends on overcoming
dispersal and recruitment limitations, the ac-
cumulation of rare shade-tolerant species, and
tree turnover (which takes decades to centu-
ries). Recovery in SC varied substantially across
sites (as indicated by wide credibility intervals;
Fig. 3D), possibly because sites vary in the
drivers of succession, such as land-use history,
the number and identity of remnant trees,
proximity to seed sources, resprouting abil-
ity, and the proportion of wind-dispersed tree
species in the community (17).
We used a chronosequence approach to
infer long-term recovery because few studies
have monitored succession over time. Our
approach assumes that all plots within a
chronosequence had similar starting condi-
tions and follow a similar recovery trajec-
tory, which is not necessarily the case (9, 30).
SF chronosequence studies that also moni-
tored dynamics over time showed that dy-
namic pathways in species diversity, SC, and
species structure generally matched chro-
nosequence trends but also showed devia-
tions for some plots (31, 32). Instantaneous
trends could show a faster increase for di-
versity and similar trends for composition and
basal area compared with chronosequence
predictions (31, 32). Hence, the patterns we
observed in this study should be corrobo-
rated by long-term studies that monitor SF
dynamics over time.
Network properties and proxies for
multidimensional recovery
We hypothesized that recovery of different
forest attributes would be positively corre-
lated, because recovery of certain attributes
(e.g., biomass) can facilitate that of others (e.g.,
soil C) or can only occur when recovery of
other attributes occurs simultaneously. We
performed network analyses of relative re-
covery of multiple attributes after 20 years.
The first network analysis was based on
pairwise correlations among all 12 attributes
and showed that recovery of attributes oc-
curred in parallel (Fig. 4A), with the highest
expected influence (i.e., many links with
other attributes) for SC, followed by the three
structural attributes and soil C (Fig. 4C). This
was also confirmed by the results of a principal
components analysis, which showed similar
associations between recovery of different
forest attributes (fig. S3).
The second network analysis was based on
partial correlations—i.e., accounting for the
variation explained by other attributes—thus
showing independent, causal links between
attributes. We focused on seven forest attrib-
utes that were measured at most study sites
(N = 74). We found two clusters of attributes
whose recovery is likely to be causally linked
(Fig. 4B). First, recovery of the three struc-
tural attributes was highly connected, because
large trees (Dmax) lead to large SH and con-
tribute disproportionally to forest biomass
(AGB). Forests with more biomass also have a
more complex structure. Second, recovery in
SR and SD were positively linked, because both
increase during succession when species arrive.
When the analysis was repeated for the
43 sites for which SC was also included, a
third cluster emerged that showed that re-
covery in SC, WD, and nitrogen fixation were
linked (fig. S4). This may be explained by
the fact that succession in SC is underlain by
concomitant changes in WD because in wet
forests, pioneer species with low WD are re-
placed by OGF species with high WD, whereas
in dry forests, pioneer species with high WD
are replaced by OGF species with low WD (24).
The clustering of forest attributes into mul-
tiple groups suggests that recovery of different
forest attributes is shaped by different drivers
or processes. For example, recovery of biodi-
versity attributes may be driven by the land-
scape context (17), land-use history, and the
availability of seed trees and dispersal vectors,
whereas recovery of structural attributes may
be driven by resource availability [i.e., water
availability, soil fertility (16), and remnant trees].
We hypothesized that AGB would be the
best predictor of multidimensional recovery
because ecosystem processes and flux rates
strongly depend upon the amount of vegeta-
tion. Instead, we found that recovery of Dmax
had the highest influence (Fig. 4D), indicating
that it is strongly linked with other forest
structural attributes. The largest tree can be
one that regenerated during succession or is
a remnant from previous land use. Remnant
trees may act as nuclei of forest regeneration
(33) and kickstart succession (34) because they
improve microclimate and soil conditions, at-
tract frugivorous seed dispersers (35), and favor
regeneration of old-growth species. This net-
work structure of forest recovery may be af-
fected by future climate change, but at this
stage, we cannot predict how.
)
%
(
y
r
e
v
o
c
e
r
e
v
l
i
t
a
e
R
100
90
80
BD
SLA
C
WD
60
N
40
NF
SH
DMAX
20
SD
SR
SC
AGB
0
Attribute group
Soil
Diversity
Function
Structure
0
20
40
60
Time (y)
80
100
120
Fig. 2. Predicted relative recovery trajectories over time for 12 forest attributes. The attributes are
related to soil (brown), plant functioning (purple), structure (green), and diversity (turquoise). Relative
recovery is expressed for each attribute as the similarity (in percentage) between the predicted age-
dependent SF value and the OGF value. For some attributes, absolute values increase over time (e.g., AGB),
whereas for other attributes, the absolute values generally decrease over time (e.g., BD) (compare
Fig. 1A). Here, we show similarity with OGF values, which, by definition, increases over time (compare
Fig. 1B). Succession often starts with some remnant trees and soil legacies, and some attributes (SLA and
WD) can never be zero, which explains why most attributes do not start at zero (see main text). Dashed
lines indicate relative recovery at 20 years, recovery at 40 years, and RT until 90% recovery toward OGF values
(see Fig. 3). Recovery trajectories are across-site median values were estimated by using Bayesian models for
each attribute (see supplementary text S1). C, soil C; N, soil N; NF, proportional basal area of nitrogen-fixing
species; SLA, community-weighted mean SLA; WD, community-weighted mean WD.
Poorter et al., Science 374, 1370–1376 (2021)
10 December 2021
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
To assess and monitor ecosystem recovery,
we aimed to identify indicators that change
continuously during succession and that are
closely correlated with recovery of other eco-
system properties and functions. To be ope-
rational, the indicators should be easy to
measure, scalable, and cost-effective to imple-
ment and use (36). To serve simultaneously
as resilience indicators, they should be slowly
changing variables that underlie ecosystem
capacity to recover (5). We identified three
clusters of forest attributes, related to struc-
ture, species diversity, and SC (Fig. 4D and
fig. S4B). Dmax and SH are robust indicators
of structural recovery; they take, respectively,
5 and 2.5 decades to recover, have a central
position in the multidimensional recovery net-
work, and can easily be measured and moni-
tored, either in the field or by remote sensing
(37). Recovery of Dmax is strongly linked to
recovery of AGB, which is more time consum-
ing to measure, but Dmax is weakly linked to
recovery of biodiversity attributes (Fig. 4B).
Hence, SR should be used as an additional
indicator, because it is more closely linked
to the recovery of other biodiversity attrib-
utes. A large number of species also ensures
that there is a large diversity in species re-
sponses to environmental conditions, which
increases the adaptive capacity of ecosys-
tems to deal with environmental change (38).
When these slower indicators have recovered,
faster attributes, such as soil and plant func-
tioning, will have recovered as well (Fig. 3D).
Resilience
We assessed the resilience of forest attributes
on the basis of the relative starting value at
agricultural abandonment (t0; resistance) and
subsequent recovery rates (l) during second-
ary succession. Aboveground attributes such as
structure and diversity had low starting values
because of the nearly complete removal of
woody vegetation for agricultural use, whereas
soil attributes had high starting values because
of belowground legacies (Fig. 3A). We found
that resistance and recovery were positively
correlated [correlation coefficient (r) = 0.78,
P = 0.0026; fig. S5), which partially explains
why some attributes recover quickly and others
slowly. All 12 attributes recovered close to
their predisturbance values within ~120 years
(Figs. 2 and 3D), which is notably fast given
that tropical forests are complex in terms of
structure, SR, evenness, and plant interactions
(6). Fast forest recovery during secondary suc-
cession can be explained by the many legacies
and the relatively productive, warm, and wet
conditions of most study sites. We show that
tropical forests are resilient to agricultural use,
provided that agricultural use has not been
too long, intense (39), or extensive and that
there is sufficient forest in the surrounding
area to provide seeds (40). Average RTs of
the 12 attributes varied from <1 to 12 decades.
To assess ecological resilience, the attributes
After 0 y
Soil
Diversity
Function
Structure
After 20 y
B
)
%
(
y
r
e
v
o
c
e
r
e
v
i
t
a
l
e
R
100
80
60
40
20
0
BD SLA C WD
N
NF
SH DMAX SD SR SC AGB
SLA BD
N
C WD SH NF
SR DMAX SD SC AGB
D
120
)
y
(
y
r
e
v
o
c
e
r
%
0
9
o
t
e
m
T
i
100
75
50
25
0
A
)
%
(
y
r
e
v
o
c
e
r
e
v
i
t
a
l
e
R
C
)
%
,
y
0
−
y
0
2
(
y
r
e
v
o
c
e
r
e
v
i
t
a
l
e
r
n
i
e
g
n
a
h
C
100
80
60
40
20
0
100
80
60
40
20
0
SR SH SD DMAX N
NF AGB SC
C WD SLA BD
BD SLA C WD
N
NF
SH SR DMAX SD AGB SC
Fig. 3. Relative recovery of SF attributes. (A) Recovery at land abandonment
(0 year). (B) Recovery at 20 years. (C) Difference in recovery between 0 and
20 years (R20y − R0y). (D) RT until 90% recovery toward OGF values. Relative
recovery is expressed as the similarity (in percentage) between the SF value and
the OGF value. Medians and 95% credible intervals are based on site predictions;
N = 21 sites for soil attributes, 31 for SLA, 46 for SC, and 77 for all the other
attributes. The credible intervals nearly correspond to the range of observations,
and in (D), it has been truncated to 120 years to increase resolution. The
attributes are related to soil, plant functioning, structure, and diversity and are
ranked from fast (left) to slow (right).
Poorter et al., Science 374, 1370–1376 (2021)
10 December 2021
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
and time frames that are considered [compare
(5)] are therefore crucial, because soil scien-
tists would consider forests to be highly re-
silient on the basis of soil legacies, whereas
conservationists would consider forests to
have low resilience on the basis of the slow
recovery of SC.
Applied implications
SFs cover large areas and provide multiple
services to local and global stakeholders (3).
Their fast multidimensional recovery has im-
portant implications for ecosystem restoration,
climate change mitigation, and biodiversity
conservation. Rapid recovery of plant func-
tioning suggests that restoration of ecosystem
functioning (such as productivity) should be
similarly rapid, because it is underpinned by
the traits used in this study (41). Rapid soil N
and C recovery indicate that natural regrowth
provides an inexpensive, nature-based solu-
tion to restore the fertility of agricultural lands.
Rapid recovery of soil C is crucial for climate
change mitigation. The soil C pool exceeds
that of biomass and is more persistent because
it is less affected by aboveground disturbances,
such as fire and clearing. Rapid recovery in SR
means that SFs form an important biodiversity
reservoir in human-modified landscapes and
should be conserved (17). This will also lead to
greater connectivity for plants and animals in
fragmented, human-modified landscapes.
Although SFs show, on average, a rapid re-
covery, there is also substantial variation across
the study region (see credibility intervals in
Fig. 3), which indicates that some areas may
show arrested succession because of a lack of
seed sources or dominance of invasive grass,
ferns, or woody species. Under such condi-
tions, management practices for assisted natu-
ral regeneration—such as weeding, controlling
invasive species, enrichment planting, and
the establishment of ecological corridors—are
needed to safeguard multidimensional recovery.
Given the local and global importance of SFs
and their substantial R20y (on average, 78%;
range, 33 to 100%; Fig. 3B), we urge the em-
brace of SFs as a low-cost, nature-based solu-
tion to meet the United Nations’ Sustainable
Development goals and the United Nations’
Decade on Ecosystem Restoration goals (42).
Enabling policies combined with careful land-
scape planning should identify areas where
SFs are best conserved and provide the most
co-benefits while minimizing socioecological
conflicts [e.g., (3, 4)]. SFs should feature pro-
minently in restoration portfolios, where older
SFs and OGFs are conserved, severely de-
graded areas are actively restored, and young
regrowth is protected from deforestation. For
example, SFs can only deliver their full con-
servation potential if they are conserved for
a sufficient amount of time, such that tree
species can attain reproductive maturity and
maintain viable populations (43). In addition,
substantial gains are made when young,
20-year-old SFs are conserved for 20 years
more, because AGB, SR, and similarity with
OGFs increase by 15 to 22% (fig. S1C). All OGFs
Fig. 4. Network analysis for relative
recovery of attributes after
20 years. (A to D) The networks are
based on Pearson’s correlations for
12 attributes (left panels) and partial
correlations for seven attributes
(right panels). (A) and (B) show
connectivity among SF attributes.
(C) and (D) show expected influence
of individual attributes on the network.
The correlation network to the left
indicates how attributes are associated
with one another and uses, for each
pairwise correlation, the maximum
number of sites possible (N = 17 to
77; table S2). The partial correlation
network indicates the direct links
between two attributes, independent
from others, and is based on a subset
of attributes available for most sites
(N = 74 sites). In (A) and (B), the line
thickness indicates the strength of
the (partial) correlation, and lines indi-
cate significant pairwise correlations
(i.e., with a 95% confidence interval
that does not overlap with zero). The
expected influence is the sum of the
partial correlations between the target
attribute and other attributes. Edge
weights and expected influence were
estimated from 10,000 bootstraps
of the empirical network, and the
bootstraps were used to calculate the
95% confidence intervals displayed
in (C) and (D). Soil attributes were
based on a smaller sample size (N = 21)
and therefore had wider credibility intervals. The spinglass algorithm identified three clusters in (A) (AGB-Dmax-SH-C-N, BD-NF, and SC-SR-SD-SLA-WD) and
two clusters in (B) (AGB-Dmax-SH and SR-SD). For the partial network, the correlation stability coefficient was 0.43 for edge weights and 0.51 for expected influence.
Attributes are colored according to their category: soil (brown), plant functioning (purple), structure (green), and diversity (turquoise).
Poorter et al., Science 374, 1370–1376 (2021)
10 December 2021
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
should be conserved because little remains;
they harbor many distinctive OGF species and
provide seed sources and dispersers to assure
landscape resilience (17, 40). To monitor pas-
sive and active restoration success, we re-
commend using simple indicators in different
phases of succession, in which SH can be used
for the first 25 years (Fig. 3D) and Dmax and SR
in the following 25 years.
Conclusions
Our analysis shows that tropical forests and
their soils are highly resilient because all at-
tributes recover within 12 decades after low- to
moderate-intensity land use. Recovery of soil
attributes (<1 decade) and plant functional at-
tributes (<2.5 decades) is very fast, followed by
recovery of structure and diversity (2.5 to 6 dec-
ades), and recovery of AGB and SC is slowest
(12 decades). Network analysis shows that re-
covery is multidimensional, with three clus-
ters of attributes related to structure, SR, and
SC. Monitoring of forest restoration could use
Dmax, SH, and SR as complementary indica-
tors of multidimensional recovery.
RE FE RENCES AND N OT ES
1.
Food and Agricultural Organization of the United Nations
(FAO), “Global forest resources assessment 2020 –
Key findings” (FAO, 2020).
2. R. L. Chazdon et al., Sci. Adv. 2, e1501639 (2016).
3. P. H. S. Brancalion et al., Sci. Adv. 5, eaav3223 (2019).
4. B. B. N. Strassburg et al., Nat. Ecol. Evol. 3, 62–70 (2019).
5. S. Carpenter, B. Walker, J. M. Anderies, N. Abel, Ecosystems 4,
765–781 (2001).
6. S. L. Pimm, Nature 307, 321–326 (1984).
7. B. Ayala-Orozco et al., Land Degrad. Dev. 29, 315–325 (2018).
J. P. Grime, Plant Strategies, Vegetation Processes, and
8.
Ecosystem Properties (Wiley, 2006).
9. N. Norden et al., Proc. Natl. Acad. Sci. U.S.A. 112, 8013–8018
(2015).
10. D. Tilman, Plant Strategies and the Dynamics and Structure of
Plant Communities (Monographs in Population Biology,
Princeton Univ. Press, 1988).
11. J. H. Connell, Science 199, 1302–1310 (1978).
12. C. D. Oliver, B. C. Larson, Forest Stand Dynamics: Updated
Edition (Wiley, 1996).
13. L. R. Walker, D. A. Wardle, R. D. Bardgett, B. D. Clarkson,
J. Ecol. 98, 725–736 (2010).
14. R. L. Chazdon, Second Growth: The Promise of Tropical Forest
Regeneration in an Age of Deforestation (Univ. Chicago
Press, 2014).
15. M. Gei et al., Nat. Ecol. Evol. 2, 1104–1111 (2018).
16. L. Poorter et al., Nature 530, 211–214 (2016).
17. D. M. A. Rozendaal et al., Sci. Adv. 5, eaau3114 (2019).
18. See supplementary materials.
19. E. Veldkamp, M. Schmidt, J. S. Powers, M. D. Corre, Nat. Rev.
Earth Environ. 1, 590–605 (2020).
20. P. A. Martin, A. C. Newton, J. M. Bullock, Proc. Biol. Sci. 280,
20132236 (2013).
21. P. H. Nye, D. J. Greenland, The Soil Under Shifting Cultivation
(Technical Communication No. 51, Commonwealth Agricultural
Bureau, 1960).
22. L. Poorter, M. van de Plassche, S. Willems, R. G. A. Boot, Plant
Biol. 6, 746–754 (2004).
23. S. A. Batterman et al., Nature 502, 224–227 (2013).
24. L. Poorter et al., Nat. Ecol. Evol. 3, 928–934 (2019).
25. B. Finegan, Trends Ecol. Evol. 11, 119–124 (1996).
26. M. F. Barros et al., For. Ecol. Manage. 482, 118881 (2021).
27. M. Lohbeck, L. Poorter, M. Martínez-Ramos, F. Bongers,
Ecology 96, 1242–1252 (2015).
28. J. W. F. Slik et al., Glob. Ecol. Biogeogr. 22, 1261–1271
(2013).
29. F. Bongers, L. Poorter, W. D. Hawthorne, D. Sheil, Ecol. Lett. 12,
798–805 (2009).
30. E. A. Johnson, K. Miyanishi, Ecol. Lett. 11, 419–431 (2008).
31. E. Lebrija-Trejos, J. A. Meave, L. Poorter, E. A. Pérez-García,
F. Bongers, Perspect. Plant Ecol. Evol. Syst. 12, 267–275
(2010).
32. R. L. Chazdon et al., Philos. Trans. R. Soc. London Ser. B 362,
273–289 (2007).
33. G. Yarranton, R. Morrison, J. Ecol. 62, 417–428 (1974).
34. A. E. N’Guessan et al., For. Ecol. Manage. 433, 325–331
(2019).
35. A. D. Manning, J. Fischer, D. B. Lindenmayer, Biol. Conserv. 132,
311–321 (2006).
36. T. H. Oliver et al., Trends Ecol. Evol. 30, 673–684 (2015).
37. T. Jucker et al., Glob. Change Biol. 23, 177–190 (2017).
38. B. Sakschewski et al., Nat. Clim. Change 6, 1032–1036
(2016).
39. C. C. Jakovac, M. Peña-Claros, T. W. Kuyper, F. Bongers,
J. Ecol. 103, 67–77 (2015).
40. V. Arroyo-Rodríguez et al., Biol. Rev. Camb. Philos. Soc. 92,
326–340 (2017).
41. L. Poorter et al., Glob. Ecol. Biogeogr. 26, 1423–1434
(2017).
42. UN Environment Assembly (UNEA), “Resolution 73/284:
United Nations Decade on Ecosystem Restoration
(2021–2030)” (UNEA, 2019); https://undocs.org/A/RES/73/
284.
43. M. van Breugel et al., PLOS ONE 8, e82433 (2013).
44. D. Craven, dylancraven/MultiDimensional Recovery, Zenodo
(2021); https://doi.org/10.5281/zenodo.4988848.
ACKN OWLED GMEN TS
This paper is a product of the 2ndFOR collaborative research
network on SFs (www.2ndFOR.org) and the sDiv working group
sUCCESS and is paper no. 7 of 2ndFOR. We thank the owners of
the SF sites for access to their forests, all the people who have
established and measured the plots, the institutions and funding
agencies that supported them (see below), and M. Aide for data
use. Funding: This research was supported by European Research
Council-ERC (Advanced Grant PANTROP 834775 to L.P.); German
Centre for Integrative Biodiversity Research (iDiv) Halle-Jena-Leipzig
(sDIV W7.20 sUCCESS to L.P., N.R., and M.v.B.) funded inter alia by the
Deutsche Forschungsgemeinschaft (DFG; FZT-118); Netherlands
Organisation for Scientific Research - NWO (ALW.OP241 to L.P., M.T.v.
d.S., and C.C.J.; ALWOP.457 to F.B. and R.Mu.; ALW 863.15.017 to
M.L.; and Veni.192.027 to M.T.v.d.S.); NWO-Fundação de Amparo á
Pesquisa do Estado de São Paulo 17418 (NEWFOR) to P.H.B., F.B., and
M.P.-C.; Agencia Nacional de Investigación y Desarrollo (FONDECYT
Regular No. 1201347) to D.C.; Conselho Nacional de Desenvolvimento
Científico e Tecnológico-CNPq (to A.C.F., I.C.G.V.; 308471/2017-2
to M.M.E.-S.; 308877/2019-5 to Y.R.F.N.; #312178/2019-0 to B.A.S.;
CNPq 308778-2017-0 to I.C.G.V.; CNPq to G.W.F.; Universal 01/2016 to
H.M.T.; 309659/2019-1 to S.C.M.; and SinBiose-REGENERA 442371/
2019-5 to C.C.J. and R.Me.); Corredor Biológico La Gamba
(COBIGA) to F.O.; Deutsche Forschungsgemeinschaft DFG (RU
1536/3-1 to N.R.); Fondo Mixto CONACYT-Gobierno del estado de
Yucatán (FOMIX YUC-2008-C06-108863 to J.M.D. and J.L.H.-S.);
Fundacão de Amparo à Pesquisa de Minas Gerais-FAPEMIG (to G.W.F.;
PPM-00627-16 to Y.R.F.N.; PPM-00726-16 to M.M.E.-S.; PPM-00623-16
to M.D.M.V.; and APQ-03348-16 to H.M.T.); Fundação de Amparo à
Pesquisa do Estado do Rio Grande do Sul-FAPERGS (2218–2551/ 12-2
to S.C.M.); Fundación Jardín Botánico de Medellín to A.I.; GEF/FONACIT
to P.M.V.; German Centre for Integrative Biodiversity Research (iDiv)
Halle-Jena-Leipzig, a research center of the German Research
Foundation (DFG – FZT 118) (iDiv-Flexpool grant nos. 34600967 and
34600970 to N.R. and S.K.); Herbario JAUM to A.I.; Rainforest
Luxemburg to F.O.; Secretaría de Educación Pública-Consejo Nacional
de Ciencia y Tecnología, Ciencia Básica (SEP-CONACYT 2015-255544
to P.B. and F.M.); Stichting Het Kronendak to H.M.T. and H.F.M.V.;
STRI, ForestGEO, Heising–Simons Foundation, HSBC Climate
Partnership, Stanley Motta, Small World Institute Fund, the Hoch family,
to J.S.H. and M.v.B.; Universidad de Antioquia to J.J.-M., Universidad
Nacional Autónoma de México, Programa de Apoyo a Proyectos de
Investigación e Innovación Tecnológica (DPAGA–PAPIIT IN218416,
DPAGA–PAPIIT IN217620 to J.A.M., E.A.P.-G. and R.Mu.; PAPIIT-UNAM
IN211417 to P.B. and F.M.); Rufford Small Grants 19426-2 to F.M.;
SENACYT Panama Grant (COL10-052) to D.H.D.; Tropenbos
Foundation to H.F.M.V.; US National Science Foundation (no. 9208031
to D.H.D. and S.J.D.; EAR-1360391 to M.v.B.; and Graduate Fellowship
to S.G.L.); Wageningen University and Research Interdisciplinary
Research and Education Fund (FOREFRONT program) to F.B. and
H.M.T.; and Yale-NUS College and MOE (through a startup grant and
grant IG16-LR004) to M.v.B. A.H.-J. was supported by the LICCI project,
funded by the European Research Council (FP7-771056-LICCI). This
work contributes to the “María de Maeztu” Programme for Units of
Excellence of the Spanish Ministry of Science and Innovation (CEX2019-
000940-M). Author contributions: L.P., B.H., D.C., C.C.J., M.T.v.d.S.,
L.A., F.B., C.E.F., R.L.C., S.K., J.A.M., R.Mu., N.N., N.R., and
M.v.B. conceived the idea; all authors but C.E.F., S.K., and N.R.
contributed data; B.H., D.C., C.C.J., M.T.v.d.S., and L.P. analyzed the
data; L.P. led the writing of the manuscript with the help of B.H.,
D.C., C.C.J., and M.T.v.d.S.; L.A., F.B., C.E.F., R.L.C., S.K., J.A.M., R.Mu.,
N.N., N.R., M.v.B., A.M.A.Z., B.A., J.L.A., P.B., P.H.S.B., E.N.B.,
H.d.F., D.H.D., G.D., S.J.D., J.M.D., S.M.D., A.C.F., C.E.F., B.F., A.H.-J.,
J.L.H.-S., P.H., A.H.-J., J.K., S.G.L., M.L., R.L.-C., M.M.-R., F.P.L.M.,
F.M., S.C.M., A.E.N., F.O., E.O.-M., E.A.P.-G., B.X.P., D.P., J.S.P.,
S.R.-B., D.M.A.R., J.R., M.T., H.M.T., E.V.d.S.B.S., H.F.M.V., H.v.d.W.,
P.M.V., and G.W.F. commented upon the results and the manuscript;
and all authors approved submission of the manuscript. Competing
interests: The authors declare no competing interests. Data and
materials availability: Data on relative recovery of 12 attributes and
the code are available at Zenodo (44).
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abh3629
Materials and Methods
Supplementary Text
Figs. S1 to S5
Tables S1 to S3
References (45–70)
10 March 2021; accepted 6 October 2021
10.1126/science.abh3629
Poorter et al., Science 374, 1370–1376 (2021)
10 December 2021
7 of 7
|
10.1126_science.abf8113
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
MICROBIOLOGY
A human apolipoprotein L with detergent-like
activity kills intracellular pathogens
Ryan G. Gaudet, Shiwei Zhu, Anushka Halder, Bae-Hoon Kim, Clinton J. Bradfield, Shuai Huang,
Dijin Xu, Agnieszka Mamiñska, Thanh Ngoc Nguyen, Michael Lazarou, Erdem Karatekin,
Kallol Gupta, John D. MacMicking*
INTRODUCTION: In the arms race between path-
ogen and host, infecting microbes often escape
extracellular defense mechanisms to exploit
the nutrient-rich intracellular environment
as a replicative niche. In humans, this is coun-
tered by the interferon-g (IFN-g) response, which
confers widespread pathogen resistance in
most nucleated cells through the transcrip-
tional induction of hundreds of interferon-
stimulated genes (ISGs) encoding putative
antimicrobial restriction factors. Remarkably,
despite the importance of IFN-g against all
taxonomic classes of intracellular pathogens,
many restriction factors elicited by this cyto-
kine remain to be characterized, as do their
molecular activities.
RATIONALE: Identified as the major human
macrophage-activating cytokine in 1983, IFN-g
in fact transcriptionally reprograms numer-
ous host cell types to eliminate infection. This
A
B
APOL3
includes nonimmune epithelial cell popula-
tions, which lack many traditional phagocytic
defenses ascribed to IFN-g stimulation, yet still
manage to mount protective cell-autonomous
immune responses. To find ISG effectors in-
volved in safeguarding mucosal and barrier
tissue types, we conducted a genome-wide
CRISPR-Cas9 screen in IFN-g–activated human
epithelial cells for their ability to restrict virulent
intracellular pathogens such as Salmonella
enterica serovar Typhimurium.
RESULTS: We identify the ISG apolipoprotein
L3 (APOL3) as a potent effector protein capable
of killing cytosol-invasive bacteria. The human
APOL family is a cluster of six genes that have
evolved rapidly under positive selection in
simian primates; however, aside from the
founding member APOL1, a secreted extra-
cellular protein that forms the trypanolytic
factor of human serum, the function of the
Human cell
Cytosol-invasive
bacterium
Bacterial targeting
Restriction
APOL3
ISGs
(GBP1)
Salmonella
Bacterial cytosol
Single-particle
analysis
APOL3
IM
OM
IFN-γ
transcriptional
program
APOL3 kills intracellular bacteria. (A) Negative-stain electron microscopy of recombinant APOL3 (bead) added to
Salmonella Typhimurium (periplasm pseudocolored yellow). Destruction of bacterial membrane (blue-bordered inset) is
triggered by APOL3 extracting lipid to form lipoproteins (burgundy-bordered inset). (B) Bacterial mutants (DwaaL) expressing
a truncated O-antigen permit passage of APOL3 through the outer membrane (OM) to the inner membrane (IM); this
passage inside cells is facilitated by synergizing ISG-encoded proteins such as GBP1 that co-target cytosol-exposed bacteria.
intracellular APOL family members is un-
known. Human cells genetically engineered
to lack APOL3 failed to control the replication
of multiple cytosol-invasive Gram-negative bacte-
ria after IFN-g activation. Such findings were
validated in primary human intestinal epithe-
lial cells, intestinal myofibroblasts, and venular
endothelium—all cellular targets not typically
considered part of the immune system. We
tracked APOL3 by live microscopy and found
that it rapidly relocated to cytosol-exposed bacte-
ria, whereas other APOL family members did
not. A combination of superresolution imaging,
bioengineered reporters, and cell-free reconsti-
tution revealed that when APOL3 targets path-
ogens inside IFN-g–activated cells, it inflicts
a lethal insult to the bacterial inner membrane
(IM). Here APOL3 synergizes with other ISG-
encoded proteins, including guanylate-binding
protein 1 (GBP1), that perturb the bacterial
O-antigen outer membrane (OM) permeabil-
ity barrier to allow APOL3 access to the IM
underneath. Using a panel of composition-
ally distinct liposome targets, we found that
APOL3 membranolytic activity toward micro-
bial rather than host endomembranes stemmed
from an ability to dissolve bacterial polyanio-
nic lipid substrates lacking cholesterol into
discoidal lipoprotein complexes; single-particle
cryo–electron microscopy found that these
complexes resembled apolipoprotein-scaffold
“nanodiscs.” Corroborating these findings in
live bacteria by native mass spectrometry, we
found that APOL3 transitioned from a par-
tially disordered lipid-free state to tightly folded
lipoprotein nanodiscs upon extracting lipid
from the IM—a process that resulted in rapid
death of the bacterium.
CONCLUSION: Detergents are highly effective
antimicrobials used to decontaminate surfaces
infected by deadly pathogens. Our results iden-
tify APOL3 as an IFN-g–stimulated host defense
protein that has evolved potent detergent-like
activity to bestow bactericidal protection in
the cytosol of human cells. APOL3 synergizes
with other host ISGs in a multipronged attack
against the double membrane of Gram-negative
bacteria—a formidable barrier that imparts re-
sistance to many classes of antibiotics. This
study reveals that antibacterial agents that
dismantle this barrier during infection nat-
urally exist inside human cells. That these agents
are encoded within the IFN-g–inducible defense
program reinforces the importance of this power-
ful antimicrobial network for cell-autonomous
immunity in humans.▪
The list of author affiliations is available in the full article online.
*Corresponding author. Email: john.macmicking@yale.edu
Cite this article as R. G. Gaudet et al., Science 373, eabf8113
(2021). DOI: 10.1126/science.abf8113
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.abf8113
Gaudet et al., Science 373, 296 (2021)
16 July 2021
1 of 1
RES EARCH
R E S E A R C H A R T I C L E
◥
MICROBIOLOGY
A human apolipoprotein L with detergent-like
activity kills intracellular pathogens
Ryan G. Gaudet1,2,3,4, Shiwei Zhu1,2,3,4, Anushka Halder5,6, Bae-Hoon Kim1,2,3,4†,
Clinton J. Bradfield1,2,3,4, Shuai Huang1,2,3,4, Dijin Xu1,2,3,4, Agnieszka Mamiñska1,2,3,4,
Thanh Ngoc Nguyen7, Michael Lazarou7, Erdem Karatekin5,8,9,10,
Kallol Gupta5,6, John D. MacMicking1,2,3,4*
Activation of cell-autonomous defense by the immune cytokine interferon-g (IFN-g) is critical to the
control of life-threatening infections in humans. IFN-g induces the expression of hundreds of host
proteins in all nucleated cells and tissues, yet many of these proteins remain uncharacterized. We
screened 19,050 human genes by CRISPR-Cas9 mutagenesis and identified IFN-g–induced apolipoprotein
L3 (APOL3) as a potent bactericidal agent protecting multiple non–immune barrier cell types against
infection. Canonical apolipoproteins typically solubilize mammalian lipids for extracellular transport;
APOL3 instead targeted cytosol-invasive bacteria to dissolve their anionic membranes into human-
bacterial lipoprotein nanodiscs detected by native mass spectrometry and visualized by single-particle
cryo–electron microscopy. Thus, humans have harnessed the detergent-like properties of
extracellular apolipoproteins to fashion an intracellular lysin, thereby endowing resident nonimmune
cells with a mechanism to achieve sterilizing immunity.
C ell-autonomous immunity operates across
all three domains of life to defend against
infection (1). This ancient form of host
defense protects against intracellular
pathogens through direct and indirect
effector mechanisms (1). In vertebrates, these
effector mechanisms can be mobilized by the
type II cytokine interferon-g (IFN-g), which
regulates the transcription of hundreds of
IFN-stimulated genes (ISGs) to help combat
bacteria, viruses, parasites, and fungi in a wide
variety of host cell types (2). Human popula-
tion genetics and animal models have firmly
established the importance of IFN-g signal-
ing in organismal defense (3, 4), yet few ISGs
with direct pathogen-neutralizing activity have
been characterized. This is especially true with-
in human mucosal or stromal cell lineages
that are historically viewed as separate from
the classical immune system. These cell lineages,
1Howard Hughes Medical Institute, Yale University School of
Medicine, New Haven, CT 06510, USA. 2Yale Systems
Biology Institute, West Haven, CT 06477, USA. 3Department
of Immunobiology, Yale University School of Medicine, New
Haven, CT 06510, USA. 4Department of Microbial
Pathogenesis, Yale University School of Medicine, New
Haven, CT 06510, USA. 5Yale Nanobiology Institute, West
Haven, CT 06477, USA. 6Department of Cell Biology, Yale
University School of Medicine, New Haven, CT 06510, USA.
7Department of Biochemistry and Molecular Biology, Monash
Biomedicine Discovery Institute, Monash University,
Melbourne 3800, Australia. 8Department of Cellular and
Molecular Physiology, Yale University School of Medicine,
New Haven, CT 06510, USA. 9Department of Molecular
Biophysics and Biochemistry, Yale University, New Haven, CT
06510, USA. 10Saints-Pères Paris Institute for the
Neurosciences, Centre National de la Recherche Scientifique
(CNRS), Université de Paris, F-75006 Paris, France.
*Corresponding author. Email: john.macmicking@yale.edu
†Present address: Rare Disease R&D Center, PRG Science and
Technology Co. Ltd., Busan, Republic of Korea.
known more for their role in shaping organ
architecture and creating tissue boundaries,
nonetheless mount protective responses when
appropriately instructed by activating sig-
nals such as IFN-g (5, 6). The mechanisms and
protein machineries involved in this nonclas-
sical or “structural” arm of immunity remain
poorly understood (6).
Discovery of human APOL3 as an
antibacterial ISG
We searched for new antimicrobial ISGs in
human epithelial (HeLa CCL2) cells, using
the virulent Gram-negative bacteria Salmo-
nella enterica serovar Typhimurium (Stm) as
an initial infection model. Here, bulk repli-
cation arises from a subpopulation (~10%) of
infected cells in which Stm escape their entry
vacuole to rapidly proliferate in the cytosol
and serve as a reservoir for dissemination (7).
Using fluorescence-activated cell sorting (FACS)
analysis, we found that IFN-g specifically con-
trolled this subpopulation, completely prevent-
ing the appearance of cells laden with cytosolic
hyper-replicating bacteria (HR epithelial cells)
without affecting slowly replicating Stm within
vacuolar (Lamp1+) compartments (SR epithelial
cells) (Fig. 1A). To identify the protective host
factors, we performed a genome-wide CRISPR-
Cas9 screen and retrieved single guide RNAs
(sgRNAs) selectively enriched in IFN-g–activated
HR cells failing to restrict Stm (Fig. 1B). ISGs
were simultaneously defined by RNA sequenc-
ing (RNA-seq) profiles from IFN-g–activated
versus unactivated Stm-infected HeLa cells.
Stringent significance thresholds (P < 0.001;
mRNA > 4-fold induced) identified two major
hits exclusive to IFN-g–activated cells (Fig. 1B
and table S1): the master IFN transcription fac-
tor STAT1 and the primate-specific apolipopro-
tein L family member APOL3, a gene whose
product we found to be robustly and specif-
ically induced by IFN-g (fig. S1A) but has not
previously been linked to bacterial infection.
Validating our results, two independent
CRISPR deletions of APOL3 (DAPOL3) ren-
dered IFN-g–primed cells unable to fully re-
strict Stm hyperreplication in the cytosol, a
deficiency restored by APOL3 cDNA comple-
mentation (Fig. 1C and fig. S1B). These defects
were not due to impaired bacterial uptake and
were only evident after cytokine priming (fig.
S1C). Forced expression of APOL3 in unprimed
cells did not have a significant effect on Stm
replication (fig. S1, D to F), indicating that
APOL3 is necessary but not itself sufficient
for bacterial control. APOL3 was required for
restriction of other cytosol-invasive bacteria
[Shigella flexneri, Burkholderia thailandensis,
or a hyper–cytosol-invasive Stm mutant (StmDsifA)
(8)] but not vacuole-residing bacteria such
as Salmonella Typhi (9) or an injectisome-
deficient Stm mutant (StmDinvA::pR1203) that
is unable to initiate vacuolar escape (Fig. 1D).
That S. flexneri was less susceptible to APOL3-
mediated restriction may hint at the presence
of resistance mechanisms for this professional
cytosol-dwelling human pathogen. APOL3 like-
wise operated in primary human intestinal epi-
thelium, intestinal myofibroblasts, and vascular
endothelium, where small interfering RNA
(siRNA) silencing of IFN-g–stimulated APOL3
expression led to a significant loss of Stm
control (Fig. 1E and fig. S2, A to C). Thus, APOL3
is an IFN-g–inducible restriction factor that
controls cytosol-invasive pathogens in human
tissue cells originating outside of the hema-
topoietic compartment.
APOL3 targets cytosol-invasive bacteria
The human APOL family is a cluster of six genes
that have evolved rapidly under positive se-
lection in primates (10). Aside from APOL1, a
secreted protein associating with high-density
lipoprotein (HDL) to form the trypanolytic
factor of human serum (11, 12), protective func-
tions for the remaining five family members
that are intracellular and lack a secretion
signal are unknown. We examined the entire
family and found that most APOL genes were
highly induced by IFN-g in a STAT1-dependent
manner (fig. S3A), yet only cells chromosomally
deficient in APOL3 failed to restrict Stm (fig.
S3B). Using live microscopic imaging to in-
vestigate its subcellular location, we found that
ectopically expressed APOL3 fused to monomeric
NeonGreen fluorescent protein (APOL3mnGFP),
but not mnGFP fused to the other family mem-
bers, rapidly relocated to Stm and proceeded
to “coat” bacteria over a ~20- to 45-min period
(Fig. 2A, fig. S3C, and movie S1). Such APOL3
Gaudet et al., Science 373, eabf8113 (2021)
16 July 2021
1 of 14
RES EARCH | R E S E A R C H A R T I C L E
A
104
103
102
101
)
o
t
u
a
(
3
-
L
F
1h
Uninfected
68.2
Infected
31.8
6h
6h + IFN-γ
Uninfected
65.3
SR
90.6
Infected
34.7
HR
9.4
Uninfected
69.9
SR
99.6
Infected
31.1
HR
0.4
102 103 104 105
102 103 104 105
102 103 104 105
StmGFP
Lamp1 StmGFP
Slow replicating (SR)
Hyper-replicating (HR)
)
h
6
(
n
o
i
t
l
a
u
p
o
P
SR
SR
HR
HR
i
10 μm
l
l
e
c
r
e
p
a
i
r
e
t
c
a
B
1000
245
100
10
5.5
1
SR
HR
HR vs. SR Enrichment (19,050 genes)
-IFN-γ
+IFN-γ
STAT1
APOL3
B
Human epithelia
lentiviral
CRISPR
library
- IFN-γ
+ IFN-γ
StmGFP
FACS
SR
HR
SR
HR
)
e
u
a
v
P
l
0
1
g
o
L
-
(
t
n
e
m
h
c
i
r
n
E
5
4
3
2
1
0
Enrichment of sgRNAs in
HR versus SR populations
-5
5
5
0
IFN-γ-induced gene expression (Log2)
10
-5
0
10
C
Single-cell Stm burden
Population Stm burden
−IFN-γ
+IFN-γ
11.8
Cytosol-invasive + IFN-γ
B. thailandensis
3
75
StmΔsifA
6h
18h
***
S. flexneri
**
8h
***
***
WT
ΔAPOL3
ΔSTAT1
***
50
25
0
**
2
1
0
Vacuole-confined + IFN-γ
S. Typhi
8h
***
ns
StmΔinvA::pR1203
4
3
8h
***
3
2
1
0
ns
2
1
0
Primary human intestinal fibroblasts (10h)
2.2
Brightfield
StmmScarlet
Overlay
Untreated
+ IFN-γ
WT
HR
9.9%
HR
0.9%
ΔAPOL3
11.2%
6.0%
ΔAPOL3
+ APOL3
9.7%
1.9%
104
103
102
101
104
103
102
101
104
103
102
101
101 102 103 104
101 102 103 104
StmGFP
)
o
t
u
a
(
3
-
L
F
30
20
10
0
30
20
10
0
30
l
)
d
o
f
(
n
o
i
t
20
a
c
10
i
l
p
e
r
14.3
m
S
t
0
0
2
4
Time (h)
6
)
T
W
:
l
e
r
(
d
a
o
l
l
a
i
r
e
t
c
a
B
20
15
10
5
0
D
E
-
γ
N
F
I
+
-
γ
N
F
I
−
siCtrl
siAPOL3
siCtrl
− IFN-γ siCtrl
+ IFN-γ siAPOL3
+ IFN-γ siCtrl
3000
2000
1000
l
l
e
w
r
e
p
i
c
o
f
R
H
ns
0
0
P
=
0
.
0
0
0
2
10
5
Time (h)
IFN-γ −
siCtrl
+
37 kD−
siA3
+
φ
APOL3
β−actin
HR Stm foci
t
e
s
n
I
10-35 μm
Fig. 1. Genome-wide identification of human APOL3 as an antibacterial
ISG. (A) FACS of HeLa cells infected with GFP-expressing Salmonella enterica
Typhimurium (StmGFP). HR and SR gates are percentages of infected cells.
Below, 3D confocal microscopy and calculated bacterial load per cell from each
population (mean ± SEM). (B) Genome-wide CRISPR-Cas9 screen schema
and gene-level enrichment scores in HR versus SR populations in the presence or
absence of IFN-g. Each gene is plotted against relative (fold) induction of its
mRNA in IFN-g–activated cells determined by RNA-seq. (C) StmGFP growth by
FACS at 6 hours (left) and gentamicin protection assays (right) in APOL3-
deficient HeLa cells (DAPOL3) genetically complemented with APOL3
(bottom row) or empty retroviral control (top two rows). Fold is given as
relative to 1-hour starting time point (input). (D) Increase in bacterial load
[relative to wild-type (WT) cells] recovered from APOL3- or STAT1-deficient
IFN-g–activated HeLa cells after the indicated time. **P < 0.01, ***P < 0.001
(one-way ANOVA); ns, not significant. (E) Human primary intestinal myofibro-
blasts treated with APOL3 siRNA or nontargeting scrambled control (siCtrl)
(immunoblot, bottom right) and infected with StmmScarlet. Shown are represen-
tative final micrographs (10 hours) and quantification of HR Stm (foci 10 to
35 mm) every hour (mean ± SD, n = 3, representative of two independent
experiments. In (C) and (D), data are means ± SEM from four independent
experiments and FACS plots representative of at least four independent
experiments. f denotes a nonspecific band. Scale bar, 75 mm.
coating occurred in unprimed cells, indicating
that cytokine activation was needed for APOL3
expression but not its relocation. Even so, once
APOL3 relocated to bacteria within IFN-g–
activated cells, it conferred antibacterial activity;
identification and mutation of a hydrophobic
patch (LAP137QSS) (fig. S3D) that prevented
bacterial targeting also abolished restriction
(Fig. 2B). Targeting was specific to cytosol-
exposed bacterium, because vacuole-confined
StmDinvA::pR1203 failed to recruit APOL3 unless
released into the cytosol with the lysosomotropic
agent L-leucyl-L-leucine methyl ester (LLOMe),
which restored both bacterial targeting and sub-
sequent restriction (Fig. 2C). Moreover, APOL3
targeted cytosolic S. flexneri, B. thailandensis,
and Listeria monocytogenes but not compart-
mentalized S. Typhi (fig. S4A), which is generally
in keeping with the susceptibility profiles of
these pathogens to APOL3 restriction. Notably,
sterile damage triggered by LLOMe was suffi-
cient to mobilize APOL3 to the lumen of rup-
tured LAMP1+ endolysosomes independent of
bacteria (fig. S4B and movie S2). It is therefore
likely that initial APOL3 targeting signals are
damage-associated molecular patterns (DAMPs)
induced by pathogenic bacteria when they
rupture their LAMP1+ vacuole to enter the cy-
tosol, akin to the defense protein Galectin-8 (13).
To assess the fate of APOL3-coated bacte-
ria, we engineered reporter strains to diag-
nose Stm fitness inside human cells. Stm
inner membrane (IM) integrity was tracked
via minD, a bacterial cell division protein that
loses its lateral membrane oscillatory behavior
when IM potential is perturbed (14). Loss of
both IM localization and oscillation (demar-
cated by minD aggregates) was significantly
elevated in APOL3-coated versus uncoated
Gaudet et al., Science 373, eabf8113 (2021)
16 July 2021
2 of 14
RES EARCH | R E S E A R C H A R T I C L E
APOL3mnGFP StmRFP
46:00
56:00
67:00
APOL3 WT
LAP137QSS
WT
ΔAPOL3
ΔAPOL3 + APOL3
ΔAPOL3 +
LAP137QSS
+IFN-γ
***
9
6
3
l
)
d
o
f
(
n
o
i
t
a
c
i
l
p
e
r
84:00
C
93:00
StmΔinvA::pR1203
−LLOMe
+LLOMe
APOL3
StmRFP
A
B
D
APOL3 StmRFP
m
S
t
0
0
3
Time (h)
6
Stm:minD-GFP IM reporter
Stm:minD-GFP
APOL3RFP
568 ApoL3
Functional IM
IM dysfunction
−IFN-γ
+IFN-γ
Bacterial coating
−
APOL3
(n = 466,450)
APOL3+ (n = 104,123)
Bacterial fitness
WT (n = 604)
ΔAPOL3
(n = 638)
***
)
%
(
s
e
t
a
g
e
r
g
g
a
20
15
10
5
0
i
D
n
m
h
t
i
w
m
S
t
E
***
ns
+IFN-γ
−IFN-γ
+IFN-γ
Stm de novo mRNA reporter
mCherry
infect arabinose
araB-
GFP
30 min
2.5 h
Stm:mCherry
Stm:araB-GFP
APOL3FLAG
Merged
o
i
t
a
r
y
r
r
e
h
C
m
P
F
G
/
P = 0.0817
APOL3−
APOL3+
P < 0.0001
−IFN-γ
+IFN-γ
3
2
1
0
IFN-γ −
−
+
+
)
%
(
d
e
t
e
g
r
a
t
m
S
t
l
)
d
o
f
(
n
o
i
t
a
c
i
l
p
e
r
3
0
2
1
R
p
:
:
A
v
n
Δ
i
m
S
t
30
20
10
−IFN-γ
+IFN-γ
0
APOL1
APOL1b
APOL2
APOL3
APOL4
APOL5
APOL6
150
100
50
0
0
−IFN-γ
+IFN-γ
WT
WT + LLOMe
ΔAPOL3
ΔAPOL3
+ LLOMe
***
3
6
Time (h)
0
3
Time (h)
6
F
−IFN-γ
+IFN-γ
APOL3
LPS
i
n
m
5
4
i
n
m
0
5
1
2D SIM
3D-SIM stack
3D-mid z
-
γ
N
F
I
i
+
n
m
0
5
1
)
%
(
m
S
t
APOL3 position:
Exterior
Interior
100
75
50
25
0
IFN-γ −
−+
+
45min 150min
APOL3GFP
IFN-γ
150'
IM
OM
Fig. 2. Human APOL3 targets and inflicts damage to cytosolic bacteria.
(A) APOL3mnGFP targeting StmRFP by live imaging in HeLa cells (movie S1).
Percentage of total Stm targeted by HA-tagged APOL family members (2 hours)
is shown at right. (B) StmRFP targeting and replication in IFN-g–primed
DAPOL3 cells complemented with the indicated APOL3 variant. (C) Deconvolved
wide-field images of APOL3mnGFP targeting vacuole-confined StmRFP
(StmDinvA::pR1203) with or without vacuole release with LLOMe; fold replication is
shown at right. (D) Inner membrane (IM) integrity as measured by minDmnGFP
aggregation within Stm in HeLa cells expressing APOL3RFP at 2 hours with
or without IFN-g. Quantification reflects aggregation in APOL3-coated versus
uncoated bacteria or total bacteria in WT versus D APOL3 cells via Fisher’s
exact test. (E) Arabinose-induced GFP in Stm targeted by APOL3FLAG in
HeLa cells with or without IFN-g. Maximal-intensity GFP/mCherry ratios are
shown (mean ± SD, n = 50). (F) Immunofluorescence and SIM of APOL3HA and
LPS on Stm with or without IFN-g at selected times. Mid-2D z-planes are
shown. Quantification of LPS penetrance (25 bacilli, mean ± SEM, n = 3)
and 3D surface rendering are shown below. Blue arrows indicate cryo-
immunogold EM staining of APOL3GFP in Stm-infected HeLa cells; OM, outer
membrane. Micrographs are representative of at least three independent
experiments. Data are means ± SEM [(A), (B), (C)] with significance by one-
way ANOVA at 6 hours. ***P < 0.001. Scale bars, 5 mm [(A), (B), (C), and (E)],
2 mm (D), 1 mm (F).
Stm and reversed in DAPOL3 cells treated
with IFN-g (Fig. 2D and movies S3 and S4).
Stm lacking the Cpx-driven IM repair path-
way (StmDcpxR:FRT) had exacerbated damage
and was more susceptible to APOL3-driven
restriction (fig. S5, A to C), underscoring the
importance of bacterial membrane repair for
resisting APOL3-dependent immunity. In
addition, a dual-reporter Stm strain respon-
sive to de novo arabinose-induced GFP ex-
pression became completely unresponsive to
this stimulus when targeted by APOL3 (Fig.
2E), indicating that these bacteria were com-
promised in their ability to respond transcrip-
tionally to external cues. These fitness defects
Gaudet et al., Science 373, eabf8113 (2021)
16 July 2021
3 of 14
RES EARCH | R E S E A R C H A R T I C L E
were accompanied by a marked transformation
of the APOL3 bacterial coat: Superresolution
structured illumination microscopy (SIM)
revealed that APOL3 breached the lipopoly-
saccharide (LPS) outer membrane (OM) to
penetrate the bacterial cytoplasm in IFN-g–
activated cells by 2.5 hours of infection (Fig. 2F
and movie S5). Cryo–immunoelectron micros-
copy confirmed this penetrance (Fig. 2F), with
clearance of APOL3-targeted bacteria within
IFN-g cells occurring shortly thereafter (fig. S5,
C and D, and movie S6).
Human APOL3 exhibits bactericidal activity
The above results suggested that APOL3 could
be directly exerting antibacterial effects. Crit-
ically, these effects were observed exclusively
on bacteria within IFN-g–activated human
cells, where, in addition to being susceptible to
OM penetration by APOL3, bacteria displayed
irregular staining for the LPS O-antigen (fig.
S5E), an essential component of the OM per-
meability barrier. Other ISGs could therefore
facilitate direct APOL3 bactericidal activity
by increasing OM permeability, which would
normally exclude such a large hydrophobic
agent. We generated recombinant rAPOL3
(15) to test this possibility directly (fig. S6A).
Cytosol-enriched Stm extracted from DAPOL3
IFN-g–activated cells were highly susceptible
to direct rAPOL3 killing, whereas Stm from
vacuoles, from unprimed cells, or grown in
broth were resistant despite equivalent APOL3
binding (Fig. 3A and fig. S6, B and C). Similar
results were obtained with Stm extracted from
APOL3-silenced, IFN-g–activated primary hu-
man intestinal myofibroblasts (fig. S6D). Live
imaging revealed that Stm released from IFN-
g–activated cells had sustained transient per-
turbations to their cell wall that enabled rAPOL3
to gain initial access to the OM (loss of peri-
plasm ssTorA-GFP) and to permeabilize the IM
(uptake of zombie-UV) (Fig. 3B). These events
were rapid, beginning 2 to 3 min after rAPOL3
exposure (fig. S6E and movies S7 to S9) and ex-
plain why IFN-g priming is required for APOL3
antibacterial activity inside human cells.
Stm sensitization in situ could be pheno-
copied in cell-free settings. Preexposure of Stm
to sublethal concentrations of five different
OM-destabilizing agents facilitated direct
ssTorA-GFP
568-rAPOL3
Zombie-UV
Merged
A
B
Stm from ΔAPOL3 cells
- IFN-γ
+ IFN-γ
)
e
t
a
s
y
a
d
%
i
l
(
y
t
i
l
i
i
b
a
v
m
S
t
100
75
50
25
100
75
50
25
Source of Stm
Lysis buffer
Vacuolar
Cytosolic
***
0
0.01 0.1 1 10
rAPOL3 (μM)
0
0.01 0.1 1 10
rAPOL3 (μM)
C
Dialysate
rAPOL3
Stm
100
Stm:ssTorA-GFP
OM
IM
GFP
ssTorA
ssTorA-GFP
Zombie-UV
100
50
t
m
S
+
3
O
P
A
f
o
%
IFN-γ
100
m
S
t
l
a
t
o
t
50
f
o
%
IFN-γ
rAPOL3
D
3
L
O
P
A
r
-
8
6
5
+
m
S
c
t
i
l
o
s
o
t
y
c
-
N
F
I
-
-
N
F
I
+
E
EDTA Lysozyme
LD50
(μM)
Low
High
)
t
u
p
n
i
%
(
y
t
i
l
i
b
a
V
i
75
50
25
0
***
***
***
***
***
***
***
***
***
EDTA
LB
C18G
PMBNP
Lysine20
E. coli
B.thailandensis
Hypotonic
L. monocytogenes
S. flexneri
G
rAPOL3 (0.15M KGl)
Cytosol
extracted
EDTA-
pulsed
3.2
>25 >25 >25
0.8
12.8 12.8 >25
0.02M KCl
1.6
6.4 >25 >25
0.15M KGl
12.8 >25 >25 >25
rAPOL3
hBD-2
RegIIIβ
ΔAH
rAPOL3
μH
Hydrophobicity
0.8
H
μ
0.0
AH1 AH2
AH3 AH4 AH5
Tm1
Tm2
i
1.0
0.0
y
t
i
c
b
o
h
p
o
r
d
y
H
1-333
ΔAH
ΔTm1
ΔTm2
ΔTms
79-176
179-333
0
50
Stm viability (%)
100
F
O-unit
O-unit
wzy
100
)
%
(
y
t
i
l
i
b
a
V
i
10
1
waaL
waaJ
waaI
waaG
hldE
GlcNAc Glc
Gal
Gal Glc
Hep Hep
Hep
Kdo
Kdo
Kdo
LipidA
Stm LPS structure
0.1
WT
ΔwaaL
Δwzy
ΔwaaJ
ΔwaaG
ΔwaaI
ΔhldEWT
Mock + Ni2+-Gold
E. coli ΔhldE
His6-rAPOL3 (2μM) + Ni2+-Gold
His6-rAPOL3 (10μM) + Ni2+-Gold
500 nm
200 nm
!""#$
inset
500 nm
inset
Stm
E. coli
Fig. 3. Human APOL3 has direct bactericidal activity. (A), Viability of Stm
extracted from DAPOL3 cells with or without IFN-g and exposed to rAPOL3
(3 hours). (B) Live imaging of cytosol-extracted Stm:ssTorA-GFP treated with
568-labeled rAPOL3 (5 mM) and membrane-impermeable Zombie-UV dye
with quantitation (n = 100). (C) Bacterial viability after pulsing with the
indicated agent followed by rAPOL3 exposure for 3 hours. (D) Median lethal
dose (LD50) at 3 hours after in vivo or ex vivo sensitization of Stm. (E) rAPOL3
domain analysis (10 mM) at 3 hours after incubation with EDTA-pulsed
Stm; Hydrophobicity and amphipathicity (mH) plot above. AH, amphipathic
helix; TM, transmembrane domain. (F) Sensitivity of Stm and E. coli LPS
truncation mutants to rAPOL3 (10 mM) in potassium gluconate (KGl). (G) EM
micrographs of E. coliDhldE exposed to His6-rAPOL3 for 5 min and detected
with 5-nm Ni2+-gold beads. Data are means ± SEM from three to five
independent experiments [(A), (B), (C), (E), and (F)] or are representative
of three independent experiments [(D) and (G)]. ***P < 0.001. Scale bar in
(B), 2 mm.
Gaudet et al., Science 373, eabf8113 (2021)
16 July 2021
4 of 14
RES EARCH | R E S E A R C H A R T I C L E
rAPOL3-induced killing in each case (Fig. 3C).
S. flexneri, B. thailandensis, Escherichia coli,
and L. monocytogenes were likewise sensi-
tized to rAPOL3 by small amounts of EDTA
or lysozyme (Fig. 3C). Relative to the bona fide
antimicrobial peptides human b-defensin–2
(hBD-2) and mouse RegIIIb (16, 17), rAPOL3
was more active on an equimolar basis by a
factor of 5 to 16, confirming its status as a
powerful antibacterial lysin (Fig. 3D). Puta-
tive transmembrane domains and amphi-
pathic helical (AH) repeats identified in silico
were required for APOL3 bactericidal activity
(Fig. 3E and fig. S7). In some cases, N-terminal
(rAPOL379–176) and C-terminal (rAPOL3179–333)
fragments harboring these motifs were more
toxic than the full-length protein, possibly be-
cause they are small enough to penetrate bac-
terial cell walls without prior damage, as seen
for L. monocytogenes (which lacks an OM) and
DH5a E. coli (possessing only a short O-antigen)
(fig. S8, A and B). However, these smaller frag-
ments were missing motifs for intracellular
trafficking and they could not fully target
bacterial pathogens inside human cells, re-
sulting in loss of killing activity. Only full-
length APOL3 could restore such activity in
situ (fig. S8, C to F).
A panel of Stm and E. coli mutants with
progressive O-antigen truncations mimicking
OM damage underscored its potency. Here,
all strains lacking a complete polymerized O-
antigen were directly killed by rAPOL3 in
cytosolic salt concentrations (Fig. 3F). This
suggests that potentiating agents inside host
cells need only to perturb the outer O-antigen
barrier to facilitate APOL3 killing because
lipid A and core sugars are vulnerable to its
attack. Notably, the trypanolytic human APOL1
ion-channel protein (12) failed to kill trun-
cated Stm mutants under these conditions
(fig. S9A), revealing mechanistic differences
with APOL3. Indeed, immuno–electron mi-
croscopy revealed that rAPOL3 localized to
large pores spanning both the OM and IM
before complete cell wall disintegration and
blebbing ensued at higher dosage (Fig. 3G and
fig. S9B). Biophysical measurements supported
EM analysis: Bacterial membrane depolariza-
tion coupled with loss of fluidity, IM integrity,
and cellular ATP in addition to cytosolic leak-
age all preceded bacteriolysis and were de-
pendent on OM permeabilization (fig. S9, C to
H). Thus, rAPOL3 exhibits potent membrano-
lytic activity upon weakening of the OM per-
meability barrier, as seen within IFN-g–activated
human cells.
APOL3 bactericidal activity
is facilitated by GBP1
We next considered the identity of host ISGs
that weaken the OM permeability barrier of
cytosolic bacteria for APOL3 killing. Galectin-8,
p62/SQSTM1, and guanylate-binding protein
1 (GBP1) are defense proteins that target cytosol-
invasive bacteria in human cells. Galectin-8
and p62/SQSTM1 restrict bacteria through
xenophagy (13, 18), whereas GBP1 belongs to a
family of IFN-g–inducible guanosine triphos-
phatases (GTPases) that establish signaling
platforms for cell-autonomous immunity and
inflammasome activation (19–21). In human
epithelium, GBP1 assists the LPS-responsive
caspase-4 inflammasome to initiate pyroptotic
cell death (22, 23). APOL3+ bacteria harbored
all of these cytosolic defense markers and their
proximal adaptors, yet they localized to differ-
ent microdomains on bacilli and targeting was
mutually independent, as shown by genetic
ablation of 16 different signaling nodes or com-
ponents, suggestive of parallel defense pathways
(fig. S10, A to C).
Notably, co-targeting was highest for GBP1,
which, in addition to recruiting caspase-4, has
also been reported to disrupt the O-antigen
barrier upon bacterial coating (24). We there-
fore considered that GBP1 may aid APOL3
killing by facilitating its penetration through
the OM. In support of this model, genetic re-
moval of GBP1 in DAPOL3 cells (DAPOL3/GBP1
double knockouts) rendered cytosol-extracted
Stm less vulnerable to killing by exogenous
rAPOL3 (Fig. 4A). In our in vitro system, treat-
ment of wild-type Stm with recombinant GBP1
(rGBP1) purified from human cells, which coated
bacteria in a GTP-dependent manner (Fig.
4B), was sufficient to sensitize wild-type Stm
to killing by rAPOL3 (Fig. 4C). This synergy
stemmed from the ability of rGBP1 to increase
bacterial OM permeability [measured by up-
take of the fluorescent dye NPN (N-phenyl-1-
naphthylamine)] and facilitate APOL3 disruption
of the IM (Sytox uptake), resulting in a loss of
cellular ATP (Fig. 4D). Cellular reconstitution
corroborated these findings. Forced expres-
sion of both APOL3 and GBP1 transgenes in
tandem (fig. S10D) was sufficient to confer
antibacterial protection even in unprimed
HeLa cells, partly mimicking the actions of
IFN-g (Fig. 4E). This agrees with SIM imaging
of the bacterial surface inside IFN-g–activated
cells where penetrating APOL3 foci were lo-
calized at regions of low LPS O-antigen inten-
sity; such regions were reduced in HeLa cells
doubly deleted for GBP1 and GBP2 (Fig. 4F).
Thus, human GBP1 is one host factor that can
sensitize bacteria to APOL3 killing, although
other factors likely exist.
We next examined the importance of this co-
operative behavior for host defense in DAPOL3/
GBP1 cells. Synergistic defects in IFN-g–dependent
bacterial restriction were observed in the dou-
ble knockout (Fig. 4G). This was independent
of the noncanonical inflammasome because
individual deletion of APOL3, unlike GBP1, did
not reduce Stm-triggered cell death, caspase-4
cleavage, or downstream interleukin (IL)–18
processing in IFN-g–activated cells (Fig. 4, G
and H). Thus, both genes operate in distinct
host defense pathways that intersect on in-
dividual bacteria; in the process of activating
the noncanonical inflammasome, GBP1 con-
tributes to the OM damage that renders Stm
vulnerable to direct killing by APOL3.
APOL3 dissolves bacterial membranes into
nanodisc-like lipoproteins
How does APOL3 discriminate and permea-
bilize bacterial membranes? We prepared
liposomes mimicking bacterial [80:20 phospha-
tidylethanolamine (PE):phosphatidylglycerol
(PG)] or mammalian [60:10:30 phosphatidyl-
choline (PC):phosphatidylserine (PS):choles-
terol] membrane composition and found the
former to be >10 times as sensitive to rAPOL3
permeabilization (Fig. 5A). A panel of com-
positionally distinct liposome targets revealed
that this selectivity arose from a preference for
acidic phospholipids naturally rich in bacterial
membranes that promoted cationic APOL3
binding and permeabilization, coupled with
an aversion to the eukaryote-restricted lipid
cholesterol, which inhibited lysis (fig. S11, A to
C). It also fitted the pH and salt dependency of
bacterial killing (fig. S11D). Liposome perme-
ation, like permeabilization of bacteria, re-
quired APOL3 TM regions and amphipathic
helices and liberated luminal reporter mol-
ecules as a function of protein concentration
irrespective of their size (0.158 to 40 kDa) or net
charge (0 to 3+) (fig. S11, E to G). Thus, rAPOL3
does not impose defined gating properties on
the type of molecule released.
Instead, we found that APOL3 dissolved an-
ionic liposomes into discoidal lipoprotein
complexes that we tracked by real-time optical
absorbance (Fig. 5, B and C) and visualized di-
rectly by negative-stain EM (Fig. 5D). Subse-
quent single-particle cryo–electron microscopy
(cryo-EM) captured these discoidal complexes
in their native state; three modular classes
arose from 431,789 particles sampled, but all
were ~45 Å in height, indicating a single lipid
bilayer bounded by different arrangements of
APOL3 (Fig. 5E and fig. S12). This bilayer con-
figuration is reminiscent of apolipoprotein-
scaffold nanodiscs and nascent HDL particles
(25). Indeed, liposome dissolution by rAPOL3
was accelerated by lipid packing defects in-
duced by temperature shift (fig. S11H) in a
manner that resembles other bona fide apo-
lipoproteins such as APOA-1 (26).
The critical solubilizing concentration (CSC)
for rAPOL3 was lower than that of a conven-
tional detergent (such as Triton X-100) by a
factor of ~40; this represents potent deter-
gent activity and is mechanistically distinct
from the antimicrobial activities of hBD-2
and APOL1, which did not trigger liposome
clarification (Fig. 5C). We used sub-CSC rAPOL3
concentrations together with unsaturated lipid
to slow the detergent-like activities of rAPOL3
Gaudet et al., Science 373, eabf8113 (2021)
16 July 2021
5 of 14
RES EARCH | R E S E A R C H A R T I C L E
Extracted Stm from
IFN-γ-activated cells
B
rGBP1RFP + GTP
ΔSTAT1
ΔAPOL3
ΔAPOL3 /GBP1
0
0.25 0.5 1
8
2
Log2 rAPOL3 (μM)
4
P<0.0001
P<0.0001
Stm
rGBP1RFP − GTP
16
Stm
Unprimed HeLa cells
(−) Dox
(+) Dox
ns
ns
ns
***
40
30
20
10
0
TRE
Empty
TRE
APOL3
TRE
GBP1
TRE
APOL3
GBP1
HeLa + IFN-γ + Stm (2h)
APOL3FLAG O-antigen
GBP1
merged
merged
A
100
l
)
e
t
a
s
a
d
%
i
50
(
y
t
i
l
i
i
b
a
v
m
S
t
E
l
)
d
o
f
(
h
6
n
o
i
t
a
c
i
l
p
e
r
m
S
t
F
WT
ΔGBP1/2
P2A
APOL3 OM
infiltration
*
WT ΔGBP1/2
m
S
t
r
e
p
i
c
o
f
g
n
i
t
a
r
t
e
n
e
P
6
4
2
0
-2
C
100
50
)
t
u
p
n
i
%
(
y
t
i
l
i
i
b
a
v
m
S
t
rGBP1
GTP
rAPOL3
G
WT
+ IFN-γ
r
u
o
h
6
mock + mock
rGBP1 + mock
mock + rAPOL3
rGBP1 + rAPOL3
Bacterial
ATP (nM)
0.46
2.67
15.42
D
y
t
i
l
i
b
a
e
m
r
e
p
M
O
)
e
c
n
e
s
e
r
o
u
l
f
N
P
N
(
107
106
105
104
104
105
106
107
IM permeability
(Sytox fluoresence)
***
−
−
+ + + +
−
−
+
+
+
−−−
+ +
−
+
ΔGBP1
ΔAPOL3
ΔAPOL3/GBP1
S
y
t
o
x
t
S
m
G
F
P
10
8
6
4
2
0
20
15
10
5
0
)
l
a
t
o
t
%
(
i
c
o
f
t
m
S
R
H
)
l
a
t
o
t
%
(
s
l
l
e
c
+
x
o
t
y
S
+IFN-γ
WT
ΔAPOL3
ΔGBP1
ΔAPOL3/GBP1
0
2
4
6
***
*
ns
*
ns
***
0
4
2
Time (h)
6
10
8
6
4
2
0
20
15
10
5
0
−IFN-γ
ns
H
Wildtype
+
+
+
−
−
−
IFN-γ
Stm
75
0
2
4
6
ns
37
50
37
20
15
37
0
2
4
Time (h)
6
ΔAPOL3/GBP1
ΔGBP1
ΔAPOL3
+
+
+
+
+
+
GBP1
φ
APOL3
C
A
S
P
4
I
L
-
1
8
(Pro)
p32
(Pro)
(IL-18)
β−actin
Fig. 4. Human GBP1 potentiates APOL3 bactericidal activity. (A) Viability
of cytosolic Stm extracted from the indicated HeLa cell genotype (+IFN-g)
and exposed to rAPOL3. (B to D) Stm treated with 5 mM recombinant human
GBP1RFP (rGBP1) (1 hour) with or without GTP and imaged by confocal
microscopy (B), washed, treated with 5 mM rAPOL3 for 1 hour, and then analyzed
by colony counting (C) or ATP (bubble size) in conjunction with both OM
and IM permeability by NPN or Sytox uptake, respectively (D). Bubbles represent
five independent experiments in technical duplicate. (E) Fold replication of
Stm in unprimed HeLa cells expressing doxycycline-inducible (TRE, Tet
response element) APOL3, GBP1, or both in tandem separated by the self-
cleavable P2A peptide. (F) IFN-g–activated HeLa cells expressing APOL3FLAG
infected with Stm for 2 hours were analyzed by SIM (mid-2D single-plane
imaging) after immunostaining for FLAG, GBP1, and Stm LPS (O-antigen).
Arrows indicate penetrating APOL3 foci and quantification (n = 50). (G and
H) IFN-g–activated HeLa cells infected with StmGFP were analyzed by live
microscopy for HR Stm (foci 10 to 35 mm) and cell death (Sytox+) (G) or
whole-cell lysates probed by immunoblot after 3 hours (H). Representative
images and immunoblots from one of three independent experiments and
quantification of total events per well (% total cells) are shown. Data are means ±
SEM from three to five independent experiments. Micrographs in [(B) and (F)]
are representative of three independent experiments. Statistics indicate
significance by one-way ANOVA [(C) and (G)], two-way ANOVA (E), unpaired
t test (F), or nonlinear regression (A). *P < 0.05, ***P < 0.001. Scale bars, 10 mm
(B), 1 mm (F), 80 mm (G).
sufficiently to capture lipid extraction and
blebbing by live confocal imaging of giant
unilamellar vesicles (GUVs) and negative-
stain EM of liposomes (fig. S13, A and B, and
movies S10 and S11). These effects were
completely lost with rAPOL3DAH that could
not undergo lipid-triggered a-helical conver-
sion for insertional disruption of the bilayer,
as measured by circular dichroism (fig. S13,
C and D). This solubilizing activity was re-
quired for bacterial restriction inside human
cells, as shown by mutating four Phe resi-
dues to Ser (4F-S) on the hydrophobic face
of AH-2 and -3 that structural homology
modeling predicted were critical to positive
curvature induction and membranolytic ac-
tivity (Fig. 5F and fig. S14, A and B). The APOL3
4F-S mutant exhibited decreased liposome
solvation and bacterial killing in vitro and
could not restore IFN-g–induced immunity
when reintroduced into DAPOL3 human cells,
despite continuing to traffic to bacteria (Fig. 5,
G and H).
Finally, we used high-energy native mass
spectrometry (nativeMS) (27, 28) to study
APOL3 structural dynamics during membrane
solubilization. In aqueous solution, rAPOL3
existed as a partly disordered “open” mono-
meric species with exposed surface area ac-
cumulating many positive charges (+16 to
+10) during ionization. Upon engaging lip-
osomes, rAPOL3 underwent a marked shift,
adopting a tightly folded “closed” conformer
of lower charge state (+7 to +5) associated with
multiple lipid adducts, confirming lipoprotein
particle assembly (Fig. 6A). This same confor-
mational change occurred with live bacteria.
Here, open rAPOL3 monomers converted to
closed monomers and dimers (44% of soluble
Gaudet et al., Science 373, eabf8113 (2021)
16 July 2021
6 of 14
RES EARCH | R E S E A R C H A R T I C L E
A
100
Tx-100
protein
50
LD50= 0.27 µM
LD50= 4.1 µM
)
%
l
(
e
s
a
e
e
r
n
e
c
a
C
l
i
Liposome:
bacterial
rAPOL3
mock
mammalian
rAPOL3
mock
0
0 200 400 600 800
1000
Time (sec)
B
e
c
n
a
b
r
o
s
b
a
e
m
o
s
o
p
L
i
1.0
0.8
0.6
0.4
0.2
0.0
DMPC
DMPC:DMPG
(3:1)
DMPG
mock
rAPOL3
0 10 20 30
10 20 30
10 20 30
Time (min)
C
e
c
n
a
b
r
o
s
b
a
e
m
o
s
o
p
L
i
E
DMPG liposomes (30 min)
rAPOL3
rAPOL1
rAPOL3 ΔAH
hBD2
TX-100
CSC ~
6.5 μM
CMC ~ 250 μM
1.0
0.8
0.6
0.4
0.2
0.0
10-2
10-1
1
10 102 103 104
Molarity (μM)
D
F
DMPC:DMPG liposomes
−rAPOL3
+rAPOL3
APOL3 lipoproteins
236,356
70,381
48,204
50 nm
G
Lipid solubilization
StmΔwzy killing
50 nm
H
i
s
d
e
102 Å
10 nm
45 Å
t
o
p
142 Å
model
APOL3
scaffold
lipid core
In situ Stm targeting (2h)
In situ Stm restriction
e
c
n
a
b
r
o
s
b
a
i
e
m
o
s
o
p
L
G
P
M
D
0.75
0.50
0.25
0.00
P = 0.002
10-6
10-5
10-4
10-3
10-2
10-1
Ctrl
WT
4F S
rAPOL3
Ctrl
WT
4F S
rAPOL3
APOL3 WT
LPS
5 μm
APOL3 4F S LPS
)
d
o
l
f
(
h
6
n
o
i
t
a
c
i
l
p
e
r
+IFN-γ
P = 0.007
8
6
4
2
m
S
t
0
WT
ΔAPOL3
pAPOL3
pA3 4F S
+
−
−
−
− −
+ +
+
−
− −
−
+
−
+
AH2/AH3 bundle
R191
K196
K73
R78
R200
F195 F81
F85
F77
APOL3
Homology prediction
Fig. 5. APOL3 dissolves anionic membranes into lipoprotein nanodiscs.
(A) Calcein leakage from “bacterial” (80:20 DOPE:DOPG) or “mammalian”
(60:10:30 DOPC/DOPS/cholesterol) liposomes (500 mM lipid) exposed to
rAPOL3 (500 nM). Vertical dashed line indicates dosage yielding 50% dye
release (LD50) in 200 s. (B and C) Turbidity of liposomes treated with rAPOL3
or indicated reagent over time (B) or after 30 min (C). (D) Negative-stain
EM of liposomes before and after addition of rAPOL3 for 30 min as in
(B). (E) Single-particle cryo-EM reconstruction of APOL3 lipoprotein nano-
discs. Isosurface representation of top three particle classes (number of
particles) is shown, with space-constrained model below. Thickness is
equivalent to a single DMPC or DMPG bilayer (45 Å). (F) Phyre2 structural
homology prediction. Inset indicates arrangement of amphipathic helices
(AH) 2 and 3, with four Phe (F) residues on the interior hydrophobic face
highlighted in yellow and exterior-facing acidic residues Arg (R) and Lys (K)
highlighted in red. (G) Liposome turbidity and viability of StmDwzy treated
with wild-type or mutant rAPOL3. The four Phe residues depicted in (F) were
mutated to Ser (S). (H) Complementation of DAPOL3 HeLa cells with
the indicated APOL3HA genotype evaluated for Stm targeting (left) and IFN-g–
dependent restriction (right). Data are means ± SEM from three or four
independent experiments [(B), (G), and (H)] or are representative of
three or more independent experiments [(A), (C), and (D)]. Statistics indicate
significance by one-way ANOVA.
rAPOL3) upon killing bacteria and were only
evident at a sufficiently high mass spectral col-
lisional activation (HCD) energy to trigger col-
lapse of the nanodisc and release the rAPOL3
scaffold (Fig. 6B) (29). EM analysis of recov-
ered rAPOL3 from these same fractions con-
firmed the transition from lipid-free to discoidal
lipoprotein form (Fig. 6C). Thus, APOL3 under-
goes marked structural changes to extract lipids
to form discoidal bacterial-human hybrid lipo-
protein complexes during killing.
Discussion
The double membranes surrounding Gram-
negative bacteria make them exceptionally
difficult to kill. Consequently, combination ther-
apies that use OM-permeabilizing agents to
facilitate passage of larger, more effective
antibiotics have emerged as a promising treat-
ment option (30–32). Our results show that
humans have evolved an analogous strat-
egy for cellular self-defense, with the natural
detergent-like effector APOL3 being given
access to the bacterial IM by other synergiz-
ing ISGs, including GBP1, that help to per-
meabilize the OM. Once inside the bacterium,
APOL3 exerts broad-spectrum membranolytic
activity through solubilization of the IM into
discoidal lipoprotein complexes. This mode of
killing appears to differ from that of canonical
extracellular antimicrobial proteins (AMPs),
which tend to form proteinaceous pores or in-
duce local membrane dysfunction (33). Thus,
APOL3 may have arisen as a host adaptation
to support intracellular killing specifically,
given that both the ionic strength and the
divalent cation concentration of the human
cytosol are incompatible with the activity of
many AMPs (34).
Our biochemical studies found that APOL3
targets anionic lipids that are highly enriched
in bacterial membranes (35). In contrast, cho-
lesterol, which is present exclusively in eukary-
otic membranes, inhibits APOL3 membranolytic
activity. This selectivity may aid discrimination
between self and non-self lipid structures,
as seen for the cytolytic T cell antimicrobial
protein granulysin, which similarly targets
Gaudet et al., Science 373, eabf8113 (2021)
16 July 2021
7 of 14
RES EARCH | R E S E A R C H A R T I C L E
A
Aqueous rAPOL3
rAPOL3 conformers
B
rAPOL3 + E. coliΔhldE
Quantification
C
rAPOL3 purified from supernatant
100
e
c
n
a
d
n
u
b
A
0
100
e
c
n
a
d
n
u
b
A
0
e
c
n
a
d
n
u
b
A
0
14+
11+
15+
10+
9+
8+
100
12+
13+
y
g
r
e
n
e
D
C
H
e
c
n
a
d
n
u
b
A
0
150eV
14+
15+
13+
“open”
+ + +
+
+
+
+
+
+
+
+
+
+
lipid extraction
+
lipid bilayer
charge state: High
12+
monomer
dimer
11+
“closed”
+
+
13+
12+
100
+
Low
0eV
+
+
monomer
dimer
Mock
Bacteria
Mock
50
40
30
20
10
0
3
L
O
P
A
r
d
e
s
o
c
%
l
HCD: Low High
50 nm
10 nm
Bacteria
rAPOL3 + liposomes
6+
7+
9+ 8+
10+
11+
12+
13+
6+
+1L
100
e
c
n
a
d
n
u
b
A
e
v
i
t
a
e
R
l
0
+2L
+3L
7000
5+
+4L
m/z
7400
4000
6000
m/z
8000
10000
11+
10+
6+/12+
/
9+
7+/14+
/
5+/10+
/
14+
8+
15+
13+
9+
4000
6000
m/z
11+
8000
10000
50 nm
Fig. 6. APOL3 extracts bacterial lipid to form lipoproteins during killing.
(A) Conformational analysis by native mass spectrometry (nativeMS) of rAPOL3
in ammonium acetate buffer (aqueous) or after incubation with DMPC/DMPG
liposomes for 30 min. Inset shows that satellite peaks correspond to successive
lipid (L) adducts. Schematic indicates the two observed charge states: lipid-free
“open” or lipid-bound “closed” monomers (single circle) or dimers (doublet).
(B) NativeMS spectra of soluble rAPOL3 after incubating with live E. coliDhldE for
1 hour. Collisional activation energy (HCD) was set to 0 eV (top) or 150 eV
(bottom). Inset shows nativeMS quantification of “closed” APOL3 conformers
before (mock) and after treatment of bacteria and analyzed at the indicated
HCD energy. (C) rAPOL3 was incubated with live E. coliDhldE as in (B), purified
from the supernatant by Ni-NTA pull-down and analyzed by negative-stain
electron microscopy. Data from [(A) to (C)] are representative of three
independent experiments.
cholesterol-poor anionic membranes (36–38).
Notably, organelles such as mitochondria
also contain some of the same anionic lipids
found in bacteria (cardiolipin, PG), whereas
endoplasmic reticulum membranes exhibit
naturally low cholesterol content (39). Thus,
additional host factors probably help to re-
strain or localize APOL3 activity in order to
prevent damage to cellular host structures. In
this regard, an APOL3 loss-of-function variant
shows signs of recent positive selection in
Africans (40), which suggests that its potent
membranolytic properties could be detrimental
in certain modern human populations if spu-
rious activation contributes to pro-inflammatory
disease. Future studies may identify other
APOL3 mutations in susceptible individuals
and may determine whether they come with
a fitness cost.
Examination of immune or microbial stimuli
eliciting APOL3 expression found that type II
IFN was the major trigger versus type I IFN
(IFN-a, IFN-b), tumor necrosis factor (TNF)–a,
IL-1b, or LPS as a Toll-like receptor 4 ligand.
Hence, it principally operates as part of the
IFN-g–inducible defense program in human
cells (2). This program enlists other IFN-g–
inducible defense factors, including the GBP
family, that have emerged as central orchestra-
tors of cell-autonomous immunity to intracel-
lular bacterial pathogens (19–21). Cooperation
between APOL3 and GBP1 was evident in loss-of-
function, gain-of-function, and cell-free experi-
mental systems. Notably, however, convergence
of these proteins on the surface of Gram-negative
bacteria yielded bifurcating outcomes. GBP1-
mediated damage to the bacterial OM not only
allowed APOL3 access to the IM for eventual
killing but also activated human caspase-4 and
IL-18 (18, 19). In contrast, APOL3 did not trig-
ger the noncanonical caspase-4 inflammasome
pathway, instead conferring protection through
direct bactericidal activity. These data reveal
distinct functional roles for these two IFN-g–
induced defense proteins within the “interfer-
ome” signature (2).
Our results elucidate the role of human
APOL3 as a potent bactericidal agent deployed
within nonimmune cells to combat cytosolic
pathogens. Our findings reinforce the growing
appreciation for the contributions made by
cells outside of the hematopoietic compart-
ment toward IFN-g–induced host resistance
(4, 5, 41). Our work also reveals the involve-
ment of lipoproteins in intracellular killing
in humans, adding to immune functions de-
scribed earlier for insects that use serum lipo-
proteins as a form of systemic, extracellular
defense (42, 43). Thus, although human APOL3
is considered a young gene that evolved rela-
tively recently (~33 million years ago) (10),
membrane solubilization may itself be an
ancient bactericidal mechanism that appears
to have been harnessed for sterilizing intra-
cellular immunity in primates.
Materials and methods
Plasmids, antibodies, and reagents
CRISPR deletions were generated using pX459.
Complementary DNAs (cDNAs) for the human
APOL genes were amplified from HeLa cells
and verified by sequencing. APOL3 isoform 2
was chosen, as this is considered the most
commonly expressed isoform (15). cDNAs were
inserted into the retroviral plasmid pMSCV-
puro with an N-terminal hemagglutinin (HA)
tag or N-terminal EGFP for stable expression in
complementation and certain imaging experi-
ments. For doxycycline-inducible expression,
APOL3 or GBP1 cDNA was cloned into MCS1
or MCS2, respectively, of pCW57-MCS1-2A-
MCS2 (Addgene) and transductants obtained
by selection in puromycin. For live and high-
resolution imaging, pmNeonGreen-C1 or pCMV-
3XFLAG encoding APOL3 was used. N-terminal
tags were used for all experiments. For recom-
binant protein expression, cDNA was inserted
into a modified pET28a vector containing an
N-terminal 6×-His tag followed by a precision
protease cleavage site. Truncation mutants were
generated by PCR with overlapping primers
flanking deletion sites. Point mutations were
inserted using a single mutated primer and
Phusion polymerase (NEB). The DAH variant of
APOL3 was obtained as a Geneblock from IDT.
For qPCR, RNA was isolated using an RNeasy
Mini Kit (Qiagen) and converted to cDNA using
PrimeScript RT master mix (Takara). Ampli-
fication was done using PowerUp SYBR Green
master mix (ThermoFisher) on a QuantStudio-5
Real Time PCR system with gene-specific primers
and analyzed using the 2-DDCt method with
GAPDH as the housekeeping gene. For labeling
of bacteria, pFPV25.1 (Addgene) encoding EGFP,
mCherry, RFP, or mScarlet was used. The dual
transcriptional reporter pFcCGi encoding a
constitutive mCherry and PBAD-GFP has been
described previously (44, 45) and was ob-
tained from Addgene. To generate the minD
reporter plasmid, minD was amplified from
Stm 1344 genomic DNA using the primers 5′-
atggcacgcattattgttgttacttcgggtaaaggg-3′ and
5′-ttatcctccgaacaggcgtttgaggaaacctttcttc-3′ and
cloned into pmNeonGreen (mnGFP)-C1 using
Gaudet et al., Science 373, eabf8113 (2021)
16 July 2021
8 of 14
RES EARCH | R E S E A R C H A R T I C L E
XhoI and EcoRI. The entire mnGFP-minD fusion
protein was cloned into the PBAD-GFP po-
sition (XbaI and SphiI) of pFcCGi, creating
pFcCmNmDi. To free the red channel for cer-
tain microscopy experiments, mCherry was re-
moved by overlapping PCR, creating pFmNmDi.
Plasmids were transformed into electrocom-
petent Stm and selected with carbenicillin
(100 mg/ml). To induce expression of mnGFP-
minD, overnight Stm was subcultured 1/33
and grown for 3.5 hours in the presence of
carbenicillin (50 mg/ml) and 0.2% L-(+)-arabinose
(Sigma) before infection. To express IM-anchored
GFP, the TorA signal sequence (ssTorA) of Stm
was ligated onto the N terminus of EGFP in
pFPV25.1 using overlapping primers. This se-
quence directs properly folded EGFP through
the twin arginine transporter pathway and
then serves as a peripheral membrane anchor
on the periplasmic face of the IM (46).
Antibodies used were anti-Flag M2 (Sigma),
anti-HA (16B12 Biolegend), anti-APOL3 (ab154869),
anti-Salmonella O Group B antiserum (BD),
anti–b-actin (ab6276), anti-Lamp1 (PA1-654A,
ThermoFisher), anti-p62 (bd 610832), anti-
Galectin 8 (sc-377133), anti-GBP1 (sc-53857),
anti-COX2 (sc-1747), anti-Mx2 (sc-271527), anti–
IL-18 (PM014; MBL), anti-IFITM3 (Proteintech;
11714-1-AP), and anti–Caspase-4 (clone 4B9;
Enzo). All lipids were purchased from Avanti
Polar Lipids Inc. Fluorescein-labeled dextrans
were from Sigma Aldrich. TbCl3+ and dipicolinic
acid (DPA) were from Biotium. Calcein was from
Life Technologies. L-Leucyl-L-leucine methyl
ester (LLOMe) hydrochloride was from Cayman
Chemicals. Zombie-UV was from Biolegend.
Bacterial strains
Bacterial strains were kindly provided by the
following groups: Salmonella enterica serovar
Typhimurium (Stm) strain 1344 and injectosome
deficient StmDinvA::pR1203 (J. Galan); Listeria mono-
cytogenes 140203S (H. Agaisse); Shigella flexneri
strain M90T (F. Randow); Stm 1344DcpxR:FRT
(J. Vogel) (47); E. coli DH5aDhldE (S. Gray-Owen)
(48) Stm UK-1 wild-type, Dwzy, DwaaL, DwaaJ,
DwaaI, and DwaaG (R. Curtiss III) (49).
Burkholderia thailandensis strain 700388 and
S. enterica serovar Typhi strain 700931 were
purchased from ATCC.
Bacterial infections
For Stm infections, overnight bacterial cul-
tures were diluted 1:33 in fresh LB, grown for
3 hours before being washed once in PBS, and
used to infect HeLa cells at 80% confluence
with an MOI of 5 unless otherwise indicated.
Plates were centrifuged for 10 min at 1000g
and incubated for 30 min at 37°C to allow
invasion. Extracellular bacteria were killed by
replacing media with fresh DMEM containing
gentamicin (100 mg/ml) for 30 min. Cells were
washed three times and incubated with gen-
tamicin (20 mg/ml) for the duration. To enu-
merate live bacteria, cells were lysed in PBS +
0.5% Triton X-100 and serial dilutions plated
on LB agar. To estimate bacterial load based
on GFP intensity, StmGFP was used and cells
were trypsinized and fixed in 4% PFA (Santa
Cruz) for 15 min. After washing, cells were re-
suspended in PBS and GFP fluorescence deter-
mined on a FACSAria (BD). Analysis was done
using Flowjo. To release bacteria from vacuoles,
LLOMe (600 mM) and Z-VAD-FMK (20 mM)
were added after 2 hours and incubated for the
duration. For Shigella infections, overnight
cultures were diluted 1:100 in tryptic soy broth
(BD) and growth to OD600 = 0.5 before infect-
ing HeLa cells at 80% confluence at MOI of 50.
Cells were then processed as for Stm infections.
For S. Typhi infections, overnight cultures were
diluted 1:20 in LB + 0.3 M NaCl, grown to
OD600 = 1.0 and processed as per Stm. For
B. thailandensis infections, overnight cultures
were diluted 1:10 and grown to OD600 = 1.0
(5 × 108 cfu/ml). Bacteria were washed once
in PBS and added to cells at MOI of 200,
centrifuged at 1000g for 10 min, and left for
1 hour at 37°C. Cells were rinsed and incubated
for 24 hours in complete medium containing
kanamycin (1 mg/ml). For L. monocytogenes
infections, bacteria were grown overnight at
30°C in brain heart infusion broth (BHI; BD)
and adjusted to OD600 = 1.0. Bacteria were
washed and used to infect HeLa cells at MOI of
50 following the protocol outlined for Shigella.
Cell culture and transfection
HeLa (CCL2) and 293T cells were purchased
from ATCC. Cells were grown in DMEM sup-
plemented with 10% (v/v) heat-inactivated
fetal bovine serum (FBS) at 37°C in a 5% CO2
incubator. Autophagy-deficient Penta-KO and
Hexa-KO cell lines have been described pre-
viously (50, 51). HUVECs from a single donor
were obtained from LONZA (CC-2517) and
maintained in EBM Basal Medium with growth
factors and used prior to passage 10. Primary
intestinal epithelial cells (CC-2931) and intes-
tinal myofibroblasts (CC-2902) were from
LONZA and maintained in SmGM Medium
with growth factors. Epithelial cells were main-
tained at 33°C as per manufacturer’s instruc-
tions and thawed directly from frozen into
96-well plates and used on days 5 to 7. All cells
were maintained in antibiotic-free media.
Lentiviral (LentiCrisprV2) or retroviral (pMSCV-
puro) transductions were done by incubat-
ing dilutions of 0.45 mm–filtered supernatants
from transfected 293T cells with polybrene
(8 mg/ml) for 24 hours. For selection of stable
transductants, puromycin (1 mg/ml) was in-
cluded. For transient transfections, TransIT-LT1
(MIRUS) was used according to manufacturer’s
instructions. To minimize toxicity in micros-
copy experiments, 200 ng of DNA was trans-
fected per 24-well cover slip. To generate stable
gene knockouts, sgRNAs were cloned into
pX459 (Addgene) per established protocols.
2-4 sgRNAs (table S2) targeting each gene
(200 ng total DNA) were transfected in 24-well
plates for 24 hours, followed by selection with
puromycin (1 mg/ml) for 48 hours. Surviving
cells were expanded into media lacking puro-
mycin for 48 hours, then subjected to limiting
dilution to obtain single colonies. Colonies were
screened first by PCR, then by Western blot, and
when appropriate the genotype of each posi-
tive clone was determined by Sanger sequenc-
ing. For siRNA knockdown, ON-TARGETplus
Human APOL3 siRNA smartpool (Dharma-
con) or nontargeting control were transfected
(20 nM) with Dharmafect 1 transfection reagent
for 48 hours as per manufacturer’s instructions.
When required, HeLa cells were stimulated
with IFN-g (500 U/ml) and primary human cells
stimulated with IFN-g (50 U/ml) for 18 hours.
Genome-wide screen
LentiCrisprV2 pooled library (GeCKO v2) was
a gift from F. Zhang (Addgene #1000000048)
(52). 25 × 106 HeLa cells were transduced on
four separate days and processed as individual
biological replicates (N = 4) throughout the
experiment. Transductants were selected with
puromycin (1 mg/ml) for 48 hours, then al-
lowed to rest without selection for an addi-
tional 48 hours. Surviving cells were split into
two groups (±IFN-g) and seeded into 6-well
plates at 80% confluency. After 24 hours, IFN-g
(500 U/ml; R&D systems) was added for an
additional 18 hours. Late log StmGFP was added
to each well at MOI of 20 and incubated for
6 hours as described for infections. Cells were
trypsinized and fixed in 4% PFA for 15 min
and analyzed within 48 hours on a FACSAria
(BD). Uninfected cells were gated out based on
comparisons with cells only control, and in-
fected cells gated into two groups, high GFP
(HR) or low GFP (SR), based on the maximal
difference obtained between the IFN-g–treated
and untreated control groups processed in
parallel. For each of the four biological repli-
cates, an average of 1 × 106 and 5 × 106 cells were
collected for the HR and SR gates, respectively.
After sorting, each group was pelleted and DNA
extracted using the PicoPure DNA Extraction
Kit (Applied Biosciences). sgRNA sequences
were amplified using Herculese II Phusion
DNA polymerase (Agilent) and amplicons pu-
rified from an 8% polyacrylamide Tris/Borate/
EDTA (TBE) gel. Amplicons were sequenced
using an Illumina HiSeq2500 (20 million reads
per sample) and analyzed using the MAGeCK
algorithm (53). Gene-level enrichment scores
(P value) for sgRNAs enriched in the HR versus
the SR populations were determined for both the
IFN-g–treated and untreated groups (table S1).
RNA sequencing
HeLa cells were stimulated with IFN-g (500 U/ml)
for 18 hours or were left untreated. Infections
Gaudet et al., Science 373, eabf8113 (2021)
16 July 2021
9 of 14
RES EARCH | R E S E A R C H A R T I C L E
with Stm were performed on individual tripli-
cates at MOI 5. After 5 hours, cells were washed
three times in sterile prewarmed DMEM, lysed
in 300 ml of RLT buffer, and processed via
RNeasy (Qiagen) kits per the manufacturer’s
protocol. RNA was checked for quality using
a denaturing MOPS gel and Nanodrop. 10 mg
of sample RNA was annealed to oligo-dT beads
followed by first- and second-strand cDNA syn-
thesis (Illumina). cDNA was then pair-end–
barcoded with Illumina Universal Adapters
and sequenced on a HiSeq4000 sequencer.
Data acquired were bin-sorted and de-barcoded
through a Sickle-Schythe Pipe. FASTA files were
then aligned to human reference genome
HsGRCh37 by Hisat and Tophat2, yielding
95% alignment. Annotated genes were quan-
titated via CufflinksV2 and subsequent data
were processed for visual display through
CummeRbund, ggplots, and ggplot2 in R.
Microscopy
HeLa cells were seeded on 12-mm high-
performance cover glass #1.5h (Thorlabs). For
live imaging, cells were seeded on four-well
chambers with #1.5 high-performance cover
glass (Cellvis). Cells were seeded 48 hours
prior to imaging to reach 80% confluency on
the day of infection and treated with IFN-g
(500 U/ml) where required for 18 to 24 hours
prior to imaging. To image bacterial infec-
tions, bacteria were added to cells as described
for infections at an MOI of 20. Images were
analyzed on a DeltaVision OMX SR microscopy
system (GE Healthcare) or a laser scanning
confocal model SP8 (Leica). For analysis of
relative positions of APOL3 and LPS, HA-
APOL3 was detected with anti-HA, followed
by anti-mouse Alexa-fluor 488 (ThermoFisher)
and LPS detected with anti-Stm LPS and Alexa-
fluor 568 or 647 anti-rabbit antibody. APOL3-
positive bacteria were identified and the
intensity profiles for LPS and APOL3 signal
were determined from linescans drawn on a
single plane slices of 2-mm z-stacks [following
alignment and structure illumination micros-
copy (SIM) reconstruction]. Images were ana-
lyzed using Fiji. Bacteria with APOL3 signal
intensity of >50% of max inside the LPS layer
were quantified at 45 min or 150 min post-
infection in the presence or absence of IFN-g.
For analysis of the bacterial response to arab-
inose post-infection, overnight Stm pFcCGi
were subcultured 1/33 in LB containing car-
benicillin (50 mg/ml) for 3 hours and used to
infect HeLa cells in 24-well plates, transfected
24 hours earlier with 200 ng pCMV-3XFLAG-
APOL3 in the presence or absence of IFN-g, at
an MOI of 20 for 15 min. Extracellular bacteria
were killed with gentamicin (100 mg/ml) for
30 min, and after three washes, replaced with
fresh media containing gentamicin (10 mg/ml)
and 0.4% L-(+)-arabinose for an additional
2 hours prior to fixation with 4% PFA (Santa
Cruz). 50 APOL3-positive or -negative bacteria
were selected at random from micrographs in
IFN-g–treated or nontreated conditions and
the maximum intensities of mCherry and GFP
signals for each bacterium were determined
using Fiji. For live high-content imaging, 5 ×
104 cells (primary intestinal epithelium; InEpC
and HeLa) or 1 × 104 primary intestinal fibro-
blasts (InMyoFib) were seeded into black
96-well clear-bottom plates 7 days (InEpC),
72 hours (InMyoFib), or 48 hours (HeLa) prior
to imaging. When required, 48 hours prior to
imaging, cells were transfected with control or
APOL3-targeting siRNA for 30 hours, then
treated with IFN-g (50 U/ml) [primary cells, or
HeLa cells (500 U/ml)]. Infections were done
with StmmScarlett or StmGFP at an MOI of 20 as
described above with the following modifica-
tions for InEpC: Bacteria were incubated for
60 min after centrifugation and 60 min after
addition of gentamicin (100 mg/ml) rather
than the usual 30 min. Cells were imaged live
at 1-hour intervals while incubating at 33°C
(InEpC) or 37°C (InMyoFib, HeLa) and 5% CO2
on an ImageXpress Pico Automated Imaging
System (Molecular Devices) at 10× magnifica-
tion. Analysis was done in an unbiased manner
using CellReporterXpress with the preconfig-
ured “Endocytosis” protocol template modi-
fied to identify the following intracellular Stm
populations: Based on preliminary experi-
ments in HeLa cells, SR Stm foci were defined
as objects between 1 and 10 mm and HR Stm
foci defined as objects between 10 and 35 mm.
A threshold intensity of 60 units above back-
ground was set to exclude nonspecific fluores-
cence. Where required, Sytox Orange was
included at 2 mM for the duration and dead
cells defined as objects 2 to 10 mm. Data was
normalized to total number of cells from
parallel wells stained with Hoechst 33342
(5 mg/ml; Invitrogen). For in vitro imaging,
cytosol-enriched Stm pFmNmDi or pFPV25.1-
ssTorA-GFP bacteria were harvested from in-
fected cells using Triton X-100 as described for
bacterial killing assays, washed three times,
and resuspended in Buffer A [50 mM MES pH
6.0, 100 mM potassium gluconate (KGl)] con-
taining 5 mM 568-labeled rAPOL3. For minD
imaging, bacteria were immediately imaged
live on 1.5% agarose pads. To image ssTorA-
GFP and simultaneously monitor membrane
integrity, Zombie-UV (1/200) was added for
5 min, bacteria were washed once, resuspended
in PBS, and imaged live on 1.5% agarose pads.
Purification of recombinant proteins
APOL proteins: Overnight cultures of BL21
(DE3) pLysS harboring the APOL3 or APOL1
expression plasmid (pET28a-6XHis-PP) were
grown in LB containing kanamycin (50 mg/ml)
and chloramphenicol (20 mg/ml). Cultures
were grown to OD600 = 0.7 in media without
chloramphenicol and induced with 1 mM IPTG
for 4 hours at 37°C. Cell pellets were lysed in
50 mM Tris pH 8.0, 5 mM EDTA and lysozyme
(100 mg/ml; Sigma) with sonication. The pres-
ence of lysozyme had no effect on the quality
or activity of the recovered protein but did
increase yield. Insoluble material was pelleted
at 20,000g and washed with lysis buffer con-
taining 0.5 M NaCl. Pellets were solubilized in
6 M guanidine hydrochloride, 50 mM potas-
sium phosphate pH 8.0, 1 mM TCEP, 10 mM
imidazole for 1 hour at room temperature with
gentle sonication and clarified by centrifuga-
tion at 40,000g for 30 min. Solubilized pro-
tein was affinity purified using Ni-NTA beads
(Qiagen) and dialyzed extensively (>4 buffer
changes of at least 1000 fold v/v) over 24 hours
into 20 mM acetic acid. For certain experi-
ments, the His-tag fusion protein was digested
with 3C protease (Genscript) in 50 mM MES
pH 6.0 overnight at 4°C. To remove the tag,
undigested protein and the protease, the re-
action and all insoluble precipitate was solu-
bilized in 6 M guanidine hydrochloride, 50 mM
potassium phosphate pH 8.0, 1 mM TCEP,
10 mM imidazole and incubated overnight
with fresh Ni-NTA beads. Flow through was
collected and dialyzed extensively into 20 mM
acetic acid. The absence of the His-tag was
confirmed using Western blot.
To refold the DAH variant, protein was first
dialyzed against 20 mM acetic acid, 250 mM
arginine hydrochloride (Sigma) for 6 hours
before dialysis into 20 mM acetic acid. All pu-
rified proteins were concentrated (>10 mg/ml)
and flash-frozen in liquid nitrogen and stored
at –80°C. Proteins were thawed once, as activity
decreased upon each freeze/thaw. rHis-APOL3
and rHis-cleaved APOL3 exhibited almost iden-
tical biological activity in both killing assays
and liposome leakage/solubilization assays.
However, rHis-APOL3 demonstrated increased
stability at higher pH, maintained stability at a
higher concentration, and was thus purified to
greater yield. Therefore, the His-tagged protein
was used when required. rHis-APOL1 was used
for bactericidal and liposome solubilization as-
says. Preliminary experiments revealed that
rAPOL3 protein stability was compromised in
high concentrations (>0.1 M) of traditional
chloride salts such as NaCl and KCl. Therefore,
a gluconate salt of potassium (KGl, the most
common ion in the cytosol) was included in
reaction buffers, unless otherwise indicated, to
maintain a near-physiological salt concentra-
tion. To prepare fluorescently labeled protein,
750 mg of protein was mixed with 333 mM AFDye
568 maleimide (Fluoroprobes) in a 300-ml re-
action volume. pH was adjusted to pH 7.0 with
1 M HEPES pH 7.4 and reaction incubated for
2 hours at room temperature. During this time,
~75% of rAPOL3 protein precipitated. Precipi-
tated protein was collected by centrifugation,
dissolved in 6 M guanidine hydrochloride,
50 mM potassium phosphate pH 8.0 and
Gaudet et al., Science 373, eabf8113 (2021)
16 July 2021
10 of 14
RES EARCH | R E S E A R C H A R T I C L E
dialyzed extensively (10 kDa MWCO) into
20 mM acetic acid to both refold the protein
and remove unincorporated dye.
Guanylate binding protein 1 (GBP1): The
coding sequence of human GBP1 (hGBP1) was
cloned into a customized vector to generate
pCMV-His10-Halo-HRV-mRFP-TEV-hGBP1.
HEK293f suspension cells (a gift from J. Rothman)
was maintained at a concentration of 0.4 ×
106 to 4 × 106 cells/ml in Expi293 expression
medium (ThermoFisher Scientific). 24 hours
prior to transfection, cells were seeded at a
concentration of 1.2 × 106 cells/ml. For trans-
fection, cells were harvested and resuspended
in fresh medium at a concentration of 2.5 × 106
cells/ml. Cells were transfected by adding pCMV-
His10-Halo-HRV-mRFP-TEV-hGBP1 to a final
concentration of 1 mg/ml in media containing
PEI at a concentration of 5 mg/ml. 24 hours
after transfection, cells were diluted 1:1 (v/v)
with fresh medium containing 4 mM valproic
acid and cultured for an additional 2 days. 2 ×
109 cells were harvested via centrifugation
(500g, 10 min), washed once in cold PBS, re-
suspended in lysis buffer (50 mM HEPES, pH
7.5, 500 mM NaCl, 1 mM MgCl2, 10% glycerol,
0.5% CHAPS, 1 mM TCEP) and lysed via son-
ication. Cells were cleared at 35,000g for 1 hour
at 4°C. Supernatant was collected and incu-
bated with 1 ml bed volume of HaloLink resin
(Promega) at 4°C overnight with gentle rota-
tion. The resin was sequentially washed twice
(10 min each) with wash buffer 1 (50 mM HEPES,
pH 7.5, 500 mM NaCl, 1 mM MgCl2, 10% gly-
cerol, 0.5% CHAPS), wash buffer 2 (50 mM
HEPES, pH 7.5, 1 M NaCl, 10% glycerol) followed
by wash buffer 1. To elute bound proteins, Halo
resin was resuspended in lysis buffer and di-
gested with homemade GST-HRV-His prote-
ase overnight at 4°C with gentle rotation. Resin
was pelleted and the HRV protease was re-
moved from the supernatant via Ni-NTA beads
by affinity chromatography (Qiagen). Flow-
through was collected, concentrated, and fur-
ther purified and buffer-exchanged via size
exclusion chromatography (Superdex 200 In-
crease; GE Healthcare) equilibrated with stor-
age buffer [20 mM HEPES (pH 7.5), 150 mM
NaCl, 1 mM MgCl2, 1 mM TCEP]. Fractions were
analyzed by SDS-PAGE, pooled, concentrated,
and flash-frozen in liquid nitrogen before stor-
ing at –80°C.
Recovery of APOL3 lipoprotein complexes:
To isolate rAPOL3-lipoprotein complexes from
bacteria, overnight E. coliDhldE was subcultured
1/20 and grown to OD600 = 0.5. 10 ml of bacteria
were washed and resuspended in Buffer A
containing 20 mM rHis-APOL3 for 2 hours at
30°C and insoluble material removed by cen-
trifugation. The pH of the supernatant was
adjusted to pH 7.2 with NaOH, and both NaCl
and imidazole were added to 200 mM and
10 mM final concentrations, respectively. The
solution was incubated with Ni-NTA beads at
room temperature for 1 hour, washed 5 times
with 10 mM Tris pH 7.2, 200 mM NaCl, and
20 mM imidazole, and eluted with 400 mM
imidazole in the same buffer. Eluates were
loaded directly onto glow-discharged copper
grids and examined by negative-stain electron
microscopy.
Bacterial killing assays
To isolate bacteria from different cellular com-
partments, DAPOL3 HeLa cells with or with-
out IFN-g (500 U/ml, 18 hours) were infected
with Stm at MOI of 20 as described for infec-
tions. After 45 min, cells were either lysed in
Buffer A containing 0.5% Triton X-100 (vacu-
olar population) or treated with 1 mM LLOMe
(Sigma) in the presence of cell death inhibitor
Z-VAD-FMK (R&D) for 2 hours before lysis
(cytosolic). Bacteria were then mixed with
rAPOL3 diluted in 20 mM acetic acid for 3 hours
and enumerated by colony counting after se-
rial dilution. To induce transient permeabili-
zation of the OM by EDTA, overnight bacterial
cultures were grown to OD600 = 0.5 in LB con-
taining 2 mM CaCl2 and 2 mM MgCl2 and
immediately washed twice in 0.1 M Tris pH
8.0, 0.75 M sucrose. Pellets were resuspended
in 350 ml of the same buffer and 700 ml of 1 mM
EDTA in H2O was added for 20 min. 50 ml of
0.5 M MgCl2 was added on ice for 5 min and
bacteria pelleted at 4°C. Bacteria were resus-
pended in Buffer A, and rAPOL3 diluted in
10 mM acetic acid was added to the indicated
concentration and incubated for 3 hours at
30°C. The final dialysate from rAPOL3 puri-
fications was used as the negative control.
Bacteria were enumerated by serial dilution
on LB agar. To induce transient OM perme-
abilization by other means, mid–log-phase
bacteria were washed and resuspended in
Buffer A supplemented with the indicated con-
centration of polymyxin B nanopeptide (PMBN;
Sigma), poly-L-lysine hydrobromide (avg. mw
20,000 Da; PKLB20, Alamanda Polymers), or
human platelet factor IV 18 (C18G; Eurogentec).
To induce hypotonic stress, bacteria were resus-
pended in Buffer A with 20 mM KCl substituted
for 100 mM KGl. Bacteria were incubated for
30 min before washing and addition of rAPOL3
in Buffer A. Bacteria incubated in LB or Buffer
A alone served as the nonpermeabilized con-
trol. After initial washes, a sample of bacteria
was plated on LB agar prior to addition of any
OM permeabilizing agent to define the in-
put. For LD50 assays, Stm were incubated for
3 hours with twofold serial dilutions of rAPOL3,
recombinant mouse RegIIIb (R&D), or human
b-defensin-2 (Biolegend) in either Buffer A or
low salt (20 mM KCl) buffer. The minimum
concentration required to kill >50% of Stm
was determined. For LPS-truncated mutants,
bacteria were grown to OD600 = 0.4 and 0.5 ml
washed once in Buffer A, resuspended to 0.5 ml
in Buffer A with 150 mM KGl and incubated
with 10 mM rAPOL3 or dialysate at 30°C for
3 hours with shaking at 250 rpm. Bacteria
were enumerated by serial dilution and colony
counting. For treatment with hGBP1, bacteria
were incubated for 1 hour with 5 mM hGBP1 in
50 mM HEPES pH 7.4, 150 mM NaCl, 5 mM
MgCl2, with or without 2 mM GTP. Bacteria
were then pelleted and resuspended in Buffer A
containing rAPOL3 and incubated at 37°C for
1 hour prior to plating.
Bacterial membrane and cytotoxicity assays
Permeability of the OM was determined using
the fluorescent dye NPN (Sigma) uptake assay.
Stm were prepared as described for the killing
assay, treated with permeabilizing agent at
the indicated concentration for 15 min, and
resuspended in Buffer A containing 10 mM NPN.
rAPOL3 or dialysate was added and incubated
for 15 min. Fluorescence (Ft15) was recorded by
SpectraMax i3X plate reader; lEx = 350 nm
and lEm = 420 nm. NPN uptake after 15 min
was calculated as (t) (%) = (Ft15 – Ft0) × 100/
(Ft100 – Ft0), where the fluorescence from un-
treated Stm was defined as Ft0 and in the pres-
ence of 5 mM EDTA and lysozyme (10 mg/ml)
as Ft100. Permeability of the IM was determined
by Sytox orange (ThermoFisher) or propidium
iodide (PI; Sigma) uptake. Stm were prepared
as described for the killing assay and treated
with 10 mM rAPOL3 or dialysate in Buffer A for
15 min (static) or for the indicated time (time
course). PI was included at 50 mM and fluores-
cence measured using a SpectraMax i3X plate
reader; lEx = 535 nm and lEm = 620 nm. PI
uptake at each time point was calculated as (t)
(%) = (Ft – Ft0) × 100/(Ft100 – Ft0). Fluorescence
from untreated Stm was defined as Ft0 and in
the presence of polymyxin B (25 mg/ml) and
0.2% SDS as Ft100. Sytox orange was included at
0.2 mM and fluorescence determined as de-
scribed for PI uptake with lEx = 545 nm and
lEm = 570 nm. IM potential of 5 × 107 Stm
was determined using the BacLight bacterial
membrane potential kit (ThermoFisher) fol-
lowing the manufacturer’s protocol. Stm were
treated as described for the killing assay, then
incubated with 5 mM CCCP, 10 mM rAPOL3
wild-type, DAH mutant, or equal volume of dial-
ysate for 1.5 hours before addition of DiOC2(3)
for 30 min. Samples were analyzed on a FACSAria
(BD). The ratio of red to green fluorescence
provides an indication of membrane potential
and was calculated for each treatment by di-
viding the mean fluorescence intensity (MFI)
for the red channel (FL-2) by the MFI for the
green channel (FL-1) after gating bacteria by
forward and side scatter. Bacterial ATP con-
tent was determined using BacTiter-Glo micro-
bial cell viability assay (Promega) following the
manufacturer’s protocol. To measure mem-
brane fluidity, mid-log E. coliDhldE were washed
and resuspended in PBS containing 0.2% glu-
cose and 10 mM Laurdan (Cayman Chemicals)
Gaudet et al., Science 373, eabf8113 (2021)
16 July 2021
11 of 14
RES EARCH | R E S E A R C H A R T I C L E
and incubated for 10 min. Bacteria were washed
3× with PBS containing 0.2% glucose and 1%
dimethylformamide (DMF). 500 nM rAPOL3
was added and fluorescence at two wavelengths
recorded over time using a SpectraMax i3X
plate reader: (i) lEx = 350 nm and lEm = 460 nm;
(ii) lEx = 350 nm and lEm = 500 nm. Background
fluorescence was subtracted from each value
and generalized polarization calculated as
GP = (I460 – I500)/(I460 + I500) where I is the
normalized intensity at each lEm wavelength.
For treatment with rGBP1, bacteria were in-
cubated for 1 hour with 5 mM rGBP1 (or mock)
in 50 mM HEPES pH 7.4, 150 mM NaCl, 5 mM
MgCl2, 2 mM GTP. Bacteria were then pelleted
and resuspended in Buffer A supplemented
with 10 mM NPN and 0.2 mM Sytox orange
containing 5 mM rAPOL3 (or mock) for 1 hour
before fluorescence at two wavelengths was
simultaneously measured using the Spectra-
Max i3X plate reader as above. Background
fluorescence of NPN and Sytox orange in
buffer alone was subtracted from each value.
ATP content was then determined as described
above.
Liposome preparation
Phospholipids were dissolved in chloroform
and mixed in a glass vial. Solvent was evap-
orated under nitrogen and dried overnight in
a vacuum. For liposome leakage assays, lipid
film was hydrated in Buffer A. For liposome
solubilization and nativeMS experiments, lip-
ids were hydrated with 20 mM ammonium
acetate. Lipids were solubilized with continual
vortexing followed by five freeze/thaw cycles.
Liposomes were generated by extrusion through
a 0.1-mm polycarbonate filter (Avanti Polar
Lipids Inc.) 30 times using a mini-extruder
device (Avanti Polar Lipids Inc.). To generate
calcein-encapsulated liposomes,
lipid was
hydrated in 50 mM MES pH 6.0, 20 mM
potassium gluconate, and 80 mM calcein. For
Tb3+-encapsulated liposomes, lipids were hy-
drated with 50 mM MES pH 6.0, 35 mM KGl,
50 mM sodium citrate, and 15 mM TbCl3. Non-
encapsulated calcein or Tb3+ was removed
using Illustra Microspin G50 columns (GE
Healthcare). For dextran liposomes, hydra-
tion was done with Buffer A and indicated
FITC-Dextran (2 mg/ml). Nonincorporated
dextran was removed by buffer exchange with
a centrifugal filter device (Amicon Ultra-15
100K MWCO, Millipore). All liposomes were
used within 24 hours.
Liposome binding, leakage, and turbidity assays
To measure liposome binding, indicated lipo-
somes (2.5 mM lipid) were incubated for 20 min
with recombinant APOL3 (rAPOL3; 1 mM pro-
tein such that liposomes were not completely
dissolved) for 20 min in Buffer A. Samples were
centrifuged for 1 hour at 120,000g in a Beckman
Optima XE-100 Ultracentrifuge at 4°C. Super-
natant (S) was collected and the pellet (P)
washed twice with 700 ml of incubation buffer
and resuspended in the same volume as su-
pernatant. Samples were analyzed by SDS
page followed by Coomassie blue staining.
To measure leakage, liposomes of the indi-
cated composition (500 mM lipid) were mixed
with rAPOL3 (500 nM or the indicated con-
centration) in Buffer A. To measure Tb3+ ef-
flux, 15 mM DPA was included in the buffer.
The excitation and emission wavelengths were:
lEx = 495 nm and lEm = 525 nm for calcein,
lEx = 270 nm and lEm = 490 nm for Tb3+/DPA
chelates, and lEx = 495 nm and lEm = 520 nm
for FITC-Dextran. Fluorescence prior to addi-
tion of protein was treated as Ft0. 5 ml of APOL3
diluted in 10 mM acetic acid was added after
~1 min, and fluorescence recorded continuously
(at 10- to 15-s intervals) using a SpectraMax i3X
plate reader (Molecular Devices). 5 ml of final
dialysate (20 mM acetic acid) was used as a
mock treatment. 10 ml of 1% Triton X-100 was
added to achieve complete dye release and
the average of the top three fluorescence val-
ues defined as Ft100. The percentage of dye ef-
flux at each time point was calculated as (t)
(%) = (Ft – Ft0) × 100/(Ft100 – Ft0). To measure
FITC-dextran efflux, liposome-protein mix-
tures were incubated for 20 min at room
temperature and released FITC-dextran was
collected in the flowthrough following cen-
trifugation through a centrifugal filter device.
Supernatant fluorescence from untreated lipo-
somes was defined as Ft0 and in the presence
of 0.1% Triton X-100 as Ft100. To measure lipo-
some turbidity, DMPG or DMPC liposomes
(2 mM lipid) were generated in 20 mM ammo-
nium acetate and mixed with 40 mM rAPOL3
(50:1 lipid:protein ratio) in 20 mM ammo-
nium acetate at the indicated temperature
and absorbance at 400 nm determined. For
the temperature transition, liposome-APOL3
mixtures were incubated at 37°C for 2 min
before being transferred to room temperature
for the duration.
Giant unilamellar vesicle (GUV) assays
79 nmol of DOPC, 20 nmol of DOPG, and
1 nmol of Cy5-labeled DOPC were mixed in
50 ml of 3:1 chloroform:methanol and spotted
onto two indium tin oxide (ITO)–coated slides
and evaporated under vacuum for 2 hours.
ITO slides were sandwiched between PFTE
spacers to create a GUV chamber and filled
with swelling buffer (50 mM MES pH 6.0,
195 mM sucrose) and sealed with lipid-free
modeling clay. Electroformation was con-
ducted by applying a sinusoidal alternating
voltage (10 Hz) increasing from 0.02 to 1.2 V
over 50 min and holding this voltage for
120 min. Vesicles were removed and subjected
to buffer exchange by adding 100 ml to 1 ml
100 mM MES pH 6.0, 150 mM KGl containing
50 mM Dylight 488 free acid (ThermoFisher)
and mixed gently by inversion. After 30 min
incubation at room temperature, vesicles were
collected from the bottom of the tube and added
to BSA-coated 20 mM glass-bottom dishes.
300 nM 568-labeled rAPOL3 was added to the
well, mixed by pipetting, and imaged using a
Nikon TiE inverted spinning disc confocal mi-
croscope or Nikon TE2000 microscope.
Circular dichroism
Spectra were taken of 3 mM rAPOL3 proteins
in 10 mM MES pH 6.0, 20 mM KGl, 1 mM CaCl2
using a Chirascan circular dichroism spectrom-
eter (Applied Photophysics). To determine the
lipid-associated spectra, 3 mM rAPOL3 proteins
were mixed with 3 mM PC/PG liposomes in
the same buffer for 20 min and insoluble mate-
rial removed by centrifugation. The amount of
soluble protein remaining was determined and
adjusted accordingly so that the lipid-bound
and lipid-free concentrations were equivalent.
Baseline spectra for buffer or liposome alone
were used as the blank. CAPITO software (54)
was used to estimate the secondary structure
based on the observed spectra. Protein content
was determined by BCA assay (ThermoFisher).
Electron microscopy
Negative-stain electron microscopy: To visual-
ize the effect of rAPOL3 addition to liposomes,
liposomes containing 75:25 DMPC/DMPG
(2 mM total lipid) were generated in 20 mM
ammonium acetate and mixed with 40 mM
rAPOL3 at 37°C for 5 min. The reaction was
transferred to room temperature for an ad-
ditional 30 min, then diluted 1/20 before load-
ing onto glow-discharged copper coated EM
grids (EMS, cat#CF400-Cu-50) and stained with
2% uranyl formate for 1 min. Grids were exam-
ined in JEOL1400 plus electron microscope
with acceleration voltage of 80 kV. To visualize
rAPOL3 on bacteria, log-phase E. coliDhldE or
StmDwaaL was incubated with 10 mM, 5 mM, or
2 mM rHis-APOL3 in 100 ml Buffer A for 5 min
at room temperature. Bacteria were pelleted
and blocked by resuspension in 360 ml 20 mM
Tris pH 7.4, 10 mM imidazole, 200 mM NaCl
containing 1.5% skim milk for 5 min at room
temperature. 40 ml of 5 nm Ni-NTA nanogold
beads (nanoprobes) was added for 10 min at
room temperature and washed three times in
20 mM Tris pH 7.4, 20 mM imidazole, 200 mM
NaCl before loading directly onto glow-discharged
copper grids. Bacteria were treated with dialy-
sate alone and processed in parallel to assess
nonspecific binding of beads to bacteria.
Cryo–immunoelectron microscopy: HeLa cells
were transduced with pMSCV-EGFP-APOL3 for
7 days, then infected with Stm as above before
fixation/rehydration steps and immunogold la-
beling with anti-GFP antibodies as described (19).
Negative-stain EM particle averaging: More
than 400 meshed copper grids coated with
carbon film (EMS, cat #CF400-Cu-50) were
Gaudet et al., Science 373, eabf8113 (2021)
16 July 2021
12 of 14
RES EARCH | R E S E A R C H A R T I C L E
glow-discharged (PELCO easiGlow Glow Dis-
charge Cleaning System) for 20 s to increase
hydrophilicity of the carbon surface. 5-ml drop-
lets containing rAPOL3 and liposomes were
transferred to the glow-discharged EM grid
and incubated for 1 min. Droplet samples were
blotted with filter paper (Whatman qualitative
filter paper, Grade 1), quickly washed with 5 ml
2% uranyl formate solution and stained on the
grid by the same solution for 1 min. Residual
staining solution was again removed by filter
paper. Negative-staining EM grids were trans-
ferred to a JEOL1400plus electron microscope
and images taken under an 80-kV electron
gun at 40,000× magnification, and captured
by a Hamamatsu ORCA HR camera, resulting
in 0.234 nm/pixel. Five subframes were auto-
matically taken at near in-focus plane and
merged into a single image stack. All images
were then preprocessed by e2proc2d.py func-
tion from eman2 packages for generating the
images with MRC format (http://blake.bcm.
edu/emanwiki/EMAN2). MRC format images
were imported into Relion3 package for com-
plete processing from 2D classification to 3D
reconstruction (55). 2705 nanodisc-like parti-
cles were manually picked from 14 micrographs
without processing motion correction and
contrast transfer function correction. Two
repeats on 2D classification were performed
to remove nonspecific particles (56). Ten con-
formational classes representing 2071 par-
ticles from a final 2D classification were used
to generate an initial model and 3D classifica-
tion. Because no significant structural variation
arose from 3D classification, 3D refinement
was performed to obtain an averaged nano-
disc structure.
Cryo-EM sample preparation: 5 mM DMPC/
DMPG (3:1) liposomes were mixed with 100 mM
rAPOL3 in 20 mM ammonium acetate for 5 min
at 37°C, then transferred to 22°C for 1 hour and
any insoluble material removed by centrifuga-
tion at 16,000g for 10 min. 4 ml was added to a
C-Flat EM grid (C-Flat, 300 mesh CF-2/2-3Cu-
50) that had been glow-discharged for 20 s
(PELCO easiGlow Glow Discharge Cleaning
System). Samples were incubated for 5 s and
blotted in a Vitrobot Mark IV (Thermo Fisher
Scientific). Blotting conditions are as follows:
5 s of blotting time, a blotting force of 8 at 90%
humidity. The grid was subsequently plunged
into liquid ethane and then transferred to
liquid nitrogen for sample screening.
Cryo-EM data collection and image process-
ing: EM grids were loaded into a 200kV cryo–
electron microscope (Thermo Scientific Glacios)
equipped with a K2 Summit direct electron
detector. Data were collected at 45,000 mag-
nification, resulting in a physical pixel size of
0.896 angstroms (Å). The stage was adjusted
such that focus ranged from 1 mm to 2 mm for
data collection and the illumination area was
set to 1 mm in diameter. Data were collected in
superresolution movie mode with a 7-s expo-
sure equaling 35 frames with a total electron
dose of 50 e– Å–2. In total, 4648 stacks were
collected in a 2-day session. Superresolution
frames with a pixel size of 0.448 Å were treated
with motion correction process by MotionCor2
(57). Parameters for processing drift correction
are as follows: -Pathc 5 5 -PixSize 0.448 -Iter
30 -FtBin 2 -FmDose 1.43 -Bft 150 -Group 3.
Each micrograph was initially screened man-
ually to remove ice contamination or aggre-
gates. During motion correction two types of
images were generated: dose-weighted images
and non–dose-weighted images. The non–dose-
weighted images were used to estimate contrast
transfer function (CTF) by Gctf (58). The CTF
fitting of each micrograph was examined by
manually checking the fitting accuracy of the
Thon ring. The dose-weighted micrographs
were imported into Relion for further particle
picking and image processing. For particle
picking, 1194 particles were manually picked
from five representative micrographs to gener-
ate a template for the auto-picking process. In
total, 502,901 particles were automatically picked
from 4155 micrographs. These particles were
extracted in a binning factor 4 for 2D classi-
fication. The initial 3D model was generated
with the stochastic gradient descent algorithm
in Relion. Multiple rounds of 3D classification
were performed to screen homogeneous parti-
cles. The predominant class consisted of 236,364
particles and was used for a final 3D recon-
struction with a 13.5-Å resolution based on gold-
standard Fourier shell correlation criterion. The
final 3D EM map was visualized and segmented
by UCSF Chimera (59). The 3D reconstruction of
APOL3 lipoprotein from the cryo-EM dataset
was fitted into the EM density map from the
negative-staining TEM dataset using the fit-
in-map function of UCSF Chimera.
Native mass spectrometry (nativeMS)
DMPC/DMPG (75:25; 2 mM lipid) liposomes
made in 20 mM ammonium acetate were
treated with 40 mM APOL3 at 37°C for 5 min
before transfer to 22°C for 1 hour. Incubation
with 20 mM ammonium acetate without lipid
served as the negative control. For analysis of
bacterial-treated rAPOL3, overnight E. coliDhldE
was diluted 1/20 and grown to OD600 = 0.5.
0.5 ml was centrifuged and washed three times
in 20 mM ammonium acetate and resuspended
in 250 ml of 20 mM ammonium acetate buffer.
rHis-APOL3 was added to 20 mM and incubated
for 1 hour before insoluble material was pelleted
and supernatant harvested and placed on ice
to limit degradation by released bacterial pro-
teases. Remaining protein content was esti-
mated by protein gel. All samples were diluted
to 5 mM and equilibrated to room temperature
for 15 min prior to analysis. NativeMS was per-
formed on a Q Exactive UHMR mass spectrom-
eter (Thermo Fisher Scientific) using in-house
nano ion-emitting capillaries. The ultrahigh
vacuum was set at 5.65 × 10–10 mbar and cap-
illary voltage 1.5 kV. Insource trapping and
higher-energy collisional dissociation (HCD)
were optimized for best-quality spectra. Rela-
tive quantitation was performed by combining
the area under curves for each charge state.
In silico protein sequence analysis
Physiochemical properties of APOL3, APOL3-
DAH, or APOE1 were calculated using Heliquest
(60). Transmembrane domains and protein
structure was predicted using Phyre2.0 (61).
Hydrophobicity and charge were visualized
by applying YRB lighting (62) in Pymol.
Experimental design and statistics
No sample size calculation or blinding was per-
formed. For quantification of micrographs,
sample size reflects both prior knowledge of
variation and the maximum number of events
that could be reasonably quantified. No data
were excluded. Samples were randomly allo-
cated into experimental groups and typically
started with common pools of cells or bacteria.
Data were analyzed by GraphPad Prism 8.0
software. Unless otherwise indicated, statisti-
cal significance was determined by t test (two-
tailed) or one-way analysis of variance (ANOVA)
(Dunnett’s multiple-comparison tests) or two-
way ANOVA (multiple comparisons).
REFERENCES AND NOTES
1.
2.
F. Randow, J. D. MacMicking, L. C. James, Cellular self-defense:
How cell-autonomous immunity protects against pathogens.
Science 340, 701–706 (2013). doi: 10.1126/science.1233028;
pmid: 23661752
J. D. MacMicking, Interferon-inducible effector mechanisms in
cell-autonomous immunity. Nat. Rev. Immunol. 12, 367–382
(2012). doi: 10.1038/nri3210; pmid: 22531325
3. K. Schroder, P. J. Hertzog, T. Ravasi, D. A. Hume, Interferon-g:
An overview of signals, mechanisms and functions. J. Leukoc.
Biol. 75, 163–189 (2004). doi: 10.1189/jlb.0603252;
pmid: 14525967
4. S.-Y. Zhang et al., Inborn errors of interferon (IFN)-mediated
immunity in humans: Insights into the respective roles of
IFN-a/b, IFN-g, and IFN-l in host defense. Immunol. Rev. 226,
29–40 (2008). doi: 10.1111/j.1600-065X.2008.00698.x;
pmid: 19161414
5. S.-Y. Zhang et al., Human inborn errors of immunity to
infection affecting cells other than leukocytes: From the
immune system to the whole organism. Curr. Opin. Immunol.
59, 88–100 (2019). doi: 10.1016/j.coi.2019.03.008;
pmid: 31121434
6. T. Krausgruber et al., Structural cells are key regulators of
organ-specific immune responses. Nature 583, 296–302
(2020). doi: 10.1038/s41586-020-2424-4; pmid: 32612232
7. L. A. Knodler et al., Dissemination of invasive Salmonella via
bacterial-induced extrusion of mucosal epithelia. Proc. Natl.
Acad. Sci. U.S.A. 107, 17733–17738 (2010). doi: 10.1073/
pnas.1006098107; pmid: 20876119
8. C. R. Beuzón et al., Salmonella maintains the integrity of its
intracellular vacuole through the action of SifA. EMBO J. 19,
3235–3249 (2000). doi: 10.1093/emboj/19.13.3235;
pmid: 10880437
9. M. Chen et al., Itaconate is an effector of a Rab GTPase
cell-autonomous host defense pathway against Salmonella.
Science 369, 450–455 (2020). doi: 10.1126/science.aaz1333;
pmid: 32703879
10. E. E. Smith, H. S. Malik, The apolipoprotein L family of
programmed cell death and immunity genes rapidly evolved in
primates at discrete sites of host-pathogen interactions.
Genome Res. 19, 850–858 (2009). doi: 10.1101/gr.085647.108;
pmid: 19299565
Gaudet et al., Science 373, eabf8113 (2021)
16 July 2021
13 of 14
RES EARCH | R E S E A R C H A R T I C L E
11. L. Vanhamme et al., Apolipoprotein L-I is the trypanosome lytic
factor of human serum. Nature 422, 83–87 (2003).
doi: 10.1038/nature01461; pmid: 12621437
12. D. Pérez-Morga et al., Apolipoprotein L-I promotes
trypanosome lysis by forming pores in lysosomal membranes.
Science 309, 469–472 (2005). doi: 10.1126/science.1114566;
pmid: 16020735
13. T. L. M. Thurston, M. P. Wandel, N. von Muhlinen, A. Foeglein,
F. Randow, Galectin 8 targets damaged vesicles for autophagy
to defend cells against bacterial invasion. Nature 482, 414–418
(2012). doi: 10.1038/nature10744; pmid: 22246324
14. H. Strahl, L. W. Hamoen, Membrane potential is important for
bacterial cell division. Proc. Natl. Acad. Sci. U.S.A. 107,
12281–12286 (2010). doi: 10.1073/pnas.1005485107;
pmid: 20566861
15. F. Fontaine et al., APOLs with low pH dependence can kill all
African trypanosomes. Nat. Microbiol. 2, 1500–1506 (2017).
doi: 10.1038/s41564-017-0034-1; pmid: 28924146
16. T. Miki, W.-D. Hardt, Outer membrane permeabilization is an
essential step in the killing of gram-negative bacteria by the
lectin RegIIIb. PLOS ONE 8, e69901 (2013). doi: 10.1371/
journal.pone.0069901; pmid: 23922847
17. R. L. Gallo, L. V. Hooper, Epithelial antimicrobial defence of the
skin and intestine. Nat. Rev. Immunol. 12, 503–516 (2012).
doi: 10.1038/nri3228; pmid: 22728527
18. Y. T. Zheng et al., The adaptor protein p62/SQSTM1 targets
invading bacteria to the autophagy pathway. J. Immunol.
183, 5909–5916 (2009). doi: 10.4049/jimmunol.0900441;
pmid: 19812211
19. B.-H. Kim et al., A family of IFN-g-inducible 65-kD GTPases
protects against bacterial infection. Science 332, 717–721
(2011). doi: 10.1126/science.1201711; pmid: 21551061
20. B.-H. Kim et al., Interferon-induced guanylate-binding proteins
in inflammasome activation and host defense. Nat. Immunol.
17, 481–489 (2016). doi: 10.1038/ni.3440; pmid: 27092805
21. A. R. Shenoy et al., GBP5 promotes NLRP3 inflammasome
assembly and immunity in mammals. Science 336, 481–485
(2012). doi: 10.1126/science.1217141; pmid: 22461501
22. M. P. Wandel et al., Guanylate-binding proteins convert
cytosolic bacteria into caspase-4 signaling platforms.
Nat. Immunol. 21, 880–891 (2020). doi: 10.1038/s41590-020-
0697-2; pmid: 32541830
23. J. C. Santos et al., Human GBP1 binds LPS to initiate assembly
of a caspase-4 activating platform on cytosolic bacteria. Nat.
Commun. 11, 3276 (2020). doi: 10.1038/s41467-020-16889-z;
pmid: 32581219
24. M. Kutsch et al., Direct binding of polymeric GBP1 to LPS
disrupts bacterial cell envelope functions. EMBO J. 39,
e104926 (2020). doi: 10.15252/embj.2020104926;
pmid: 32510692
25. T. H. Bayburt, Y. V. Grinkova, S. G. Sligar, Self-Assembly of
Discoidal Phospholipid Bilayer Nanoparticles with Membrane
Scaffold Proteins. Nano Lett. 2, 853–856 (2002). doi: 10.1021/
nl025623k
26. W. K. Surewicz, R. M. Epand, H. J. Pownall, S. W. Hui, Human
apolipoprotein A-I forms thermally stable complexes with
anionic but not with zwitterionic phospholipids. J. Biol. Chem.
261, 16191–16197 (1986). doi: 10.1016/S0021-9258(18)66697-9;
pmid: 3097001
27. K. Gupta et al., Identifying key membrane protein lipid
interactions using mass spectrometry. Nat. Protoc. 13,
1106–1120 (2018). doi: 10.1038/nprot.2018.014
pmid: 29700483
28. K. Gupta et al., The role of interfacial lipids in stabilizing
membrane protein oligomers. Nature 541, 421–424 (2017).
doi: 10.1038/nature20820; pmid: 28077870
29. J. E. Keener et al., Chemical additives enable native mass
spectrometry measurement of membrane protein oligomeric
state within intact nanodiscs. J. Am. Chem. Soc. 141,
1054–1061 (2019). doi: 10.1021/jacs.8b11529;
pmid: 30586296
30. J. M. Stokes et al., Pentamidine sensitizes Gram-negative
pathogens to antibiotics and overcomes acquired colistin
resistance. Nat. Microbiol. 2, 17028 (2017). doi: 10.1038/
nmicrobiol.2017.28; pmid: 28263303
31. M. Vaara, Polymyxins and Their Potential Next Generation as
Therapeutic Antibiotics. Front. Microbiol. 10, 1689 (2019).
doi: 10.3389/fmicb.2019.01689; pmid: 31404242
32. T. Clairfeuille et al., Structure of the essential inner
membrane lipopolysaccharide-PbgA complex. Nature 584,
479–483 (2020). doi: 10.1038/s41586-020-2597-x;
pmid: 32788728
33. K. A. Brogden, Antimicrobial peptides: Pore formers or
metabolic inhibitors in bacteria? Nat. Rev. Microbiol. 3,
238–250 (2005). doi: 10.1038/nrmicro1098; pmid: 15703760
34. N. Mookherjee, M. A. Anderson, H. P. Haagsman, D. J. Davidson,
Antimicrobial host defence peptides: Functions and clinical
potential. Nat. Rev. Drug Discov. 19, 311–332 (2020).
doi: 10.1038/s41573-019-0058-8; pmid: 32107480
35. C. Sohlenkamp, O. Geiger, Bacterial membrane lipids: Diversity
in structures and pathways. FEMS Microbiol. Rev. 40, 133–159
(2016). doi: 10.1093/femsre/fuv008; pmid: 25862689
36. W. A. Ernst et al., Granulysin, a T cell product, kills bacteria by
altering membrane permeability. J. Immunol. 165, 7102–7108
(2000). doi: 10.4049/jimmunol.165.12.7102; pmid: 11120840
37. H. Barman et al., Cholesterol in negatively charged lipid
bilayers modulates the effect of the antimicrobial protein
granulysin. J. Membr. Biol. 212, 29–39 (2006). doi: 10.1007/
s00232-006-0040-3; pmid: 17206515
38. S. Stenger et al., An antimicrobial activity of cytolytic T cells
mediated by granulysin. Science 282, 121–125 (1998).
doi: 10.1126/science.282.5386.121; pmid: 9756476
39. G. van Meer, D. R. Voelker, G. W. Feigenson, Membrane lipids:
Where they are and how they behave. Nat. Rev. Mol. Cell Biol.
9, 112–124 (2008). doi: 10.1038/nrm2330; pmid: 18216768
40. A. Rausell et al., Common homozygosity for predicted loss-of-
function variants reveals both redundant and advantageous
effects of dispensable human genes. Proc. Natl. Acad.
Sci. U.S.A. 117, 13626–13636 (2020). doi: 10.1073/
pnas.1917993117; pmid: 32487729
41. R. G. Gaudet, C. J. Bradfield, J. D. MacMicking, Evolution of
Cell-Autonomous Effector Mechanisms in Macrophages versus
Non-Immune Cells. Microbiol. Spectr. 4, 10.1128/microbiolspec.
MCHD-0050-2016 (2016). doi: 10.1128/microbiolspec.MCHD-
0050-2016; pmid: 28087931
Genome Biol. 15, 554 (2014). doi: 10.1186/s13059-014-0554-4;
pmid: 25476604
54. C. Wiedemann, P. Bellstedt, M. Görlach, CAPITO—A web
server-based analysis and plotting tool for circular dichroism
data. Bioinformatics 29, 1750–1757 (2013). doi: 10.1093/
bioinformatics/btt278; pmid: 23681122
55. J. Zivanov et al., New tools for automated high-resolution cryo-
EM structure determination in RELION-3. eLife 7, e42166
(2018). doi: 10.7554/eLife.42166; pmid: 30412051
56. J. R. Gallagher, A. J. Kim, N. M. Gulati, A. K. Harris, Negative-
Stain Transmission Electron Microscopy of Molecular
Complexes for Image Analysis by 2D Class Averaging.
Curr. Protoc. Microbiol. 54, e90 (2019). doi: 10.1002/cpmc.90;
pmid: 31518065
57. S. Q. Zheng et al., MotionCor2: Anisotropic correction of beam-
induced motion for improved cryo-electron microscopy.
Nat. Methods 14, 331–332 (2017). doi: 10.1038/nmeth.4193;
pmid: 28250466
58. K. Zhang, Gctf: Real-time CTF determination and correction.
J. Struct. Biol. 193, 1–12 (2016). doi: 10.1016/j.jsb.2015.11.003;
pmid: 26592709
59. E. F. Pettersen et al., UCSF Chimera—A visualization
system for exploratory research and analysis. J. Comput.
Chem. 25, 1605–1612 (2004). doi: 10.1002/jcc.20084;
pmid: 15264254
60. R. Gautier, D. Douguet, B. Antonny, G. Drin, HELIQUEST: A web
server to screen sequences with specific a-helical properties.
Bioinformatics 24, 2101–2102 (2008). doi: 10.1093/
bioinformatics/btn392; pmid: 18662927
61. L. A. Kelley, S. Mezulis, C. M. Yates, M. N. Wass,
M. J. E. Sternberg, The Phyre2 web portal for protein modeling,
prediction and analysis. Nat. Protoc. 10, 845–858 (2015).
doi: 10.1038/nprot.2015.053; pmid: 25950237
42. L. Kamareddine, J. Nakhleh, M. A. Osta, Functional Interaction
62. D. Hagemans, I. A. E. M. van Belzen, T. Morán Luengo,
between Apolipophorins and Complement Regulate the
Mosquito Immune Response to Systemic Infections. J. Innate
Immun. 8, 314–326 (2016). doi: 10.1159/000443883;
pmid: 26950600
S. G. D. Rüdiger, A script to highlight hydrophobicity and
charge on protein surfaces. Front. Mol. Biosci. 2, 56 (2015).
doi: 10.3389/fmolb.2015.00056; pmid: 26528483
43. A. Zdybicka-Barabas, M. Cytryńska, Involvement of
AC KNOWLED GME NTS
apolipophorin III in antibacterial defense of Galleria mellonella
larvae. Comp. Biochem. Physiol. B 158, 90–98 (2011).
doi: 10.1016/j.cbpb.2010.10.001; pmid: 20959145
44. R. Figueira, K. G. Watson, D. W. Holden, S. Helaine,
Identification of salmonella pathogenicity island-2 type III
secretion system effectors involved in intramacrophage
replication of S. enterica serovar typhimurium: Implications
for rational vaccine design. mBio 4, e00065 (2013).
doi: 10.1128/mBio.00065-13; pmid: 23592259
45. S. Helaine et al., Internalization of Salmonella by macrophages
induces formation of nonreplicating persisters. Science 343,
204–208 (2014). doi: 10.1126/science.1244705;
pmid: 24408438
46. A. J. Karlsson et al., Engineering antibody fitness and
function using membrane-anchored display of correctly folded
proteins. J. Mol. Biol. 416, 94–107 (2012). doi: 10.1016/
j.jmb.2011.12.021; pmid: 22197376
47. Y. Chao, J. Vogel, A 3′ UTR-Derived Small RNA Provides the
Regulatory Noncoding Arm of the Inner Membrane Stress
Response. Mol. Cell 61, 352–363 (2016). doi: 10.1016/
j.molcel.2015.12.023; pmid: 26805574
48. R. G. Gaudet et al., Cytosolic detection of the bacterial
metabolite HBP activates TIFA-dependent innate immunity.
Science 348, 1251–1255 (2015). doi: 10.1126/science.aaa4921;
pmid: 26068852
49. Q. Kong et al., Effect of deletion of genes involved in
lipopolysaccharide core and O-antigen synthesis on virulence
and immunogenicity of Salmonella enterica serovar
typhimurium. Infect. Immun. 79, 4227–4239 (2011).
doi: 10.1128/IAI.05398-11; pmid: 21768282
50. M. Lazarou et al., The ubiquitin kinase PINK1 recruits
autophagy receptors to induce mitophagy. Nature 524,
309–314 (2015). doi: 10.1038/nature14893; pmid: 26266977
51. T. N. Nguyen et al., Atg8 family LC3/GABARAP proteins are
crucial for autophagosome-lysosome fusion but not
autophagosome formation during PINK1/Parkin mitophagy and
starvation. J. Cell Biol. 215, 857–874 (2016). doi: 10.1083/
jcb.201607039; pmid: 27864321
52. O. Shalem et al., Genome-scale CRISPR-Cas9 knockout
screening in human cells. Science 343, 84–87 (2014).
doi: 10.1126/science.1247005; pmid: 24336571
53. W. Li et al., MAGeCK enables robust identification of essential
genes from genome-scale CRISPR/Cas9 knockout screens.
We thank J. Nikolaus, A. Tunaru, M. Braun, K. Nelson, M. Llaguno,
and X. Liu for experimental advice and technical help. Funding:
Supported by National Institute of Allergy and Infectious Diseases
grants R01AI068041-14 and R01AI108834-07 (J.D.M.); National
Institute of Neurological Disorders and Stroke grant R01NS113236
(E.K.); and National Health and Medical Research Council grants
GNT1106471 and GNT1160315 and Australian Research Council
grants FT1601100063 and DP200100347 (M.L.). R.G.G. is an HHMI
Helen Hay Whitney Foundation Fellow. J.D.M. is an Investigator
of the Howard Hughes Medical Institute. Author contributions:
J.D.M. and R.G.G. conceived the study, designed experiments,
and wrote the manuscript. R.G.G. performed most experiments
with significant contributions by other authors. Specifically, S.Z.
undertook negative-stain and single-particle EM imaging plus
lipoprotein particle averaging; A.H. conducted nativeMS and
identified APOL3-LP adducts; B.-H.K. generated CRISPR-Cas9
knockout human cell lines and maintained bacterial mutants; C.J.B.
generated CRISPR-Cas9 knockout human cell lines and performed
RNA-seq analysis; D.X. and A.M. initially supervised and
collected high-content and superresolution microscopic images,
respectively; S.H. generated and FPLC-purified recombinant human
GBP1; T.N.N. and M.L. generated and validated CRISPR-Cas9
knockout human cell lines; E.K. facilitated and interpreted GUV
experiments; and K.G. helped plan, execute, and interpret nativeMS
experiments. All authors discussed the results and commented
on the manuscript. Competing interests: The authors declare
there are no competing interests. Data and materials availability:
All data are available in the main text or the supplementary
materials. The cryo-EM density map for the APOL3 lipoprotein
nanodisc is available in the Electron Microscopy Databank (EMDB)
with accession code EMD-24144.
SUPPLEMENTARY MATERIALS
science.sciencemag.org/content/373/6552/eabf8113/suppl/DC1
Figs. S1 to S14
Tables S1 and S2
Movies S1 to S11
View/request a protocol for this paper from Bio-protocol.
20 November 2020; resubmitted 29 April 2021
Accepted 3 June 2021
10.1126/science.abf8113
Gaudet et al., Science 373, eabf8113 (2021)
16 July 2021
14 of 14
|
10.1126_science.abi4506
|
RES EARCH
CORONAVIRUS
Chimeric spike mRNA vaccines protect against
Sarbecovirus challenge in mice
David R. Martinez1*, Alexandra Schäfer1, Sarah R. Leist1, Gabriela De la Cruz2, Ande West1,
Elena N. Atochina-Vasserman3, Lisa C. Lindesmith1, Norbert Pardi3, Robert Parks4, Maggie Barr4,
Dapeng Li4, Boyd Yount1, Kevin O. Saunders4, Drew Weissman3, Barton F. Haynes4,
Stephanie A. Montgomery5, Ralph S. Baric1*
The emergence of severe acute respiratory syndrome coronavirus (SARS-CoV) in 2003 and
SARS-CoV-2 in 2019 highlights the need to develop universal vaccination strategies against the
broader Sarbecovirus subgenus. Using chimeric spike designs, we demonstrate protection against
challenge from SARS-CoV, SARS-CoV-2, SARS-CoV-2 B.1.351, bat CoV (Bt-CoV) RsSHC014, and a
heterologous Bt-CoV WIV-1 in vulnerable aged mice. Chimeric spike messenger RNAs (mRNAs) induced
high levels of broadly protective neutralizing antibodies against high-risk Sarbecoviruses. By contrast,
SARS-CoV-2 mRNA vaccination not only showed a marked reduction in neutralizing titers against
heterologous Sarbecoviruses, but SARS-CoV and WIV-1 challenge in mice resulted in breakthrough
infections. Chimeric spike mRNA vaccines efficiently neutralized D614G, mink cluster five, and the
UK B.1.1.7 and South African B.1.351 variants of concern. Thus, multiplexed-chimeric spikes can prevent
SARS-like zoonotic coronavirus infections with pandemic potential.
A novel severe acute respiratory syndrome
coronavirus (SARS-CoV) emerged in 2003
and caused more than 8000 infections
and ~800 deaths worldwide (1). In 2012,
the Middle East respiratory syndrome
coronavirus (MERS-CoV) emerged in Saudi
Arabia (2), with multiple outbreaks that
have resulted in at least ~2600 cases and
900 deaths (3). In December 2019, another
novel human SARS-like virus from the genus
Betacoronavirus and subgenus Sarbecovirus
emerged in Wuhan China, designated SARS-
CoV-2, causing the ongoing COVID-19 pan-
demic (4, 5).
Bats are known reservoirs of SARS-like
coronaviruses (CoVs) and harbor high-risk
“preemergent” SARS-like variant strains, such
as WIV-1-CoV and RsSHC014-CoV, which are
able to use human ACE2 (angiotensin-converting
enzyme 2) receptors for entry, replicate effici-
ently in human primary airway epithelial cells,
and may escape existing countermeasures (6, 7).
Given the high pandemic potential of zoonotic
and epidemic Sarbecoviruses (8), the develop-
ment of countermeasures such as broadly ef-
fective vaccines, antibodies, and drugs is a
global health priority (9–11).
1Department of Epidemiology, University of North Carolina at
Chapel Hill, Chapel Hill, NC, USA. 2Lineberger Comprehensive
Cancer Center, University of North Carolina School of
Medicine, Chapel Hill, NC, USA. 3Infectious Disease Division,
Department of Medicine, Perelman School of Medicine,
University of Pennsylvania Perelman School of Medicine,
Philadelphia, PA, USA. 4Duke Human Vaccine Institute, Duke
University School of Medicine, Durham, NC, USA.
5Department of Laboratory Medicine and Pathology,
University of North Carolina School of Medicine, Chapel Hill,
NC, USA.
*Corresponding author. Email: david.rafael.martinez@gmail.com
(D.R.M.); rbaric@email.unc.edu (R.S.B.)
Sarbecovirus spike proteins have immuno-
genic domains: the receptor binding domain
(RBD), the N-terminal domain (NTD), and
the subunit 2 (S2) (12, 13). RBD, NTD, and to
a lesser extent S2 are targets for potent
neutralizing and non-neutralizing antibodies
elicited to SARS-CoV-2 and MERS-CoV spike
(12, 14–19). Passive immunization with SARS-
CoV-2 NTD-specific antibodies protect naïve
mice from challenge, demonstrating that the
NTD is a target of protective immunity (12, 19, 20).
However, it remains unclear whether vaccine-
elicited neutralizing antibodies can protect
against in vivo challenge with heterologous
epidemic and bat coronaviruses. We gener-
ated nucleoside-modified mRNA-lipid nano-
particle (LNP) vaccines expressing chimeric
spikes that contain admixtures of different
RBD, NTD, and S2 modular domains from
zoonotic, epidemic, and pandemic CoVs and
examined their efficacy against homologous
and heterologous Sarbecovirus challenge in
aged mice.
Results
Design and expression of chimeric spike
constructs to cover pandemic and zoonotic
SARS-related coronaviruses
Sarbecoviruses exhibit considerable genetic
diversity (Fig. 1A), and SARS-like bat CoVs
(Bt-CoVs) are recognized threats to human
health (6, 8). Because potent neutralizing
antibody epitopes exist in each of the modu-
lar structures on CoV spikes (21), we hypoth-
esized that chimeric spikes that encode NTD,
RBD, and S2 domains into “bivalent” and “tri-
valent” vaccine immunogens have the po-
tential to elicit broad protective antibody
responses against clades I to III Sarbecovi-
ruses. We designed four sets of chimeric spikes.
Chimera 1 included the NTD from clade II
Bt-CoV Hong Kong University 3-1 (HKU3-1),
the clade I SARS-CoV RBD, and the clade III
SARS-CoV-2 S2 (Fig. 1B). Chimera 2 included
SARS-CoV-2 RBD and SARS-CoV NTD and S2
domains (11). Chimera 3 included the SARS-
CoV RBD and SARS-CoV-2 NTD and S2, where-
as chimera 4 included the RsSHC014 RBD and
SARS-CoV-2 NTD and S2. We also generated a
monovalent SARS-CoV-2 spike furin knock-
out (KO) vaccine, partially phenocopying the
Moderna and Pfizer mRNA vaccines in hu-
man use, and a negative control norovirus
GII capsid vaccine (Fig. 1, B and C). We generated
these chimeric spikes and control spikes as
lipid nanoparticle-encapsulated, nucleoside-
modified mRNA vaccines with LNP adjuvants
(mRNA-LNP), as described previously (22).
This mRNA LNP stimulates robust T follic-
ular helper cell activity, germinal center B cell
responses, durable long-lived plasma cells, and
memory B cell responses (23, 24). We verified
their chimeric spike expression in human
embryonic kidney (HEK) cells (fig. S1B). To
confirm that scrambled coronavirus spikes
are biologically functional, we also designed
and recovered several high-titer recombi-
nant live viruses of RsSHC014/SARS-CoV-2
NTD, RBD, and S2 domain chimeras that
included deletions in nonessential, accessory
open reading frame 7 (ORF7) and ORF8 that
encoded nanoluciferase (fig. S1C). SARS-CoV-2
ORF7 and -8 antagonize innate immune sig-
naling pathways (25, 26), and deletions in
these ORFs are associated with attenuated
disease in humans (27, 28).
Immunogenicity of mRNAs expressing chimeric
spike constructs against coronaviruses
We next sought to determine whether simul-
taneous immunization with mRNA-LNP expres-
sing the chimeric spikes of diverse Sarbecoviruses
was a feasible strategy to elicit broad bind-
ing and neutralizing antibodies. We immu-
nized aged mice with the chimeric spikes
formulated to induce cross-reactive responses
against multiple divergent clades I to III
Sarbecoviruses, a SARS-CoV-2 furin KO spike,
and a GII.4 norovirus capsid negative control.
Group 1 was primed and boosted with chime-
ric spikes 1, 2, 3, and 4 (fig. S1A). Group 2 was
primed with chimeric spikes 1 and 2 and
boosted with chimeric spikes 3 and 4 (fig. S1A).
Group 3 was primed and boosted with chim-
eric spike 4 (fig. S1A). Group 4 was primed and
boosted with the monovalent SARS-CoV-2
furin KO spike (fig. S1A). Last, group 5 was
primed and boosted with a norovirus capsid
GII.4 Sydney 2011 strain (fig. S1A). We then
examined the binding antibody responses
by means of enzyme-linked immunosorbent
assay (ELISA) against a diverse panel of CoV
spike proteins that included epidemic, pan-
demic, and zoonotic coronaviruses.
Martinez et al., Science 373, 991–998 (2021)
27 August 2021
1 of 8
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Genetic design of chimeric Sarbecovirus spike vaccines. (A) Genetic diversity of pandemic
and bat zoonotic coronaviruses. HKU3-1 is shown in yellow, SARS-CoV is shown in light blue, RsSHC014
is shown in orange, and SARS-CoV-2 is shown in purple. (B) Spike chimera 1 includes the NTD from
HKU3-1, the RBD from SARS-CoV, and the rest of the spike from SARS-CoV-2. Spike chimera 2 includes
the RBD from SARS-CoV-2 and the NTD and S2 from SARS-CoV. Spike chimera 3 includes the RBD
from SARS-CoV and the NTD and S2 SARS-CoV-2. Spike chimera 4 includes the RBD from RsSHC014 and
the rest of the spike from SARS-CoV-2. SARS-CoV-2 furin KO spike vaccine and is the norovirus capsid
vaccine. (C) Table summary of chimeric spike constructs.
Mice in groups 1 and 2 generated the highest-
magnitude responses to SARS-CoV Toronto
Canada isolate (Tor2), RsSHC014, and HKU3-1
spike as compared with group 4 (Fig. 2, A, G,
and H). Whereas mice in group 2 generated
lower-magnitude binding responses to both
SARS-CoV-2 RBD (Fig. 2C) and SARS-CoV-2
NTD (Fig. 2D), mice in group 1 generated
similar-magnitude binding antibodies to SARS-
CoV-2 D614G (in which aspartic acid at posi-
tion 614 is replaced with glycine) as compared
with that of mice immunized with the SARS-
CoV-2 furin KO spike mRNA-LNP (Fig. 2B). Mice
in groups 1 and 2 generated similar-magnitude
binding antibody responses against SARS-
CoV-2 D614G, Pangolin GXP4L, and RaTG13
spikes (Fig. 2, B, E, and F) compared with those
of mice from group 4. Mice in group 1 and
group 4 elicited high-magnitude levels of
hACE2 blocking responses, as compared with
those of groups 2 and 3 (Fig. 2J). Because bind-
ing antibody responses after boost mirrored
the trend of the after-prime responses, it is
likely that the second dose is boosting im-
munity to the vaccine antigens in the prime
(Fig. 2). Last, we did not observe cross-binding
antibodies against common-cold CoV spike
antigens from HCoV-HKU1, HCoV-NL63, and
HCoV-229E in most of the vaccine groups (fig.
S2, A to D), but we did observe low binding
levels against more distant group 2C MERS-
CoV (Fig. 2I) and other Betacoronaviruses such
as group 2A HCoV-OC43 in vaccine groups 1
and 2 (fig. S2B). These results suggest that
chimeric spike vaccines elicit broader and
higher-magnitude binding responses against
pandemic and bat SARS-like viruses as com-
pared with those of monovalent SARS-CoV-2
spike vaccines.
Neutralizing antibody responses against live
Sarbecoviruses and variants of concern
We then examined the neutralizing antibody
responses against SARS-CoV, Bt-CoV RsSHC014,
Bt-CoV WIV-1, and SARS-CoV-2 including var-
iants of concern using live viruses as previously
described (Fig. 3, A to D) (17). Group 4 SARS-
CoV-2 S mRNA–vaccinated animals mounted
a robust response against SARS-CoV-2; how-
ever, responses against SARS-CoV, RsSHC014,
and WIV-1 were 18-, >300- or 116-fold decreased,
respectively (Fig. 3, A to D, and fig. S3, G and H).
By contrast, aged mice in group 2 showed a
42- and twofold increase in neutralizing titer
against SARS-CoV and WIV1 and less than
onefold decrease against RsSHC014 relative to
SARS-CoV-2 neutralizing titers (Fig. 3, A to D,
and fig. S3, C and D). Mice in group 3 elicited
thee- and sevenfold increased neutralizing
titers against SARS-CoV and RsSHC014 yet
showed a threefold decrease in WIV-1 neu-
tralizing titers relative to SARS-CoV-2 (Fig. 3,
A to D, and fig. S3, E and F). Last, mice in
group 1 generated the most balanced and
Martinez et al., Science 373, 991–998 (2021)
27 August 2021
2 of 8
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Human and animal coronavirus spike binding and hACE2-blocking responses in
chimeric and monovalent SARS-CoV-2 spike-vaccinated mice. Serum antibody ELISA
binding responses were measured in the five different vaccination groups. Before immunization,
after prime, and after boost binding responses were evaluated against Sarbecoviruses,
MERS-CoV, and common-cold CoV antigens including (A) SARS-CoV Toronto Canada (Tor2)
S2P, (B) SARS-CoV-2 S2P D614G, (C) SARS-CoV-2 RBD, (D) SARS-CoV-2 NTD, (E) Pangolin
GXP4L spike, (F) RaTG13 spike, (G) RsSHC014 S2P spike, (H) HKU3-1 spike, (I) MERS-CoV
spike, and (J) hACE2 blocking responses against SARS-CoV-2 spike in the distinct immunization
groups. Blue squares indicate mice from group 1, orange triangles indicate mice from group 2,
green triangles indicate mice from group 3, red rhombuses indicate mice from group 4,
and upside-down triangles indicate mice from group 5. Statistical significance for the binding
and blocking responses is reported from a Kruskal-Wallis test after Dunnett’s multiple
comparison correction. *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001.
highest neutralizing titers, which were 13-
and 1.2-fold increased against SARS-CoV and
WIV-1 and less than onefold decreased against
RsSHC014 relative to the SARS-CoV-2 neutral-
izing titers (Fig. 3, A to D, and fig. S3, A and B).
The serum of mice from groups 1 and 4 neu-
tralized the dominant D614G variant with
similar potency as that of the wild-type D614
nonpredominant variant, and both groups had
similar neutralizing antibody responses against
the UK B.1.1.7 and the mink cluster 5 variant
as compared with the D614G variant (Fig. 3,
E and F). Despite the significant but small
reduction in neutralizing activity against the
B.1.351 variant of concern (VOC), we did not
observe a complete ablation in neutralizing
activity in either group. Mice from groups 1
and 2 elicited lower binding and neutralizing
responses to SARS-CoV-2 as compared with
those of group 4, perhaps reflecting a decreased
amount of mRNA vaccine incorporated into
multiplexed formulations; the monovalent vac-
cines may drive a more focused B cell response
to SARS-CoV-2, whereas chimeric spike anti-
gens lead to more breadth against distant
Sarbecoviruses. Thus, both monovalent SARS-
CoV-2 vaccines and multiplexed chimeric
Martinez et al., Science 373, 991–998 (2021)
27 August 2021
3 of 8
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Live Sarbecovirus neutralizing antibody responses in vaccinated
mice. Neutralizing antibody responses in mice from the five different vaccination
groups were measured by using nanoluciferase-expressing recombinant viruses.
(A) SARS-CoV neutralizing antibody responses from baseline and after boost
in the distinct vaccine groups. (B) SARS-CoV-2 neutralizing antibody responses
from baseline and after boost. (C) RsSHC014 neutralizing antibody responses
from baseline and after boost. (D) WIV-1 neutralizing antibody responses from
baseline and after boost. (E) The neutralization activity in groups 1 and 4 against
SARS-CoV-2 D614G, South African B.1.351, UK B.1.1.7, and mink cluster 5
variant. (F) Neutralization comparison of SARS-CoV-2 D614G versus South
African B.1.351, versus UK B.1.1.7, and versus mink cluster 5 variant. Statistical
significance for the live-virus neutralizing antibody responses is reported from a
Kruskal-Wallis test after Dunnett’s multiple comparison correction. *P < 0.05,
**P < 0.01, ***P < 0.001, and ****P < 0.0001.
spikes elicit neutralizing antibodies against
newly emerged SARS-CoV-2 variants, and mul-
tiplexed chimeric spike vaccines outperform
the monovalent SARS-CoV-2 vaccines in terms
of breadth against multiclade Sarbecoviruses.
In vivo protection against heterologous
Sarbecovirus challenge
To assess the ability of the mRNA-LNP vac-
cines to mediate protection against previously
epidemic SARS-CoV, pandemic SARS-CoV-2,
and Bt-CoVs, we challenged the different
groups and observed the mice for signs of
clinical disease. Mice from group 1 or group 2
were completely protected from weight loss
and lower- and upper-airway virus replication
as measured with infectious virus plaque assays
after 2003 SARS-CoV mouse-adapted (MA15)
challenge (Fig. 4, A, B, and C). Similarly, these
two vaccine groups were also protected against
SARS-CoV-2 mouse-adapted (MA10) challenge.
By contrast, group 3 showed some protection
against SARS-CoV MA15–induced weight loss
but not against viral replication in the lung or
nasal turbinates. Group 3 was fully protected
against SARS-CoV-2 MA10 challenge. By con-
trast, group 5 vaccinated mice developed severe
disease, including mortality in both SARS-CoV
MA15 and SARS-CoV-2 MA10 infections (fig.
S5, B and C). Monovalent SARS-CoV-2 mRNA
vaccines were highly efficacious against SARS-
CoV-2 MA10 challenge but failed to protect
against SARS-CoV MA15–induced weight loss
and replication in the lower and upper respi-
ratory tract (Fig. 4, A, B, and C), suggesting
that SARS-CoV-2 mRNA-LNP vaccines are not
likely to protect against future SARS-CoV emer-
gence events. Mice from groups 1 to 4 were
completely protected from weight loss and
lower airway SARS-CoV-2 MA10 replication
(Fig. 4, D, E, and F). Using both a Bt-CoV
RsSHC014 full-length virus and a more viru-
lent RsSHC014-MA15 chimera in mice (6), we
also demonstrated protection in groups 1 to
3 against RsSHC014 replication in the lung
and nasal turbinates (fig. S4) but not in mice
that received the SARS-CoV-2 mRNA vac-
cine. Group 5 control mice challenged with
RsSHC014-MA15 developed disease, includ-
ing mortality (fig. S5D). Group 3 mice, which
received a SARS-CoV-2 NTD/RsSHC014 RBD/
SARS-CoV-2 S2, were fully protected against
both SARS-CoV-2 and RsSHC014 challenge,
whereas group 4 mice were not, demonstrat-
ing that a single NTD and RBD chimeric spike
can protect against more than one virus com-
pared with a monovalent spike.
Martinez et al., Science 373, 991–998 (2021)
27 August 2021
4 of 8
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. In vivo protection
against Sarbecovirus
challenge after mRNA-LNP
vaccination. (A) Percent
starting weight from the dif-
ferent vaccine groups of mice
challenged with SARS-CoV
MA15. (B) SARS-CoV MA15
lung viral titers in mice from
the distinct vaccine groups.
(C) SARS-CoV MA15 nasal
turbinate titers. (D) Percent
starting weight from the
different vaccine groups of
mice challenged with SARS-
CoV-2 MA10. (E) SARS-CoV-2
MA10 lung viral titers in
mice from the distinct vac-
cine groups. (F) SARS-CoV-2
MA10 nasal turbinate titers.
(G) Percent starting weight
from the different vaccine
groups of mice challenged with
WIV-1. (H) WIV-1 lung viral
titers in mice from the distinct
vaccine groups. (I) WIV-1 nasal
turbinate titers. (J) Percent
starting weight from the dif-
ferent vaccine groups of mice
challenged with SARS-CoV-2
B.1.351. (K) SARS-CoV-2
B.1.351 lung viral titers in mice
from the distinct vaccine
groups. (L) SARS-CoV-2
B.1.351 nasal turbinate titers.
The vaccines used in the
different groups are denoted
at bottom. Statistical signifi-
cance for weight loss is
reported from a two-way
analysis of variance (ANOVA)
after Dunnett’s multiple
comparison correction. For
lung and nasal turbinate titers,
statistical significance is
reported from a one-way
ANOVA after Tukey’s multiple
comparison correction. *P <
0.05, **P < 0.01, ***P <
0.001, and ****P < 0.0001.
We then performed a heterologous challenge
experiment with the bat preemergent WIV-1-
CoV (7). Mice from groups 1 and 2 were fully
protected against heterologous WIV-1 challenge,
whereas mice that received the SARS-CoV-2
mRNA vaccine had breakthrough replication
in the lung (Fig. 4, G, H, and I). We also chal-
lenged with a virulent form of SARS-CoV-2
VOC B.1.351, which contains deletions in the
NTD and mutations in the RBD, and observed
full protection in vaccine groups 1, 2, and 4
compared with that in controls, whereas break-
through replication was observed in group 3,
further indicating the importance of the NTD
in vaccine-mediated protection (Fig. 4, J, K,
and L). The reduced protection against the
B.1.351 variant containing NTD deletions
indicates that the NTD is a clear target of
Martinez et al., Science 373, 991–998 (2021)
27 August 2021
5 of 8
RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. Lung pathology in vaccinated
mice after SARS-CoV and SARS-CoV-2
challenge. (A) Hematoxylin and eosin
4 days after infection lung analysis of
SARS-CoV MA15–challenged mice from
the different groups: group 1, chimeras 1
to 4 prime and boost; group 2,
chimeras 1 and 2 prime and 3 and 4;
group 3, chimera 4 prime and boost,
SARS-CoV-2 furin KO prime and
boost, and norovirus capsid prime and
boost. (B) Lung pathology quantitation in
SARS-CoV MA15–challenged mice from
the different groups. Macroscopic lung
discoloration score, microscopic ALI
score, and DAD in day 4 after infection
lung tissues are shown. (C) Hematoxylin
and eosin 4 days after infection lung
analysis of SARS-CoV-2 MA10–challenged
mice from the different groups. (D) Lung
pathology measurements in SARS-CoV-2
MA10–challenged mice from the different
groups. Macroscopic lung discoloration
score, microscopic ALI score, and DAD in
day 4 after infection lung tissues are
shown. Statistical significance is reported
from a one-way ANOVA after Dunnet’s
multiple comparison correction. *P <
0.05, **P < 0.01, ***P < 0.001, and
****P < 0.0001.
Martinez et al., Science 373, 991–998 (2021)
27 August 2021
6 of 8
RES EARCH | R E S E A R C H A R T I C L E
protective immunity and that its inclusion in
vaccination strategies, as opposed to RBD-
alone vaccines, may be required to achieve
full protection. Moreover, the SARS-CoV-2
mRNA vaccine protected against SARS-CoV-2
B.1.351 challenge in aged mice despite a re-
duction in the neutralizing activity against
this VOC.
Lung pathology and cytokines in mRNA-LNP–
vaccinated mice challenged with epidemic and
pandemic coronaviruses
To quantify the pathological features of acute
lung injury (ALI) in mice, we used a tool from
the American Thoracic Society (ATS). We sim-
ilarly scored lung tissue sections for diffuse
alveolar damage (DAD), the pathological hall-
mark of ALI (29, 30). We observed significant
lung pathology with both the ATS and DAD
scoring tools in groups 4 and 5 vaccinated
animals. By contrast, multiplexed chimeric
spike vaccine formulations in groups 1 and 2
provided complete protection from lung pa-
thology after SARS-CoV MA15 challenge (Fig.
5, A and B). Mice immunized with the SARS-
CoV-2 mRNA vaccine that showed break-
through infection with SARS-CoV MA15
developed similar lung inflammation as that of
control vaccinated animals, potentially sug-
gesting that future outbreaks of SARS-CoV
may cause disease even in individuals vacci-
nated with SARS-CoV-2. Because eosinophilic
infiltrates have been observed in vaccinated,
2003 SARS-CoV–challenged mice previously
(31), with immunohistochemistry we analyzed
lung tissues in protected versus infected ani-
mals with SARS-CoV MA15 for eosinophilic
infiltrates (fig. S6). Groups 1 and 2 contained
rare, scattered eosinophils in the interstitium.
Group 3 showed bronchus-associated lymph-
oid tissue. By contrast, group 4 and group 5
contained frequent perivascular cuffs with
prevalent eosinophils. All groups challenged
with SARS-CoV-2 MA10 were protected against
lung pathology compared with the norovirus
capsid-immunized control group, supporting the
hypothesis that the SARS-CoV-2 NTD present
in the chimeric spike from group 3 is sufficient
for protection (Fig. 5, C and D).
We measured lung proinflammatory cyto-
kines and chemokines in the different vaccina-
tion groups. Groups 1 and 2 had baseline levels
of macrophage-activating cytokines and chemo-
kines, including interleukin-6 (IL-6), chemokine
(C-C motif) ligand 2 (CCL2), IL-1a, granulocyte
colony-stimulating factor (G-CSF), and CCL4,
compared with group 5 after SARS-CoV MA15
challenge (fig. S7A). Group 3 and group 4
showed high and indistinguishable levels of
IL-6, CCL2, IL-1a, G-CSF, and CCL4 compared
with those of group 5 mice after SARS-CoV
MA15 challenge. After SARS-CoV-2 MA10 chal-
lenge, group 4 and group 1 showed the lowest
levels of IL-6 and G-CSF relative to that in
group 5 controls (fig. S7B), and we only ob-
served significant reductions in CCL2, IL-1a,
and CCL4 lung levels in groups 3 and 4 com-
pared with the group 5 control, despite full
protection from both weight loss and lower-
airway viral replication.
Discussion
The Moderna and Pfizer/BioNTech SARS-CoV-2
mRNA-LNP vaccines were safe and efficacious
against SARS-CoV-2 infections in large phase 3
efficacy human clinical trials (32–34), but there
is a growing concern regarding VOCs such as
South African B.1.351, which is five- to sixfold
more resistant to vaccine-elicited polyclonal
neutralizing antibodies (35). We sought to rep-
licate the mRNA platform to formulate chi-
meric vaccines that specifically target distant
Sarbecovirus strains. A caveat of including
multiple chimeric spikes in a single shot is
the potential formation of heterotrimers not
present in the intended vaccine formulation.
Chimera 4, which contains the RsSHC014 RBD
and SARS-CoV-2 NTD and S2, elicited binding
and neutralizing antibodies, and mice were fully
protected from Bt-CoV RsSHC014 and SARS-
CoV-2 challenge, whereas SARS-CoV-2 full
length did not fully protect against RsSHC014,
suggesting that CoV spike vaccines can be de-
signed to maximize their display of protec-
tive epitopes and indicates that NTD/RBD/S2
chimeric spikes may enhance protection rela-
tive to monovalent spikes. Because the NTD,
RBD, and S2 contain epitopes that are targeted
by protective antibodies (17, 19, 36), modular
chimeric spikes may provide a way to design
CoV spikes to elicit protective immunity against
three Sarbecoviruses as compared with a single
Sarbecovirus by a monovalent spike. The lack
of protection against WIV-1 and SARS-CoV and
only partial protection against RsSHC014 chal-
lenge in SARS-CoV-2 immunized mice indicates
the need for the development of universal vac-
cination strategies that can achieve broader
coverage against preemergent bat SARS-CoV–
like and SARS-CoV-2–like viruses. Despite the
lower-magnitude antibody responses against
SARS-CoV-2 in the chimeric spike groups, a
clear advantage of our chimeric spike vaccines
is the clear breadth of protection against mul-
ticlade Sarbecoviruses and SARS-CoV-2 vari-
ants compared with that from the monovalent
SARS-CoV-2 vaccine. Although other strategies
exist, including multiplexing mosaic Sarbeco-
virus RBDs (37) and RBDs on nanoparticles
(38), chimeric spike mRNA-LNP vaccination
can achieve broad protection by using exist-
ing manufacturing technologies and are por-
table to other high-risk emerging coronaviruses
such as group 2C MERS-CoV–related strains.
Thus, chimeric spikes can clearly protect against
more than one Sarbecovirus, but it is possible
that multiplexed full-length spikes may pro-
tect against Sarbecoviruses.
As previously reported with RNA recombi-
nant viruses, live Sarbecoviruses lacking ORF7/
ORF8 but containing distinct SARS-CoV-2 anti-
genic domains were viable, reaffirming the
known interchangeability and functional plas-
ticity of the CoV spike (21, 39, 40). Our dem-
onstration of cross-protection against multiple
Sarbecovirus strains in mice lends support to
the hypothesis that universal vaccines against
group 2B CoVs are likely achievable. Moving for-
ward, it will be important to determine whether
other combinations of chimeric mRNA-LNP
vaccines from other SARS-like viruses are pro-
tective, elicit broad T cell responses, prevent
the rapid emergence of escape viruses, elicit
protective responses in nonhuman primate
models of Sarbecovirus pathogenesis, and can
boost Sarbecovirus protective breadth in SARS-
CoV-2–vaccinated or convalescent individuals.
REFERENCES AND NOTES
J. D. Cherry, P. Krogstad, Pediatr. Res. 56, 1–5 (2004).
1.
2. A. M. Zaki, S. van Boheemen, T. M. Bestebroer,
A. D. Osterhaus, R. A. Fouchier, N. Engl. J. Med. 367,
1814–1820 (2012).
3. C. I. Paules, H. D. Marston, A. S. Fauci, JAMA 323, 707–708 (2020).
4. P. Zhou et al., Nature 579, 270–273 (2020).
5. Coronaviridae Study Group of the International Committee
on Taxonomy of Viruses, Nat. Microbiol. 5, 536–544 (2020).
6. V. D. Menachery et al., Nat. Med. 21, 1508–1513 (2015).
7. V. D. Menachery et al., Proc. Natl. Acad. Sci. U.S.A. 113,
3048–3053 (2016).
8. B. Hu et al., PLOS Pathog. 13, e1006698 (2017).
9. T. P. Sheahan et al., Sci. Transl. Med. 9, eaal3653 (2017).
10. C. G. Rappazzo et al., Science 371, 823–829 (2021).
11. D. R. Martinez et al., A broadly neutralizing antibody protects
against SARS-CoV, pre-emergent bat CoVs, and SARS-CoV-2
variants in mice. bioRxiv [Preprint] 28 April 2021.
doi:10.1101/2021.04.27.441655.
12. W. N. Voss et al., Science 372, 1108–1112 (2021).
13. C. Graham et al., Immunity 54, 1276–1289.e6 (2021).
14. L. Premkumar et al., Sci. Immunol. 5, eabc8413 (2020).
15. L. Liu et al., Nature 584, 450–456 (2020).
16. L. Dai et al., Cell 182, 722–733.e11 (2020).
17. D. Li et al., bioRxiv [Preprint] 2 January 2021.
doi:10.1101/2020.12.31.424729.
18. A. R. Shiakolas et al., Cell Rep. Med. 2, 100313 (2021).
19. M. McCallum et al., Cell 184, 2332–2347.e16 (2021).
20. N. Suryadevara et al., Cell 184, 2316–2331.e15 (2021).
21. M. M. Becker et al., Proc. Natl. Acad. Sci. U.S.A. 105,
19944–19949 (2008).
22. D. Laczkó et al., Immunity 53, 724–732.e7 (2020).
23. P. Zhou et al., A protective broadly cross-reactive human
antibody defines a conserved site of vulnerability on
beta-coronavirus spikes. bioRxiv [Preprint] 31 March 2021.
doi:10.1101/2021.03.30.437769.
24. K. Lederer et al., Immunity 53, 1281–1295.e5 (2020).
25. D. E. Gordon et al., Science 370, eabe9403 (2020).
26. J. Y. Li et al., Virus Res. 286, 198074 (2020).
27. Y. C. F. Su et al., mBio 11, e01610-20 (2020).
28. B. E. Young et al., Lancet 396, 603–611 (2020).
29. T. P. Sheahan et al., Nat. Commun. 11, 222 (2020).
30. M. E. Schmidt et al., PLOS Pathog. 14, e1006810 (2018).
31. M. Bolles et al., J. Virol. 85, 12201–12215 (2011).
32. L. A. Jackson et al., N. Engl. J. Med. 383, 1920–1931 (2020).
33. E. E. Walsh et al., N. Engl. J. Med. 383, 2439–2450 (2020).
34. L. R. Baden et al., N. Engl. J. Med. 384, 403–416 (2021).
35. K. Wu et al., N. Engl. J. Med. 384, 1468–1470 (2021).
36. M. M. Sauer et al., Nat. Struct. Mol. Biol. 28, 478–486 (2021).
37. A. A. Cohen et al., Science 371, 735–741 (2021).
38. K. O. Saunders et al., Nature (2021).
39. L. R. Banner, J. G. Keck, M. M. Lai, Virology 175, 548–555 (1990).
40. J. G. Keck, L. H. Soe, S. Makino, S. A. Stohlman, M. M. Lai,
J. Virol. 62, 1989–1998 (1988).
AC KNOWLED GME NTS
Funding: D.R.M. is currently supported by a Burroughs Wellcome
Fund Postdoctoral Enrichment Program Award and a Hanna H. Gray
Fellowship from the Howard Hugues Medical Institute and was
Martinez et al., Science 373, 991–998 (2021)
27 August 2021
7 of 8
RES EARCH | R E S E A R C H A R T I C L E
supported by an NIH NIAID T32 AI007151 and an NIAID F32
AI152296. This research was also supported by funding from the
Chan Zuckerberg Initiative awarded to R.S.B. This project was
supported by the North Carolina Policy Collaboratory at the
University of North Carolina at Chapel Hill, with funding from the
North Carolina Coronavirus Relief Fund established and appropriated
by the North Carolina General Assembly. This project was funded
in part by the National Institute of Allergy and Infectious Diseases
(NIAID), NIH, US Department of Health and Human Services awards
U01 AI149644, U54 CA260543, AI157155, and AI110700 to R.S.B.;
AI124429 and a BioNTech SRA to D.W. and E.N.A.-V.; as well as
an animal models contract from the NIH (HHSN272201700036I).
Animal histopathology services were performed by the Animal
Histopathology and Laboratory Medicine Core at the University of
North Carolina, which is supported in part by an NCI Center
Core Support Grant (5P30CA016086-41) to the UNC Lineberger
Comprehensive Cancer Center. We thank B. L. Mui and Y. K. Tam
from Acuitas Therapeutics, Vancouver, BC V6T 1Z3, Canada, for
supplying the LNPs. Author contributions: D.R.M. and R.S.B.
conceived the study. D.R.M. and R.S.B. designed experiments. D.R.M.,
A.S., S.R.L., and A.W. performed laboratory experiments. N.P. and
K.O.S. provided critical reagents. D.R.M., A.S., S.R.L., G.D.l.C., A.W.,
E.N.A.-V., L.C.L., N.P., R.P., M.B., D.L., B.Y., K.O.S., D.W., B.F.H., and
S.A.M. analyzed data and provided critical insight. D.R.M. wrote
the first draft of the paper. D.R.M., A.S., S.R.L., G.D.l.C., A.W.,
E.N.A.-V., L.C.L., N.P., N.P., R.P., M.B., D.L., B.Y., K.O.S., D.W., B.F.H.,
S.A.M., and R.S.B. read and edited the paper. Funding acquisition:
D.R.M. and R.S.B. All authors reviewed and approved the manuscript.
Competing interests: The University of North Carolina at Chapel
Hill has filed provisional patents for which D.R.M. and R.S.B. are
co-inventors (US provisional application no. 63/106,247 filed
on 27 October 2020) for the chimeric vaccine constructs and their
applications described in this study. Data and materials availability:
The amino acid sequences of the chimeric spike constructs are
included in table S1. mRNA sequences are deposited in GenBank with
the following accession nos: chimera 1, MZ393687; chimera 2,
MZ393688; chimera 3, MZ393689; and chimera 4, MZ393690.
Materials generated as part of this study are available from R.S.B.
upon request. This work is licensed under a Creative Commons
Attribution 4.0 International (CC BY 4.0) license, which permits
unrestricted use, distribution, and reproduction in any medium,
provided the original work is properly cited. To view a copy of this
license, visit https://creativecommons.org/licenses/by/4.0/.
This license does not apply to figures/photos/artwork or other
content included in the article that is credited to a third party; obtain
authorization from the rights holder before using such material.
SUPPLEMENTARY MATERIALS
science.sciencemag.org/content/373/6558/991/suppl/DC1
Materials and Methods
Figs. S1 to S7
Table S1
References (41–43)
MDAR Reproducibility Checklist
11 March 2021; accepted 15 June 2021
10.1126/science.abi4506
Martinez et al., Science 373, 991–998 (2021)
27 August 2021
8 of 8
|
10.1126_science.abh2858
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
NEUROSCIENCE
Astrocyte Ca2+-evoked ATP release regulates
myelinated axon excitability and conduction speed
Jonathan Lezmy*, I. Lorena Arancibia-Cárcamo, Tania Quintela-López, Diane L. Sherman,
Peter J. Brophy, David Attwell*
INTRODUCTION: Astrocytes support neuronal
function throughout the central nervous sys-
tem. In the gray matter, they regulate synapse
number during development, remove synap-
tically released neurotransmitters to terminate
their action and prevent excitotoxicity, control
the extracellular potassium concentration to pre-
vent hyperexcitability, regulate blood flow to en-
sure an adequate energy supply, provide lactate
to neurons for energy, and respond to rises of
intracellular calcium concentration ([Ca2+]i) by
releasing adenosine triphosphate (ATP) and
other gliotransmitters that act on neuronal re-
ceptors to modulate information processing.
However, their role is unclear in the white mat-
ter, which transmits information rapidly between
gray matter areas using axons wrapped with
capacitance-reducing myelin (although they have
been suggested to regulate myelination during
development and during normal function).
RATIONALE: Recently, it has been suggested
that learning and memory may reflect not only
changes in synaptic function in the gray matter,
but also changes in white matter function. In
particular, neural circuit function might be reg-
ulated by changes in the conduction speed of
myelinated axons that result in an altered ar-
rival time of action potentials at a distant
neuron. These speed changes might be brought
about by alterations of the properties of the
passively conducting myelinated internodes or
of the intervening excitable nodes of Ranvier,
where the action potential is generated. We
applied immunohistochemistry to assess how
astrocytes interact with myelinated axons, neu-
ronal stimulation and light-evoked calcium un-
caging in astrocytes to evoke Ca2+-dependent
release of gliotransmitters, and electrophysiol-
ogy and pharmacology to characterize how
astrocyte-released substances might affect the
axon initial segment (AIS) and nodes of Ranvier
of myelinated neurons. Measurements of con-
duction velocity and computer modeling al-
lowed us to interpret the results.
RESULTS: Astrocytes closely approach the axons
of myelinated neurons in layer V of the cerebral
cortex that enter the corpus callosum. Uncaging
Ca2+ within astrocytes or stimulating spike trains
in neurons evoked a rise of astrocyte [Ca2+]i that
triggered the release of ATP-containing vesicles
from these cells. This evoked an inward current
in the AIS and nodes of Ranvier of the pyram-
idal neurons. Pharmacology showed that this
was mediated by the activation of Gs-linked
adenosine A2a receptors (A2aRs), implying that
the released ATP was converted to adenosine
by extracellular enzymes. The A2aRs raise the
intracellular concentration of cyclic AMP,
which activates hyperpolarization-activated cy-
clic nucleotide–gated (HCN) channels mediating
the inward hyperpolarization-activated current
(Ih) and thus depolarizes the cell. In the AIS, the
activation of A2aRs alters excitability and hence
action potential generation, whereas in the
nodes of Ranvier, it decreases the conduction
speed of the action potential along the axon.
CONCLUSION: As in the gray matter, astrocyte
[Ca2+]i regulates the release of ATP into the
extracellular space in the white matter. After
conversion to adenosine, this regulates the ex-
citability and conduction speed of myelinated
axons. The changes in excitability at the AIS
will lead to changes in the relationship be-
tween the synaptic input and action potential
output of the cell. The altered conduction speed
of the myelinated axon may change neural
circuit function by changing the action po-
tential arrival time at the cell’s output synap-
ses, thus altering the integration of signals in
postsynaptic neurons. Variations in astrocyte-
derived adenosine level can occur between
wake and sleep states, and the extracellular
adenosine concentration rises during energy
deprivation conditions. These changes in aden-
osine level could thus control white matter
information flow and neural circuit function.▪
The list of author affiliations is available in the full article online.
*Corresponding author. Email: j.lezmy@ucl.ac.uk (J.L.);
d.attwell@ucl.ac.uk (D.A.)
Cite this article as J. Lezmy et al., Science 374, eabh2858
(2021). DOI: 10.1126/science.abh2858
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.abh2858
Location
Function
Mechanism
Cortex
Pyramidal neuron
Astrocyte
Ca2+
ATP release
Corpus callosum
AIS
y
t
i
l
i
b
a
t
i
c
x
e
Δ
Astrocyte
Δ speed
Myelin sheath
Node of Ranvier
ATP
Ado
Inward
current
A2aR
HCN2
cAMP
Node of Ranvier
Astrocytes regulate myelinated axon excitability and conduction speed. For cortical neurons with myelinated axons crossing the corpus callosum (left),
astrocytes regulate AIS excitability and axonal conduction speed (middle). Increases of astrocyte [Ca2+]i release ATP, which, after conversion to adenosine (Ado)
extracellularly, activates A2aRs that raise the intracellular cyclic AMP concentration and thus generate an inward current through HCN2 channels in the AIS and nodes
of Ranvier (right). Image was created with BioRender.com.
Lezmy et al., Science 374, 300 (2021)
15 October 2021
1 of 1
RES EARCH
R E S E A R C H A R T I C L E
◥
NEUROSCIENCE
Astrocyte Ca2+-evoked ATP release regulates
myelinated axon excitability and conduction speed
Jonathan Lezmy1*, I. Lorena Arancibia-Cárcamo1,2, Tania Quintela-López1, Diane L. Sherman3,
Peter J. Brophy3, David Attwell1*
In the brain’s gray matter, astrocytes regulate synapse properties, but their role is unclear for the
white matter, where myelinated axons rapidly transmit information between gray matter areas.
We found that in rodents, neuronal activity raised the intracellular calcium concentration ([Ca2+]i)
in astrocyte processes located near action potential–generating sites in the axon initial segment
(AIS) and nodes of Ranvier of myelinated axons. This released adenosine triphosphate, which was
converted extracellularly to adenosine and thus, through A2a receptors, activated HCN2-containing
cation channels that regulate two aspects of myelinated axon function: excitability of the AIS
and speed of action potential propagation. Variations in astrocyte-derived adenosine level between
wake and sleep states or during energy deprivation could thus control white matter information
flow and neural circuit function.
A strocytes support neuronal function
throughout the central nervous system.
In the gray matter, they regulate synapse
number during development, remove
synaptically released neurotransmitters,
control the extracellular [K+] ([K+]o), regulate
blood flow, and provide lactate to neurons
for energy (1). However, their role is unclear in
the white matter, which transmits information
rapidly between gray matter areas using my-
elinated axons. Wrapping of these axons with
myelin by oligodendrocytes reduces the axon
capacitance and thus confers a high conduc-
tion speed for the action potential, which, once
generated at the axon initial segment (AIS) (2),
is maintained as it propagates by sodium in-
flux at the nodes of Ranvier occurring between
the myelinated internodes. Because oligoden-
drocytes isolate almost all of the axon from the
extracellular space, they are thought to be the
main mediators of [K+]o control and energy
provision to axons (3, 4), making the role of
astrocytes uncertain despite astrocyte processes
occurring close to nodes of Ranvier (5).
Labeling for glial fibrillary acidic protein
(GFAP) revealed astrocytes throughout the gray
matter and white matter (Fig. 1A). On patch-
clamping layer V cortical pyramidal cells or
oligodendrocytes, we observed that astrocyte
processes were aligned and intimately asso-
ciated with the myelinated axon and its inter-
nodal sheaths (Fig. 1, B and C, and fig. S1, A to
F). We used double whole-cell patch-clamping
1Department of Neuroscience, Physiology and Pharmacology,
University College London, London WC1E 6BT, UK. 2UK
Dementia Research Institute, Francis Crick Institute, London
NW1 1AT, UK. 3Centre for Discovery Brain Sciences, University
of Edinburgh, Edinburgh EH16 4SB, UK.
*Corresponding author. Email: j.lezmy@ucl.ac.uk (J.L.);
d.attwell@ucl.ac.uk (D.A.)
to load an astrocyte with a Ca2+-sensing dye
(Fluo-4) and to depolarize a pyramidal cell to
stimulate action potentials. When the pyram-
idal cell was driven to fire briefly (30 Hz for
1 s), the astrocyte intracellular calcium con-
centration ([Ca2+]i) rose more in processes near
the neuronal dendrites [within 12.7 ± 2.5 s from
start of stimulus to peak fractional change in
the fluorescence signal (DF/F); mean DF/F =
0.20 ± 0.03, n = 5 neuron-astrocyte pairs] than
in processes near the axon (DF/F = 0.03 ± 0.02,
P = 0.023; Fig. 1E). However, if prolonged
spiking was evoked (10 s), [Ca2+]i rose in both
locations (Fig. 1, D and E, and movie S1; DF/F =
0.19 ± 0.08 and 0.17 ± 0.05, respectively, P =
0.63, n = 3). Uncaging Ca2+ in the astrocyte
soma also raised astrocyte [Ca2+]i (Fig. 1F, DF/
F = 1.06 ± 0.51 near dendrite, DF/F = 1.05 ± 0.53
near axon, P = 0.86, n = 6), and neuronal
activity evoked a further [Ca2+]i rise super-
imposed on this (DF/F = 0.39 ± 0.09 near den-
drite, DF/F = 0.01 ± 0.06 near axon, P = 0.018).
In the gray matter, astrocytes can modulate
neuronal function by releasing adenosine
triphosphate (ATP) (6). We imaged putative
vesicles containing ATP in astrocytes around
myelinated axons of layer V pyramidal neurons
using quinacrine (7). Uncaging Ca2+ in the as-
trocyte soma evoked a Ca2+ wave that propa-
gated into processes along the axon (movie S2)
and triggered the loss of 43% of quinacrine-
labeled vesicles (Fig. 1, G to I; P = 0.0053), a loss
that was not seen when the two-photon uncag-
ing excitation was insufficient (see the materials
and methods, fig. S2) to evoke a detectable rise
of [Ca2+]i nor when assessing ATP vesicle num-
ber in areas 5 mm away (Fig. 1, J and K). The
resulting ATP release into the extracellular so-
lution was detected using luciferin-luciferase
(see the materials and methods). Puffing 1 mM
ATP, but not artificial cerebrospinal fluid
(aCSF) from a pipette into the sensing solution
evoked a large increase of luciferin lumines-
cence (Fig. 2A), and uncaging Ca2+ in an as-
trocyte soma evoked a similar response that
was not seen when the uncaging illumination
was too weak to evoke a detectable [Ca2+]i rise
(Fig. 2, B and C).
On release, ATP is rapidly hydrolyzed to aden-
osine by ecto–adenosine triphosphatases (ecto-
ATPases) located on microglia and astrocytes
(8). Microglia and astrocytes associate with both
the AIS and nodes of Ranvier (9, 10). Adenosine
receptors have previously been reported at
synapses. We used immunohistochemistry to
identify adenosine receptors on layer V neu-
ron myelinated axons. Neither A1 nor A2b re-
ceptors were detected (fig. S1, G and H), but
A2a receptors (A2aRs) were found at 92% of
AISs (Fig. 2D, E and H) and at 85% of nodes
of Ranvier (Fig. 2, F to H). Overview images of
the cortex (fig. S3D) showed A2aR–expressing
AISs leading toward the corpus callosum, im-
plying that these were excitatory neurons. A2aRs
were detected in fewer nodes of the cerebel-
lar white matter, which contains excitatory
mossy and climbing fiber axons as well as in-
hibitory Purkinje cell axons (70%; fig. S3F), and
were absent from AISs of cerebellar Purkinje
cells (fig. S3E), implying a neuron type–specific
expression of A2aRs in myelinated axons. A2aRs
raise the cyclic AMP level, which can affect cell
excitability by promoting the opening of the
hyperpolarization-activated cyclic nucleotide–
gated (HCN) type channels (11, 12) present in
axons (13–15). We detected HCN2 channel sub-
units at 51% of AISs and at 64% of nodes of
Ranvier (Fig. 2, I to M). HCN1 subunits were
observed in pyramidal cell somata and in
nodes of Ranvier (fig. S1, G to I). Adenosine
receptors and HCN channels have not previ-
ously been reported at the node of Ranvier.
To examine the effect of activation of these
A2aRs, we patch-clamped layer V pyramidal
neurons and applied adenosine (100 mM from
a puffer pipette) or another agonist for A2aRs
(CGS 21680, 0.5 mM) to the AIS (Fig. 3A).
Activating A2aRs in this way depolarized the
cell by 6 to 7 mV (Fig. 3, A, B, and E; aCSF had
no effect). The action potential response to
small injected currents was increased in fre-
quency (Fig. 3, C, F, and G), but at high injec-
ted currents was reduced, presumably as a
result of increased Na+ channel inactivation
(Fig. 3, D, F, and H). Hyperpolarizing the cell
evoked a time-dependent inward current
increase (Fig. 3I), which was mediated by Ih
(HCN) channels, because it was blocked by
the Ih blocker ZD7288 (fig. S4). Applying
CGS 21680 increased the amplitude of Ih tail
currents, reflecting an increase in magnitude
of the fully activated conductance of 51% in
6 cells (P = 0.029), and shifted the 50% point
of the activation curve derived from these
Lezmy et al., Science 374, eabh2858 (2021)
15 October 2021
1 of 10
RES EARCH | R E S E A R C H A R T I C L E
A
GFAP
200 m
L I
L II/III
L IV
L V
L VI
White
matter
(WM)
B
Axon + GFAP + Caspr
C
Oligodendrocyte + GFAP
m
5
m
2
Layer V
M
W
m
0
2
Layer V
M
W
20 m
m
5
m
5
D
G
Dendrite
10 m
Axon
E
/
F
F
1
.
0
Astro. [Ca2+] near dendrite
Astro. [Ca2+] near axon
0.3
P=0.023
F
Astro. [Ca2+] near dendrite
Astro. [Ca2+] near axon
2
P=0.858
Brief
1 s
Prolonged
F
F
/
0
0.3
P=0.634
Ca2+ uncaging
in astrocyte
F
F
/
0
P=0.018
Astrocyte
Astrocyte
H
1
0
m
V
F
F
/
0
/
F
F
Spikes
in neuron
2
.
20 s0
0.5
F
F
/
0
Astrocyte process + Quinacrine
Control
Astro. Ca2+
uncaging
I
2
m
0.4
P = 0.0053
1 m
J
2
m
0.4
K
2
m
0.4
1 m
Adjacent
areas
10 m
i
l
/
s
e
c
s
e
v
P
T
A
0.3
0.2
0.1
8
8
0
Control Astro.
Ca2+ u.
l
i
/
s
e
c
s
e
v
P
T
A
0.3
0.2
0.1
6
0
Control
l
i
/
s
e
c
s
e
v
P
T
A
6
2P laser
(No Ca2+ u.)
0.3
0.2
0.1
11
11
0
Control Astro.
Ca2+ u.
Fig. 1. Astrocyte processes near myelinated axons release ATP in response to [Ca2+]i rises. (A) GFAP
labeling of mouse coronal slice showing astrocytes in gray matter and white matter. (B) Patch-clamp–loaded
Alexa Fluor 594 labels axon of layer V pyramidal cell. Expanded internodal (top) and nodal (bottom, identified
by Caspr labeling, node is at arrow) regions reveal astrocytes at both locations. (C) Patch-clamped
oligodendrocyte in layer V with insets below showing GFAP labeling around myelinated internodes. (D) Dendrite
and axon of patch-clamped rat layer V pyramidal cell near an astrocyte loaded with the Ca2+ sensor Fluo-4.
(E) [Ca2]i response in astrocyte processes near dendrite (black) and axon (red) to neuron depolarization with
500 pA for 1 s (top, n = 5) or 10 s (bottom, n = 3) to evoke spiking (gray). (F) [Ca2+]i response in astrocyte
processes to Ca2+ uncaging and to brief neuronal spike trains (n = 6). Top bar chart shows the Ca2+ response just
before spiking; bottom bar chart shows Ca2+ response after neuron spiking evoked by 500 pA for 1 s. (G) Patch-
clamped rat astrocyte loaded with Ca2+ cage NP-EGTA and Rhod-2 to measure [Ca2]i; the region imaged for (H)
is shown. (H) Quinacrine-labeled puncta in astrocyte process are depleted on uncaging Ca2+ (Rhod 2 may not enter
the smallest astrocyte processes, explaining why some puncta appear outside the astrocyte). (I to K) Quantification
of ATP vesicles present per square micrometer of astrocyte process before and after uncaging [Astro. Ca2+ u.] (I);
before and after excitation that did not evoke a [Ca2]i rise in the astrocyte [No Ca2+ u.; see fig. S2] (J); and in
regions outside the astrocyte (5 mm away) before and after uncaging (K). Numbers of processes shown on bars.
Processes came from 12 cells. Data shown in (A) to (C) are from mice; those in (D) to (K) are from rats.
currents in the depolarizing direction by +8.2 mV
(P = 0.045) so that more Ih is activated in the
physiological range (Fig. 3, I to K). This shift
was mimicked by including 50 mM cAMP in
the patch pipette (shifted by +8.6 mV in 6 cells,
P = 0.026), and in these cells, CGS 21680 evoked
no further depolarizing shift (–1.77 mV in 6 cells,
P = 0.087; Fig. 3, J to L).
To assess the function of A2aRs at the nodes
of Ranvier, we patch-clamped layer V pyram-
idal cells. Dye filling revealed an enlarged bleb
at the end of the axon where it was cut during
the brain-slicing process. This allowed patch-
clamping of the bleb (see the materials and
methods), which was ~3 (mean 3.2 ± 0.4 in 5
cells) nodes away from the soma (Fig. 4A).
In Thy1-Caspr-GFP mice, the nodes of layer V
pyramidal neurons could be seen during the
experiment, from paranodal GFP fluorescence
and because axon branches often occur at nodes,
and postrecording immunohistochemistry re-
vealed A2aRs at the AIS and nodes (Fig. 4A).
Puffing CGS 21680 onto a node (Fig. 4A) did
not significantly depolarize the pyramidal cell
soma (0.99 ± 0.46 mV in 5 cells, P = 0.1), im-
plying that the puffed drug did not reach the
soma or AIS. Injecting current (500 pA for
5 msec, repeated at 2 Hz) into the soma
evoked an action potential. This evoked a de-
layed action potential in the axonal bleb, which
became more delayed when CGS 21680 was
puffed at an intervening node (Fig. 4B; >100
responses were averaged). Plots of rate of change
of signal (d(signal)/dt) versus signal (Fig. 4C)
were used to estimate when the action poten-
tials in each location showed an accelerating
onset phase (soma depolarization >1 mV/msec
and a bleb inward current increase more nega-
tive than –0.5 pA/msec, shown as dashed lines
in Fig. 4B). The latency of the response at the
bleb increased from 108 ± 31 msec in control
conditions to 320 ± 89 msec when CGS 21680
was puffed (P = 0.03, paired t test, n = 5 cells;
Fig. 4D) and the spike width recorded at
the bleb increased by 0.28 msec from 0.30 ±
0.10 msec to 0.58 ± 0.14 msec (P = 0.015, paired
t test, n = 5).
To convert this latency increase to a change
of conduction speed, we need to assume where
in the AIS the action potential is initiated and
how fast it propagates backward to the soma
compared with forward to the bleb (see the
materials and methods). If the spike starts at
the middle of the AIS and the forward speed
is twice the backward speed (16), then CGS
21680 reduced the forward speed from 1.21 ±
0.23 to 0.43 ± 0.08 m/sec (64% reduction; filled
circles in Fig. 4E). Alternatively, if the spike
starts at the end of the AIS and the forward
speed is three times faster than the backward
speed (16), then the speed is reduced from
0.68 ± 0.16 to 0.31 ± 0.13 m/sec (54% reduction;
open circles in Fig. 4E). Thus, for a range of as-
sumptions, activation of nodal A2aRs produces
Lezmy et al., Science 374, eabh2858 (2021)
15 October 2021
2 of 10
RES EARCH | R E S E A R C H A R T I C L E
A
ATP puff
B
2P laser
P = 0.0006
C
2P laser
(No Ca2+ u.)
Astro. Ca2+
uncaging
0.6
0.4
0.2
aCSF puff
50 s
D
NaNav
Merged
Merged
+ DAPIDAPI
30 s
E
y
t
i
s
n
e
n
t
I
F
1
0.5
0
0
Nav
A2aR
H
10
20
30
40
Caspr + NaNav
Caspr
7
0
2P laser
(No Ca2+ u.)
7
Astro.
Ca2+u.
)
%
(
i
n
o
s
s
e
r
p
x
e
R
a
2
A
100
80
60
40
20
0
25
134
AIS
Nodes
I
NaNav
HCN2
HCN2
Merged
Merged
+ DAPIDAPI
G
y
t
i
s
n
e
n
t
I
J
y
t
i
s
n
e
n
t
1
0.5
0
0
1
0.5
Caspr
Nav
A2aR
2
4
6
8
10
Nav
HCN2
M
I
0
0
10
20
30
40
K
L
y
t
i
s
n
e
t
n
I
1
0.5
0
0
HCN2
Caspr + HCN2
Caspr
Caspr
HCN2
2
4
6
8
10
)
%
(
i
n
o
s
s
e
r
p
x
e
2
N
C
H
100
80
60
40
20
0
49
98
AIS
Nodes
Fig. 2. ATP release from astrocytes may target adenosine receptors on myelinated axons of
layer V pyramidal neurons. (A) Detection of ATP puffed into extracellular solution using luciferin-luciferase.
(B) Two-photon excitation uncaging Ca2+ in astrocytes evoked a luciferin-luciferase signal, unless the
excitation failed to raise astrocyte [Ca2+]i (No Ca2+ u.; see fig. S2). (C) Quantification of experiments in
(B) in seven cells. (D to H) A2aRs are present in the AIS (D), where they overlap with Nav expression [mean
of 14 Nav and 25 A2aR profiles (E)] and at the node of Ranvier (F), where they overlap with Nav and are
flanked by Caspr labeling [mean of 48 Caspr, 48 A2aR and 24 Nav profiles (G)]. (H) Percentage of 25 AISs
and 134 nodes that express A2aRs. (I to M) HCN2 channel subunits are present in the AIS (I), where
they overlap with Nav expression [mean of 19 NaV and 16 HCN2 profiles (J)], and at the node of Ranvier
(K), where they overlap with Nav and are flanked by Caspr labeling [mean of 53 Caspr and HCN2
profiles (L)]. (M) Percentage of 49 AISs and 98 nodes that express HCN2. Data shown in (A) to (E),
(I), and (J) are from rats; those in (F), (G), (K), and (L) are from mice; and those in (H) and (M)
combine rat AIS and mouse node data.
a very significant reduction of conduction
velocity (P = 0.0017 comparing the ratio of
the conduction velocities).
To explore how A2aRs and Ih affect ex-
citability and conduction velocity, we used a
MATLAB model of myelinated axons in the
corpus callosum (17), which was adapted to
mimic either (i) a soma attached to a trun-
cated axon with three internodes and a termi-
nal bleb (Fig. 5A) or (ii) an infinitely long axon
(see the materials and methods). The AIS and
nodes of Ranvier contained voltage-gated Na+
and K+ channels to generate action potentials,
whereas the soma and bleb lacked these chan-
nels. We modeled the Ih current in the absence
of A2aR activation as being present in the soma
and proximal AIS only and unaffected by the
puff application of A2aR-activating drugs at
the distal AIS (and the [cAMP] rise that they
produce), consistent with the presence in the
somatodendritic compartment (fig. S1G) of Ih
channels composed of HCN1 subunits (18),
which are relatively insensitive to cAMP (19).
The increase in Ih evoked by A2aR activation
in the AIS was modeled as the addition of an
extra current mediated by cAMP-activated
HCN2 channels located in the distal AIS. The
maximum conductance of this extra current
was set to increase the total maximum Ih con-
ductance (measured at the soma) by 51% as
observed (Fig. 3I), and the midpoint of the ac-
tivation curve was set at –80 mV to reproduce
the positive shift for the total Ih apparent ac-
tivation curve in Fig. 3J. Adenosine-activated
Ih was also added to the nodes of Ranvier at a
density 1.41-fold higher than in the distal AIS,
as measured immunohistochemically (see the
materials and methods).
Adding the adenosine-activated conductance
to the distal AIS in the model depolarized the
soma by 5.9 mV, as seen experimentally (Figs.
3E and 5B). It also increased the firing evoked
by a small injected current (Fig. 5, C and D)
while decreasing that evoked by a large cur-
rent (Fig. 5, C and D), also as seen experimen-
tally (Fig. 3, C to H). However, the simulated
depression of action potential amplitude at
large currents was larger than that observed
experimentally. These changes were largely
a result of the Ih added to the AIS, because
adding it solely to the nodes had only a minor
effect on the resting potential at the soma (Fig.
5B). Adding Ih to the nodes of Ranvier de-
creased the conduction speed (Fig. 5E), as ob-
served when puffing CGS 21680 onto the nodes
(Fig. 4E). The predicted reduction was smaller
than that observed experimentally, presumably
because of the mild depolarization generated
at the nodes (+3.8 mV). Using the infinite
cable model, adding the adenosine-activated
Ih depolarized the nodes by +11.4 mV, reduced
the predicted conduction speed by 48% from
2.23 to 1.17 m/s (Fig. 5F), and increased the axo-
nal spike width by 0.22 ms (mainly by increasing
Lezmy et al., Science 374, eabh2858 (2021)
15 October 2021
3 of 10
C
Control
CGS puff to AIS
V
m
0
1
100 ms
Control
CGS puff to AIS
V
m
0
1
100 ms
RES EARCH | R E S E A R C H A R T I C L E
A
B
CGS 21680
puff to AIS
- 80 mV
V
m
5 s5
E
15
)
P = 0.017
P = 0.0005
V
m
(
10
t
s
e
r
V
Δ
5
0
t
u
p
n
i
w
o
L
)
A
p
0
0
3
(
10 μm
-80 mV
D
t
u
p
n
i
h
g
H
i
)
A
p
0
0
7
(
-80 mV
F
)
z
H
(
e
t
a
r
g
n
i
r
i
F
50
40
30
20
10
0
Control
CGS 21680
aCSF CGS Adenosine
0
200 400 600 800
Current (pA)
I
Control
-56 mV
CGS 21680
-146 mV
A
p
0
0
4
100 ms
J
)
d
e
s
i
l
1
0.5
a
m
r
o
n
(
l
i
a
t
I
0
-140
Ctrl
CGS
cAMP
cAMP
+CGS
-120 -100 -80 -60
V (mV)
G
)
z
H
(
e
t
a
r
g
n
i
r
i
F
H
Low input
25 P = 0.008
20
15
10
5
26 26
0
Control CGS
High input
50 P = 0.0017
40
30
20
10
18 18
0
Control CGS
K
L
P = 0.045
P = 0.087
)
V
m
(
2
/
1
V
-90
-100
-110
-120
-90
-100
-110
-120
Ctrl CGS
cAMP
cAMP
+CGS
Fig. 3. A2a receptors, through cAMP and Ih, modulate excitability at the AIS. (A) Patch-clamped layer V
pyramidal cells loaded with Alexa Fluor 594. Lower pipette puffing adenosine at distal AIS contains Alexa
Fluor 594 to delineate the region affected by adenosine. (B) Depolarization of soma evoked by puffing the
A2aR agonist CGS 21680 (0.5 mM). (C and D) Voltage response to 300 pA (C) or 700 pA (D) injected current
with and without CGS 21680 application. (E) Mean resting potential change when puffing aCSF (n = 6),
0.5 mM CGS 21680 (n = 12), or 100 mM adenosine (n = 6) onto the AIS. (F) Firing rate averaged over 1 s as
a function of injected current, in control conditions or with CGS 21680 applied to the AIS (n = 9; SEM shown
faint). (G and H) Firing rate change for “low” [averaged over 100 to 300 pA, 26 current steps from
nine cells (G)] or “high” (700 to 900 pA, 18 current steps from eight cells (H)] input currents. (I) Specimen
currents on stepping from –56 mV to various pulse potentials and then to –136 mV to evoke tail currents,
allowing construction of the activation curve in control conditions and with CGS 21680 puffed at the AIS.
(J) Ih activation curves for normal conditions, puff application of CGS 21680 (n = 6), and both with 50 mM
cAMP included in the patch pipette (n = 6). (K) V1/2 values (50% activation voltage of fitted Boltzmann
curve) before (Ctrl) and during CGS 21680 application. (L) As in (K) but with cAMP in patch pipette. All data
are from rats.
Na+ channel inactivation at the resting poten-
tial because this was mimicked by decreasing
the peak Na+ conductance ~twofold), similar to
the 0.28 msec reported experimentally above.
Because astrocyte [Ca2+]i rises evoked the
release of ATP (Fig. 1, G to K, and Fig. 2, A to
C), which is expected to be converted to adeno-
sine extracellularly and to act on A2aRs at the
AIS and nodes of Ranvier, we tested whether
uncaging Ca2+ in astrocytes modulated the ex-
citability and conduction speed of cortical
layer V pyramidal cell myelinated axons. Ca2+
was uncaged in an astrocyte with processes
running near the AIS of a pyramidal cell (Fig.
6A and movie S3). Within 70 s of uncaging the
Ca2+, the resting potential of the neuron was
depolarized (Fig. 6B) and the action potential
response of the neuron was changed (Fig. 6C),
exhibiting a higher firing rate to low injected
currents (P = 0.0175 when comparing firing
rates normalized to the maximum rate evoked
at high injected current) and a lower firing
rate to high injected currents (P = 0.0136). These
changes, which are similar to those seen
when activating AIS A2aRs (Fig. 3, B to H), were
not mediated by glutamate release from astro-
cytes (fig. S5), were blocked by superfusion of
the A2aR blocker ZM 241385 (100 nM) or by
intracellular dialysis of the neuron with the Ih
blocker ZD7288 (20 mM), and were not seen
if the uncaging illumination failed to evoke a
[Ca2+]i rise (Fig. 6, B to F).
Uncaging Ca2+ in astrocytes, which were
shown by post hoc immunostaining to have
processes close to nodes of Ranvier (Fig. 6, G
and H), evoked a decrease of axonal conduc-
tion speed (Fig. 6I) in experiments monitoring
the action potential propagation from the
soma to a bleb ~4 (mean 3.8 ± 0.5 in 5 cells)
nodes along the axon. These data are similar
to those obtained by puffing an A2aR agonist
onto the nodes (Fig. 4) and, as for those exper-
iments, the decrease of conduction speed cal-
culated was dependent on the exact site of
action potential initiation within the AIS
and its forward and backward propagation
speeds (Fig. 6I). Introducing ZD7288 (20 mM)
into the targeted axons to block Ih (HCN) chan-
nels prevented the speed reduction evoked
by astrocyte Ca2+ activity (fig. S6A). In unmy-
elinated axons and some other neuronal types,
axonal HCN channels can speed, rather than
slow, the action potential (13, 14) (see fig. S6
legend). Inducing robust neuronal activity
(30 Hz for 1 min) to raise astrocyte [Ca2+]i
(Fig. 1E) also induced a decrease of axonal
conduction speed (P = 0.014), which was abol-
ished by superfusing the A2aR blocker ZM
241385 (100 nM) (fig. S6B).
These data reveal a new modulation of neu-
ronal circuit function by glial cells. Astrocyte
modulation of neuronal synaptic function in
the gray matter has been accepted as a major
determinant of neuronal function (6, 20, 21),
but the role of white matter astrocytes is less
clear. Our data, summarized in fig. S7, dem-
onstrate that astrocytes modulate the excit-
ability of the AIS of excitatory neurons and
regulate the conduction speed of myelinated
axons. This regulation is a result of astrocytes
raising the concentration of adenosine near
the AIS or nodes of Ranvier (of a single axon
or possibly several axons at once), which can
increase or decrease AIS excitability and de-
creases axon conduction speed. Adenosine re-
lease can occur when axonal action potential
firing raises [Ca2+]i in adjacent astrocytes (22);
this triggers the release of ATP from vesicles
(Figs. 1 and 2), which is subsequently converted
to adenosine by ecto-ATPases expressed by mi-
croglia, astrocytes, and oligodendrocyte line-
age cells (8). Neuronal activity in cortical layer V
Lezmy et al., Science 374, eabh2858 (2021)
15 October 2021
4 of 10
RES EARCH | R E S E A R C H A R T I C L E
A
B
Alexa 594
r- A2aR
Caspr-GFP
Alexa 488
Layer V soma
Node
Branch point node
1 m
AIS
20 m
1 m
First paranode
C
Somatic spike
Axonal spike
Axonal spike
V
m
0
2
A
p
1
A
p
1
t = 160 s
t = 340 s
1 ms
)
s
m
V
m
/
(
t
d
/
V
d
)
s
m
A
p
(
/
t
d
/
I
d
)
s
m
A
p
(
/
t
d
/
I
d
pipette
40
20
0
-45
V (mV)
-35
5
0
-5
-0.1
2
-1
I (pA)
0.2
-4
-0.1
I (pA)
0.1
800
800
600
600
400
400
200
200
0
0
D
n
o
x
a
-
a
m
o
S
)
s
(
l
y
a
e
d
e
k
p
s
i
E
)
s
/
m
(
d
e
e
p
s
n
o
i
t
c
u
d
n
o
C
1.5
1
0.5
0
Bleb
P = 0.0306
Control CGS
P = 0.0017
Control CGS
Fig. 4. A2aRs in the node of Ranvier modulate conduction velocity. (A) Myelinated axon in Thy1-
Caspr-GFP mouse filled with Alexa Fluor 594 and patch-clamped at the cell soma and end-of-axon bleb.
(B) Average of >100 evoked action potentials in the soma and bleb in control conditions and while puffing
0.5 mM CGS 21680 at a node of Ranvier. Dashed lines show times of initiation of action potential derived
from threshold values of rate of change of voltage (dV/dt) or current (dI/dt) with time (see text).
(C) Phase plane plots showing times indicated in (B) (dots). (D) Response latency in bleb. (E) Conduction
velocities derived making assumptions discussed in the main text (closed circles assume spike starts
at the middle of the AIS and the forward speed is twice the backward speed; open circles assume
spike starts at the end of the AIS and the forward speed is three times faster than the backward speed).
Data in (A) to (C) are from mice; those in (D) and (E) combine data from rats and mice (neither
the initial speed nor the percentage change evoked by CGS 21680 differed significantly between rats
and mice, P = 0.43 and 0.15, respectively).
gray matter readily raised [Ca2+]i in astrocyte
processes near neuronal dendrites, presum-
ably as a result of classical neurotransmitters
or other signaling molecules acting on astro-
cytes, but more intense neuronal activity (or
conceivably a large number of neurons firing
at a low rate) was needed to raise it in the
processes near axons (Fig. 1, E and F). Thus,
the overall level of neuronal activity may
modulate the conduction speed of white
matter axons.
Changes in the ATP level in astrocytes or
neurons may also alter the extracellular aden-
osine concentration as a result of intracellu-
lar interconversion of ATP and adenosine and
export or import of adenosine across the cell
membrane by the equilibrative nucleoside
transporter. Across a wide area of neocortex,
the intracellular ATP level decreases during
neuronal activity, on passing from the awake
state to non–rapid eye movement (non-REM)
sleep, and on passing from non-REM sleep to
REM sleep (23). In the basal forebrain, the ac-
cumulation of extracellular adenosine when
awake has been proposed to generate pressure
to sleep (24), and this adenosine is likely to be
generated from ATP released by astrocytes
(25, 26). Although the arousal-modulating ef-
fects of adenosine have often been attributed
to it acting on presynaptic A1 receptors to sup-
press glutamate release at excitatory synapses,
genetic evidence suggests that in fact these
effects are generated by A2aRs (27). Thus, the
effects of astrocyte-derived adenosine on AIS
excitability and myelinated axon conduction
speed that we have characterized may be cru-
cial contributors to mediating changes between
wake and sleep states.
GABA, dopamine, and serotonin receptors reg-
ulate neuronal excitability at the AIS (15, 28, 29).
Our data demonstrate that the node of Ranvier
can have its electrical function rapidly altered
by substances released from surrounding cells.
We have shown this for astrocyte-released
ATP/adenosine, but oligodendrocyte precur-
sor cells and microglia could theoretically ex-
ert similar effects either by releasing ATP or
adenosine directly onto spike generation sites
or by responding to astrocytic ATP release by
releasing other gliotransmitters onto the axons.
Neural circuits depend crucially on action po-
tential arrival time for their function (30), and
cognition depends on oscillations of neural
firing probability that in turn depend on pro-
pagation time in myelinated neurons (31–33).
Thus, the decrease of the conduction speed of
myelinated axons that is evoked by adenosine
may change information processing and cog-
nition, when the adenosine level rises either
during a prolonged awake period or in re-
sponse to complex motor behavior and path-
ological conditions. Indeed, behavioral effects
linked to exploratory activity, aggression and
anxiety (34, 35), and pathological effects in
epilepsy, Parkinson’s disease, Alzheimer’s dis-
ease, depression, and autism have all been
linked to impairments of adenosine signaling
(36, 37). Our results make the testable predic-
tion that, when adenosine levels rise during
prolonged wakefulness or in complex motor
behavior and pathological conditions, the
conduction speed of myelinated axons should
decrease, resulting in delays to the action po-
tential arrival time at downstream synapses
and possible impairments of coincidence de-
tection or oscillatory firing generation.
Materials and Methods
Animals
Sprague-Dawley rats or transgenic mice at
postnatal days 28 to 32, housed on a 12-hour
light/12-hour dark cycle were used in all ex-
periments. At this age, myelination is ad-
vanced, but cells are still suitable for stable
patch-clamp recordings. Animals were killed
to make brain slices 4 hours after the room
light was turned on. Each experiment was per-
formed on brain slices from at least three ani-
mals and at least one of each sex. Expression
of A2aRs and HCN channels in sites of spike
generation appeared the same in mouse and
rat and in males and females. Animal proce-
dures were carried out in accordance with the
guidelines of the UK Animals (Scientific Proce-
dures) Act 1986 and subsequent amendments
(under Project License 70/8976). Thy1-Caspr-GFP
mice were generated as previously described (38).
Acute brain slice preparation
Coronal cortical slices (300 mm thick) were pre-
pared on a vibratome in ice-cold solution con-
taining (in mM): 93 N-methyl-D-glucamine
Lezmy et al., Science 374, eabh2858 (2021)
15 October 2021
5 of 10
RES EARCH | R E S E A R C H A R T I C L E
A
C
t
u
p
n
i
w
o
L
t
u
p
n
i
h
g
H
i
E
Active membranes
Passive membranes
Myelinated membranes
Nodes
Internodes
Bleb
0.15
0.3
Soma
Prox. Distal
AIS
gIh (mS/mm2)=
0
V
m
0
1
20 ms
Ih in distal AIS
Ih in nodes
B
)
V
m
(
a
m
o
s
t
a
t
s
e
r
V
- 75
- 80
- 85
0
0.04
0.08 0.12
gIh (mS/mm2)
D
Low input
High input
t
a
e
t
a
r
g
n
i
r
i
F
)
x
a
m
%
(
a
m
o
s
100
75
50
25
0
0.1
gIh at distal AIS (mS/mm2)
0.2
0
n
o
i
t
c
u
d
n
o
c
l
a
n
o
x
A
)
s
/
m
(
d
e
e
p
s
1.8
1.7
1.6
1.5
1.4
1.3
- 70
- 75
- 80
V
r
e
s
t
a
t
n
o
d
e
s
(
m
V
)
0.1 0.2 0.3 0.4 0.5
- 85
0
gIh at nodes (mS/mm2)
F
n
o
i
t
c
u
d
n
o
c
l
a
n
o
x
A
2.5
2
1.5
1
)
s
/
m
(
d
e
e
p
s
0.5
0
0.05
gIh at nodes (mS/mm2)
0.1
0.15
- 65
- 70
- 75
- 80
- 85
V
r
e
s
t
a
t
n
o
d
e
s
(
m
V
)
Fig. 5. Computational modeling predicts the adenosine-evoked decrease of axonal conduction speed.
(A) Schematic diagram of the model of the experiment in Fig. 4. (B) Adding different densities of adenosine-
sensitive maximal Ih conductance ((cid:1)gIh) to the distal AIS evokes a larger soma depolarization than adding it to the
nodes of Ranvier. Vertical dashed line shows measured maximal conductance (0.11 mS/mm2). (C) Voltage
response at soma to injecting 20 pA (top row) or 180 pA (bottom row) with three different levels of (cid:1)gIh added
(as indicated) to the distal AIS. As observed experimentally (Fig. 3, C to H), at low injected current, the action
potential frequency is increased; at high injected current, it is decreased (because of the simulated decrease of
action potential amplitude, we defined an action potential as occurring if the voltage crossed –50 mV). (D) Firing rate
as a function of (cid:1)gIh for simulations as in (C). (E) Action potential speed from the first to the last node in (A) and
average resting potential of the three nodes as a function of (cid:1)gIh added to each node (vertical dashed line shows the
estimated physiological value of 0.1565 mS/mm2; see the materials and methods). (F) Predictions of infinite axon
model for conduction speed and node resting potential as a function of (cid:1)gIh.
(NMDG) chloride, 2.5 KCl, 30 NaHCO3, 10 MgCl2,
1.2 NaH2PO4, 25 glucose, 0.5 CaCl2, 20 HEPES,
5 sodium ascorbate, 3 sodium pyruvate, and
1 kynurenic acid. The slices were incubated at
37°C in this solution for 20 min, then trans-
ferred to a similar solution containing (in mM):
93 NaCl, 1 MgCl2, and 2 CaCl2 instead of the
NMDG chloride, MgCl2, and CaCl2, and incu-
bated at room temperature until use. Experi-
ments were performed in aCSF containing (in
mM): 125 NaCl, 3 KCl, 26 NaHCO3, 2 MgCl2,
2 CaCl2, 1.25 NaH2PO4, and 10 glucose heated
to 37°C. All the solutions were gassed with
95% O2 and 5% CO2. In some experiments,
50 mM ZD7288, 100 nM ZM 241385, 20 mM D-
AP5, 10 mM DNQX, 50 mM MSPG, and 1 mM
NPS 2390 (all from Tocris Bioscience) were
added to the aCSF.
Patch-clamping of neuronal somata
Experiments were performed with an Olym-
pus BX51WI microscope under an Olympus
LUMPlanFI 40× lens. Layer V pyramidal cells
were identified by their location and morphol-
ogy. Microelectrodes with resistances of 5 to
6 MW were pulled from borosilicate glass capil-
laries (Harvard Apparatus) and filled with an
intracellular solution containing (in mM): 145
K-gluconate, 2 MgCl2, 0.5 H2-EGTA, 2 MgATP,
0.2 Na2GTP, and 10 HEPES, pH adjusted to
7.2 with KOH. Somata were patch-clamped in
whole-cell configuration, and the signals were
amplified using a Multiclamp 700B (Molecular
Devices), filtered at 4 kHz, and digitized at
50 kHz. The series resistance was <20 MW,
was compensated by ≥70%, and did not vary
by >10% during experiments. The mean input
resistance in whole-cell mode, after compen-
sation for series resistance, was 130.6 ± 25.9 MW
in control conditions and 84.8 ± 23.8 MW
during CGS 21680 application to the AIS (n =
6, P = 0.049). Recordings were acquired with
Clampex and analyzed with Clampfit soft-
ware (Molecular Devices). In current-clamp
mode, bridge balance and pipette capacitance
were corrected. To generate graphs of firing
rate versus injected current, current steps in
increments of 100 pA were injected for 1 s.
All membrane potentials were corrected for a
pipette liquid junction potential of –16 mV. To
analyze currents activated by hyperpolariza-
tion, tail currents were analyzed at –136 mV
obtained after 1 s pretest pulses in 10-mV in-
crements from –56 to –146 mV. The tail cur-
rents analyzed were primarily generated by
HCN channels (Ih current) because they were
significantly diminished by adding 50 mM
ZD7288 (Tocris Bioscience) (fig. S4). Tail cur-
rents were fitted with single exponential curves
that were extrapolated back to the end of the
pretest pulses. The resulting tail amplitude
values were fit to a function:
I = Imax + (Imin – Imax)/
{1 + exp[k(V – V1/2)]}
(1)
where Imax is the predicted tail amplitude at
–136 mV for a large depolarizing voltage step,
Imin is the tail amplitude for a large hyperpo-
larizing step, V1/2 is the voltage midpoint of
the activation curve, and k is a factor defining
the slope of the activation curve. The con-
ductance conferred by the maximally activated
current is given by
Gmax = (Imax – Imin)/[Vrev – (–136 mV)]
(2)
where the reversal potential of Ih was taken
as –23 mV (13). In some experiments, 50 mM
cAMP (Sigma-Aldrich) or 20 mM ZD7288 (Tocris
Bioscience) was added to the intracellular so-
lution. ZD7288 was previously shown to block
Ih when applied intracellularly (15).
Patch-clamp recording from axonal blebs
Experiments were performed using a Zeiss
LSM780 two-photon confocal microscope with
a W Plan-Apochromat 20× objective. Blebs are
patchable structures formed at the cut end of
axons at the surface of slices (39). Alexa Fluor
594 (ThermoFisher Scientific) was allowed time
to diffuse from the soma to the bleb. Action
potentials were recorded as a current in voltage-
clamp mode. To avoid phototoxicity, expo-
sures to excitation lights were briefly used only
to trace the axon and monitor the puffing area
(see below). High-resolution images of whole
axons were acquired after completion of the
electrophysiology experiments. Exposure to
the excitation lasers did not significantly affect
Lezmy et al., Science 374, eabh2858 (2021)
15 October 2021
6 of 10
RES EARCH | R E S E A R C H A R T I C L E
A
Alexa 594
Layer V soma
B
15
)
V
m
(
10
t
s
e
r
V
5
0
D
)
z
H
(
t
e
a
r
g
n
i
r
i
F
60
40
20
0
0
m
0
2
n
i
4
-
o
u
F
l
e
t
y
c
o
r
t
s
a
P = 0.0001
P = 0.0002
P < 0.0001
+Intra.
ZD
+ZM
C
50
40
30
20
10
)
z
H
(
e
a
r
t
g
n
i
r
i
F
Astro. Ca2+ uncaging
E
ZM
ZM+Astro. Ca2+ u.
60
0
2P laser
(No Ca2+ u.)
Intra. ZD
Intra. ZD
+Astro. Ca2+ u.
0
F
40
20
0
0
1000 1500
500
Current (pA)
1000 1500
500
Current (pA)
Control
Astro. Ca2+ uncaging
P = 0.0136
200 400 600 800
Current (pA)
40
30
20
10
0
0
Control
2P laser
(No Ca2+ u.)
400
800
Current (pA)
Alexa 594
Fluo-4
Alexa 488
50 m
Layer V soma
Astrocyte
Bleb
G
H
GFAP + Caspr
10 m
GFAP + Caspr
3 m
5 m
I
)
s
/
m
(
d
e
e
p
s
n
o
i
t
c
u
d
n
o
C
4
4
3
3
2
2
1
1
0
0
P = 0.0038
Control Astro. Ca2+
uncaging
Fig. 6. Ca2+ concentration rises in astrocytes regulate pyramidal cell excitability and axonal conduc-
tion speed. (A) Patch-clamped L5 pyramidal neuron (with Alexa Fluor 594 in the left pipette, red) and a
periaxonal astrocyte filled with the Ca2+ cage NP-EGTA and Fluo-4 (right pipette, green). [Ca2+]i rises in
astrocyte processes near the axon (see movie S3). (B) Resting potential (Vrest) depolarized after astrocyte
Ca2+ uncaging but not when blocking A2aRs with superfused ZM 241385 (100 nM) or Ih channels with ZD7288
(20 mM in the pipette) nor when two-photon laser excitation failed to raise [Ca2+]i (No Ca2+ u.; see fig. S2)
(Astro. Ca2+ uncaging: n = 6, +ZM: n = 4, +Intra. ZD: n = 5, 2P laser without uncaging: n = 3; one-way ANOVA,
P < 0.0001). (C) Neuronal firing rate evoked by injecting 1 s current steps in 100 pA increments before
(black) and after astrocytic Ca2+ uncaging (red). (D to F) As in (C) but with ZM 241385 [n = 4 (D)], or
ZD7288 [n = 4 (E)], or with illumination that failed to uncage Ca2+ [n = 3 (F)]. (G) Live imaging of an L5
pyramidal neuron patch-clamped at the soma (left pipette, loading red Alexa Fluor 594) and the axon end
(right pipette, green). A perinodal astrocyte (near axon branches) was patch-filled with NP-EGTA and Fluo-4.
(H) Middle: high-resolution image of the dashed box in (G). Left and right images show the areas in the
dashed boxes of the middle image after immunostaining for GFAP and Caspr. Nodes flanked by Caspr (green)
are close to GFAP-positive astrocyte processes (cyan). (I) Estimated axonal conduction speed before and
after astrocyte Ca2+ uncaging (see text associated with Fig. 4E for assumptions made). Data in (A) to (H) are
from rats; (I) combines data from rats and mice (neither the initial speed nor the percentage change evoked
by Ca2+ uncaging differed significantly between rats and mice, P = 0.08 and P = 0.93, respectively).
the waveform of action potentials. In Fig. 6, D
and E, where the effects of adenosine receptors
or HCN channels were blocked, there was no
change in action potential width measured
at the soma. For the first spike evoked by a
500-pA step, the full width at half-maximum
with ZM 241385 was 1.15 ± 0.13 msec before
and 1.18 ± 0.14 msec after Ca2+ uncaging (n = 4;
not significantly different, P = 0.12); with
ZD7288, it was 1.25 ± 0.17 msec before and
1.21 ± 0.13 msec after Ca2+ uncaging (n = 4;
not significantly different, P = 0.44). Similarly,
there was no difference in the action potential
width measured at the bleb with ZD7288 in-
side the neurons (fig. S6A); the full width at
half-maximum (the maximum being the am-
plitude of the inward current peak) was 0.53 ±
0.17 msec before and 0.53 ± 0.10 msec after Ca2+
uncaging (n = 5; not significantly different, P =
0.99). Microelectrodes with resistances of 8 to
9 MW were filled with aCSF and 100 mM Alexa
Fluor 488 (ThermoFisher Scientific). Blebs were
patch-clamped in cell-attached configuration,
and the signals were filtered at 2 kHz and
digitized at 50 kHz. Action potentials at the
soma were evoked and recorded, and action
currents at the axon bleb were recorded sim-
ultaneously. To accurately determine the delay
between the somatic and axonal signals, at
least 100 somatic spikes were evoked. First
derivatives of the somatic and axonal signals
were generated and then time-shifted to align
the peaks of the somatic spikes. The signals
were then averaged, and the soma-axon delay
was measured as the difference between the
onset of the axonal spike (time point recorded
before a current drop of >0.5 pA/ms) and the
somatic spike (time point recorded before an
increase of >1 mV/ms).
Local application of drugs to axons
CGS 21680 (500 nM; Cayman Chemical) or
100 mM adenosine (Sigma-Aldrich) were puff-
applied onto the AIS or nodes of Ranvier. The
drugs were added to aCSF with 100 mM Alexa
Fluor 594 to monitor the spread of the drugs.
Two or three MW pipettes were filled with this
solution and placed 10 mm away from the tar-
geted axon. Positive pressure was either ap-
plied with a microinjector (PMI-100, Dagan)
for 20 ms at 10 psi or controlled manually with
a syringe. In both cases, positive pressure was
calibrated and live monitored to eject drug
over a radius of ≤20 mm (from the tip of the
pipette set at saturating intensity to the edge of
detectable fluorescence). The flow of the per-
fusion was set to wash out the puffed drug in
the direction away from the targeted neuron.
The AIS was detected by adding 100 mM Alexa
Fluor 594 into the intracellular solution, and
its diffusion into the axon was live monitored.
The pipette puffing onto the AIS was placed
near the distal AIS 25 to 30 mm away from the
soma. The pipette puffing onto the nodes was
placed near an axonal branch (detected from
the Alexa Fluor 594 fill) or green fluorescent
protein (GFP) signal detected along the axon
of a Thy1-Caspr-GFP mouse. Caspr was post
hoc immunolabeled to confirm the presence
of nodes at the puffing sites.
Patch-clamping, calcium uncaging, and imaging
of astrocytes
Astrocytes were detected by their morphology
and with the selective marker sulforhodamine
101 (SR 101). Before the experiments, the
slices were incubated with 1 mM SR101 (Tocris
Bioscience) added to the slicing solution for
20 min at 37°C. The presence of astrocytes was
confirmed by post hoc fixing slices and im-
munolabeling for GFAP. Microelectrodes with
resistances of 7 to 8 MW were filled with a K-
methylsulfate intracellular solution containing
(in mM): 100 KMeSO4, 50 KCl, 2 MgCl2, 4 MgATP,
Lezmy et al., Science 374, eabh2858 (2021)
15 October 2021
7 of 10
RES EARCH | R E S E A R C H A R T I C L E
0.3 Na2GTP, and 10 HEPES. Fluo-4 (100 mM;
ThermoFisher Scientific) was added to im-
age calcium changes, and 5 mM NP-EGTA
(ThermoFisher Scientific) was added to allow
uncaging of calcium in single astrocytes. Im-
ages were acquired at 2 Hz with a confocal
argon laser, and NP-EGTA photolysis was
obtained by applying two-photon excitation
at 720 nm to the soma (10 iterations, 2.55 ms
pixel dwell time). Laser intensity was set at
10 mW and, if needed, was increased grad-
ually until a calcium concentration rise was
evoked (fig. S2). Intensities just below those
needed to evoke a detectable [Ca2+]i rise were
used as control experiments to check that two-
photon illumination alone was not respon-
sible for the release of ATP and adenosine
generation observed. In the experiments ex-
amining calcium activity at astrocyte processes,
images were acquired at the plane of the axon
or of a dendrite contacting astrocyte processes.
Regions of interest (ROIs) were selected and
[Ca2+]i changes were measured as DF/F after
background subtraction. At the start of the
experiments, the baseline astrocyte Ca2+ activ-
ity was recorded for 10 s, and, over this brief
period, no spontaneous Ca2+ transients were
observed.
Immunohistochemistry
Brains from Thy1-Caspr-GFP mice were perfusion-
fixed in 4% paraformaldehyde (PFA) in 0.01 M
phosphate-buffered saline (PBS), and fixed tis-
sue was then cut into 70-mm-thick slices. Alter-
natively, 300-mm-thick acute slices from mouse
and rat brains were immersion-fixed in a solu-
tion containing 4% PFA, 4% sucrose, and 0.1 M
PBS. The slices were permeabilized with 0.2%
Triton X-100 (Sigma-Aldrich) in a blocking so-
lution (10% goat serum in 0.01 M PBS) for
1 hour. The slices were then incubated over-
night at 4°C with the following primary anti-
bodies, as required: rabbit anti-A2aR (Abcam,
ab3461, 1:100), mouse anti-A2aR [Millipore,
05-717, which has been validated as giving no
labeling in A2aR KO tissue (40), 1:200], rabbit
anti-A1R (Abcam, ab82477, 1:100), rabbit anti-
A2bR (Cohesion, CPA3755, 1:100), mouse anti-
Ankyrin G (Neuromab, N106/36, 1:500), mouse
anti-panNav (Sigma-Aldrich, K58/35, 1:100),
mouse anti-Caspr (Neuromab, K65/35, 1:100),
chicken anti-GFP (Millipore, AB16901, 1:1000),
rabbit anti-GFAP (Millipore, AB5804, 1:500),
rabbit anti-HCN1 (Alomone, APC 056, 1:100),
and rabbit anti-HCN2 (Alomone, APC 030, 1:100).
The slices were then incubated for 2 hours at
room temperature with the following sec-
ondary antibodies, as required (ThermoFisher
Scientific, 1:500): anti-chicken or anti-mouse
Alexa Fluor 488, anti-rabbit or anti-mouse Alexa
Fluor 546, and anti-rabbit Alexa Fluor 647. The
slices were incubated with 4′,6-diamidino-2-
phenylindole nuclear stain (1:50,000 in PBS
from a 5 mg/ml stock concentration) for 10 min
and mounted in Dako fluorescent mounting
medium. Slices were washed with 0.01 M PBS
three times with 10 min between each step.
Negative control experiments were also per-
formed to check for any labeling caused by
unspecific binding of the secondary antibod-
ies. For this, we followed the same protocol
but omitted the incubation with the primary
antibodies. Slices were imaged using a Zeiss
LSM700 confocal microscope with a Zeiss Plan-
Apochromat 63× oil immersion lens, and im-
ages were acquired with ZEN Microscope
Software (Zeiss). Z-stacks of 1-mm intervals
were imaged, and their maximum intensity
was projected using ImageJ (FIJI). A line was
drawn along the axons to plot intensity pro-
files. To plot node of Ranvier profiles, the
width of the line was adjusted to fit the width
of the Caspr-labeled paranodes. To analyze
GFAP staining running parallel to axons, the
width of the line was adjusted to 5 mm and
centered on the axon.
Quantification of putative
ATP-containing vesicles
Before the experiments, the slices were incu-
bated with 20 mM quinacrine dihydrochloride
(Sigma-Aldrich) added to the slicing solution
for 25 min at 37°C. Quinacrine is a green se-
lective marker for ATP-containing vesicles (7)
or lysosomes (41). Calcium was uncaged and
imaged from single astrocytes as described
above using the red Ca2+ indicator Rhod-2
(50 mM, ThermoFisher Scientific) instead of
Fluo-4. Images of astrocyte processes with
ATP vesicles were acquired before and after
Ca2+ uncaging with two-photon excitation at
720 nm at the soma. Adjacent areas were also
imaged. ROIs located a similar distance away
from the astrocyte soma but without any de-
tectable dye-filled astrocyte processes within a
5 mm perimeter minimum were selected. With
ImageJ, a constant threshold was applied to
the images and particle analysis was used to
quantify the number of ATP vesicles. The mean
Feret’s diameter of ATP vesicles was 0.56 ± 0.03
mm (n = 2227), similar to previous reports (42).
Detection of extracellular ATP
Changes in extracellular ATP level were de-
tected with a luciferin/luciferase–based chem-
iluminescence assay emitting green light.
Perfusion of brain slices was halted and 50-ml
aliquots containing 12 mg/ml luciferin (Sigma-
Aldrich) and 5 mg/ml luciferase (Sigma-Aldrich)
were added to a 1-ml bath chamber filled with
aCSF at room temperature (because luciferase
is not stable at higher temperatures). Calcium
was uncaged and imaged from single astro-
cytes as described above using the red Ca2+
indicator Rhod-2 (50 mM) instead of Fluo-4. A
field of view containing an astrocyte and its
processes was selected and ATP-derived bio-
luminescence was collected in darkness using
highly sensitive GaAsP detectors with a Zeiss
LSM780 two-photon/confocal microscope.
Estimating axonal conduction velocity
We estimated axonal conduction velocities on
the basis of parameters acquired experimen-
tally from dual soma and axon bleb patch-
clamp recordings, live confocal imaging, and
post hoc immunostaining of layer V pyramidal
neurons. In accordance with previous studies,
two assumptions were made: (i) spikes initiate
at the distal half of the AIS (the AIS being
defined as being from the edge of the soma to
the start of the first Caspr-labeled paranode)
because in the distal AIS Na+ channels are
found in high density, have a low voltage thresh-
old, and are isolated from the capacitive load
of the soma (43, 44) and (ii) forward conduc-
tion speed is two to three times faster than
backward conduction speed because of the
lower Na+ channel density in the proximal AIS
and the capacitive load of the soma (16, 44).
The axonal conduction speed can be then cal-
culated as follows:
xbleb (cid:2) 2xsoma
T
xbleb (cid:2) 3xsoma
T
or v ¼
v ¼
ð3Þ
where v is the speed from the spike initiation
site to the axon bleb, T is the delay time from
the somatic signal to the bleb signal (obtained
from dual patch-clamp recordings), and xbleb
and xsoma are the distances from the spike in-
itiation site to the recording sites on the axon
bleb or on the soma, respectively (obtained
from confocal imaging by assuming a posi-
tion for spike initiation).
Computer simulations of infinite corpus
callosal axon
As in our previous studies (17, 45, 46), action
potential conduction along myelinated axons
was simulated using MATLAB. Electrophysio-
logical parameters were based on the finite
impedance double-cable model (model C) of
Richardson et al. (47), except that the mem-
brane capacitance was taken as the physiolog-
ically measured value of 0.9 mF/cm2 (48). The
differential equations of the model were de-
rived and solved as in the myelinated axon
model of Halter and Clark (49), in which the
axon is divided into compartments represent-
ing the node, paranode, and internode. The
MATLAB code, including all parameters and
equations, is available (see the Acknowledg-
ments). Conduction speed simulations of long
callosal-like axons were performed as de-
scribed previously (17), where the parameters
used were based on experimental data. The
speed was measured between nodes 20 and
30 in a uniform axon containing 51 nodes
and 50 internodes of constant lengths (nodes:
1.5 mm, internodes: 81.7 mm) and diameters
(nodes: 0.64 mm, internodes: 0.73 mm). The
periaxonal space thickness, which has an
Lezmy et al., Science 374, eabh2858 (2021)
15 October 2021
8 of 10
RES EARCH | R E S E A R C H A R T I C L E
important effect on conduction speed (17, 50),
was set at 15 nm except in the paranodes (the
1.9-mm end parts of the internodes), where the
width was reduced to 0.0123 nm (17). As-
suming a myelin wrap periodicity of 15.6 nm,
five myelin wraps were needed to set the
g-ratio close to 0.8. Each node expressed fast
voltage-gated Na+ (NaV) channels, persistent
NaV channels, and slow voltage-gated potassium
(KV) channels at fully activated conductance
densities of 10, 0.01, and 0.4 mS/mm2, respective-
ly, as well as adenosine-gated Ih channels at a
density of either 0 (for no A2aR activation) or
0.1565 (to mimic A2aR activation) mS/mm2.
This density was experimentally derived as
follows: the maximum Ih conductance (derived
from tail current analysis as in Fig. 3, I and J)
evoked in the ~10-mm length of the distal AIS
was divided by the area of the distal AIS to
obtain a fully-activated conductance density of
0.111 mS/mm2. Immunohistochemical labeling
of A2aRs then showed their density at nodes
of Ranvier to be a factor of 1.41 higher than
in the distal AIS (calculated by integrating
the background-subtracted fluorescent sig-
nal across a small length w of the axon and
dividing by p•w•d, where d is the diameter of
the axon), implying a density at the nodes of
0.1565 mS/mm2. The node resting potential in
the absence of adenosine was set to –82 mV by
adjusting the magnitude of a leak conductance
(to 0.113 mS/mm2) with a reversal potential of
–84 mV (the reversal potential of the other K+
currents in the model). The leak current rep-
resents K+ leak channels such as TRAAK present
in the nodes of Ranvier (51, 52).
Computer simulations of patch-clamped neuron
with soma, AIS, and three internodes
This model was then modified so that the
dimensions of a specimen patch-clamped and
imaged neuron acquired for the present study
(Fig. 4) were implemented (Fig. 5A). The mod-
el included “nodes” representing the soma, the
proximal AIS, the distal AIS, three nodes of
Ranvier, and a terminating bleb. To represent
the area and thus the capacitance of the real
soma, the diameter and length of the cylindri-
cal “node” representing the soma were both
set to 18 mm. The proximal and distal AIS had
lengths of 21.9 and 10 mm, respectively, and
diameters (measured experimentally as the
average over each of these two zones) of 1.046
and 0.525 mm respectively. The soma and pro-
ximal AIS, and the proximal and distal AIS,
were separated by extremely small (10−7 mm
long) “internodes” (which are essential for the
alternating node-internode model to function
in MATLAB; not shown in Fig. 5A) with dia-
meters of 1.046 and 0.79 mm, respectively. Dis-
tal to the AIS are three internodes of lengths
24.91, 52.44, and 39.20 mm and diameters of
0.51, 0.56, and 0.53 mm respectively, each fol-
lowed by nodes of Ranvier of lengths 1.126,
0.915, and 0.414 mm, and diameters of 0.445,
0.452, and 0.401 mm, respectively. The third
node of Ranvier was followed by an extremely
small (10−7 mm long) “internode” of diameter
0.4 mm and then a terminal bleb of length
1.64 mm and diameter 3 mm. Each node of
Ranvier expressed the same conductance as
in the infinite cable model. The proximal and
distal AIS compartments expressed fast NaV,
persistent NaV, and slow KV channels at the
same densities as in the infinite cable model,
and in addition a low threshold KV channel
with a maximum conductance of 0.4 mS/mm2,
which was needed to allow the generation of
repetitive action potential trains in response
to current injected at the soma. Adenosine-
insensitive Ih conductance (with a half-maximum
activation voltage of –103.5 mV, as found on
average in experiments as in Fig. 3J) was ex-
pressed in the soma and proximal AIS (as seen
immunohistochemically for HCN1 channels)
at a density of 0.0033 mS/mm2 (set so as to
reproduce the mean Ih maximal conductance
seen in the absence of adenosine in Fig. 3, I
and J), whereas adenosine-sensitive Ih was
present in the distal AIS at a density of
0.111 mS/mm2 and in the nodes of Ranvier at a
density 1.41-fold higher (see above). The den-
sity used for the distal AIS was set to increase
the maximum Ih conductance measured at the
soma by 51% as observed during activation
of A2aRs in the distal AIS (Fig. 3, I and J), and
the midpoint of its activation curve was at
–80 mV to reproduce the positive shift for
the total Ih apparent activation curve in Fig.
3J. The terminal bleb expressed no voltage-
gated conductances. To match the model to
the experimental data, we wanted to repro-
duce the 5.8-mV depolarization recorded at
the soma when puffed adenosine activated
Ih in the distal AIS (Fig. 3E). We found that if
we set the resting potential to –82 mV in the
soma and AIS using a leak conductance with a
reversal potential of –84 mV (as in the nodes),
then this introduced too much resting con-
ductance, reducing the depolarization evoked
at the soma when A2aRs were activated in the
distal AIS to only +1.55 mV. Accordingly, in the
soma, AIS, and terminal bleb, we used a no-
minal leak reversal potential of –4000 mV,
converting the leak into an essentially voltage-
independent (~zero conductance) current, sim-
ilar conceptually to the constant outward cur-
rent generated by the Na+/K+ pump. The
adenosine-sensitive Ih (when activated in the
distal AIS) then generated a depolarization of
5.9 mV at the soma. To investigate the effect of
the adenosine-evoked Ih current, the action
potential speed was measured between the
first and the last nodes of Ranvier when cur-
rent was injected briefly (1 nA for 100 msec) at
the soma. The model reproduced reasonably
well the assumptions that we made to assess
the conduction speed in our experiments: In
the simulations, spikes evoked by current in-
jection at the soma were initiated in the distal
AIS “node”, the forward conduction speed was
faster than the backward speed (from distal
AIS to bleb: 1.33 m/s; from distal AIS to soma:
0.23 m/s), and the conduction speed from the
distal AIS to the bleb (1.33 m/s) was in the
range estimated from the experiments in Figs.
4E and 6I (1.31 to 2.23 m/s).
Code availability
The computer code used for the simula-
tions described above has been deposited in
Zenodo (53).
Statistics
Statistical analyses and graphs were performed
with GraphPad Prism 6. Data are presented as
mean ± SEM. Data normality was assessed with
D’Agostino & Pearson omnibus or Kolmogorov-
Smirnov tests. Comparisons of normally distrib-
uted two groups were made using two-tailed
Student t tests. When treatments involved
more than two independent groups, statisti-
cal comparisons were performed with one-
way ANOVA and post hoc Tukey’s multiple
comparisons. Data that were not normally dis-
tributed were analyzed with Mann-Whitney
test or Wilcoxon matched-pairs signed rank
test. Assessment of whether the slope of linear
regressions differed significantly from zero
was obtained using the t statistic for the slope.
P values within any figure panel were adjusted
for multiple comparisons.
Gender and species effects
Mice were used in experiments when it was
essential to have the nodes labeled with Caspr-
GFP. Rats were also used because they are
more generally available at all ages in our uni-
versity. We observed no differences between
the two species, for example, in
(i) the presence of A2aRs and HCN channels
at nodes of Ranvier and the AIS;
(ii) the initial speed and the percentage
change evoked by CGS 21680 puff onto nodes,
which did not differ significantly between rats
and mice (P = 0.43 and P = 0.15, respectively);
and
(iii) the initial speed and the percentage
change evoked by Ca2+ uncaging in perinodal
astrocytes, which did not differ significantly
between rats and mice (P = 0.08 and P = 0.93,
respectively).
Animals of both sexes were used because of
increasing awareness of the fact that gender
can play a role in determining function. For
major results, we found no differences be-
tween the two sexes, for example, in
(i) the presence of A2aRs and HCN channels
at nodes of Ranvier and the AIS;
(ii) the initial resting membrane potential
and the change evoked by CGS 21680 puff
onto the AIS, which did not differ significantly
Lezmy et al., Science 374, eabh2858 (2021)
15 October 2021
9 of 10
RES EARCH | R E S E A R C H A R T I C L E
between males and females (P = 0.69 and 0.75
respectively); and
(iii) the initial speed and the percentage
change evoked by CGS 21680 puff onto nodes,
which did not differ significantly between
males and females (P = 0.3 and P = 0.45,
respectively).
RE FE RENCES AND N OT ES
1. N. J. Allen, Astrocyte regulation of synaptic behavior.
Annu. Rev. Cell Dev. Biol. 30, 439–463 (2014). doi: 10.1146/
annurev-cellbio-100913-013053; pmid: 25288116
2. M. H. P. Kole et al., Action potential generation requires a high
sodium channel density in the axon initial segment.
Nat. Neurosci. 11, 178–186 (2008). doi: 10.1038/nn2040;
pmid: 18204443
3. V. A. Larson et al., Oligodendrocytes control potassium
accumulation in white matter and seizure susceptibility. eLife
7, e34829 (2018). doi: 10.7554/eLife.34829; pmid: 29596047
4. A. S. Saab, I. D. Tzvetanova, K.-A. Nave, The role of myelin
and oligodendrocytes in axonal energy metabolism. Curr. Opin.
Neurobiol. 23, 1065–1072 (2013). doi: 10.1016/
j.conb.2013.09.008; pmid: 24094633
5. C. Ffrench-Constant, R. H. Miller, J. Kruse, M. Schachner,
M. C. Raff, Molecular specialization of astrocyte processes at
nodes of Ranvier in rat optic nerve. J. Cell Biol. 102, 844–852
(1986). doi: 10.1083/jcb.102.3.844; pmid: 2419343
6. A. Araque et al., Gliotransmitters travel in time and space.
Neuron 81, 728–739 (2014). doi: 10.1016/
j.neuron.2014.02.007; pmid: 24559669
7. S. Forveille et al., Quinacrine-mediated detection of
intracellular ATP. Methods Enzymol. 629, 103–113 (2019).
doi: 10.1016/bs.mie.2019.05.050; pmid: 31727236
8. T. Pfeiffer, D. Attwell, Brain’s immune cells put the brakes on
neurons. Nature 586, 366–367 (2020). doi: 10.1038/
d41586-020-02713-7; pmid: 32999439
9. K. Baalman et al., Axon initial segment-associated microglia.
J. Neurosci. 35, 2283–2292 (2015). doi: 10.1523/
JNEUROSCI.3751-14.2015; pmid: 25653382
10. J. Zhang, X. Yang, Y. Zhou, H. Fox, H. Xiong, Direct contacts of
microglia on myelin sheath and Ranvier’s node in the
corpus callosum in rats. J. Biomed. Res. 33, 192–200 (2019).
pmid: 30403198
11. D. DiFrancesco, P. Tortora, Direct activation of cardiac
pacemaker channels by intracellular cyclic AMP. Nature 351,
145–147 (1991). doi: 10.1038/351145a0; pmid: 1709448
12. G. Vargas, M. T. Lucero, Dopamine modulates inwardly
rectifying hyperpolarization-activated current (Ih) in cultured
rat olfactory receptor neurons. J. Neurophysiol. 81, 149–158
(1999). doi: 10.1152/jn.1999.81.1.149; pmid: 9914276
13. N. Byczkowicz et al., HCN channel-mediated neuromodulation
can control action potential velocity and fidelity in central
axons. eLife 8, e42766 (2019). doi: 10.7554/eLife.42766;
pmid: 31496517
21. B. S. Khakh, Astrocyte-neuron interactions in the striatum: Insights
on identity, form, and function. Trends Neurosci. 42, 617–630
(2019). doi: 10.1016/j.tins.2019.06.003; pmid: 31351745
22. N. Hamilton et al., Mechanisms of ATP- and glutamate-
mediated calcium signaling in white matter astrocytes. Glia 56,
734–749 (2008). doi: 10.1002/glia.20649; pmid: 18293404
23. A. Natsubori et al., Intracellular ATP levels in mouse cortical
excitatory neurons varies with sleep-wake states. Commun. Biol. 3,
491 (2020). doi: 10.1038/s42003-020-01215-6; pmid: 32895482
24. T. Porkka-Heiskanen et al., Adenosine: A mediator of the
sleep-inducing effects of prolonged wakefulness. Science 276,
1265–1268 (1997). doi: 10.1126/science.276.5316.1265;
pmid: 9157887
25. M. M. Halassa et al., Astrocytic modulation of sleep
homeostasis and cognitive consequences of sleep loss.
Neuron 61, 213–219 (2009). doi: 10.1016/
j.neuron.2008.11.024; pmid: 19186164
26. L. I. Schmitt, R. E. Sims, N. Dale, P. G. Haydon, Wakefulness
affects synaptic and network activity by increasing
extracellular astrocyte-derived adenosine. J. Neurosci. 32,
4417–4425 (2012). doi: 10.1523/JNEUROSCI.5689-11.2012;
pmid: 22457491
27. Z. L. Huang et al., Adenosine A2A, but not A1, receptors
mediate the arousal effect of caffeine. Nat. Neurosci. 8,
858–859 (2005). doi: 10.1038/nn1491; pmid: 15965471
28. W. Wefelmeyer, D. Cattaert, J. Burrone, Activity-dependent
mismatch between axo-axonic synapses and the axon initial
segment controls neuronal output. Proc. Natl. Acad. Sci. U.S.A.
112, 9757–9762 (2015). doi: 10.1073/pnas.1502902112;
pmid: 26195803
29. K. J. Bender, C. P. Ford, L. O. Trussell, Dopaminergic
modulation of axon initial segment calcium channels regulates
action potential initiation. Neuron 68, 500–511 (2010).
doi: 10.1016/j.neuron.2010.09.026; pmid: 21040850
30. P. König, A. K. Engel, W. Singer, Integrator or coincidence
detector? The role of the cortical neuron revisited. Trends
Neurosci. 19, 130–137 (1996). doi: 10.1016/S0166-2236(96)
80019-1; pmid: 8658595
31. E. Salinas, T. J. Sejnowski, Correlated neuronal activity and the
flow of neural information. Nat. Rev. Neurosci. 2, 539–550
(2001). doi: 10.1038/35086012; pmid: 11483997
32. C. Adaikkan, L.-H. Tsai, Gamma entrainment: Impact on
neurocircuits, glia, and therapeutic opportunities. Trends
Neurosci. 43, 24–41 (2020). doi: 10.1016/j.tins.2019.11.001;
pmid: 31836315
33. S. Pajevic, P. J. Basser, R. D. Fields, Role of myelin plasticity in
oscillations and synchrony of neuronal activity. Neuroscience
276, 135–147 (2014). doi: 10.1016/j.neuroscience.2013.11.007;
pmid: 24291730
34. C. Ledent et al., Aggressiveness, hypoalgesia and high blood
pressure in mice lacking the adenosine A2a receptor. Nature
388, 674–678 (1997). doi: 10.1038/41771; pmid: 9262401
35. M. P. Kaster et al., Caffeine acts through neuronal adenosine
A2A receptors to prevent mood and memory dysfunction
triggered by chronic stress. Proc. Natl. Acad. Sci. U.S.A.
112, 7833–7838 (2015). doi: 10.1073/pnas.1423088112;
pmid: 26056314
14. F. C. Roth, H. Hu, An axon-specific expression of HCN channels
36. Y.-J. Liu et al., Research progress on adenosine in central
catalyzes fast action potential signaling in GABAergic
interneurons. Nat. Commun. 11, 2248 (2020). doi: 10.1038/
s41467-020-15791-y; pmid: 32382046
15. K. W. Ko, M. N. Rasband, V. Meseguer, R. H. Kramer,
N. L. Golding, Serotonin modulates spike probability in the
axon initial segment through HCN channels. Nat. Neurosci. 19,
826–834 (2016). doi: 10.1038/nn.4293; pmid: 27110919
16. Y. Yu, Y. Shu, D. A. McCormick, Cortical action potential
backpropagation explains spike threshold variability and
rapid-onset kinetics. J. Neurosci. 28, 7260–7272 (2008).
doi: 10.1523/JNEUROSCI.1613-08.2008; pmid: 18632930
I. L. Arancibia-Cárcamo et al., Node of Ranvier length as a
potential regulator of myelinated axon conduction speed. eLife
6, e23329 (2017). doi: 10.7554/eLife.23329; pmid: 28130923
17.
18. A. Lörincz, T. Notomi, G. Tamás, R. Shigemoto, Z. Nusser,
Polarized and compartment-dependent distribution of HCN1 in
pyramidal cell dendrites. Nat. Neurosci. 5, 1185–1193 (2002).
doi: 10.1038/nn962; pmid: 12389030
19. B. J. Wainger, M. DeGennaro, B. Santoro, S. A. Siegelbaum,
G. R. Tibbs, Molecular mechanism of cAMP modulation
of HCN pacemaker channels. Nature 411, 805–810 (2001).
doi: 10.1038/35081088; pmid: 11459060
20. N. B. Hamilton, D. Attwell, Do astrocytes really exocytose
neurotransmitters? Nat. Rev. Neurosci. 11, 227–238 (2010).
doi: 10.1038/nrn2803; pmid: 20300101
nervous system diseases. CNS Neurosci. Ther. 25, 899–910
(2019). doi: 10.1111/cns.13190; pmid: 31334608
37. D. J. Hines, P. G. Haydon, Astrocytic adenosine: From synapses
to psychiatric disorders. Philos. Trans. R. Soc. Lond. B Biol. Sci.
369, 20130594 (2014). doi: 10.1098/rstb.2013.0594;
pmid: 25225088
38. V. Brivio, C. Faivre-Sarrailh, E. Peles, D. L. Sherman,
P. J. Brophy, Assembly of CNS nodes of Ranvier in myelinated
nerves is promoted by the axon cytoskeleton. Curr. Biol. 27,
1068–1073 (2017). doi: 10.1016/j.cub.2017.01.025;
pmid: 28318976
39. Y. Shu, A. Hasenstaub, A. Duque, Y. Yu, D. A. McCormick,
Modulation of intracortical synaptic potentials by presynaptic
somatic membrane potential. Nature 441, 761–765 (2006).
doi: 10.1038/nature04720; pmid: 16625207
40. A. G. Orr et al., Astrocytic adenosine receptor A2A and
Gs-coupled signaling regulate memory. Nat. Neurosci.
18, 423–434 (2015). doi: 10.1038/nn.3930; pmid: 25622143
41. Z. Zhang et al., Regulated ATP release from astrocytes through
lysosome exocytosis. Nat. Cell Biol. 9, 945–953 (2007).
doi: 10.1038/ncb1620; pmid: 17618272
42. C. Cserép et al., Microglia monitor and protect neuronal
43. W. Hu et al., Distinct contributions of Na(v)1.6 and Na(v)1.2 in
action potential initiation and backpropagation. Nat. Neurosci.
12, 996–1002 (2009). doi: 10.1038/nn.2359; pmid: 19633666
44. G. Baranauskas, Y. David, I. A. Fleidervish, Spatial mismatch
between the Na+ flux and spike initiation in axon initial
segment. Proc. Natl. Acad. Sci. U.S.A. 110, 4051–4056 (2013).
doi: 10.1073/pnas.1215125110; pmid: 23341597
45. M. C. Ford et al., Tuning of Ranvier node and internode properties in
myelinated axons to adjust action potential timing. Nat. Commun.
6, 8073 (2015). doi: 10.1038/ncomms9073; pmid: 26305015
46. Y. Bakiri, R. Káradóttir, L. Cossell, D. Attwell, Morphological and
electrical properties of oligodendrocytes in the white matter of
the corpus callosum and cerebellum. J. Physiol. 589, 559–573
(2011). doi: 10.1113/jphysiol.2010.201376; pmid: 21098009
47. A. G. Richardson, C. C. McIntyre, W. M. Grill, Modelling the
effects of electric fields on nerve fibres: Influence of the myelin
sheath. Med. Biol. Eng. Comput. 38, 438–446 (2000).
doi: 10.1007/BF02345014; pmid: 10984943
48. L. J. Gentet, G. J. Stuart, J. D. Clements, Direct measurement
of specific membrane capacitance in neurons. Biophys. J. 79,
314–320 (2000). doi: 10.1016/S0006-3495(00)76293-X;
pmid: 10866957
49. J. A. Halter, J. W. Clark Jr.., A distributed-parameter model of
the myelinated nerve fiber. J. Theor. Biol. 148, 345–382 (1991).
doi: 10.1016/S0022-5193(05)80242-5; pmid: 2016898
50. K. M. Young et al., Oligodendrocyte dynamics in the healthy adult
CNS: Evidence for myelin remodeling. Neuron 77, 873–885
(2013). doi: 10.1016/j.neuron.2013.01.006; pmid: 23473318
51. H. Kanda et al., TREK-1 and TRAAK are principal K+ channels at the
nodes of Ranvier for rapid action potential conduction on
mammalian myelinated afferent nerves. Neuron 104, 960–971.e7
(2019). doi: 10.1016/j.neuron.2019.08.042; pmid: 31630908
52. S. G. Brohawn et al., The mechanosensitive ion channel TRAAK is
localized to the mammalian node of Ranvier. eLife 8, e50403
(2019). doi: 10.7554/eLife.50403; pmid: 31674909
53. Code used in the simulations for: J. Lezmy, I. L. Arancibia-Cárcamo,
T. Quintela-López, D. L. Sherman, P. J. Brophy, D. Attwell,
Astrocyte Ca2+-evoked ATP release regulates myelinated axon
excitability and conduction speed, Zenodo (2021);
doi: 10.5281/zenodo.5177076
54. Code used to simulate the action potential in myelinated axons for:
J. Lezmy, I. L. Arancibia-Cárcamo, T. Quintela-López,
D. L. Sherman, P. J. Brophy, D. Attwell, Astrocyte Ca2+-evoked ATP
release regulates myelinated axon excitability and conduction
speed, Github (2021); https://github.com/AttwellLab/
MyelinatedAxonModel.
AC KNOWLED GME NTS
We thank A. Gourine, M. Häusser, C. Madry, D. Rusakov, M. Shah,
H. Shiina, and V. Vyazovskiy for comments on the manuscript.
Funding: This work was supported by European Research Council
(BrainEnergy) and Wellcome Trust Investigator Awards (099222/Z/
12/Z and 219366/Z/19/Z) to D.A., an EMBO fellowship (ALTF
430-2019) to J.L., and a Wellcome Trust Investigator Award (107008)
to P.J.B. This research was funded in part by grants 099222/Z/12/Z
and 219366/Z/19/Z from the Wellcome Trust, a cOAlition S
organization. The authors will make the Author Accepted Manuscript
(AAM) version available under a CC BY public copyright license.
Author contributions: J.L. performed all electrophysiological and
live imaging experiments. T.Q. and J.L. performed the
immunohistochemistry experiments. L.A., J.L., and D.A. performed
computer simulations. D.L.S. and P.J.B. produced the Thy1-Caspr-GFP
mouse. J.L. and D.A. conceived the project, analyzed the data, and
wrote the first draft of the manuscript. All other authors commented on
the manuscript. Competing interests: The authors declare no
competing financial interests. Data and materials availability: All data
are available in the manuscript or the supplementary materials. Thy1-
Caspr-GFP mice are available under a material transfer agreement with
the University of Edinburgh from P.J.B. and D.L.S. Code used for
simulations is freely available (53, 54).
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abh2858
Figs. S1 to S7
Movies S1 to S3
Reference (55)
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
function through specialized somatic purinergic junctions.
Science 367, 528–537 (2020). doi: 10.1126/science.aax6752;
pmid: 31831638
2 March 2021; resubmitted 5 July 2021
Accepted 1 September 2021
10.1126/science.abh2858
Lezmy et al., Science 374, eabh2858 (2021)
15 October 2021
10 of 10
|
10.1126_science.abg0879
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
SPLICEOSOME
Structure of the activated human minor spliceosome
Rui Bai*, Ruixue Wan*†, Lin Wang, Kui Xu, Qiangfeng Zhang, Jianlin Lei, Yigong Shi†
INTRODUCTION: Precursor messenger RNA (pre-
mRNA) splicing, which involves the removal of
noncoding introns and the ligation of the coding
exons, is achieved by the spliceosome. An intron
is defined by a 5′ splice site (5′SS), a branch point
sequence (BPS), and a 3′ splice site (3′SS). Most
introns belong to the U2 type and are removed
by the major spliceosome that contains the
U2 small nuclear RNA (snRNA). A very small
percentage of introns are the U12 type, which
is characterized by a distinct 5′SS and BPS.
The U12-type genes, present in all major
eukaryotic taxa, play an essential role in
development. The U12-type intron is removed
by the minor spliceosome, which contains five
snRNAs: U11, U12, U4atac, U5, and U6atac. Of
these snRNAs, only U5 is shared between the
major and minor spliceosomes.
Because of its scarcity in cells, the minor
spliceosome has represented a challenge for
biochemical studies. Before now, there was no
structural information or published protocol
on the purification of the minor spliceosome.
It was unclear how many U12-specific protein
components are present in each major func-
tional state of the minor spliceosome and how
they may facilitate the splicing reaction. We
did not even know whether rules derived from
the major spliceosome could apply to the
minor spliceosome.
RATIONALE: To address these questions, we
need to both develop a protocol for the assembly
and purification of the minor spliceosome and
determine the structure of the minor spliceo-
some at resolutions that give sufficiently de-
tailed structural features. Structural comparison
between the major and minor spliceosomes
may reveal valuable information on protein com-
ponents specific to either spliceosome, U12-type
intron recognition, snRNA conformation, active
site configuration, and regulatory mechanisms.
RESULTS: We replaced the U2-type 5′SS, BPS,
and 3′SS of the intron in the MINX pre-mRNA
with those of the U12-type, generating a MINX-
U12 pre-mRNA. In an in vitro splicing assay,
MINX-U12 undergoes splicing to produce ligated
exons under conditions in which the major
spliceosomes are inactivated in the nuclear
extract. These results demonstrate the pres-
ence of active minor spliceosome in the extract.
We then truncated MINX-U12 such that the
binding site for the ATPase/helicase PRP2 is
absent. This strategy, which was designed to
enrich the activated minor spliceosome (Bact
complex) by preventing its remodeling by
PRP2, proved to be successful. We purified the
human minor Bact complex and determined its
cryo–electron microscopy (cryo-EM) structure
at an average resolution of 2.9 Å.
Structure of the minor Bact complex at 2.9 Å
SF3B
U12 core
U12
snRNA
SCNM1
CRIPT
PPIs
U6atac
snRNA
pre-mRNA
Splicing
Factors
NTC
ARMC7
RBM48
PPIL2
U5 snRNP
NTR
U5 snRNA
Structure of the activated human minor spliceosome (Bact complex). (Left) The minor Bact complex comprises U12
snRNP, U5 snRNP, U6atac snRNA, the truncated MINX-U12 pre-mRNA, nineteen complex (NTC), NTC-related (NTR), the
retention and splicing (RES) complex, two prolyl peptidyl isomerase (PPIase)–like proteins (PPIs), and nine splicing factors.
(Right) Based on EM maps, the minor spliceosome contains five newly identified proteins—SCNM1, RBM48, ARMC7, CRIPT,
and PPIL2—which together with snRNAs are highlighted in the foreground. All other components are in the background.
Although the overall organization of the
RNA elements in the human minor Bact com-
plex closely resembles that in the major Bact
complex, there are notable local differences.
Compared with U6 snRNA in the major
spliceosome, U6atac snRNA lacks the 5′ stem
loop, and the U12/U6atac duplex lacks helix II.
Notably, the 3′-end sequences of U6atac snRNA
form a characteristic 3′ stem loop that is placed
in approximately the same location as that of
the U2/U6 helix II in the major Bact complex.
The distinct 5′SS and BPS of the U12-type intron
are recognized through extensive base-pairing
interactions by U6atac and U12 snRNA, respec-
tively. Two catalytic metals, M1 and M2, are
already loaded in the splicing active site center
and poised for catalysis of the branching reaction.
The EM maps allow for the identification of
five previously unidentified proteins—SCNM1,
RBM48, ARMC7, CRIPT, and PPIL2—that ap-
pear to play key roles in the minor spliceo-
some. SCNM1 mimics the SF3a complex of the
major spliceosome. The N-terminal domain of
SCNM1 shares sequence homology with SF3a66
of the SF3a complex. Similarly to SF3a66, the
N-terminal domain of SCNM1 binds the BPS/U12
duplex and the proteins SF3b155, SF3b145, and
CDC5L, whereas the N terminus inserts into the
active site to interact with 5′SS, U6atac snRNA,
and the splicing factor RNF113A. The C-terminal
domain of SCNM1 functionally mimics SF3a60
(another component of the SF3a complex). The
RBM48-ARMC7 complex binds the g-monomethyl
phosphate cap of the guanine nucleotide at
the 5′ end of U6atac snRNA through extensive
interactions. The splicing factor CRIPT binds
RNF113A and stabilizes U12 small nuclear
ribonucleoprotein (snRNP) through interactions
with SF3b14b, SF3b145, and SF3b155. The U-box
protein PPIL2 in an extended conformation
interacts with a number of proteins and RNA,
stabilizing the overall conformation of U5 snRNP.
The N terminus of PPIL2 specifically recognizes
loop I of U5 snRNA. Additionally, PRP2 and its
coactivator SPP2 are bound to the minor Bact
complex, which suggests similar roles for PRP2
and SPP2 in the minor spliceosome as those
observed in the major spliceosome.
CONCLUSION: The cryo-EM structure of the
human minor Bact complex reveals a number
of previously unknown features, including the
identification of minor spliceosome–specific
proteins. This structure serves as a framework
for the mechanistic understanding of the
functions of the minor spliceosome.▪
The list of author affiliations is available in the full article online.
*These authors contributed equally to this work.
†Corresponding author. Email: wanruixue@westlake.edu.cn
(R.W.); syg@westlake.edu.cn (Y.S.)
Cite this article as R. Bai et al., Science 371, eabg0879
(2021). DOI: 10.1126/science.abg0879
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.abg0879
Bai et al., Science 371, 1220 (2021)
19 March 2021
1 of 1
RES EARCH
R E S E A R C H A R T I C L E
◥
SPLICEOSOME
Structure of the activated human minor spliceosome
Rui Bai1,2,3*, Ruixue Wan1,2,3*†, Lin Wang4, Kui Xu4, Qiangfeng Zhang4, Jianlin Lei4,5, Yigong Shi1,2,3,4†
The minor spliceosome mediates splicing of the rare but essential U12-type precursor messenger
RNA. Here, we report the atomic features of the activated human minor spliceosome determined
by cryo–electron microscopy at 2.9-angstrom resolution. The 5′ splice site and branch point
sequence of the U12-type intron are recognized by the U6atac and U12 small nuclear RNAs (snRNAs),
respectively. Five newly identified proteins stabilize the conformation of the catalytic center: The
zinc finger protein SCNM1 functionally mimics the SF3a complex of the major spliceosome, the
RBM48-ARMC7 complex binds the g-monomethyl phosphate cap at the 5′ end of U6atac snRNA,
the U-box protein PPIL2 coordinates loop I of U5 snRNA and stabilizes U5 small nuclear
ribonucleoprotein (snRNP), and CRIPT stabilizes U12 snRNP. Our study provides a framework for
the mechanistic understanding of the function of the human minor spliceosome.
P recursor mRNA (pre-mRNA) splicing,
which involves the removal of non-
coding introns and the ligation of the
coding exons, is achieved by the spliceo-
some (1–3). An intron is defined by
three sequence elements: the 5′ splice site
(5′SS), the branch point sequence (BPS), and
the 3′ splice site (3′SS) (4). Most introns con-
form to the U2 type and are removed by the
major spliceosome that contains the U2 small
nuclear RNA (snRNA) (4–7). A very small percent-
age of introns belong to the U12 type (8, 9),
which were initially identified to have the
dinucleotides AT at the 5′ end of their 5′SS
and AC at their 3′SS (10, 11). Compared with
the U2 type, the U12-type introns are charac-
terized by a highly conserved 5′SS and BPS
(12–16). The U12-type genes play an essential
role in development and are present in all
major eukaryotic taxa including fungi, plants,
and animals (17–21). Pre-mRNA splicing in-
volving U12-type introns is mediated by the
minor spliceosome (22, 23), which contains
five snRNAs: U11, U12, U4atac, U5, and U6atac
(24–27). Of these snRNAs, only U5 is shared
between the major and minor spliceosomes.
Because of its scarcity in cells (8, 9, 28), the
minor spliceosome represents a challenge for
1Key Laboratory of Structural Biology of Zhejiang Province,
School of Life Sciences, Westlake University, Xihu District,
Hangzhou 310024, Zhejiang Province, China. 2Westlake
Laboratory of Life Sciences and Biomedicine, Xihu District,
Hangzhou 310024, Zhejiang Province, China. 3Institute of
Biology, Westlake Institute for Advanced Study, Xihu District,
Hangzhou 310024, Zhejiang Province, China. 4Beijing
Advanced Innovation Center for Structural Biology and
Beijing Frontier Research Center for Biological Structure,
Tsinghua-Peking Joint Center for Life Sciences, School of
Life Sciences, Tsinghua University, Beijing 100084, China.
5Technology Center for Protein Sciences, Ministry of
Education Key Laboratory of Protein Sciences, School of Life
Sciences, Tsinghua University, Beijing 100084, China.
*These authors contributed equally to this work.
†Corresponding author. Email: wanruixue@westlake.edu.cn
(R.W.); syg@westlake.edu.cn (Y.S.)
biochemical studies. To date, there is no
published protocol on the purification of the
minor spliceosome. Our knowledge about the
protein components of the minor spliceo-
some has been mostly derived from mass
spectrometry (MS) analyses of the purified
U11/U12 di–small nuclear ribonucleoprotein
(di-snRNP) (29) and immunoprecipitation
studies on the tri-snRNP (30). There is no
structural information on the minor spliceo-
some, and it is unclear how many U12-specific
components are present in each major func-
tional state of the minor spliceosome. Our lack
of understanding of the minor spliceosome
contrasts sharply with our comprehensive
knowledge of the major spliceosome, for which
the major functional states from Saccharomyces
cerevisiae and Homo sapiens have been struc-
turally characterized by cryo–electron micros-
copy (cryo-EM) since 2015 (31–34).
In this study, we report the in vitro assembly
and purification of the activated human minor
spliceosome (i.e., the minor Bact complex). In
the major Bact complex, the splicing active
site has been formed, but productive docking
of the BPS is yet to happen (35, 36). We
determine the cryo-EM structure of the minor
Bact complex at an average resolution of 2.89 Å,
revealing a wealth of information.
Reconstitution and purification of the human
minor Bact complex
The in vitro splicing assay using HeLa nuclear
extract was established specifically to study
the minor spliceosome 25 years ago (27). The
U2-type 5′SS, BPS, and 3′SS of the intron in
the MINX pre-mRNA were replaced by those
of the U12 type, which resulted in the MINX-
U12 pre-mRNA. The specificity of MINX-U12
was confirmed in our in vitro splicing assay
modified from the established protocol (27)
(fig. S1A). Relying on endogenous ribonuclease
(RNase) H activity in the nuclear extract, two
sets of antisense DNA oligonucleotides were
used to digest either U1/U2/U6 snRNAs (groups
2 and 5) or U11/U12/U6atac snRNAs (groups 3
and 6). With application of the anti-U1/U2/
U6 oligos, the splicing of MINX was severely
diminished (fig. S1A, groups 1 and 2), whereas
the splicing of MINX-U12 was unaffected
(groups 4 and 5). Conversely, with application
of anti-U11/U12/U6atac oligos, the splicing of
MINX was unaffected (fig. S1A, groups 1 and
3), whereas the splicing of MINX-U12 became
undetectable (groups 4 and 6).
Next, we truncated MINX-U12 18 nucleo-
tides downstream of the BPS to generate the
MINX-U12D pre-mRNA, which lacks the
binding site for the adenosine triphosphatase
(ATPase)/helicase Prp2. As previously shown
for the major Bact complex (35–37), such a pre-
mRNA disables Prp2-mediated spliceosome
remodeling, thereby favoring the accumula-
tion of the minor Bact complex. To facilitate
purification, three tandem MS2 binding sites
were inserted between the 5′SS and BPS of
MINX-U12D. After incubation of the HeLa
nuclear extract with MINX-U12D, the minor
Bact complex was purified through affinity
chromatography and glycerol gradient cen-
trifugation (fig. S1B). Chemical cross-linking
was used to stabilize the minor spliceosome.
Relying on these strategies, we were able to
obtain the minor Bact complex as confirmed by
the presence of U12, U5, and U6atac snRNAs
(fig. S1C). The spliceosomal particles appear
intact on the cryo-EM micrograph (fig. S1D).
To facilitate protein identification in subse-
quent cryo-EM analyses, the purified sample
was analyzed by MS (table S1).
Overall structure of the human minor
Bact complex
In total, 20,390 micrographs were recorded
using a Gatan K3 detector mounted on a Titan
Krios microscope, which yielded 3.5 million
auto-picked particles and good two-dimensional
(2D) averages (fig. S1E). Through multiple
cycles of multireference 3D classifications,
101,443 particles were selected to yield a
reconstruction of the minor Bact complex at an
average resolution of 2.89 Å (Fig. 1A; fig. S2;
fig. S3, A to D; and table S2). The core region
of the spliceosome reaches 2.6 Å (fig. S3B).
Improved reconstruction for the SF3b com-
plex and the 5′-end region of U6atac snRNA
was obtained through focused 3D classifica-
tion and refinement (fig. S2). The quality of
the EM maps allows atomic modeling of the
human minor Bact complex and identification
of many previously unrecognized features
(figs. S4 to S8).
The final atomic model contains 45 pro-
teins, three snRNAs (U12, U5, and U6atac),
and the MINX-U12D pre-mRNA, amounting
to a molecular mass of 1.7 MDa (Fig. 1A). All
45 proteins are detected by MS (table S1). The
Bai et al., Science 371, eabg0879 (2021)
19 March 2021
1 of 12
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Cryo-EM structure of the activated human minor spliceosome (Bact complex). (A) Overall
structure of the human minor Bact complex. Two surface views are shown. Spliceosomal components are
tabulated below the images. Proteins that are found only in the structure of the major, but not minor, Bact
complex are indicated by dashed gray boxes. Compared with the major Bact complex (35, 36), the minor
spliceosome contains four newly identified proteins ARMC7, CRIPT, RBM48, and SCNM1. Although PPIL2
(colored dark purple) is also a component of the major Bact complex (35), its atomic model is reported here.
(B) Overall structural comparison between the minor and major Bact complexes. The orientation and scale are
the same for the two complexes. Identical protein components are shown in the background with faded
colors. RNA elements and protein components that differ in the two complexes are color coded. Structural
images in this figure were prepared using ChimeraX (80).
minor Bact complex comprises U12 snRNP,
U5 snRNP, U6atac snRNA, MINX-U12D pre-
mRNA, two components of the nineteen com-
plex (NTC) (CDC5L and SYF3/CRNKL1), three
components of the NTC-related (NTR) (CWC15/
AD002, SKIP, and PLRG1/PRL1), the retention
and splicing (RES) complex (SNIP1, RBMX2,
and BUD13), two candidate prolyl peptidyl
isomerases (PPIases) (PPIL2 and CWC27/NY-
CO-10), and nine splicing factors (ARMC7,
CRIPT, CWC22, GPKOW, PRP2, RBM48,
RNF113A, SRm160, and SRm300) (Fig. 1A).
Among the splicing factors, GPKOW (Spp2
in S. cerevisiae) is the coactivator of the
DEAH-box ATPase/helicase PRP2. U12 snRNP
consists of the core (U12 snRNA and U12-Sm
ring), the SF3b complex, and a newly identi-
fied protein, SCNM1.
Compared with the major Bact complex
(35, 36, 38, 39), the cryo-EM structure of the
minor Bact complex contains four newly iden-
tified proteins: SCNM1 of U12 snRNP and
three splicing factors, ARMC7, CRIPT, and
RBM48 (Fig. 1B). All four proteins are clearly
defined by EM density maps (figs. S6 to S8).
U11/U12-65K is a known component of U12
snRNP (29, 40), but the EM density around
the U11/U12-65K region is poor and inadequate
for judgment (fig. S6A). Nonetheless, U11/U12-
65K shows up in the MS analysis (table S1).
Although PPIL2 is also a component of the
major Bact complex (35, 41), it is atomically
modeled only in our current structure. As
will be detailed later, the location and putative
function of these proteins in the minor
spliceosome provide an explanation for the
departure of a number of proteins that are
specific to the major spliceosome.
RNA elements in the human minor
Bact complex
We modeled 63 nucleotides in U12 snRNA
(1 to 51 and 73 to 84), 90 nucleotides in U5
snRNA (8 to 71, 85 to 103, and 110 to 116), 61
nucleotides in U6atac snRNA (1 to 56 and
61 to 65), 12 nucleotides in 5′ exon (−12 to −1),
and nucleotides 1 to 19 and 192 to 223 in the
intron of MINX-U12D (Fig. 2, A and B; figs. S4
and S5; and table S3). The active site center
is composed of the intramolecular stem loop
(ISL) of U6atac snRNA, helix I of the U12/
U6atac duplex, loop I of U5 snRNA, and six
metal ions. Helix Ib of the U12/U6atac duplex
forms a catalytic triplex with G19, A20, and U46
of U6atac (fig. S4, B to D). Notably, the 3′-end
sequences of U6atac snRNA form an inter-
nal stem loop (named 3′ stem loop) that is
placed in approximately the same location
as that of the U2/U6 helix II in the major Bact
complex (35). The g-monomethyl phosphate
cap at the 5′ end of U6atac snRNA is un-
ambiguously defined by the EM map (Fig.
2A and fig. S4C).
The overall organization of the RNA ele-
ments in the minor Bact complex resembles
that in the major Bact complex (35), with
notable local differences (Fig. 3A). Compared
with U6 snRNA, U6atac lacks the 5′ stem loop
but has a distinct 3′ stem loop. Unlike that in
the major spliceosome, the U12/U6atac duplex
lacks helix II because there are no nucleotides
5′ to helix Ib in U12 snRNA (Fig. 2B). Recog-
nition of the highly conserved U12-type 5′SS
(5′-AUAUCCUUU-3′) by U6atac involves six
canonical Watson-Crick base pairs and one
Hoogsteen base pair, which together offer
more-stringent sequence specificity than that
in the major Bact complex (35) (Fig. 3B and fig.
S5A). Similarly, the highly conserved U12-type
BPS (5′-UCCUUAACUC-3′) is recognized by
U12 snRNA through nine Watson-Crick base
pairs, leaving only the nucleophile-containing
adenine base unpaired (Fig. 3C and fig. S5B).
By contrast, the BPS in the major Bact com-
plex is recognized by five Watson-Crick base
pairs (35).
RNA elements in the active site adopt a
nearly identical conformation to those in the
major spliceosome (Fig. 3D). Unlike the major
Bact complex (35, 36), both catalytic metal ions
M1 and M2 are loaded in the active site of the
minor Bact complex (Fig. 3E and fig. S5, C and
Bai et al., Science 371, eabg0879 (2021)
19 March 2021
2 of 12
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Organization and structure of the RNA elements in the human minor Bact complex. (A) Overall
structure of the RNA elements. U12, U5, and U6atac snRNAs are colored marine, orange, and green,
respectively. The 5′ exon and intron of MINX-U12D are colored red and violet, respectively. The g-monomethyl
phosphate cap of U6atac snRNA is shown in stick representation. The catalytic center comprises the
ISL of U6atac snRNA, helix I of the U12/U6atac duplex, and loop I of U5 snRNA. (B) Summary of the
base-pairing interactions among the RNA elements. Helix Ib of the U12/U6atac duplex forms a catalytic
triplex with three nucleotides of U6atac. Canonical Watson-Crick and noncanonical base-pairing interactions
are identified by solid lines and dots, respectively. Structural images in this and the following figures
were prepared using PyMol (81).
D). M1, which stabilizes the leaving group
during the branching reaction, is coordinated
by two phosphate groups from G44 and U46
of U6atac. M2, which activates the nucleo-
phile, is coordinated by three phosphate
groups from A26, G27, and U46. M1 and M2
are located 4.0 and 3.9 Å away from the
phosphate group of the first nucleotide (A1)
of 5′SS, respectively (Fig. 3E). The distance
between M1 and the 3′ oxygen atom from the
scissile phosphodiester bond is 3.3 Å. There-
fore, the catalytic metals are poised for catalysis
of the branching reaction in the minor Bact
complex. Coordination of M2 in the minor
Bact complex is identical to that in the major
spliceosome (Fig. 3F), in which M2 is fixed
by the phosphates from A53, G54, and U74 of
U6 snRNA.
An essential role of SCNM1 in the
minor spliceosome
U12 snRNP differs from U2 snRNP in both
composition and overall organization (Fig. 4A).
U2-A′ and U2-B′′ in the U2 core are known
to be undetected in the U12 snRNP (29) and
are probably replaced by U11/U12-65K given
its homology to U2-B′′ (40). In fact, the C-
terminal portion of U11/U12-65K exhibits struc-
tural homology with the N-terminal portion of
U2-B′′ (40) (fig. S6A). Consistent with previous
studies (29), the SF3a complex, which bridges
SF3b and the U2 core in the major spliceosome,
is absent in the minor spliceosome. The N-
terminal and C-terminal domains (N- and C-
domains) of the newly identified component
SCNM1 are located on two opposing sides of
the U12 snRNP, spanning a distance of ~80 Å
(Fig. 4A). A newly identified splicing factor
CRIPT, which contains two zinc finger (ZnF)
motifs, stabilizes the U12 snRNP through
simultaneous interactions with its three com-
ponents (SF3b14b, SF3b145, and SF3b155) (Fig.
4B). CRIPT also directly binds to the splicing
factor RNF113A, an ortholog of the yeast splic-
ing factor Cwc24.
SCNM1, a modifier of the severity of sodium
channelopathy in mice (42, 43), function-
ally mimics the SF3a complex of the major
spliceosome. SCNM1 comprises an N-domain
(residues 1 to 75), a C-domain (residues 183 to
226), and an intervening flexible linker (resi-
dues 76 to 182) (Fig. 4C and fig. S6C). Both
N- and C-domains of SCNM1 are highly con-
served in vertebrates (fig. S9). The N-domain
contains a C2H2-type ZnF and shares sequence
homology with that of SF3a66 from the SF3a
complex (Fig. 4D). Similar to SF3a66, the
SCNM1 ZnF interacts with the BPS/U12 du-
plex and the proteins SF3b155, SF3b145, and
CDC5L (Fig. 4E). The functional mimicry is
underscored by structural similarity between
the ZnF of SCNM1 and that of SF3a66, which
occupy the same corresponding location in the
minor and major Bact complexes (Fig. 4F).
Just like SF3a66 (35, 36, 38), the N terminus
of SCNM1 inserts into the active site to interact
with the 5′SS, U6atac snRNA, and RNF113A
(Fig. 4G). Similar to that in the major spliceo-
some (35, 36, 38), the adenine nucleotide at the
5′ end of 5′SS (A1) is sandwiched by the aromatic
side chains of Phe213 and Lys218 of RNF113A.
The aromatic side chain of Phe3 from SCNM1
is located next to the backbone of U2 and
between the nucleobases of A1 and A3 from
5′SS. The positively charged side chains of
Lys4 and Arg5 from SCNM1 donate H bonds
to the backbone phosphates of U24 and A26
from the U6atac snRNA. These interactions
stabilize the recognition of 5′SS by RNF113A
and help maintain the conformation of
U6atac. Notably, all three N-terminal residues
of SCNM1 are invariable in other vertebrates
(fig. S9).
Bai et al., Science 371, eabg0879 (2021)
19 March 2021
3 of 12
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Conservation and varia-
tion of the RNA structures
between the minor and major
Bact complexes. (A) Overall
structural comparison of the RNA
elements between the minor and
major Bact complexes. The RNA
elements from the minor spliceo-
some are color coded, whereas
those from the major spliceosome
are shown in gray. Compared with
U6 snRNA (35), U6atac lacks the
5′ stem loop and cannot form
helix II with U12, but forms a
distinct 3′ stem loop. (B) Recog-
nition of 5′SS by U6atac (left)
involves more base-pairing inter-
actions than that by U6 snRNA
(35) (right). (C) Recognition of
BPS by U12 (left) involves more
base-pairing interactions than that
by U2 snRNA (35) (right). (D) The
active site conformation in the
minor Bact complex is nearly
identical to that in the major Bact
complex (35). Shown here is an
overlay of the active site RNA
elements from both complexes.
(E) Coordination of the catalytic
metals in the minor Bact complex.
Unlike the major Bact complex
(35), both catalytic metals M1 and
M2 are loaded in the minor Bact
complex. (F) Coordination of the
catalytic metals in the major Bact
complex [PDB code 6FF7 (35)].
The orientation is the same as
that in (E). M1 is yet to be loaded.
The C-domain of SCNM1 interacts with
SF3b130 and SF3b145 (Fig. 4H). The spatial
location occupied by the C-domain of SCNM1
overlaps with that occupied by SF3a60, a com-
ponent of the SF3a complex. In the major Bact
complex, SF3a60 stabilizes the SF3b complex
by interacting with SF3b130 and SF3b145 (35, 36).
Therefore, the C-domain of SCNM1 can be
regarded as a functional mimicry of SF3a60.
Among the three components of the SF3a
complex, SF3a60 connects SF3a66 to SF3a120,
which interacts with U2-A′ and the Sm ring
in the major Bact complex (35, 36). SF3a120 is
absent in the minor spliceosome. Together,
the above analyses explain the absence of the
SF3a complex in the minor spliceosome.
PPIL2 stabilizes loop I and U5 snRNP
U5 snRNP is the only shared snRNP between
the minor and major spliceosomes (16, 22, 23).
In the minor Bact complex, U5 snRNP is
stabilized by the newly identified component
PPIL2 (Fig. 5A). PPIL2 is one of eight nuclear
cyclophilins in human cells (44) and is con-
served from Schizosaccharomyces pombe to
humans (fig. S10A). PPIL2 is thought to
contain a U-box E3 ubiquitin ligase domain
at its N-terminal half and a PPIase domain at
its C-terminal half. Our EM reconstruction
allows for the identification of four discrete
fragments in the N-terminal half of PPIL2
(Fig. 5, A and B, and fig. S7A). The quality
of the EM density for the C-terminal half is
insufficient for sequence assignment, but
its appearance allows for the docking of the
crystal structure of the PPIL2 PPIase domain
[Protein Data Bank (PDB) code 1ZKC (45)].
The structure of PPIL2 comprises fragment
I (residues 2 to 26), fragment II (residues 33 to
163), fragment III (residues 191 to 225), frag-
ment IV (residues 239 to 263), and a PPIase
domain (residues 271 to 457) (Fig. 5B). Each
fragment appears to play a specific, important
role. Fragment I reaches into the catalytic cen-
ter and interacts with loop I of U5 snRNA, the
N-domain of PRP8, the N terminus of SNU114,
and the WD40 domain of the NTR compo-
nent, PLRG1 (Fig. 5C). The N terminus of
PPIL2 inserts into the groove on the back side
Bai et al., Science 371, eabg0879 (2021)
19 March 2021
4 of 12
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Structure of U12 snRNP in the human minor Bact complex. (A) Overall structure of U12
snRNP. U12 snRNP consists of the core (U12 snRNA and Sm ring), the SF3b complex, and the newly
identified component SCNM1 (82). A newly identified splicing factor CRIPT is also shown. (B) A close-up
view of the cysteine-rich PDZ-binding protein CRIPT and its surrounding proteins. CRIPT interacts
with SF3b155, SF3b145, SF3b14b, and the splicing factor RNF113A. (C) A schematic diagram of the
sequence features of SCNM1. The interacting proteins and RNA elements are tabulated below the
sequences. (D) Sequence alignment between the N-domains of SCNM1 and SF3a66. Conserved
sequences are boxed, and identical residues are shaded red. (E) The N-domain of SCNM1 interacts
with the BPS/U12 duplex and the proteins SF3b145, SF3b155, and CDC5L. (F) The C2H2-type ZnF of
SCNM1 closely resembles that of SF3a66. Shown here is a structural overlay. The RNA and protein
components are color coded in the minor spliceosome and gray in the major spliceosome (36).
(G) A close-up view of the N terminus of SCNM1 and its interactions with surrounding proteins
and RNA elements. The N terminus of SCNM1 is inserted between RNF113A and U6atac. (H) The
C-domain of SCNM1 interacts with SF3b130 and SF3b145 in the minor Bact complex. The spatial
location of the SCNM1 C-domain overlaps with that of SF3a60 in the major Bact complex (36).
Single-letter abbreviations for the amino acid residues are as follows: A, Ala; C, Cys; D, Asp; E,
Glu; F, Phe; G, Gly; H, His; I, Ile; K, Lys; L, Leu; M, Met; N, Asn; P, Pro; Q, Gln; R, Arg; S, Ser; T, Thr;
V, Val; W, Trp; and Y, Tyr.
of the duplex between loop I and the 5′ exon.
Six contiguous polar or charged residues of
PPIL2 (Lys3-Arg4-Gln5-His6-Gln7-Lys8) make
specific interactions to loop I (Fig. 5D and
fig. S7B). Lys3, Arg4, Gln5, and Gln7 may make
direct H bonds to the bases of A44, U33/U34,
C45/C45, and U35 of loop I, respectively. His6,
Lys8, and Tyr12 may mediate H bonds to the
phosphate groups of U42, G37, and U34. These
interactions likely stabilize the conformation
of loop I, which may help anchor the 5′ exon.
Notably, the N-terminal residues of PPIL2
are nearly invariable in other organisms except
S. pombe (fig. S10A).
Unexpectedly, fragment II of PPIL2 con-
tains two U-box E3 ubiquitin-ligase domains:
one as predicted (46) (U-box I, residues 41 to
91) and the other previously unknown (U-box
II, residues 102 to 158) (Fig. 5B and fig. S7C).
Structures of the two U-box domains can be
superimposed with a pair-wise root mean
square deviation (RMSD) of 1.86 Å over 50
aligned residues (fig. S7D). Close examination
reveals conservation of key residues between
U-boxes I and II (fig. S10B). U-box I contacts
Prp8 and U-box II interacts with SNU114 (Fig.
5E). Fragments III and IV of PPIL2 adopt an
extended conformation and interact with
PRP8, SNU114, the NTC component CDC5L,
and the NTR component SKIP, likely stabi-
lizing the association of U5 snRNP with NTC
and NTR (Fig. 5F). Because NTC and NTR
are required for the formation of the splicing
active site, the multifaceted interactions by
PPIL2 may contribute to not just the assembly
of the minor spliceosome but also the splicing
reaction itself.
The RBM48-ARMC7 complex binds the 5′ end
of U6atac snRNA
The RNA binding motif (RRM) that contains
protein RBM48 is essential for the splicing of
U12-type introns in both plants and animals
(47, 48). It interacts with the protein ARMC7
(armadillo repeat containing 7) (47–49). RBM48-
deficient human cell lines and maize endo-
sperms display aberrant splicing of the U12-type
introns, which may lead to developmental de-
fects (47, 48). However, the role of RBM48 in
U12-dependent splicing remains unknown. In
our structure, RBM48 and ARMC7 form a com-
plex, which associates with the U6atac/5′SS du-
plex and binds the characteristic g-monomethyl
phosphate cap of U6atac (26) (Fig. 6A and fig.
S8). Additionally, RBM48 interacts with the
N-terminal a-helix in the N-domain of PRP8,
which further binds to the top face of U5-40K.
The g-monomethyl phosphate cap of the
guanine nucleotide (G1) at the 5′ end of U6atac
snRNA is recognized by RBM48 (Fig. 6B). The
guanine base stacks against the aromatic side
chain of Tyr39 and is specifically recognized by
two H bonds from the side chains of Tyr121 and
Glu44. The b-phosphate of the cap is recognized
Bai et al., Science 371, eabg0879 (2021)
19 March 2021
5 of 12
RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. Multifaceted roles of PPIL2 in the human minor Bact complex. (A) PPIL2 forms at least five distinct
interfaces with the minor Bact complex and spans a distance of >100 Å. The quality of the EM density allows for
the unambiguous identification of four discrete fragments I through IV in the N-terminal half of PPIL2, but not
in the C-terminal half. (B) Schematic diagram of the sequence features of PPIL2. PPIL2 comprises four fragments
and a PPIase domain. The interacting proteins and RNA elements are tabulated below the sequences. (C) A
close-up view of fragment I of PPIL2 and its surrounding environment. This fragment interacts with loop I of U5
snRNA, the N-domain of PRP8, the WD40 domain of PLRG1, and the N terminus of SNU114. (D) A close-up
view of the interface between the N terminus of fragment I and loop I of U5 snRNA. The N terminus of PPIL2 is
inserted into the groove on the back side of the duplex between loop I and the 5′ exon. (E) A close-up view of
fragment II of PPIL2 and its surrounding environment. Notably, this fragment contains two U-box E3 ubiquitin-
ligase domains: one as predicted (46) (U-box I; light purple) and the other previously unknown (U-box II; dark
purple). U-box I contacts Prp8, and U-box II interacts with SNU114. (F) A close-up view of fragments III and IV of
PPIL2. These two fragments adopt an extended conformation and interact with SKIP, PRP8, CDC5L, and SNU114.
by two H bonds from the side chains of Tyr39
and Asp108. By contrast, the a- and g-phosphates
are each coordinated by a single H bond from
the main chain amide groups of Ala35 and
His15, respectively (Fig. 6B). All key residues
involved in the recognition of the 5′ cap are
invariable in all organisms examined (fig. S11).
Additionally, the side chain of Asn39 from
ARMC7 contacts the phosphate group of U2
from U6atac.
The nucleotides at the 5′ end of U6atac are
also stabilized by additional interactions. The
aromatic side chain of Tyr16 from ARMC7
stacks against the nucleobases of G3 and
U4 from U6atac (Fig. 6C). The guanidinium
group of Arg31 from RBM48 donates two H
bonds to the phosphate groups of U4 and
U5. As a whole, coordination of the 5′-end
sequences of U6atac differs from that in the
major spliceosome (Fig. 6D). By contrast to
the minor spliceosome, the 5′-end sequences
of U6 snRNA form a stem loop in the major
Bact complex (35, 36). RBM22 and BUD31
from the NTR and the splicing factor PRP17
stabilize the 5′ stem loop and downstream
sequences of U6 snRNA. RBM22 also directly
contacts the extended intron sequences from
the U6/5′SS duplex. Compared with the minor
Bact complex, U5-40K along with the N-terminal
helix of PRP8 in the major Bact complex is
rotated by ~45° to facilitate association of
RBM22, BUD31, and PRP17.
Discussion
In this study, we report the in vitro assembly
and purification of the human minor Bact
complex. The key to doing so was to obtain
active minor spliceosome. To achieve this goal,
we first reconstituted an in vitro splicing assay
using HeLa cell extract and demonstrated that
the U12-type intron can be removed indepen-
dently of the major snRNAs (fig. S1A). To
ensure enrichment of the minor Bact com-
plex, we truncated MINX-U12 18 nucleotides
downstream of the BPS to eliminate the Prp2
binding site. Notably, this design disables Prp2-
mediated spliceosome remodeling but has no
effect on Prp2 binding to the spliceosome.
These strategies proved effective, resulting
in the purification of the minor Bact complex
and subsequent determination of its cryo-EM
structure. Notably, the newly identified pro-
tein components in the cryo-EM structure
were all detected by MS (table S1). This general
approach may be adapted for the isolation of
other major functional states of the minor
spliceosome.
The U12-type introns were initially thought
to have the dinucleotide AT the 5′ end of the
5′SS and AC at the 3′SS (10, 11), hence the names
U4atac and U6atac (26). Subsequent studies
found the dinucleotides GT and AG to be more
prevalent in 5′SS and 3′SS, respectively (14, 15).
Nonetheless, because the dinucleotide AT
is more complementary to sequences in U11
snRNA, we chose to have ATATCCTTT as the
5′SS to favor the minor splicing pathway.
However, as was first observed in the yeast
Bact complex (38, 39), the first three nucleo-
tides at the 5′ end of the 5′SS remain unpaired
because sufficient latitude must be given to
the 5′-end nucleotide for the branching reac-
tion. This structural feature, together with the
fact that the U12-type introns have lengthy
Bai et al., Science 371, eabg0879 (2021)
19 March 2021
6 of 12
RES EARCH | R E S E A R C H A R T I C L E
Fig. 6. The RBM48-ARMC7 complex binds the 5 end of U6atac snRNA in the minor Bact complex.
(A) The RBM48-ARMC7 complex binds the 5′ end of U6atac snRNA. The g-monomethyl phosphate cap at the
5′ end of U6atac is shown in stick representation. (B) A close-up view of the recognition of the guanine
nucleotide (G1) at the 5′ end of U6atac. G1 is recognized by RBM48. (C) A close-up view of the stacking
interactions among the aromatic side chain of Tyr17 from ARMC7 and the bases of G3 and U4 from U6atac.
(D) Variations in the coordination of the 5′-end sequences from U6 and U6atac. In the major Bact complex
[PDB code 5Z56 (36)] (right), the 5′-end sequences of U6 form a stem loop. RBM22, BUD13, and PRP17
recognize the 5′ stem loop and its downstream sequences. U6atac lacks the 5′ stem loop, and its 5′ end is
bound by the RBM48-ARMC7 complex (left).
consensus sequences for 5′SS and BPS, strongly
favors the recognition of U12-type introns by
the minor spliceosome (12, 15–17).
Thirteen proteins of the major Bact complex,
including five NTC and NTR components, are
absent in the structure of the minor Bact
complex (Fig. 1, A and B). Three proteins of
the SF3a complex are substituted by SCNM1,
and U2-A′ and U2-B′′ could be replaced by
U2-65K as previously observed. For the NTR
components RBM22 and BUD31, their func-
tion has been partially replaced by the RBM48-
ARMC7 complex, which in addition to binding
the 5′ cap of U6atac may also facilitate U12-type
splicing by stabilizing the intron sequences
(Fig. 7). The explanation for the absence of the
other proteins remains to be investigated. The
absence of PRPF19 and SYF1 in the minor Bact
complex can be explained by dynamic confor-
mation or assembly.
Five proteins (ARMC7, CRIPT, RBM48,
SCNM1, and PPIL2) were structurally identi-
fied in the minor Bact complex (Fig. 7). Among
them, PPIL2 is also present in the major pre-
Bact complex (50). It is possible that CRIPT
may also be a component of the major spliceo-
some because of its unblocked location.
SCNM1, RBM48, and ARMC7 appear to be
components specific to the minor spliceo-
some (Figs. 1 and 7). These newly identified
proteins remain poorly understood, but our
study defines these five proteins as hallmarks
of the minor Bact complex. Given its predom-
inant interactions with the SF3b complex,
SCNM1 likely dissociates during the Bact-to-
B* transition. The other three proteins may
continue to associate with the minor spliceo-
some in other major functional states.
SCNM1 was first identified as a key modifier
of the severity of sodium channelopathy in
mice through its modulation of the propor-
tion of correctly spliced transcripts of the
Scn8a (Nav1.6) gene (42, 43). Because of a 5′SS
mutation in Scn8a, two neighboring exons are
skipped in 90% of the mRNA, which leads to
channelopathy in mice (51). Notably, the in-
tron between these two exons belongs to the
U12 type. A C-terminal truncation or an in-
ternal deletion in SCNM1 increases exon skip-
ping to 95% of the mRNA, resulting in lethality
(51, 52). The channelopathy phenotype could be
rescued by the transgenic expression of wild-
type SCNM1 (43). In our structure, SCNM1 is
identified as a key component of the minor
spliceosome, with its N- and C-domains each
playing a vital role. The pathogenic C-terminal
truncation of SCNM1 results in the removal of
the residues 187 to 230, thereby eliminating the
C-domain. The internal truncation (affecting
residues 133 to 196) compromises the integ-
rity of the C-domain and no longer allows a
functionally required distance between the
N- and C-domains.
RBM48 is thought to be an essential splicing
factor for U12-type introns, as its mutations
cause abnormality in genome-wide U12-type
splicing (47, 48). Together with ARMC7, RBM48
binds the 5′ end of U6atac snRNA, which lacks
the 5′ stem loop compared with U6 snRNA
in the major spliceosome. This interaction
also stabilizes six unpaired nucleotides at the
5′ end of U6atac and orients the extended
sequences of the intron downstream of the
U6atac/5′SS duplex in the minor Bact complex.
In this sense, the RBM48-ARMC7 complex ap-
pears to serve a similar role to that of the splicing
factor Prp17 and the NTR components RBM22
and BUD31 in the major Bact complex (35, 36).
PPIL2 is thought to be enriched only in the
Bact complex (41). This observation is con-
sistent with the structural finding that, during
the B-to-Bact transition, the N-domains of
PRP8 and SNU114, as a single bloc, are rotated
30° relative to the core domain of Prp8 (32),
Bai et al., Science 371, eabg0879 (2021)
19 March 2021
7 of 12
RES EARCH | R E S E A R C H A R T I C L E
Fig. 7. Structural differences between the human minor and major Bact complexes. Cartoon diagrams
of the structures around the active site of minor and major Bact complexes. In the minor spliceosome
(left), the last five nucleotides UCCUU of 5′SS are recognized by the AAGGA box of U6atac snRNA. The first
adenine nucleobase in 5′SS is sandwiched by Phe213 and Lys218 from RNF113A. The N terminus of SCNM1
inserts into the active site, stabilizing 5′SS recognition by RNF113A. In the major Bact complex (right),
5′SS recognition by U6 snRNA involves many fewer base pairs, and recognition of the first guanine nucleotide
by RNF113A is stabilized by SF3a66. Additionally, SF3a60 in the major spliceosome is replaced by the
C terminus of SCNM1 in the minor Bact complex. In these regards, SCNM1 in the minor spliceosome mimics
the SF3a complex in the major spliceosome. This is likely true not only in the minor Bact complex but also
during spliceosome assembly. In the minor spliceosome, the RBM48-ARMC7 complex binds the 5′ end of
U6atac snRNA, which lacks the 5′ stem loop compared with U6 snRNA in the major spliceosome. In the
minor Bact complex, the N terminus of PPIL2 binds loop I of U5 snRNA and interacts with several components
from U5 snRNP, NTC and NTR. By contrast, the role of PPIL2 in the major Bact complex is yet to be defined.
Structural features of the minor spliceosome that are similar to those of the major spliceosome are not
discussed here (for example, in both cases, the HEAT repeat protein SF3b155 serves as the central scaffold
of U2 snRNP and wraps around the U2/BPS duplex).
which creates a surface cleft for the recruit-
ment of PPIL2. Unexpectedly, PPIL2 actually
contains two U-boxes, and both are bound in
the surface cleft between SNU114 and the
core domain of PRP8. The interface between
U-box II and SNU114 in our structure is some-
what similar to that reported recently between
the predicted U-box (U-box I) and SNU114 in
the major pre-Bact complex (50). Docking of
the PPIase domain of PPIL2 in our minor Bact
complex resembles that observed in the major
pre-Bact complex (50). During the Bact-to-C
transition of the major spliceosome, the bind-
ing site for the PPIase domain of PPIL2 is oc-
cupied by the NTC component SYF2 (53), which
is again consistent with the reported enrichment
of PPIL2 in only the Bact complex (41).
The U12-type introns are present in genes
that play an essential role in cells. These genes
control DNA replication and repair, transcrip-
tion, RNA processing and translation, cyto-
skeletal organization, vesicular transport, and
voltage-gated ion channel activity (54). Usually
only one U12-type intron is present in each
gene, but it is likely the rate-limiting step in
splicing because of the scarcity of the minor
spliceosome. Notably, although U12-type introns
are present in all major eukaryotic taxa, they
are absent in the popular model organisms
S. cerevisiae and Caenorhabditis elegans (17, 55).
In this regard, investigation of regulators that
are specific to the minor spliceosome, such as
SCNM1, RBM48, and ARMC7, may hold the
key to understanding the evolution, function,
and disease relevance of the U12-type genes.
Although the minor spliceosome has been
known of for decades, relatively little infor-
mation is known about its composition, func-
tional states, catalysis, and regulation, especially
when compared with the major spliceosome.
Similarly, little is known about its remodeling
and the associated ATPase/helicases. The fact
that Prp2 and Spp2 are bound to the minor
Bact complex may constitute unequivocal evi-
dence that the Prp2/Spp2 complex in the minor
spliceosome functions similarly to that in the
major spliceosome. Identically to what has
been observed in the major Bact complex (56),
Prp2 directly interacts with Rse1 and Hsh155
in the minor spliceosome, and this interaction
is strengthened by Spp2. Scrutiny of the structural
features of the human minor Bact complex may
reveal additional findings about the minor
spliceosome. This study provides a framework
for the mechanistic understanding of the func-
tion of the minor spliceosome.
Materials and methods
Preparation of the pre-mRNA
The U12-type pre-mRNA for the in vitro splic-
ing assay was modified from the MINX gene,
in which the 5′SS, BPS, and 3′SS are replaced
by consensus sequences of the U12-type in-
tron. The 5′SS, BPS, and 3′SS have sequences
5′-AUAUCCUUU-3′, 5′-UCCUUAACUC-3′, and
5′-CAC-3′, respectively. The altered MINX pre-
mRNA (referred to as MINX-U12) comprises a
57-nucleotide (nt) 5′ exon, a 228-nt intron, a
51-nt 3′ exon, and three tandem MS2-binding
RNA aptamers at the 3′ end of the 3′ exon. To
enrich the minor Bact complex in our in vitro
assembly assay (37), we further modified
MINX-U12 by deleting all sequences beyond
nucleotide 18, counting from the 3′ end of the
BPS. We then inserted three tandem MS2-
binding RNA aptamers between the 5′SS and
the BPS. The resulting pre-mRNA, referred
to as MINX-U12D, was used for spliceosome
assembly and purification. The DNA templates
for in vitro transcription were generated using
polymerase chain reaction (PCR), and the RNA
substrates were synthesized using the method
of T7 runoff transcription.
In vitro splicing assay and reverse transcription
polymerase chain reaction (RT-PCR)
Nuclear extract from HeLa S3 cell lines was
prepared for in vitro splicing as described (57).
The in vitro splicing reaction, assembled in a
20-ml volume, was performed in the presence
of 1 nM pre-mRNA substrate and 50% nuclear
extract, in the buffer containing 20 mM HEPES-
KOH, pH 7.9, 65 mM KCl, 2 mM adenosine
5′-triphosphate (ATP), 20 mM creatine phosphate,
and 3 mM MgCl2. To examine the specificity and
activity of MINX-U12, we depleted U1, U2, and
U6 snRNAs (so as to examine the effect on U2-
type splicing), or U11, U12, and U6atac snRNAs
(so as to examine the effect on U12-type splic-
ing) in the nuclear extract using endogenous
RNase H at 30°C for 30 min before the splic-
ing reaction (fig. S1A). This was accomplished
by incubating the reaction with 1 mM anti-
sense DNA oligonucleotides for each of the
U1, U2, and U6 snRNAs (anti-U1 DNA oligo:
5′-CAGGTAAGTAT-3′; anti-U2 DNA oligo: 5′-
GAACAGATACTACACTTGA-3′; anti-U6 DNA
oligo: 5′-CTTCTCTGTATCGTTCCAATTTTA-
GTAT-3′), or for each of the U11, U12, and
U6atac snRNAs (anti-U11 DNA oligo: 5′-CA-
CGACAGAAGCCCTTTT-3′, 5′-TTCCGCACGCA-
GAGCAATCG-3′, 5′-GCTTCCGAAATCTCTTG-
ATG-3′, and 5′-GGGCGCCGGGACCAACGATC-3′;
anti-U12 DNA oligo: 5′-CCTTACTCATAAGTT-
TAAGGCA-3′, 5′-GTGAGGATTCG-GGCGTCAC-
CCC-3′, 5′-AGGCATCCCGCAAAGTAGGCGGG-3′,
Bai et al., Science 371, eabg0879 (2021)
19 March 2021
8 of 12
RES EARCH | R E S E A R C H A R T I C L E
and 5′-CCTTGA-GGGCGACCTTTACCCGC-3′;
anti-U6 DNA oligo: 5′-CTAACCTTCTCTCCT-
TTCATACAAC-3′, 5′-CGATGGTTAGATGCCACG
AAG-3′, 5′-GAGGGCCTCTTCCATCCTTGTC-3′,
and 5′-CAATGCCTTAACCG TATGCGTG-3′).
The splicing reaction mixture was incubated
at 30°C for varying time points of 0, 30, 60,
and 120 min, followed by proteinase K diges-
tion. RNA from the in vitro splicing assay was
extracted using phenol:chloroform:isopentanol
at a volume ratio of 25:24:1 (Coolaber Science
& Technology). Reverse transcription was per-
formed using High-Capacity cDNA Reverse
Transcription Kit (Applied Biosystems) and
random hexamers. The RT-PCR products were
resolved on 2% (w/v) agarose gel and stained by
GoldView (Beijing SBS Genetech Co., Ltd.)
(fig. S1A).
Assembly and purification of the human minor
Bact complex
The protocol for assembly and purification
of the minor Bact complex was modified from
that for the major Bact complex (35, 36) (fig.
S1B). Briefly, the splicing reaction was per-
formed in a volume of 40 ml or multiples
thereof, containing 20 mM HEPES-KOH, pH
7.9, 65 mM KCl, 2 mM ATP, 20 mM creatine
phosphate, and 3 mM MgCl2, in the presence
of 10 nM pre-mRNA, 450 nM MS2-MBP, and
50% splicing extract. The pre-mRNA was pre-
bound to MS2-MBP for 30 min on ice. The
reaction mixture was incubated for 4 hours
at 30°C and then centrifuged at 3000 g for
15 min to remove aggregates. The supernatant
was loaded onto amylose resin (NEB), allowed
to flow through the resin three times, and
washed using the G50K buffer [10 mM
HEPES-KOH, pH 7.9, 50 mM KCl, 1.5 mM
MgCl2, 0.01% NP40, and 5% (v/v) glycerol].
The spliceosomal complexes were eluted using
30 mM maltose. The eluent was loaded onto
a 10 to 30% glycerol gradient with 0 to 0.01%
glutaraldehyde (Sigma), and centrifuged at
23,000 rpm for 10 hours at 4°C in a SW41Ti
rotor (fig. S1B). Fractions of 1-ml aliquot
each were quenched by 50 mM Tris (pH 7.6),
and then examined on denaturing RNA gels
(fig. S1C). The fractions that contained the
minor Bact complex were pooled and dia-
lyzed against Buffer D (20 mM HEPES-KOH,
pH 7.9, 50 mM KCl, 1.5 mM MgCl2, 0.01% NP40)
to remove glycerol before sample prepara-
tion for electron microscopy (EM). The dia-
lyzed sample was concentrated for cryo-EM
studies.
Northern blot
Samples of the purified minor Bact complex
were heated at 95°C for 10 min, separated on
10% polyacrylamide gels containing 8M
urea. RNAs were transferred onto Hybond-N+
membranes (Amersham) in 1xTBE (Coolaber
Science & Technology) for 1 hour at 400 mA.
Membranes were fixed by UV irradiation and
hybridized to 32P-end-labeled DNA oligo-
nucleotide probes. The U12 snRNA specific
probes are derived from a published study
(24) and include: 5′-CCTTACTCATAAGTT-
TAAGGCA-3′, 5′-GTGAGGATTCGGGCGT-
CACCCC-3′, 5′-AGGCATCCCGCAAAGTAGG-
CGGG-3′, and 5′-CCTTGAGGGCGACCTT-
TACCCGC-3′. The specific probes of U6atac
snRNA include: 5′-CTAACCTTCTCTCCTTT-
CATACAAC-3′, 5′-CGATGGTTAGATGCCAC-
GAAG-3′, 5′-GAGGGCCTCTTCCATCCTTGTC-3′,
and 5′-CAATGCCTTAACCGTATGCGTG-3′. The
specific probes of U5 snRNA include: 5′-GC-
GATCTGAAGAGAAACCAGAG-3′, 5′-CTTGC-
CAAAGCAAGGCCTC-3′, 5′-GGGTTAAGACTC-
AGAGTTGTTC-3′, and 5′-CTCCACGGAAAT CT-
TTAGTAAAAGGC-3′. Membranes were ex-
posed to phosphorimager screens (Amersham
Biosciences) (fig. S1C). The exposure time is
~48 hours for the DNA probes targeting U12,
U6atac, and U5 snRNAs. The phosphorimager
screen was scanned by Personal Molecular
Imager (BioRad).
MS analysis
About 40 ml of the minor spliceosome sam-
ple was mixed with 10 ml of 5×SDS sample
loading buffer (GenScript Biotech, China)
supplemented with 150 mM dithiothreitol.
The sample was incubated at 95°C for 5 min,
and resolved using a 4 to 12% gradient SDS–
polyacrylamide gel electrophoresis (SDS-PAGE)
gel. The proteins were subjected to in-gel pro-
teolytic digestion as described (58). Peptides
were purified using Pierce C18 Spin Tips
(Thermo Fisher, USA) before liquid chroma-
tography coupled to tandem mass spectrom-
etry (LC-MS/MS) analysis using Ultimate
3000 nanoLC system coupled with Q Exactive
HF-X Hybrid Quadrupole-Orbitrap (Thermo
Fisher Scientific, San Jose, CA). About 500 ng
peptides were separated over the course of
90 min using a linear LC gradient of 3 to
28% (buffer A: 2% acetonitrile, 0.1% formic
acid; buffer B: 98% acetonitrile, 0.1% formic
acid) at a flow rate of 300 nl/min. The top
20 peptides were subjected to MS2 analy-
sis. MS2 spectra were acquired at the resolu-
tion of 30,000 (at m/z 200) in the orbitrap
using an AGC target of 1 × 105, and a max-
imum IT of 80 ms. Dynamic exclusion was
applied with a repeat count of 1 and an ex-
clusion time of 25 s. The resultant mass spec-
trometric data were analyzed using pFind (59)
(version 3.1.5) against the H. sapiens FASTA
database downloaded from UniProtKB (ver-
sion from 27 April 2020), which contains
20,365 reviewed protein sequences. Cysteine
carbamidomethyl was set as fixed modification
and methionine oxidation was set as variable
modification. A summary of mass spectromet-
ric analysis for the human minor Bact complex
is listed in table S1.
EM data acquisition and processing of the minor
Bact complex
Grids for cryo-EM data collection were carried
out essentially as described (60). Briefly, the
Quantifoil R1.2/1.3 grids coated with homemade
continuous carbon film of ~4-nm thickness were
used for cryo-EM specimen preparation. Cryo-
EM grids were prepared using Vitrobot Mark
IV (FEI Company) at 8°C and 100% humidity.
Four-microliter aliquots of the sample at a
concentration of ~0.2 mg/ml were applied to
glow-discharged grids, blotted for 1.5 s using
the standard vitrobot filter paper (Ø55/20 mm,
Ted Pella, Inc.), and plunged into liquid ethane
cooled by liquid nitrogen. The glow-discharged
grids were prepared for 30 s using the low
setting of the Plasma Cleaner (Harrick, Plasma
Cleaner PDC-32G).
The grids were loaded onto a FEI Titan
Krios electron microscope equipped with a
GIF Quantum energy filter (slit width 20 eV)
and operating at 300 kV with a nominal mag-
nification of 81,000×. Images were recorded
using a Gatan K3 detector (Gatan Company)
in the super-resolution mode, with a pixel size
of 0.5371 Å (fig. S1D). Each image was dose-
fractionated to 32 frames with a dose rate of
22.54 counts per second per physical pixel
(~19.531 e–/s per square angstrom) and a total
exposure time of 2.56 s. Total electron dose for
each image is ~50 e–/Å2. AutoEMation2 was
used for all data collection (61). All 32 frames
in each stack were aligned and summed using
the whole-image motion correction program
MotionCor2 (62) and binned to a pixel size of
1.0742 Å. The defocus value of each image was
set from 1.3 to 2.5 mm and was determined by
Gctf (63).
EM data processing for the human minor
Bact complex
In total, 3,500,973 particles were automatically
picked by DeepPicker (64) from 20,390 micro-
graphs (fig. S2). These particles were extracted
using a pixel size of 4.2968 Å and subjected
to three parallel guided multireference global
search 3D classification using RELION3.0 (65).
The volumes representing the human Bact com-
plex [Electron Microscopy Data Bank (EMDB)
ID: EMD-6891 (36)], 17S U2 snRNP [EMDB:
EMD-10689 (66)], tri-snRNP [EMDB: EMD-
6581 (67)], B complex [EMDB: EMD-9624 (68)],
C complex [EMDB: EMD-6864 (53)], and four
bad classes were low-pass filtered to 35 Å and
used as the initial references (fig. S2). The
particles from the best classes were merged
and the duplicated particles were removed
as described (69). Another subsequent three
parallel local search 3D classification (30 itera-
tions, T = 4, local angular search range = 15°)
was performed to further identify good par-
ticles, using the same references above (round
1). The resulting 1,217,469 particles were re-
centered and re-extracted with a pixel size of
Bai et al., Science 371, eabg0879 (2021)
19 March 2021
9 of 12
RES EARCH | R E S E A R C H A R T I C L E
2.1484 Å. After one round of refinement, the
resulting map is low-pass filtered to 15, 25,
35, and 60 Å. Together with the refinement
map, we obtained five references with a reso-
lution gradient. Three parallel guided multi-
reference local 3D classifications (30 iterations,
T = 4, local angular search range = 15°) using
the five references above were performed,
yielding 101,443 good particles. These particles
were further recentered and re-extracted using
a pixel size of 1.0742 Å, yielded a reconstruction
at an average resolution of 3.2 Å for the entire
human minor Bact complex. After CTF refine-
ment by RELION3.0 and Bayesian polishing
by RELION3.1, the resolution was further
improved to 2.89 Å (figs. S2 and S3A). To im-
prove the map quality for the SF3B region,
focused refinement with a soft mask on the
SF3b complex was performed (fig. S2). Another
round of focused 3D refinement with the mask
for one of the WD40 domains from SF3b130
further improved the map quality for this
region. For the 5′ end of U6atac snRNA
(RBM48-ARMC7 region), focused 3D classi-
fication (30 iterations, T = 4, local angular
search range = 15°) was performed with a soft
mask around this region. Then, 66,297 par-
ticles were selected and subjected to focused
3D refinement using the same mask as 3D
classification.
In the final EM map of the human minor
Bact complex, the local resolution reaches 2.6 Å
in the core region (fig. S3B). The angular dis-
tribution of the particles used for the final re-
construction of the human minor Bact complex
is reasonable (fig. S3C). The refinement of the
atomic coordinates did not suffer from severe
overfitting (fig. S3D). The EM maps show
fine features for individual component of the
human minor Bact complex (figs. S4 to S8)
and allow atomic modeling of five previously
unidentified proteins (PPIL2, SCNM1, CRIPT,
RBM48, and ARMC7) (figs. S6 to S8).
Reported resolutions were calculated on the
basis of the FSC 0.143 criterion, and the FSC
curves were corrected using high-resolution
noise substitution methods (70). Before visu-
alization, all density maps were corrected for
the modulation transfer function (MTF) of the
detector and sharpened by applying a negative
B-factor that was estimated using automated
procedures (71). Local resolution variations were
estimated using RELION.
Model building and refinement of the human
minor Bact complex
We combined homology modeling, rigid dock-
ing of known structures, de novo modeling,
and manual adjustment with an AI-guided
deep natural network method (table S3). Most
of the individual protein components were
identified using the atomic coordinates of the
human Bact complex [PDB codes 5Z56, 5Z58
(36), and 6FF7 (35)]. Structures of most com-
ponents (which include U5 snRNP, U6atac
snRNA, U12 snRNA and U12 Sm ring of U12
snRNP, SF3b complex of U12 snRNP, RES
complex, two components of NTC, three com-
ponents of NTR and CWC27, PRP2, RNF113A,
SRm160, CWC22, and SRm300) were docked
into the density map and manually adjusted
using COOT (72). Position of the C terminus
of U11/U12-65K, which shares homology
with U2-B′′, was predicted according to the
location of U2-B′′ using the crystal structure
3EGN (73). However, the EM density for this
region is inadequate for identification of
U11/U12-65K. The location of the N terminus
of GPKOW (Spp2 in S. cerevisiae) was identi-
fied according to the 2.5-Å cryo-EM structure
of the S. cerevisiae Bact complex (PDB code
7DCO) (56).
After these steps, a number of EM density
patches still remained unaccounted for, in-
cluding the regions around loop I of U5 snRNA,
SF3b155, SF3b130, and the 5′ end of U6atac.
The N-terminal ZnF region of SF3a66 (homo-
log of S. cerevisiae Prp11) clearly did not fit
into the well-resolved density map. To identify
the unknown protein, we used an improved
version of the A2-Net deep neural network
method (74) to recognize the side chains of
amino acids from EM density maps and to con-
nect adjacent amino acids to generate protein
fragments. Together with MS data, this practice
revealed protein candidates. The sequences of
the selected protein candidates and the density
map were then fed into the deep neural net-
work to generate the atomic model. Models of
PPIL2, SCNM1, CRIPT, and RBM48 were gen-
erated using this approach. Human RBM48
was reported to interact with ARMC7 (47, 49).
Next to RBM48, there was an EM density lobe
that closely resembles the overall shape of the
armadillo repeat helices. An initial model for
ARMC7 was generated using the protein struc-
ture prediction web server trRosetta (75). The
predicted structure of ARMC7 was docked into
the density map and manually adjusted using
COOT (72).
The entire model of the human minor Bact
complex was refined against the 2.89-Å map
using PHENIX (76) in real space with sec-
ondary structure and geometry restraints.
Model overfitting was monitored by refin-
ing the model in one of the two independent
maps and testing the refined model against
the other map (77) (fig. S3D). Model quality
was assessed using the Molprobity scores and
the Ramachandran plots (table S2). Molprobity
scores were calculated as described (78). All
EM map images were created using UCSF
Chimera (79).
RE FERENCES AND NOTES
1. E. Brody, J. Abelson, The “spliceosome”: Yeast pre-messenger
RNA associates with a 40S complex in a splicing-dependent
reaction. Science 228, 963–967 (1985). doi: 10.1126/
science.3890181; pmid: 3890181
2. P. J. Grabowski, S. R. Seiler, P. A. Sharp, A multicomponent
complex is involved in the splicing of messenger RNA
precursors. Cell 42, 345–353 (1985). doi: 10.1016/S0092-8674
(85)80130-6; pmid: 3160482
3. D. Frendewey, W. Keller, Stepwise assembly of a pre-mRNA
splicing complex requires U-snRNPs and specific intron
sequences. Cell 42, 355–367 (1985). doi: 10.1016/S0092-8674
(85)80131-8; pmid: 3160483
4. C. B. Burge, T. Tuschl, P. A. Sharp, in The RNA World,
Second Edition, R. F. Gesteland, T. R. Cech, J. F. Atkins,
Eds. (Cold Spring Harbor Laboratory Press, 1999),
pp. 525–560.
5. M. J. Moore, C. C. Query, P. A. Sharp, in The RNA World,
R. F. Gesteland, J. F. Atkins, Eds. (Cold Spring Harbor
Laboratory Press, 1993), pp. 303–357.
J. P. Staley, C. Guthrie, Mechanical devices of the
spliceosome: Motors, clocks, springs, and things. Cell 92,
315–326 (1998). doi: 10.1016/S0092-8674(00)80925-3;
pmid: 9476892
6.
7. M. C. Wahl, C. L. Will, R. Lührmann, The spliceosome:
Design principles of a dynamic RNP machine. Cell 136,
701–718 (2009). doi: 10.1016/j.cell.2009.02.009;
pmid: 19239890
8. A. Levine, R. Durbin, A computational scan for U12-
dependent introns in the human genome sequence. Nucleic
Acids Res. 29, 4006–4013 (2001). doi: 10.1093/nar/
29.19.4006; pmid: 11574683
9. N. Sheth et al., Comprehensive splice-site analysis using
comparative genomics. Nucleic Acids Res. 34, 3955–3967
(2006). doi: 10.1093/nar/gkl556; pmid: 16914448
10. I. J. Jackson, A reappraisal of non-consensus mRNA splice
sites. Nucleic Acids Res. 19, 3795–3798 (1991). doi: 10.1093/
nar/19.14.3795; pmid: 1713664
11. S. L. Hall, R. A. Padgett, Conserved sequences in a class of
rare eukaryotic nuclear introns with non-consensus splice
sites. J. Mol. Biol. 239, 357–365 (1994). doi: 10.1006/
jmbi.1994.1377; pmid: 8201617
12. P. A. Sharp, C. B. Burge, Classification of introns: U2-type or
U12-type. Cell 91, 875–879 (1997). doi: 10.1016/S0092-8674
(00)80479-1; pmid: 9428511
13. W. Y. Tarn, J. A. Steitz, Pre-mRNA splicing: The discovery of a
new spliceosome doubles the challenge. Trends Biochem. Sci.
22, 132–137 (1997). doi: 10.1016/S0968-0004(97)01018-9;
pmid: 9149533
14. Q. Wu, A. R. Krainer, Splicing of a divergent subclass of AT-AC
introns requires the major spliceosomal snRNAs. RNA 3,
586–601 (1997). pmid: 9174094
15. R. C. Dietrich, R. Incorvaia, R. A. Padgett, Terminal intron
dinucleotide sequences do not distinguish between U2- and
U12-dependent introns. Mol. Cell 1, 151–160 (1997).
doi: 10.1016/S1097-2765(00)80016-7; pmid: 9659912
16. A. A. Patel, J. A. Steitz, Splicing double: Insights from the
second spliceosome. Nat. Rev. Mol. Cell Biol. 4, 960–970
(2003). doi: 10.1038/nrm1259; pmid: 14685174
17. C. B. Burge, R. A. Padgett, P. A. Sharp, Evolutionary
fates and origins of U12-type introns. Mol. Cell 2, 773–785
(1998). doi: 10.1016/S1097-2765(00)80292-0;
pmid: 9885565
18. M. Lynch, A. O. Richardson, The evolution of spliceosomal
introns. Curr. Opin. Genet. Dev. 12, 701–710 (2002).
doi: 10.1016/S0959-437X(02)00360-X; pmid: 12433585
19. A. G. Russell, J. M. Charette, D. F. Spencer, M. W. Gray,
An early evolutionary origin for the minor spliceosome.
Nature 443, 863–866 (2006). doi: 10.1038/nature05228;
pmid: 17051219
20. M. Dávila López, M. A. Rosenblad, T. Samuelsson,
Computational screen for spliceosomal RNA genes aids in
defining the phylogenetic distribution of major and minor
spliceosomal components. Nucleic Acids Res. 36,
3001–3010 (2008). doi: 10.1093/nar/gkn142;
pmid: 18390578
21. S. Bartschat, T. Samuelsson, U12 type introns were lost
at multiple occasions during evolution. BMC Genomics
11, 106 (2010). doi: 10.1186/1471-2164-11-106;
pmid: 20149226
22. C. L. Will, R. Lührmann, Splicing of a rare class of
introns by the U12-dependent spliceosome. Biol. Chem. 386,
713–724 (2005). doi: 10.1515/BC.2005.084;
pmid: 16201866
23. J. J. Turunen, E. H. Niemelä, B. Verma, M. J. Frilander, The
significant other: Splicing by the minor spliceosome.
WIREs RNA 4, 61–76 (2013). doi: 10.1002/wrna.1141;
pmid: 23074130
Bai et al., Science 371, eabg0879 (2021)
19 March 2021
10 of 12
RES EARCH | R E S E A R C H A R T I C L E
24. K. A. Montzka, J. A. Steitz, Additional low-abundance human
small nuclear ribonucleoproteins: U11, U12, etc. Proc. Natl.
Acad. Sci. U.S.A. 85, 8885–8889 (1988). doi: 10.1073/
pnas.85.23.8885; pmid: 2973606
25. S. L. Hall, R. A. Padgett, Requirement of U12 snRNA for in vivo
splicing of a minor class of eukaryotic nuclear pre-mRNA
introns. Science 271, 1716–1718 (1996). doi: 10.1126/
science.271.5256.1716; pmid: 8596930
26. W. Y. Tarn, J. A. Steitz, Highly diverged U4 and U6 small
nuclear RNAs required for splicing rare AT-AC introns. Science
273, 1824–1832 (1996). doi: 10.1126/science.273.5283.1824;
pmid: 8791582
27. W. Y. Tarn, J. A. Steitz, A novel spliceosome containing U11,
U12, and U5 snRNPs excises a minor class (AT-AC) intron in
vitro. Cell 84, 801–811 (1996). doi: 10.1016/S0092-8674(00)
81057-0; pmid: 8625417
28. Y.-T. Yu, E. C. Scharl, C. M. Smith, J. A. Steitz, in The RNA
World, Second Edition, R. F. Gesteland, T. R. Cech, J. F. Atkins,
Eds. (Cold Spring Harbor Laboratory Press, 1999),
pp. 487–524.
29. C. L. Will et al., The human 18S U11/U12 snRNP contains a
set of novel proteins not found in the U2-dependent
spliceosome. RNA 10, 929–941 (2004). doi: 10.1261/
rna.7320604; pmid: 15146077
30. C. Schneider, C. L. Will, O. V. Makarova, E. M. Makarov,
R. Lührmann, Human U4/U6.U5 and U4atac/U6atac.U5
tri-snRNPs exhibit similar protein compositions. Mol. Cell. Biol.
22, 3219–3229 (2002). doi: 10.1128/MCB.22.10.3219-
3229.2002; pmid: 11971955
31. R. Wan, R. Bai, Y. Shi, Molecular choreography of pre-mRNA
splicing by the spliceosome. Curr. Opin. Struct. Biol. 59,
124–133 (2019). doi: 10.1016/j.sbi.2019.07.010; pmid:
31476650
32. R. Wan, R. Bai, X. Zhan, Y. Shi, How Is Precursor Messenger
RNA Spliced by the Spliceosome? Annu. Rev. Biochem. 89,
333–358 (2020). doi: 10.1146/annurev-biochem-013118-
111024; pmid: 31815536
33. C. Plaschka, A. J. Newman, K. Nagai, Structural Basis of
Nuclear pre-mRNA Splicing: Lessons from Yeast. Cold Spring
Harb. Perspect. Biol. 11, a032391 (2019). doi: 10.1101/
cshperspect.a032391; pmid: 30765413
34. B. Kastner, C. L. Will, H. Stark, R. Lührmann, Structural Insights
into Nuclear pre-mRNA Splicing in Higher Eukaryotes. Cold
Spring Harb. Perspect. Biol. 11, a032417 (2019). doi: 10.1101/
cshperspect.a032417; pmid: 30765414
35. D. Haselbach et al., Structure and Conformational Dynamics
of the Human Spliceosomal Bact Complex. Cell 172,
454–464.e11 (2018). doi: 10.1016/j.cell.2018.01.010;
pmid: 29361316
36. X. Zhang et al., Structure of the human activated spliceosome
in three conformational states. Cell Res. 28, 307–322 (2018).
doi: 10.1038/cr.2018.14; pmid: 29360106
37. S. Bessonov et al., Characterization of purified human Bact
spliceosomal complexes reveals compositional and
morphological changes during spliceosome activation and first
step catalysis. RNA 16, 2384–2403 (2010). doi: 10.1261/
rna.2456210; pmid: 20980672
38. C. Yan, R. Wan, R. Bai, G. Huang, Y. Shi, Structure of a yeast
activated spliceosome at 3.5 Å resolution. Science 353,
904–911 (2016). doi: 10.1126/science.aag0291;
pmid: 27445306
39. R. Rauhut et al., Molecular architecture of the Saccharomyces
cerevisiae activated spliceosome. Science 353, 1399–1405
(2016). doi: 10.1126/science.aag1906; pmid: 27562955
40. H. Benecke, R. Lührmann, C. L. Will, The U11/U12 snRNP 65K
protein acts as a molecular bridge, binding the U12 snRNA and
U11-59K protein. EMBO J. 24, 3057–3069 (2005).
doi: 10.1038/sj.emboj.7600765; pmid: 16096647
41. D. E. Agafonov et al., Semiquantitative proteomic analysis of
the human spliceosome via a novel two-dimensional gel
electrophoresis method. Mol. Cell. Biol. 31, 2667–2682 (2011).
doi: 10.1128/MCB.05266-11; pmid: 21536652
42. L. K. Sprunger, A. Escayg, S. Tallaksen-Greene, R. L. Albin,
M. H. Meisler, Dystonia associated with mutation of the
neuronal sodium channel Scn8a and identification of the
modifier locus Scnm1 on mouse chromosome 3. Hum. Mol.
Genet. 8, 471–479 (1999). doi: 10.1093/hmg/8.3.471;
pmid: 9949206
43. D. A. Buchner, M. Trudeau, M. H. Meisler, SCNM1, a putative
RNA splicing factor that modifies disease severity in mice.
Science 301, 967–969 (2003). doi: 10.1126/science.1086187;
pmid: 12920299
44. C. Rajiv, T. L. Davis, Structural and Functional Insights into
Human Nuclear Cyclophilins. Biomolecules 8, 161 (2018).
doi: 10.3390/biom8040161; pmid: 30518120
45. T. L. Davis et al., Structural and biochemical characterization of
the human cyclophilin family of peptidyl-prolyl isomerases.
PLOS Biol. 8, e1000439 (2010). doi: 10.1371/journal.
pbio.1000439; pmid: 20676357
46. S. Hatakeyama, M. Yada, M. Matsumoto, N. Ishida,
K. I. Nakayama, U box proteins as a new family of ubiquitin-
protein ligases. J. Biol. Chem. 276, 33111–33120 (2001).
doi: 10.1074/jbc.M102755200; pmid: 11435423
47. F. Bai et al., RNA Binding Motif Protein 48 Is Required
for U12 Splicing and Maize Endosperm Differentiation.
Plant Cell 31, 715–733 (2019). doi: 10.1105/tpc.18.00754;
pmid: 30760564
48. A. E. Siebert, J. Corll, J. P. Gronevelt, L. Levine, L. M. Hobbs,
C. Kenney, R. Davenport, A. M. Settles, W. B. Barbazuk,
R. J. Westrick, G. J. Madlambayan, S. Lal, Evolutionary
Conservation and Essential Function of Human and Maize RNA
Binding Motif Protein 48 (RBM48) in U12-Type Intron Splicing.
bioRxiv 2020.07.18.209528 [Preprint]. 19 July 2020.
doi: 10.1101/2020.07.18.209528
49. T. Hart et al., High-Resolution CRISPR Screens Reveal Fitness
Genes and Genotype-Specific Cancer Liabilities. Cell 163,
1515–1526 (2015). doi: 10.1016/j.cell.2015.11.015;
pmid: 26627737
50. C. Townsend et al., Mechanism of protein-guided folding of the
active site U2/U6 RNA during spliceosome activation. Science
370, eabc3753 (2020). doi: 10.1126/science.abc3753;
pmid: 33243851
51. J. A. Kearney et al., Molecular and pathological effects of a
modifier gene on deficiency of the sodium channel Scn8a
(Nav1.6). Hum. Mol. Genet. 11, 2765–2775 (2002).
doi: 10.1093/hmg/11.22.2765; pmid: 12374766
52. D. A. Buchner, M. Trudeau, A. L. George Jr.., L. K. Sprunger,
M. H. Meisler, High-resolution mapping of the sodium channel
modifier Scnm1 on mouse chromosome 3 and identification of
a 1.3-kb recombination hot spot. Genomics 82, 452–459
(2003). doi: 10.1016/S0888-7543(03)00152-6;
pmid: 13679025
53. X. Zhan, C. Yan, X. Zhang, J. Lei, Y. Shi, Structure of a human
catalytic step I spliceosome. Science 359, 537–545 (2018).
doi: 10.1126/science.aar6401; pmid: 29301961
54. B. Verma, M. V. Akinyi, A. J. Norppa, M. J. Frilander, Minor
spliceosome and disease. Semin. Cell Dev. Biol. 79, 103–112
(2018). doi: 10.1016/j.semcdb.2017.09.036; pmid: 28965864
55. H. W. Mewes et al., Overview of the yeast genome. Nature 387,
7–8 (1997). doi: 10.1038/387s007; pmid: 9169865
56. R. Bai et al., Mechanism of spliceosome remodeling by the
ATPase/helicase Prp2 and its coactivator Spp2. Science 371,
eabe8863 (2021). doi: 10.1126/science.abe8863;
pmid: 33243853
57. J. D. Dignam, R. M. Lebovitz, R. G. Roeder, Accurate
transcription initiation by RNA polymerase II in a soluble
extract from isolated mammalian nuclei. Nucleic Acids Res. 11,
1475–1489 (1983). doi: 10.1093/nar/11.5.1475; pmid: 6828386
58. T. Zhu et al., DPHL: A DIA Pan-human Protein Mass
Spectrometry Library for Robust Biomarker Discovery.
Genomics Proteomics Bioinformatics 18, 104–119 (2020).
doi: 10.1016/j.gpb.2019.11.008; pmid: 32795611
59. L. H. Wang et al., pFind 2.0: A software package for peptide
and protein identification via tandem mass spectrometry.
Rapid Commun. Mass Spectrom. 21, 2985–2991 (2007).
doi: 10.1002/rcm.3173; pmid: 17702057
60. R. Bai, R. Wan, C. Yan, J. Lei, Y. Shi, Structures of the fully
assembled Saccharomyces cerevisiae spliceosome before
activation. Science 360, 1423–1429 (2018). doi: 10.1126/
science.aau0325; pmid: 29794219
61. J. Lei, J. Frank, Automated acquisition of cryo-electron
micrographs for single particle reconstruction on an FEI Tecnai
electron microscope. J. Struct. Biol. 150, 69–80 (2005).
doi: 10.1016/j.jsb.2005.01.002; pmid: 15797731
62. S. Q. Zheng et al., MotionCor2: Anisotropic correction of
beam-induced motion for improved cryo-electron microscopy.
Nat. Methods 14, 331–332 (2017). doi: 10.1038/nmeth.4193;
pmid: 28250466
63. K. Zhang, Gctf: Real-time CTF determination and correction.
J. Struct. Biol. 193, 1–12 (2016). doi: 10.1016/j.jsb.2015.11.003;
pmid: 26592709
64. F. Wang et al., DeepPicker: A deep learning approach for fully
automated particle picking in cryo-EM. J. Struct. Biol. 195,
325–336 (2016). doi: 10.1016/j.jsb.2016.07.006;
pmid: 27424268
65. J. Zivanov et al., New tools for automated high-resolution cryo-
EM structure determination in RELION-3. eLife 7, e42166
(2018). doi: 10.7554/eLife.42166; pmid: 30412051
66. Z. Zhang et al., Molecular architecture of the human 17S U2
snRNP. Nature 583, 310–313 (2020). doi: 10.1038/s41586-
020-2344-3; pmid: 32494006
67. D. E. Agafonov et al., Molecular architecture of the human U4/
U6.U5 tri-snRNP. Science 351, 1416–1420 (2016). doi: 10.1126/
science.aad2085; pmid: 26912367
68. X. Zhan, C. Yan, X. Zhang, J. Lei, Y. Shi, Structures of the
human pre-catalytic spliceosome and its precursor
spliceosome. Cell Res. 28, 1129–1140 (2018). doi: 10.1038/
s41422-018-0094-7; pmid: 30315277
69. R. Wan, R. Bai, C. Yan, J. Lei, Y. Shi, Structures of the
Catalytically Activated Yeast Spliceosome Reveal the
Mechanism of Branching. Cell 177, 339–351.e13 (2019).
doi: 10.1016/j.cell.2019.02.006; pmid: 30879786
70. S. Chen et al., High-resolution noise substitution to measure
overfitting and validate resolution in 3D structure
determination by single particle electron cryomicroscopy.
Ultramicroscopy 135, 24–35 (2013). doi: 10.1016/j.
ultramic.2013.06.004; pmid: 23872039
71. P. B. Rosenthal, R. Henderson, Optimal determination of
particle orientation, absolute hand, and contrast loss
in single-particle electron cryomicroscopy. J. Mol. Biol. 333,
721–745 (2003). doi: 10.1016/j.jmb.2003.07.013;
pmid: 14568533
72. P. Emsley, K. Cowtan, Coot: Model-building tools for molecular
graphics. Acta Cryst. D60, 2126–2132 (2004). doi: 10.1107/
S0907444904019158; pmid: 15572765
73. C. Netter, G. Weber, H. Benecke, M. C. Wahl, Functional
stabilization of an RNA recognition motif by a noncanonical
N-terminal expansion. RNA 15, 1305–1313 (2009).
doi: 10.1261/rna.1359909; pmid: 19447915
74. K. Xu, Z. Wang, J. Shi, H. Li, Q. C. Zhang, A2-Net: Molecular
Structure Estimation from Cryo-EM Density Volumes. Proc.
AAAI Conf. Artif. Intell. 33, 1230–1237 (2019); .doi: 10.1609/
aaai.v33i01.33011230
75. J. Yang et al., Improved protein structure prediction using
predicted interresidue orientations. Proc. Natl. Acad. Sci. U.S.A.
117, 1496–1503 (2020). doi: 10.1073/pnas.1914677117;
pmid: 31896580
76. P. V. Afonine et al., Real-space refinement in PHENIX for cryo-
EM and crystallography. Acta Cryst. D74, 531–544 (2018).
doi: 10.1107/S2059798318006551; pmid: 29872004
77. A. Amunts et al., Structure of the yeast mitochondrial large
ribosomal subunit. Science 343, 1485–1489 (2014).
doi: 10.1126/science.1249410; pmid: 24675956
78. I. W. Davis et al., MolProbity: All-atom contacts and structure
validation for proteins and nucleic acids. Nucleic Acids Res. 35,
W375–W383 (2007). doi: 10.1093/nar/gkm216;
pmid: 17452350
79. E. F. Pettersen et al., UCSF Chimera—A visualization
system for exploratory research and analysis. J. Comput.
Chem. 25, 1605–1612 (2004). doi: 10.1002/jcc.20084;
pmid: 15264254
80. T. D. Goddard et al., UCSF ChimeraX: Meeting modern
challenges in visualization and analysis. Protein Sci. 27, 14–25
(2018). doi: 10.1002/pro.3235; pmid: 28710774
81. W. L. DeLano, The PyMOL Molecular Graphics System (2002);
https://pymol.org/2/.
82. V. M. Howell et al., Evidence for a direct role of the disease
modifier SCNM1 in splicing. Hum. Mol. Genet. 16, 2506–2516
(2007). doi: 10.1093/hmg/ddm206; pmid: 17656373
AC KNOWLED GME NTS
We thank M. Yuan and G. Zhao from Y. Qi’s laboratory for
technical assistance with Northern blot experiments; X. Yi and
T. Guo for technical support of the mass spectrometric
analyses; F. Yang and X. Li for technical support during EM data
collection; and C. Yan for help during model building. We thank
the Tsinghua University Branch of the China National Center
for Protein Sciences (Beijing) for providing the facility support.
The computation was completed on the Explorer 100 cluster
system of Tsinghua National Laboratory for Information Science
and Technology. Funding: This work was supported by funds
from the National Natural Science Foundation of China
(31930059), the China Postdoctoral Science Foundation
(2019M662120 to R.B.), the National Postdoctoral Program for
Innovative Talents of China (BX20200303 to R.B.), and start-up
funds from Westlake University (to Y.S.). Author contributions:
R.B., R.W., and Y.S. conceived of the project. R.B. and R.W.
designed the experiments. R.B. and L.W. prepared the Hela
Bai et al., Science 371, eabg0879 (2021)
19 March 2021
11 of 12
RES EARCH | R E S E A R C H A R T I C L E
nuclear extract, and R.B. and R.W. purified the spliceosome and
prepared the samples for cryo-EM. R.B., R.W., and J.L. collected
and processed the EM data. R.W. calculated the EM map and
built the atomic model. K.X. and Q.Z. helped identify the newly
identified proteins using the A2-Net deep neural network
method. All authors contributed to data analysis. Y.S., R.W., and
R.B. wrote the manuscript. Y.S. supervised the project.
Competing interests: The authors declare no competing
interests. Data and materials availability: The atomic
coordinates have been deposited in the Protein Data Bank with
the accession code 7DVQ. The EM maps have been deposited in
the Electron Microscopy Data Bank with the accession codes
EMD-30875, EMD-30878, and EMD-30879. Materials are
available from the corresponding authors on request.
Figs. S1 to S11
Tables S1 to S3
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
SUPPLEMENTARY MATERIALS
science.sciencemag.org/content/371/6535/eabg0879/suppl/DC1
9 December 2020; accepted 18 January 2021
Published online 28 January 2021
10.1126/science.abg0879
Bai et al., Science 371, eabg0879 (2021)
19 March 2021
12 of 12
|
10.1126_science.abi4882
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
MICROBIOLOGY
Spatial transcriptomics of planktonic and sessile
bacterial populations at single-cell resolution
Daniel Dar, Nina Dar, Long Cai*, Dianne K. Newman*
INTRODUCTION: Microbial populations display
heterogeneous gene expression profiles that
result in phenotypic differences between in-
dividual bacteria. This diversity can allow
populations to survive under uncertain and
fluctuating conditions such as sudden anti-
biotic exposure, divide costly functions across
different subpopulations, and enable interac-
tions between different phenotypes. In addition
to the temporal phenotypic heterogeneity seen
in planktonic cultures, microbial populations
and communities often exist in multicellular
biofilms that exhibit considerable heterogene-
ity at the microscale, both in the local phys-
icochemistry that individuals experience and
in the species composition in their neighbor-
hoods. Phenotypic and microscale variation
represent central features of microbial pop-
ulations, but the landscape of possible cellular
states, their spatiotemporal regulation, and
their roles in many biological phenomena are
still largely unknown.
RATIONALE: The microscale heterogeneity that
defines microbial life can play important roles
in community organization and function, in-
cluding in antibiotic resistance and virulence.
However, our understanding of these basic
features has been limited by our ability to
capture this heterogeneity at the relevant
spatiotemporal scales. Overcoming these lim-
itations could lead to new insights into the
inner workings of microbial assemblages.
RESULTS: We developed par-seqFISH (parallel
sequential fluorescence in situ hybridization),
a high-throughput method that captures gene
expression profiles of individual bacteria while
also preserving their physical context within
spatially structured environments. We applied
this approach to the study of Pseudomonas
aeruginosa, a model biofilm-forming bacte-
rium and an opportunistic human pathogen.
Focusing on a set of 105 marker genes repre-
senting key aspects of P. aeruginosa physiology
and virulence, we explored the transcriptional
profiles of >600,000 bacteria across dozens
of growth conditions. We uncovered a diverse
set of metabolic- and virulence-related cellular
states and quantified their temporal dynamics
during population growth. In addition to re-
cording gene expression, we also demonstrated
that par-seqFISH captures cell biological pa-
rameters such as cell size and can be further
integrated with specific dyes to measure fea-
tures such as chromosome copy in the same
cells. Applying par-seqFISH to developing
P. aeruginosa biofilms, we exposed the bio-
geographic context of cellular states, provid-
ing new insights into the spatial expression of
various genes. These included, among other
things, mutually exclusive expression patterns
of flagella and type IV pili genes and a local-
ized induction of pyocins, which are involved
in kin selection and extracellular DNA release.
Looking more closely, we found that pyocin-
encoding transcripts strongly localized to the
bacterial cell poles. In early biofilms, we iden-
tified extensive microscale phenotypic structur-
ing in which anaerobic metabolic processes
such as denitrification appeared to locally
influence the microenvironment through by-
product production. In more mature biofilms,
we detected larger-scale partitions into zones
of differential metabolic activities and viru-
lence factor biosynthesis.
CONCLUSION: Transcriptome imaging using
par-seqFISH captures the microscale pheno-
typic variation of free-living and sessile bacterial
populations. We report extensive heterogene-
ity in growing P. aeruginosa populations and
demonstrate that individual multicellular bio-
films can contain coexisting but separated
subpopulations with distinct physiological
activities. This multiplexed and spatially re-
solved method offers a generalizable tool for
studying bacterial populations in space and
time directly in their native contexts. Future
applications in natural and clinical samples
will provide insights into the conditions ex-
perienced by microbes in complex environments
and the coordinated physiological responses
that emerge in turn and reshape them.▪
The list of author affiliations is available in the full article online.
*Corresponding author. Email: dkn@caltech.edu (D.K.N.);
lcai@caltech.edu (L.C.)
Cite this article as D. Dar et al., Science 373, eabi4882 (2021).
DOI: 10.1126/science.abi4882
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.abi4882
Single-cell transcriptional profiling of planktonic
and biofilm populations with par-seqFISH
Exploring the spatial context of cellular states
High replicative
capacity
Flagella biosynthesis
Pyocin induction
Aerobic
metabolism
Dentrification and
fermentation
Starvation-induced
oxidase
Exoprotease producers
Dar et al., Science 373, 758 (2021)
13 August 2021
Transcriptome imaging
using par-seqFISH reveals
the dynamics and spatial
organization of transcrip-
tional programs in
P. aeruginosa populations
at single-cell resolution.
Transcriptional states of
individual bacterial cells
were identified using
clustering analysis (left).
The detected cellular states
are depicted in different
colors. Cell metabolic
states can be directly
mapped to their native
biofilm context to identify
emerging microenvironment
dynamics (right).
1 of 1
RES EARCH
R E S E A R C H A R T I C L E
◥
MICROBIOLOGY
Spatial transcriptomics of planktonic and sessile
bacterial populations at single-cell resolution
Daniel Dar12†, Nina Dar1, Long Cai1*, Dianne K. Newman12*
Capturing the heterogeneous phenotypes of microbial populations at relevant spatiotemporal scales
is highly challenging. Here, we present par-seqFISH (parallel sequential fluorescence in situ hybridization), a
transcriptome-imaging approach that records gene expression and spatial context within microscale
assemblies at a single-cell and molecule resolution. We applied this approach to the opportunistic
pathogen Pseudomonas aeruginosa, analyzing about 600,000 individuals across dozens of conditions
in planktonic and biofilm cultures. We identified numerous metabolic- and virulence-related transcriptional
states that emerged dynamically during planktonic growth, as well as highly spatially resolved
metabolic heterogeneity in sessile populations. Our data reveal that distinct physiological states can
coexist within the same biofilm just several micrometers away, underscoring the importance of the
microenvironment. Our results illustrate the complex dynamics of microbial populations and present a
new way of studying them at high resolution.
L ife exists in context. Cells within micro-
bial populations and communities are
typically closely associated with one an-
other in multicellular biofilms, whether
found within infected tissues, attached
to surfaces, or forming assemblages in the
deep sea (1, 2). Natural microbiota and infec-
tious bacteria generally exist in biofilm ag-
gregates that are several dozen micrometers
across and can contain many interacting spe-
cies (3–5). Despite the ubiquity of the biofilm
lifestyle in both natural and manmade habi-
tats, understanding what life is like within a
biofilm for individual microbes has proven
challenging. Whereas single-cell–level activ-
ities have been tracked at high spatial reso-
lution using a variety of approaches in diverse
contexts (6–8), we have been unable to resolve
the hundreds, if not thousands, of concurrent
activities that characterize microbial life at rel-
evant spatiotemporal scales. What we under-
stand about microbial life literally has been
limited by our ability to see.
Despite this limitation, it has become clear
in recent years that extreme phenotypic het-
erogeneity defines the microbial experience
(9, 10). This is as true for isogenic populations
as it is for complex biofilm communities. Clone-
mates sampled from the same environment
often display substantial differences that are
thought to result from stochastic gene ex-
pression and variable environmental factors
1Division of Biology and Biological Engineering, California
Institute of Technology, Pasadena, CA, USA. 2Division
of Geological and Planetary Sciences, California Institute of
Technology, Pasadena, CA, USA.
*Corresponding author. Email: dkn@caltech.edu (D.K.N.);
lcai@caltech.edu (L.C.)
†Present address: Department of Plant and Environmental
Sciences, Weizmann Institute of Science, Rehovot 76100, Israel.
(9, 11, 12). The detection of phenotypic diver-
sity even in seemingly well-mixed environ-
ments such as chemostats (11, 13) also serves
as a powerful reminder that life at the micro-
scale may inhabit far more diverse niches than
are readily apparent. Phenotypic diversity has
been rationalized as providing microbes with
a fitness advantage in an unpredictable world
(9, 14). In addition, specialized functions have
been proposed to underpin collective inter-
actions such as division of labor (9, 15–17).
However, little is still known about the range
of possible cellular phenotypic states and their
roles in most biological processes.
What triggers such phenotypic plasticity?
And are there underlying “rules” that govern
any patterns that may exist at the microscale?
In sessile communities, clonal or multispe-
cies, biological activities give rise to changing
chemical gradients that create a range of local
microenvironments (18, 19). Furthermore, spa-
tial organization enables different conflicting
metabolic states or species to coexist through
physical separation, increasing the potential
for diversity and allowing for new interac-
tions to emerge (10, 20–23). Indeed, natural
communities often contain many interacting
species that assemble into intricate spatial
structures. These microscale assemblies can
promote interactions between species and
represent a key ecosystem feature (23, 24).
However, a wide gulf, limited by technology,
still separates such observations from a coher-
ent conceptual framework to explain the rules
governing microbial ecology.
Recent advances in imaging methods pro-
vide a means to chart the physical associations
between different species in natural environ-
ments (4, 25–27), but interpreting these maps
remains challenging without additional func-
tional information on the physiological states
and activities of relevant community mem-
bers. By contrast, recent adaptations of single-
cell RNA sequencing (RNA-seq) approaches
to free-living bacteria provide a powerful
means of exploring their phenotypic land-
scape (28–30). However, these approaches do
not preserve the spatial context of analyzed
cells and are therefore limited in their capacity
to address single and multispecies biofilms.
Thus, a major gap exists in our ability to ac-
count for both spatial and functional complexity,
limiting progression toward a high-resolution
understanding of microbial life.
Single-molecule fluorescence in situ hy-
bridization (FISH)–based technologies have
been used to measure gene expression directly
within native tissues, recording both spatial
and functional information. These methods
can shed important light on single-cell hetero-
geneity but are traditionally limited to mea-
suring the expression of only a few genes at
a time (31–34). In addition to this limited
throughput, single-gene measurements do
not provide a means to capture coordinated
cellular responses, the molecular “fingerprint”
of multiple biological activities that underpin
distinct physiological states. Recent advances
in combinatorial mRNA labeling and sequen-
tial FISH (seqFISH) allow for hundreds or even
thousands of genes to be analyzed within the
same sample at a submicrometer resolution
(35–37). Until now, seqFISH has been used
in mammalian systems to expose the physi-
cal organization of cell states within tissues
(35–39). We reasoned that the high spatial
resolution of these modern transcriptome-
imaging techniques also had the potential
to illuminate the microscale organization of
microbial populations and communities.
In this study, we adapted and further de-
veloped seqFISH for studying bacteria, mea-
suring the expression of hundreds of genes
within individual cells while also capturing
their spatial context. We used Pseudomonas
aeruginosa planktonic and biofilm populations
to demonstrate how different cellular func-
tions are coordinated in time and space. Our
proof-of-concept work illustrates how the abil-
ity to observe transcriptional activities at the
microscale permits insights into the spatio-
temporal regulation and coordination of critical
life processes, enabling hitherto unrecognized,
transient physiological states to be identified
and new hypotheses to be generated. These
findings represent the tip of the iceberg, and
the opportunities for discovery enabled by this
approach promise to reveal new insights about
the rules governing microbial ecology.
A sequential mRNA-FISH framework for
studying bacterial gene expression
Combinatorial mRNA labeling requires that
each measured mRNA molecule be individually
Dar et al., Science 373, eabi4882 (2021)
13 August 2021
1 of 16
RES EARCH | R E S E A R C H A R T I C L E
resolved. This is much more challenging in
bacteria because of the small size of their
cells, as many different mRNA molecules oc-
cur in close proximity and cannot be resolved
using standard fluorescence microscopy.
We therefore used a nonbarcoded seqFISH
approach (40).
In seqFISH, target mRNAs are first hybridized
with a set of primary, nonfluorescent probes,
which are flanked by short sequences uniquely
assigned per gene (Fig. 1A). Specific genes can
be turned ON through a secondary hybridiza-
tion with short, fluorescently labeled “readout”
probes complementary to the gene-specific
flanking sequences (Fig. 1A). Several genes
can be measured at once using a set of readout
probes labeled with different fluorophores.
These short, fluorescent readout probes can
be efficiently stripped and washed away from
the sample without affecting the primary
probes (41) (Fig. 1A). Thus, once expression is
measured, fluorescence can be turned OFF
and a new set of genes can be measured by
introducing a new set of readout probes (Fig.
1B). This two-step design allows for potentially
hundreds of genes to be measured sequen-
tially, one after the other in the same sample,
using automated microscopy (Fig. 1B). The in-
dividual gene mRNA-FISH data can be com-
bined into spatially resolved multigene profiles
at the single-bacterium level (Fig. 1B).
Because of the diffraction limit and the small
size of bacteria, mRNA-FISH fluorescent signals
(appearing as spots within cells) can contain
overlapping mRNA molecules that cannot be
spatially resolved using standard microscopes.
Therefore, counting the number of spots within
a bacterial cell severely underestimates ex-
pression levels. This problem can be overcome
by integrating the fluorescence intensity per
spot, which scales linearly with the number
of mRNAs. Fluorescence intensity can be con-
verted to discrete mRNA counts by measuring
the characteristic intensity of a single tran-
script. This analog-to-digital conversion ap-
proach has been shown to provide a wide
dynamic range in bacteria (33, 42).
We developed seqFISH for the study of
P. aeruginosa, an opportunistic human path-
ogen and a severe cause of morbidity and
mortality in cystic fibrosis patients (43, 44).
A
2-step mRNA-FISH
B
sequential mRNA-FISH in bacteria
si
si
si
si
mRNA binding region (Pi) is flanked
by a unique secondary sequence (Si).
Pi
si
si
si
Secondary probe Si specifically binds Si
si = {
{
Primary + Secondary
“readout” probes
1
2
3
k
Secondary probe hybridization
“OFF”
“ON”
Probe stripping and removal
C
Parallel seqFISH: multiplexing bacterial conditions
Individually label conditions
Single-cell demultiplexing
and expression analysis
Pool samples
seqFISH
A
B
C
16S rRNA
16S rRNA
16S rRNA
Condition A
Condition B
Condition C
Fig. 1. Parallel and sequential mRNA-FISH in bacteria. (A) seqFISH probe
design scheme. Primary probes contain unique sequences (Si) that are read by
secondary probes (colored wands). Each gene is read by a unique probe and
its fluorescence can be turned ON or OFF. (B) mRNA-FISH applied sequentially
to the same sample. In each cycle, a new set of secondary readout probes are
introduced. Raw fluorescence data are shown on the right, and the detected
local spot maxima are shown in the spot detection image. Merged spots for
many genes are shown in shuffled colors. (C) Combinatorial labeling can be
used to encode species taxonomy using 16S rRNA or to enable the parallel study
of bacteria grown in different conditions.
Dar et al., Science 373, eabi4882 (2021)
13 August 2021
2 of 16
RES EARCH | R E S E A R C H A R T I C L E
We generated a probe library targeting a set
of 105 marker genes that capture many core
physiological aspects of this pathogen (tables
S1 and S2). These included genes involved in
biosynthetic capacity (ribosome and RNA-
polymerase subunits), anerobic physiology
(fermentation and denitrification pathways),
stress responses (oxidative and nutrient limi-
tation), cellular signaling [cyclic diguanylate
monophosphate (c-di-GMP)], biofilm matrix
components, motility (flagella and T4P), all
major quorum-sensing (QS) systems, as well
as multiple antibiotic resistance and core
virulence factors. In addition, to control for
false positives, we designed probes targeting
three different negative control genes that do
not exist in P. aeruginosa (fig. S1).
Parallel and sequential mRNA-FISH in single
bacterial cells
To test our bacterial seqFISH approach, we first
studied P. aeruginosa grown in well-understood
batch culture conditions. We performed a
growth curve experiment in lysogeny broth
(LB) medium, in which key parameters such
as cell density, growth rate, and oxygen levels
change in a predictable manner. We collected
11 time points representing the lag phase, ex-
ponential growth phase, and stationary phase,
and imaged the expression of 105 genes within
them simultaneously for 2 days (Fig. 2A). In-
dependent imaging of these 11 samples in a
serial manner would have taken ~3 weeks of
automated microscopy time.
To perform simultaneous imaging, we de-
veloped a multiplexing method that enables
parallel seqFISH (par-seqFISH) experiments.
We designed a set of primary probes targeting
the 16S ribosomal RNA (rRNA) (Ribo-Tags),
which contain unique combinations of flank-
ing sequences (barcodes) that serve as the
“readout” in a seqFISH run (Fig. 1C and table
S3). In principle, this multiplexing approach
can be applied to studying combinations of
different species or for pooling bacteria from
different growth conditions (Fig. 1C). We val-
idated the latter application by individually
labeling the 16S rRNAs of each of the 11 growth
curve samples with unique Ribo-Tags. The
samples were pooled, collectively hybridized
with the 105-gene-probe library, and subjected
to sequential hybridizations to measure gene
expression and to decode cell identity (Fig. 2B).
We acquired expression profiles for >50,000 in-
dividual P. aeruginosa cells, >91.8% of which
were unambiguously decoded and assigned
to the condition from which they originated
(Fig. 2B). We estimated the false-positive de-
coding rate to be 0.04% (one in 2500 cells) by
counting the number of hits for barcodes left
out of the experiment, demonstrating both
high efficiency and accuracy for par-seqFISH.
In addition to acquiring mRNA expression
profiles, our imaging-based platform permits
concurrent tracking of key information such as
cell size and shape and can be combined with
functional stains, markers, and/or immuno-
fluorescence measurements (45). This opens up
the possibility of correlating particular expres-
sion profiles at the single-cell level with integra-
tive physiological or cell biological parameters.
We applied a 4′,6-diamidino-2-phenylindole
(DAPI) stain as a part of the par-seqFISH ex-
periment and used DAPI fluorescence to esti-
mate the nucleoid size and chromosome copy
per cell. Comparing cells at different stages of
growth showed that both nucleoid size (esti-
mating cell size) and chromosome number
distributions followed identical trends, in
agreement with the P. aeruginosa literature
(46) (Fig. 2, C and D). We also estimated ribo-
some abundance using 16S rRNA fluorescence.
The distribution of this metric differed signifi-
cantly from that of the chromosome parameters,
displaying contrasting intensities at different
stages of the lag phase, increased variability
at the deep stationary phase, and a delay in
signal decline during the shift from exponen-
tial growth to the stationary phase (Fig. 2E).
Conversely, the total number of mRNAs per
cell (estimated by our 105 genes) differentiated
each time point along the growth curve, reach-
ing a maxima and minima at the fastest and
slowest growth rates, respectively (Fig. 2F).
These data further support the accuracy of our
par-seqFISH multiplexing approach and dem-
onstrate the unique ability of this method to
integrate single-cell gene expression with global
parameters.
To determine whether our expression pro-
files faithfully captured known physiological
processes that occur during culture develop-
ment, we grouped the cells according to their
decoded conditions and calculated their aver-
age gene expression profiles. We found a tem-
porally resolved expression pattern associated
with different stages of growth (Fig. 2G). For
example, genes representing high replicative
and/or biosynthetic capacity, such as those
involved in RNA and protein biosynthesis,
reached their peak expression during the max-
imal division rate but decreased between 90-
and 250-fold during the stationary phase (Fig.
2G). By contrast, stress factors involved in sta-
tionary phase adaptation and nutrient limita-
tion peaked at low division rates and higher
cell densities (Fig. 2G). QS signal production,
receptor expression, and target activation
reflect the known hierarchical QS-regulatory
network (47). The expression of anaerobic me-
tabolism genes occurred in two stages: (i) early
induction of the fermentation and nitrate-nitrite
reduction genes in the entry to stationary phase,
in which hypoxic conditions emerge, followed
by (ii) expression of the remaining denitrifi-
cation pathway at lower predicted oxygen lev-
els (48) (Fig. 2G). Furthermore, the shift from
aerobic to anaerobic metabolism was accom-
panied by sequential exchanges in terminal
oxidase identities, from ccoN1 to ccoN2 and
finally ccoN4, concomitantly with the induction
of phenazine biosynthesis (49, 50) (Fig. 2G).
Repeated mRNA measurements of the same
genes in independent and spaced hybridization
rounds were well correlated, both in average
expression and at the single-bacterium level
(Pearson’s R = 0.86, 0.89 and 0.9, for sigX,
rpsC, and rpoS, respectively). In addition, the
three negative control genes had an average
false-positive rate of 0.002 transcripts per cell
(fig. S1). We also found a good correlation be-
tween par-seqFISH and previous RNA-seq ex-
periments conducted under similar conditions
(51) (Pearson’s R = 0.79 to 0.84), as well as a
strong correlation between close time points
along the growth curve (fig. S2). Together, these
results further validate the accuracy of our
multiplexing method and demonstrate that
our marker genes capture diverse transcrip-
tional states across a wide range of physio-
logical conditions.
Transient emergence of physiologically
distinct subpopulations during LB growth
Phenotypic diversity in clonal populations can
generate distinct subpopulations that adjust
to dynamic environmental changes and spe-
cialize in different tasks at different times,
setting a fertile ground for bet-hedging behav-
iors and complex interactions (9, 15, 52). The
single-cell resolution and high sensitivity of
seqFISH has the potential to shed light on this
important yet largely unexplored aspect of mi-
crobial life.
We applied uniform manifold approximation
and projection (UMAP) dimensionality reduc-
tion and unsupervised clustering to identify
distinct transcriptional cell states in our single-
cell expression data (29, 53). The cell clusters
detected by this analysis charted the pheno-
typic landscape in LB growth from the perspec-
tive of our chosen marker genes. Analyzing the
11 time points together, we detected 20 clusters
(representing different subpopulations) with
diverse predicted functional capabilities. These
included, among others, differential replicative
capacity, exoproduct biosynthesis, and viru-
lence factor production (Fig. 3, A and B). We
found that the sampled populations of most
of the growth conditions were partitioned into
multiple coexisting subgroups with distinct
expression profiles (Fig. 3A, fig. S2, and table
S4). Our data suggest that the degree of dis-
persion within this expression space (estimat-
ing phenotypic diversity) varies significantly
between conditions and is elevated during the
stationary phase (fig. S3 and table S4).
Our growth condition–specific analysis re-
vealed intriguing dynamics during lag phase
progression. It could be expected that lag phase
cultures will follow a steady ribosome accumu-
lation as the cells progress toward exponential
Dar et al., Science 373, eabi4882 (2021)
13 August 2021
3 of 16
RES EARCH | R E S E A R C H A R T I C L E
A
2.5x
10
l
m
/
s
l
l
e
c
2.5x
10
2.5x
10
OD_1.8
OD_2.1
OD_1.5
OD_1.2
OD_0.85
OD_0.45
1:100 dilution
OD_0.2
OD_0.06
OD_0.03
0
0.5
1
1.5
2
2.5
OD_3.2
B
OD = 2.1
OD = 2.1
0
2
4
6
8
10
12
14
16
Time post inoculaton (h)
ns
OD = 3.2
Cell length
OD = 1.2
OD = 0.06
OD = 0.85
OD = 2.1
OD = 1.5
OD = 0.2
OD = 1.8
OD = 0.03
OD = 1.2
ns
OD = 1.2
OD = 2.1
OD = 2.1
OD = 0.06
C
)
m
µ
(
h
t
g
n
e
l
i
d
o
e
c
u
N
l
3.5
3.0
2.5
2.0
1.5
1.0
0.5
D
l
l
e
c
/
s
e
m
o
s
o
m
o
r
h
C
l
l
e
c
/
0
1
e
m
o
s
o
b
R
i
l
l
e
c
/
s
A
N
R
m
E
F
OD
0.03
0.03
0.06
0.2
0.85
0.45
ns
1.2
1.5
1.8
2.1
3.2
Chromosomes
G
4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
OD
0.03
0.03
0.06
0.2
0.45
0.85
1.2
1.5
1.8
2.1
3.2
Ribosomes
0.03
0.03
0.06
0.2
0.45
0.85
1.2
1.5
1.8
2.1
3.2
Total mRNAs
14
12
10
8
6
4
2
0
OD
250
200
150
100
50
0
OD
0.03
0.03
0.06
0.2
0.45
0.85
1.2
1.5
1.8
2.1
3.2
0
0.2
0.4
0.6
0.8
1
Cell density
Ribosome (rpsC)
RNAP (rpoA)
ATP synthase (atpA)
RNA processing (rne, rho)
Oxidative stress (sodB)
Proteases (clpX, lon)
Antibiotic resistance (mexB)
House-keeping sigma (rpoD)
T3SS (pscC, exoT)
Terminal oxidase (ccoN1)
Quorum sensing (lasI)
Arginine fermentation (arcA)
Nitrate reduction (narG)
Terminal oxidase (ccoN2)
Quorum sensing (pqsR, pqsC)
Quorum sensing (rhlI)
Hydrogen cyanide (hcnC)
Stationary phase sigma (rpoS)
Denitrification (nirS, norB, nosZ)
Phenazines (phzE, phzM)
Terminal oxidase (ccoN4)
Quorum sensing (lasR, rhlR)
Quorum sensing (pqsH)
Extracellular proteases (lasB)
Rhamnolipids (rhlA)
Siderophore (pchF)
M
a
x
m
a
i
l
g
r
o
w
t
h
r
a
t
e
H
y
p
o
x
a
i
S
a
t
t
i
o
n
a
r
y
Normalized
expression
Fig. 2. par-seqFISH of an LB growth curve experiment. (A) Sampled LB
growth curve. Collected time points are indicated with gray circles. Magnification
shows the sampled lag phase. The presented colony-forming units per milliliter
were estimated using OD600 values [OD600 = 1.0 reporting on ~109 cfu/ml (104)].
The OD600 values are indicated over each time point. (B) Demultiplexed bacteria and
their mRNAs. The merged, raw Ribo-Tag 16S rRNA fluorescence is shown for a
representative region. Different barcodes (16S combinations) appear as different color
combinations that identify the condition the cells were in when they were originally
collected before sample pooling (indicated with the corresponding OD600 value).
Ellipses fitted to the segmented cell boundaries are shown. The mRNA spots (fitted
position of maximal intensity) for all genes per cell are shown in unique colors per
gene. Each spot may represent more than one mRNA copy. (C to F) Condition-specific
distributions of nucleoid length, chromosome copy, ribosome levels, and total mRNAs
detected across our gene set. Distributions contain all demultiplexed cells per
condition and are significantly different from their previous time point unless
otherwise noted (Wilcoxon, P < 0.001). (G) Heatmap showing average gene
expression normalized to the maximal value for each gene across all conditions.
Highlighted gene groups and their functions are indicated on the right.
Dar et al., Science 373, eabi4882 (2021)
13 August 2021
4 of 16
RES EARCH | R E S E A R C H A R T I C L E
A
2
P
A
M
U
C
2
P
A
M
U
16
8
3
1
13
Leiden
5
14
9
17
10
20
6
19
7
11
15
2
4
18
12
D
2
P
A
M
U
UMAP 1
Lag 30 min (OD = 0.03)
3
16
8
1
13
UMAP 1
B
B
Ribosome biogenesis (rpsC)
Stationary phase marker (rpoS)
Maximal replicative capacity
Phenazine biosynthesis
Medium replicative (Lag phase)
High exoprotease producers
Low mRNA stationary cells
Fermentation + HCN producers
Maximal replicative + T3SS
Arginine fermentation (arcA)
Pyochelin biosynthesis (pchF)
Stationary + T3SS
Siderophore producers
Biofilm matrix component
Flagella biosynthesis
Triclosan efflux pump
1
2
3
4
5
6
8
18
12
13
15
20
Lag 60 min (OD = 0.03)
3
16
8
1
13
UMAP 1
F
2
P
A
M
U
E
2
P
A
M
U
Rapid growth (OD = 0.06)
3
16
8
1
13
UMAP 1
Maximal growth (OD = 0.2)
3
16
8
1
13
UMAP 1
G
Biofilm matrix protein (cdrA)
H
Phosphate-binding protein (pstS)
I
T3SS structural protein (pscC)
J
T3SS effector (exoT)
13
13
8
8
18
18
Fig. 3. Single-bacterium analysis revealing physiologically distinct dynamic
subpopulations. (A) UMAP analysis using cells from all 11 time points. Identified
clusters are shown in different colors and are indexed by group size. Specific
groups and their enriched functions are shown on the right. (B) Gene expression
overlays for four genes that report on metabolic state, stationary phase
progression, and exoproduct biosynthesis. The color map shows the normalized
expression scaled to unit variance. Cells from all 11 time points are displayed in
the plot. (C to F) Density scatter plots of cells from individual conditions in a
magnification of the UMAP [dashed box in (A)]. The clusters are indicated by
their index. (G to J) Gene expression overlays shown as in (B).
growth and maximal ribosome content (54).
However, we found a relative decline in the
average rRNA levels: Early lag phase pop-
ulations (30 min after dilution) had a higher
signal than the more advanced lag culture
(60 min after dilution; Fig. 2E). These differ-
ences appeared to be rooted in the transient
emergence and disappearance of an early lag
subpopulation with exceptionally high levels
of 16S rRNA (cluster 13, comprising 34.6% of
the population in early lag; Fig. 3, C to F; fig. S3;
and table S4). In agreement with the deviation
in the rRNA signal, this subpopulation also
showed a proportional increase in total mRNA
counts. However, its size and chromosome
copy distributions were not elevated (fig. S4,
cluster 13 versus cluster 3).
Beyond illuminating the extent of hetero-
geneity in seemingly well-mixed cultures and
classifying subpopulations into particular types,
seqFISH can also directly connect global cell-
specific parameters such as ribosome levels
or cell shape to particular gene expression
signatures. For example, a closer examination
of the metabolically hyperactive subpopula-
tion revealed a 186-fold enrichment in cdrA
expression relative to the rest of the popula-
tion (Fig. 3G). The cdrA gene encodes a major
adhesive protein component of the P. aeruginosa
biofilm matrix (55, 56). Expression of cdrA is
commonly used as a reporter for c-di-GMP
levels, a key signaling molecule involved in
surface attachment (16). This subpopulation
Dar et al., Science 373, eabi4882 (2021)
13 August 2021
5 of 16
RES EARCH | R E S E A R C H A R T I C L E
also displays a 30-fold enrichment in pstS
expression, which encodes for the phosphate-
binding component of the pstSCAB phosphate
uptake system (Fig. 3H). PstS has been prev-
iously detected in extracellular appendages
of P. aeruginosa and has been suggested to
provide an adhesion phenotype to intestinal
epithelial cells (57). In support of this non-
canonical role, pstS was recently suggested
to confer a similar adherence phenotype in
Acinetobacter baumannii, another human
pathogenic bacterium (58).
A second example from our dataset of the
type of fine-grained information that seqFISH
can provide comes from the temporal expres-
sion of genes involved in virulence factor pro-
duction. Single-cell variation in virulence factor
production has been suggested as a mecha-
nism for division of labor during infection (17).
P. aeruginosa uses a variety of virulence factors
to overcome the host immune response (44),
including the type 3 secretion system (T3SS)
that translocates toxins (effectors) directly into
host cells (59). Our gene set monitors two T3SS
structural genes, pscC and pcrD, and two main
effector genes, exoT and exoY, all of which are
encoded in different operons (60). We detected
two different types of subpopulations with en-
riched T3SS-related genes, suggesting a unique
division of cells into virulent and avirulent
states (Fig. 3, I and J). The first group tran-
siently appears during exponential growth
and constitutes 8 to 30% of the population
(Fig. 3, C to F, I, and J, and table S4). This
group expresses both the secretion system
genes (86-fold enrichment) and the effector
genes (28-fold). By contrast, the second group
appears three or four divisions later, close to
the replicative minima at stationary phase,
and occupies only ~2.7% of cells (table S4).
This subpopulation is strongly enriched for
the two effector genes (average 26-fold; Fig. 3,
I and J) but only mildly so for the secretion
system genes (sixfold) compared with the
earlier group.
We can potentially reconcile these observa-
tions as follows: P. aeruginosa has been shown
to contain one to three T3SS units per cell
under inducing conditions (61). Thus, succes-
sive divisions after T3SS expression will result
in rapid dilution of the T3SS+ group. Assuming
that the inheritance of the T3SS and effectors is
uncoupled, then T3SS+ stationary phase cells
are likely to lose their effectors during division
and thus are predicted to be “inactive.” An in-
triguing hypothesis is that P. aeruginosa in-
vests in the costly T3SS+ subpopulation during
“times of plenty” (rapid growth) and specifically
expresses the effectors at stationary to “reload”
and maintain this subpopulation after division-
based dilution, just before growth arrest. To-
gether, these examples underscore the power
of seqFISH to suggest hypotheses that can be
tested going forward.
Spatial transcriptomics at a single-cell
resolution in P. aeruginosa biofilms
Although much can be learned by applying
seqFISH to planktonic cultures, in many con-
texts, bacteria exist in biofilms (1, 2). Varia-
tion in local environmental conditions and the
effect of spatially confined metabolic activities
in biofilm populations can promote the emer-
gence of chemically distinct microenviron-
ments and phenotypes (10, 19). We reasoned
that seqFISH’s capacity to record transcrip-
tional activities with micrometer resolution
would be particularly useful in shedding light
on these processes.
The P. aeruginosa biofilm mode of life is
particularly important in chronic infections
such as those residing in the airways of in-
dividuals with cystic fibrosis (62, 63). Accord-
ingly, having used LB medium to validate
bacterial seqFISH, we switched to synthetic
cystic fibrosis sputum medium (SCFM) for our
biofilm studies (64). Briefly, bacteria were in-
cubated in coverslip-attached microwells and
the medium was replaced every several hours.
Using biofilms that were allowed to develop
for 10 or 35 hours, we imaged hundreds of
aggregates ranging in size from several bacteria
to tens of thousands of tightly bound members
(Fig. 4, A and B). As a reference for cellular
physiological states, we also performed a plank-
tonic growth curve experiment in SCFM. We
applied par-seqFISH multiplexing to image
10 time points matching those sampled in the
planktonic LB experiment. We found a simi-
lar degree of heterogeneity in SCFM- and LB
medium–grown populations (Fig. 4D). We ex-
tracted the physical coordinates of individual
bacterial cells within microaggregates, acquir-
ing a microscale spatial expression profile for
~365,000 surface-attached bacteria (Fig. 4, A
and B). In addition, we collected single-cell ex-
pression data for ~218,000 planktonic cells.
A basic question that we sought to answer
is what is the extent to which transcriptional
responses are specific to the biofilm lifestyle?
We performed a joint UMAP analysis using
both biofilm and planktonic samples (Fig.
4C). These different modes of growth cluster
into independent groups in expression space,
reflecting their significant physiological dif-
ferences (Fig. 4D). Ribosome and RNA poly-
merase subunit expression in the planktonic
experiment correlated strongly with growth
rate, as observed in LB medium (Fig. 4D).
Examining these marker genes in the biofilm-
derived cells placed the average replicative
capacity of the 10-hour (10h) and 35h biofilm
populations approximately equal to those of
the early-middle and late-stationary planktonic
populations, respectively (Fig. 4F). Expression
of the stationary phase master regulator rpoS
further supported this classification (Fig. 4G).
However, biofilm cells also have distinctive
expression profiles that distinguish them from
liquid cultures. For example, the matrix com-
ponent gene cdrA was uniformly expressed in
both the 10h and 35h biofilms but repressed in
most planktonic cells (Fig. 4E). In addition,
compared with stationary liquid cells, our data
indicate that early biofilms (10h) have higher
expression of sigX (5.1-fold), a transcription
factor recently implicated in biofilm formation
(65); mexB (>4.5-fold), of the mexA-mexB-oprM
antibiotic efflux system; and an increase in the
3′-5′ exonuclease polynucleotide phosphorylase
(pnp) (7.5-fold). Comparing the 35h biofilm
with stationary cells, we found a 3.3-fold in-
crease in the extracellular protease lasB but
reduced expression of other proteases such
lasA (3-fold lower), as well as aprA and the
rhamnolipid biosynthesis gene rhlA (~10-fold
lower). These genes are QS regulated, and our
liquid cultures expressed both lasA and rhlA
at later time points than lasB, suggesting that
these differences may reflect the age of the
biofilm rather than features that define the
biofilm state per se.
In situ analysis of biofilm-specific functions
The above data demonstrate that seqFISH can
capture both cell states and their physical posi-
tion directly within intact biofilms, providing
an opportunity to examine known and new
processes that contribute to biofilm develop-
ment from a quantitative and highly spatially
resolved perspective. To illustrate this, we fo-
cused on the expression patterns of represent-
ative genes known to define critical stages in
biofilm development, such as attachment, mat-
uration, and exclusion of competitors.
Motility systems such as the flagella and the
type 4 pilus (T4P) are a major determinant of
surface colonization and subsequent biofilm
formation (66–68). Recent work identified an
asymmetric division process coined “touch-
seed-and-go,” in which flagellated mother cells
first attach to a surface and then produce un-
flagellated daughter cells that contain the T4P.
This c-di-GMP–dependent phenotypic diversi-
fication enables the mother “spreader” cell to
spawn multiple adherent “seed” populations
(69). This is thought to be mainly regulated
by surface sensing (69). However, how such
motility-based division of labor affects the
organization of biofilms at stages beyond sur-
face attachment remains unknown.
We examined the spatial expression patterns
of the major flagellum and T4P components,
fliC and pilA, respectively, in the early sur-
face colonization experiment (10h biofilm). An
abundant “checkerboard”–like pattern was
evident, in which cells expressed high levels
of either fliC or pilA but generally not both
(Fig. 5A). We found that the highly express-
ing fliC+ and pilA+ subpopulations (>3 SDs
above the population mean) represented a
total of ~4% of all cells in our experiment
(2% for each subgroup), yet the double-positive
Dar et al., Science 373, eabi4882 (2021)
13 August 2021
6 of 16
RES EARCH | R E S E A R C H A R T I C L E
A
B
C
2
P
A
M
U
E
10h surface colonization
16S rRNA (zoom-in)
mRNA FISH (zoom-in)
30 µm
5 µm
35h surface colonization
16S rRNA (zoom-in)
mRNA FISH (zoom-in)
30 µm
5 µm
Planktonic + Biofilms (10h + 35h)
Leiden
D
Planktonic (growth curve)
Surface colonization (10h)
Surface colonization (35h)
2
1
10
12
9
6
5
8
16
18
19
7
17
4
3
13
11
14
15
UMAP 1
I
II
III
I
II
III
Exponential
Early-mid stat
Late stat
2
6
9
Maximal replicative capacity
Exoprotease production
Phenazine biosynthesis
12
Multidrug efflux (amrB)
1
7
16
18
Aerobic metabolism
Denitrification & fermentation
Carbon starvation
Pyocin induction
3
13
14
15
Exoprotease producers (lasB)
Flagella biosynthesis
Type 6 Secretion System
Starvation induced oxidase
Biofilm matrix (cdrA)
F
Ribosome biogenesis (rpsC)
G
Stationary phase marker(rpoS)
H
Extracellular protease (lasB)
Fig. 4. Spatial transcriptomics in P. aeruginosa biofilms at a single-cell
resolution. (A) Representative field of view collected during a 10h surface
colonization experiment showing cells using 16S rRNA fluorescence (gray).
Magnification (orange box) shows the cell segmentation masks depicted
as white ellipses. The 16S rRNA signal and mRNA-FISH data for several genes
are shown in different colors. (B) A 35h experiment field is shown in an
identical manner to (A). Scale bar length is annotated within the figure.
(C) Joint UMAP cluster analysis of biofilm and planktonic experiments. Planktonic
cells are shown for all time points collected. (D) UMAP scatter plots showing
cells from either planktonic or biofilm experiments as indicated. At the bottom, a
highlighted set of UMAP clusters associated with each experiment is annotated
with enriched functions. (E to H) UMAP overlay with specific gene data. The color
map shows the normalized expression scaled to unit variance. Cells from the
liquid experiment and both the 10h and 35h biofilms are displayed together.
Dar et al., Science 373, eabi4882 (2021)
13 August 2021
7 of 16
RES EARCH | R E S E A R C H A R T I C L E
A
10h surface colonization
B
35h surface colonization
C
Liquid growth experiment
fliC
pilA
fliC
pilA
fliC
pilA
10 µm
25 µm
5 µm
D
10h surface colonization
E
R2-pyocins
1
2
3
30 µm
G
1
5 µm
F
t
n
e
m
h
c
i
r
n
e
A
N
R
m
40
35
30
25
20
15
10
5
0
Zoom-in
Pyocins
rpoA
10
15
20
25
40
30
20
10
0
5
0
50
100
150
200
250
300
Neighborhood size (# cells)
16S rRNA
DNA-damage (recA)
R2-pyocins
Merged
2
1 µm
3
1 µm
Fig. 5. Spatial expression patterns for motility- and pyocin-related genes.
(A and B) Representative regions from the 10h and 35h biofilm experiments.
Cells are shown using 16S rRNA fluorescence (gray) and overlaid with raw mRNA-
FISH fluorescence for different genes as indicated. (C) Planktonic cells from the
paired liquid experiments. Cells are shown using DAPI and gene expression as
indicated. (D and E) 10h aggregate showing R2-pyocin expression. (F) Enrichment of
R2-pyocin mRNA near strong induction sites (cell with 99.5th percentile pyocin
expression). The x-axis shows the number of cells closest to an induction site that
were analyzed (neighborhood size; center cell was excluded); the y-axis shows
the enrichment in each neighborhood relative to the total population. A non-pyocin
control gene, rpoA, is shown. (G) Examples of mRNA R-pyocin transcript and
ribosome polar localization as indicated in the panel legends.
subpopulation (fliC+/pilA+) only constituted
0.07%. This pattern occurred uniformly across
most aggregates, both in small groups (tens
of cells) and in large sets containing thousands
of cells. Conversely, the older 35h biofilms
showed lower expression of pilA but contained
a sparse but uniform distribution of fliC+ cells,
suggesting that biofilm-associated bacteria
invest in a costly motility apparatus despite
being spatially confined (Fig. 5B). Examining
the expression of fliC and pilA in our paired
planktonic experiment, we found a similar
mutually exclusive pattern (~2% of both single-
positive groups and ~0.15% of the double-
positive cells) (Fig. 5C). Thus, in contrast to
Dar et al., Science 373, eabi4882 (2021)
13 August 2021
8 of 16
RES EARCH | R E S E A R C H A R T I C L E
the current model, our planktonic control ex-
periment suggests that the asymmetric dis-
tribution of motility systems is unlikely to be
directly regulated by surface sensing (Fig. 5C);
such a conclusion would not be possible with-
out the means to compare transcriptional ac-
tivities at the single-cell level.
Beyond initial surface attachment, bacteria
must establish a strong foothold for colony
development and also outcompete resident
microbes. One strategy that potentially ad-
dresses both needs is the use of phage tail–like
bacteriocins, which are broadly called tailocins
(70). These elements are thought to be adapted
from prophages and are applied as narrow-
spectrum toxins for kin exclusion (70, 71).
However, in contrast to antibiotics, these phage
tail–like structures are released into the envi-
ronment through explosive lysis events that
kill the producer and spray the toxin locally to
inhibit nearby competitors (72, 73). This event
also releases extracellular DNA that integrates
into the biofilm matrix, structurally support-
ing biofilm maturation (72, 74). How this
“sacrificial” process is regulated within devel-
oping biofilms is not well understood.
Our UMAP analysis identified a subpopula-
tion (cluster 18; Fig. 4C) exhibiting >1000-fold
enrichment in expression of the R2-pyocin
operon (P. aeruginosa tailocin), represented by
the PA14_08150 gene. This UMAP cluster was
enriched by about fourfold in 10h biofilm–
derived cells (0.45% of the entire population),
suggesting that pyocin induction is up-regulated
during surface attachment. Furthermore, we
found an 11-fold higher expression of the DNA-
repair gene recA, in agreement with its role in
inducing pyocin expression (75). Visualizing the
expression of the pyocin producers, we found
that induction events were spread across var-
ious microaggregate regions but often appeared
in local clusters (Fig. 5, D and E). Indeed, we
found a ~37-fold average spatial enrichment
in pyocin expression in the immediate vicinity
of strong induction sites compared with the
general population (Fig. 5F). This enrichment
decayed rapidly as a function of neighborhood
size, suggesting a highly localized effect (Fig. 5F).
In addition to reporting multigene expres-
sion profiles, seqFISH also reports the physical
position of measured mRNA molecules at a
submicrometer resolution. During this anal-
ysis, we observed that R2-pyocin transcript
fluorescence generally appeared as two spots.
Upon closer examination, we discovered that
this mRNA was strongly localized to the two
cell poles (Fig. 5G). The 16S rRNA fluorescence
signal in these pyocin producers showed iden-
tical polarization, a rare pattern not observed
in neighboring noninducing cells (Fig. 5G).
These data suggest that ribosomes and the
R2-pyocin transcript are mobilized after induc-
tion and spatially colocalize. By contrast, the
expression of recA did not follow this pattern,
suggesting a pyocin-specific effect (Fig. 5G). A
recent study discovered an identical polar
localization for two different Pseudomonas
protegens R-tailocins at the protein level (73).
Together, these data hint at a potentially evo-
lutionary conserved, RNA-dependent mecha-
nism for R-tailocin protein polar localization.
We hypothesize that the spatially correlated
ribosomal enrichment may provide efficient
local translation and particle accumulation
before cell lysis.
Temporal evolution of metabolic heterogeneity
during biofilm development.
Beyond resolving transcriptional activities that
contribute to biofilm developmental processes,
seqFISH can also reveal how biofilm cells meta-
bolically respond to subtle changes in their
local microenvironment. Chemical heteroge-
neity is a key feature of spatially structured
environments, and metabolic heterogeneity
characterizes mature biofilms (10, 18, 19, 76).
However, until now, it has been impossible to
capture the development of fine-grained meta-
bolic structure across multiple suites of genes
at different times.
To map biofilm metabolic development, we
focused on genes for which regulation and
functions are well understood. In particular,
we focused on catabolic genes with products
that enable energy conservation under differ-
ent oxygen concentrations. Oxygen is a central
and dynamic factor that influences metabolic
activity in bacterial biofilms (10, 19, 77, 78).
Local oxygen availability can vary significantly
within structured environments and is biotically
shaped within biofilms (18, 77, 79). P. aeruginosa
can survive under anaerobic conditions by fer-
menting different substrates and/or denitrify-
ing (50, 80, 81). Accordingly, monitoring the
expression of these catabolic genes and others
that are co-regulated with them provides a
means of tracking local oxygen availability
and its dynamic effects on biofilm metabolic
coordination.
How quickly and over what spatial scales
do biofilm cells metabolically differentiate?
Following the uspL gene, which was strongly
induced during hypoxic conditions and cor-
related with anaerobic fermentation and de-
nitrification genes in our planktonic growth
experiments, we observed unexpectedly het-
erogeneous responses to oxygen depletion over
just a few micrometers in young (10h) biofilms
(Fig. 6A). uspL expression was strongly spa-
tially correlated with multiple anaerobic mark-
ers (fig. S5), indicating that this gene reports
on local anaerobic activities. A closer exami-
nation of these putative hypoxic sites showed
a frequent anticorrelation of uspL with mul-
tiple genes that were otherwise uniformly
expressed in 10h biofilms, appearing as co-
localized but reversed expression patches
(Fig. 6B). Among the anticorrelated functions
were the tricarboxylic acid (TCA) cycle gene
sucC and replicative capacity genes such as
those encoding RNA polymerase and ribo-
some subunits (Fig. 6B and figs. S4 and S5).
However, exceptions to this anticorrelation
were also observed (fig. S6).
Can the metabolic heterogeneity revealed
by oxygen-responsive marker genes provide an
entry point for the discovery of more nuanced
cellular responses at the microscale? Our spa-
tial correlation analysis revealed an intrigu-
ing association between anaerobic metabolism
genes, such as those in the denitrification path-
way (narG-nirS-norB-nosZ), and the oxidative
stress response genes katA, katB, and sodM,
which encode for the inducible catalases and
an Mn-dependent superoxide dismutase, re-
spectively (82–84) (Fig. 6, C and D, and figs. S4
and S5). Nitrite-respiring P. aeruginosa pro-
duce the highly toxic intermediate nitric oxide
(NO) (85). Indeed, KatA was recently demon-
strated to play a role in protection from NO-
associated stress (84), suggesting that these
subaggregate regions correspond to micro-
environments with high NO levels. In agree-
ment with this hypothesis, we found that the
stress response pattern was also spatially cor-
related with heat-shock protease expression,
including the membrane protease ftsH, which
was found to play an important role in survi-
val under anoxic conditions (86) (Fig. 6E and
fig. S4). These data highlight how contrast-
ing physiological states can be established
just a few micrometers away early in biofilm
development.
We hypothesize that these coordinated ex-
pression patterns for particular genes reflected
the spatiometabolic distribution of distinct
physiological “states” across the biofilm. To
test this hypothesis, we conducted a targeted
UMAP analysis using only the 10h biofilm cells
(fig. S7). We identified two main anaerobic sub-
populations corresponding to denitrification-
and fermentation-dominated metabolic states
and representing 11.8 and 7.2% of all cells in
the experiment, respectively (fig. S7). In addi-
tion, we detected a smaller subpopulation of
denitrifying cells (2.4% of cells) with a 5.3-fold
average increase in the oxidative stress fac-
tors katB, sodM, and ahpF, the latter of which
encodes for an alkyl hydroperoxide reductase
(87). Relative to the main denitrifying sub-
group, stressed cells had lower expression of
the denitrification pathway (about fourfold)
and a more than twofold reduction in replica-
tive capacity marker levels (rpoA, rpsC, and
atpA), in support of a potentially damaged
state. Projecting these single-cell metabolic
states over their respective biofilm positions
showed a strong overlap with the above pre-
dicted hypoxic pockets, supporting our hy-
pothesis and revealing that multiple metabolic
states can coexist in the same patch (Fig. 6F
and fig. S5).
Dar et al., Science 373, eabi4882 (2021)
13 August 2021
9 of 16
RES EARCH | R E S E A R C H A R T I C L E
A
Universal stress protein (anaerobic)
B
Succinyl-CoA synthetase (TCA cycle)
C
uspL
22
sucC
Oxidative stress response
katA katB sodM
1
4
3
D
25 µm
Denitrification pathway
nirS norB nosZ
E
Heat-shock proteases
htpX ftsH clpX lon
F
UMAP cluster overlay
Denitrification
Fermentation
High oxidative stress
25 µm
Fig. 6. Oxygen availability shapes microscale metabolic heterogeneity in biofilms. (A to E) Representative 10h biofilms. Cells are shown using 16S rRNA FISH
fluorescence (gray) and overlaid with raw mRNA-FISH fluorescence for different genes as indicated in each panel. White circles highlight regions of interest.
(F) Cells painted according to their UMAP-derived metabolic state as indicated in the panel legends (also see fig. S7, clusters 0, 8, 12, and 15), showing colocalization
of multiple metabolic states within a given region.
Given the extent of transcriptional hetero-
geneity manifest in young biofilms, we won-
dered whether such heterogeneity would
persist as the biofilms aged. We speculated
that the higher cell densities and more com-
mitted spatial structuring of mature biofilms
might favor larger-scale metabolic zonation.
We therefore examined the spatial expression
patterns in a 35h biofilm experiment.
In contrast to the spatial variation in aerobic
and anaerobic metabolic processes seen in 10h
biofilms, 35h biofilms had an ~50-fold lower
average expression of the denitrification path-
way genes nar-nirs-norB-nosZ. Indeed, these
genes are known to be repressed by the las
and rhl QS systems, indicating P. aeruginosa
is programmed to shut down denitrification
at high cell densities (80, 88). However, in
addition to this complete and co-regulated
pathway, P. aeruginosa also encodes an inde-
pendent periplasmic nitrate reductase (nap)
(89). Unexpectedly, the napA gene was ex-
pressed in a spatially uniform manner but at a
low level in the 35h aggregates, a pattern that
was closely shared with the uspL gene, and
these two genes together were expressed in
20.3% (±5.5%) of the measured cells within
each individual aggregate (Fig. 7A and fig.
S7). NapA has been implicated in maintaining
redox homeostasis under oxygen limitation
(78), and the uspL paralog uspK was shown
to play a role in survival under such condi-
tions (86, 90). At first, these results seemed to
suggest that as an aggregate cell mass grows,
survival physiology on average dominates over
growth-promoting processes. However, we also
found substantial and large-scale spatial heter-
ogeneity in certain genes, such as those en-
coding the replicative capacity markers, which
were highly expressed in 17.7% (±10.9%) of ag-
gregate cells (Fig. 7B and fig. S7), and lasB,
which encodes a QS-regulated extracellular
protease and is expressed at similarly high
levels in 43.5% (±6.1%) of the cells (Fig. 7C and
fig. S7). These data suggest that a single 35h
microaggregate can contain regions with dis-
tinct physiological states and virulence-related
activities. Finally, the fact that metabolism dy-
namically shapes the microenvironment leads
to the prediction that differences in local nu-
trient availability will be reflected in hetero-
geneous transcriptional activities over small
spatial scales (10). We saw evidence of this
phenomenon in our data when focusing, for
example, on carbon metabolism. Where repli-
cative capacity appeared to be high and carbon
was presumably replete, we saw coexpression
of the TCA cycle gene sucC (Fig. 7, B to D).
However, when carbon is limiting, bacteria
can use the glyoxylate shunt (GS), which
bypasses the oxidative decarboxylation steps
of the TCA. The GS provides an alternative
metabolic pathway for using acetate and fatty
acids as carbon sources (91, 92). In the GS, car-
bon flux is redirected by isocitrate lyase, which
competes with the TCA enzyme isocitrate de-
hydrogenase for isocitrate. Because isocitrate
dehydrogenase has a much lower Michaelis con-
stant (Km), it must be enzymatically inactivated
Dar et al., Science 373, eabi4882 (2021)
13 August 2021
10 of 16
RES EARCH | R E S E A R C H A R T I C L E
A
Survival metabolism
B
Energetic state
napA
uspL
rpoA
atpA
25 µm
C
Virulence factor biosynthesis
D
Carbon limitation
lasA
lasB
sucC
aceA
coxA
Fig. 7. Functional zonation in a single microaggregate. A P. aeruginosa 35h aggregate. Bacteria are shown
using 16S rRNA FISH fluorescence (gray) and are overlaid with raw mRNA-FISH fluorescence for different
genes as described in the panel legends.
by phosphorylation for the carbon flux to be
redirected to the GS (92). However, little is still
known about the transcriptional regulation of
these pathways (93). Our gene set contains
both the GS gene aceA and the downstream
TCA cycle gene sucC. Although these genes
are often coexpressed, we found that only the
GS marker aceA was expressed in the pre-
dicted lower-energetic-capacity biofilm zones
(Fig. 7D and fig. S7), suggesting that these
subregions experience carbon limitation. In
support of this hypothesis, these regions also
expressed the tightly regulated terminal oxidase
gene coxA, which is transcriptionally induced
by carbon starvation, a condition in which it
promotes survival (86, 94) (Fig. 7D and fig. S7).
Together, aceA- and coxA-expressing cells cov-
ered up to 43% of an aggregate cell mass in our
experiment. This is just one example of the
type of coherent spatiometabolic stratification
pattern that seqFISH can reveal at any given
moment in time.
Discussion
Until now, our ability to capture the dynamic
metabolic activities of microbial populations
and communities at small spatial scales has
been limited to tracking just a few parameters.
This technical limitation has restricted our
ability to observe and understand the features
that define these ubiquitous associations. Our
analysis of P. aeruginosa populations has shown
that par-seqFISH can reveal a high degree of
transcriptional heterogeneity spanning mul-
tiple dimensions, from the subcellular to the
microscale. Moreover, by tracking the tem-
poral and spatial dynamics of cellular states
in subpopulations, our results demonstrate
that spatial transcriptomics can provide new
insights into how bacteria sustain functional
diversity.
The high temporal and spatial resolution
enabled by par-seqFISH permitted us to make
unexpected discoveries. For example, in plank-
tonic cultures, we observed the short-lived tem-
poral emergence of two T3SS+ populations: a
large group, which appeared in the exponen-
tial phase and expressed all of the needed T3SS
components, and a second, ~10-fold smaller
group, which emerged during the midstation-
ary phase and expressed the effectors but not
the secretion system. The estimated three or
four divisions that separated these subpopu-
lation correlated with their size differences,
suggesting that these two subpopulations could
represent the same T3SS+ population, just at
different stages of growth. We hypothesize that
the specific expression of effector genes in the
stationary T3SS+ subpopulation serves to re-
plenish the effectors lost by the diluting effect
of cell divisions. If true, then this would mean
that P. aeruginosa not only generates hetero-
geneous subpopulations but can also actively
maintain their functional capabilities. Such
an observation would not have been possible
without the ability to measure the expression
of many genes within the same cell.
In P. aeruginosa biofilms, despite marked
levels of metabolic heterogeneity, coherent co-
expression patterns also emerged. We found a
strong spatial correlation between denitrifica-
tion genes and oxidative stress factors, sug-
gesting that local denitrification results in NO
toxicity. This hypothesis is based on the ex-
pression of the inducible peroxidase katA,
which is known to be up-regulated by NO under
anaerobic conditions and to alleviate NO tox-
icity (84). We also observed overlapping induc-
tion of other factors such as katB (83) and the
superoxide dismutase sodM (82), suggesting
that they may also play protective roles. These
patterns were highly spatially confined, sug-
gesting that NO toxicity did not propagate to
neighboring cells, even those just a few micro-
meters beyond. However, it remains unclear
how such hydrogen peroxide– and superoxide-
detoxifying enzymes protect cells from NO. It
is known that NO interacts with relevant oxi-
dants to produce reactive nitrogen species such
as peroxynitrite (95). Therefore, perhaps these
oxidative stress–response factors act by limit-
ing the pool of oxidants available for reactive
Dar et al., Science 373, eabi4882 (2021)
13 August 2021
11 of 16
RES EARCH | R E S E A R C H A R T I C L E
nitrogen species production. Various reactive
nitrogen species cause diverse types of cellu-
lar damage, including the chemical modifi-
cation of proteins, specifically cysteine and
tyrosine residues (95). Our data point toward
elevated expression of cellular proteases in
NO-stressed regions. We therefore suggest
that these proteases act to detoxify cells by
eliminating damaged proteins, a hypothesis
that remains to be tested in future studies.
The par-seqFISH multiplexing approach,
which we developed to increase the through-
put of seqFISH for single-cell analysis, could
be applied in other ways and in both synthetic
and natural communities. For example, be-
cause par-seqFISH is based on 16S rRNA labels
(Ribo-Tags), it could in principle be used to
encode bacterial taxonomy. Recently, a con-
ceptually similar and exciting method for
combinatorial labeling of taxonomy was in-
troduced in a biogeographical study of the
human microbiome (25). In principle, the par-
seqFISH strategy could be readily extended to
capture a similar or higher level of taxonomic
complexity and add the currently missing fea-
ture of mRNA expression. A critical next step
will be to develop methods to chart the envi-
ronmental conditions that contextualize ex-
pression patterns observed in any given case.
Extension of this approach to natural and clin-
ical samples could provide important insights
into the conditions experienced by microbes
in more complex environments and the coor-
dinated physiological responses that emerge
in turn.
Materials and methods
Bacterial strains and growth conditions
P. aeruginosa strain UCBPP-PA14 was grown
aerobically with shaking at 250 rpm in LB
medium (Difco) or on LB agar plates at 37°C.
SCFM was made as previously described (64).
For the growth curve experiments, an over-
night LB culture was washed twice using fresh
growth medium (either LB or SCFM) and then
diluted 1:100 into 100 ml of prewarmed fresh
medium. The cultures were grown at 37°C with
shaking at 250 rpm and collected at various
time points, as indicated in Fig. 2A. The SCFM
samples were collected at cell densities iden-
tical to those in the LB experiment except
that the optical density at 600 nm (OD600) =
3.2 sample was omitted. Collected samples
were immediately fixed in ice-cold 2% parafor-
maldehyde (PFA), incubated on ice for 1.5 hours
in the dark, and then washed twice with 1×
phosphate-buffered saline (PBS). Samples were
resuspended in 70% EtOH and incubated at
–20°C for 24 hours to permeabilize the cells.
Surface colonization was performed by wash-
ing and diluting an LB overnight culture 1:100
into fresh SCFM and dispensing 100 ml into
coverslip-attached open incubation chambers
(Electron Microscopy Sciences, #70333-42). The
coverslips were incubated in Parafilm-sealed
sterile petri dishes at 37°C, and the medium
was gently exchanged every 4 hours. A damp
Kimwipe was placed in the petri dish to con-
trol medium evaporation. During the overnight
stage of the 35h experiment, the medium was
exchanged only once after 8 hours. Biofilm ex-
periments were collected by gently exchang-
ing the SCFM with 100 ml of ice-cold 2% PFA
solution and incubating the sample at 4°C for
1.5 hours. The samples were washed twice with
1× PBS, resuspended in 70% EtOH, incubated
overnight at 4°C, and prepared for seqFISH the
following day as described below.
seqFISH probe design and library generation
Primary probes were designed as 30-nucleotide
(nt) stretches in a GC range of 45 to 65%. Probe
sequences containing more than four consec-
utive base repeats were removed. The remain-
ing probes were compared with the reference
genome using BLAST, and any probe with
nonspecific binding of at least 18 nucleo-
tides was discarded. Negative control genes
were selected from the P1 phage genome
(NC_005856.1) using the same criteria. Each
selected gene was covered by 12 to 20 non-
overlapping probes randomly selected from
the gene probe set. The probes were designed
as a 30-nt mRNA-binding region flanked
by overhangs composed of four repeats of
the secondary hybridization sequence (com-
plementary to a designated fluorescent read-
out probe; table S2). Thus, it is estimated
that during secondary hybridization, each
mRNA was covered by 48 to 80 fluorescent
readout probes (i.e., 12 to 20 × 4), consistent
with previous mRNA-FISH experiments in
bacteria (33, 42).
A library of 1763 probes targeting 105
P. aeruginosa genes and three negative con-
trols was designed (tables S1 and S2). Addi-
tional flanking sequences were added to the
primary probe sequences to enable library am-
plification by polymerase chain reaction (PCR)
(forward 5′- TTTCGTCCGCGAGTGACCAG-3′
and reverse 5′-CAACGTCCATGTCGGGATGC-
3′). The primary probe set was purchased as
oligoarray complex pool from Twist Bioscience
and constructed as previously described (36)
(table S2). Briefly, a set of nine PCR cycles was
used to amplify the designated probe sequences
from the oligo pool. The amplified PCR pro-
ducts were purified using the QIAquick PCR
Purification Kit (Qiagen, #28104) according
to the manufacturer’s instructions. The PCR
products were used as the template for in vitro
transcription (New England Biolabs, #E2040S),
followed by reverse transcription (Thermo
Fisher Scientific, #EP7051). Then, the single-
stranded DNA probes were alkaline hydrolyzed
with 1 M NaOH at 65°C for 15 min to degrade
the RNA templates, followed by 1 M acetic acid
neutralization. Next, to clean up the probes,
ethanol precipitation was performed to re-
move stray nucleotides, phenol–chloroform
extraction was performed to remove protein,
and Zeba Spin Desalting Columns (7 K mo-
lecular weight cutoff) (Thermo Fisher Scien-
tific, #89882) were used to remove residual
nucleotides and phenol contaminants. Read-
out probes were designed as previously de-
scribed and ordered from Integrated DNA
Technologies (36).
Ribo-Tag probes were designed to target the
same region in the 16S rRNA gene according
to the criteria described above, but with
28-nt binding regions. Each probe sequence
was flanked with two secondary sequences
selected out a set of six that were dedicated to
multiplexing (table S3). An additional 16S rRNA
probe was generated as a standard between all
multiplexed samples and was hybridized to an
independent region of the 16S rRNA (table S3).
This probe provided an additional reference
and was used to register images from different
channels (see below).
Coverslip functionalization
Coverslips were cleaned with a plasma cleaner
on a high setting (Harrick Plasma, #PDC-001)
for 5 min, followed by immersion in 1% bind–
silane solution (GE, #17-1330-01) made in pH
3.5 10% (v/v) acidic ethanol solution for 30 min
at room temperature. The coverslips were
washed with 100% ethanol three times and
dried in an oven at >90 °C for 30 min. The
coverslips were then treated with 100 mg ml−1
poly-D-lysine (Sigma-Aldrich, #P6407) in water
for at least 1 hour at room temperature, fol-
lowed by three rinses with water. Coverslips
were air-dried and kept at –20°C for no longer
than 2 weeks before use.
par-seqFISH
Independent fixed samples were individually
hybridized with 16S rRNA labels, washed, and
then pooled into a single mixture that was hy-
bridized with the gene probe library and pre-
pared for imaging. Approximately 108 cells were
collected from each sample into a microcen-
trifuge, pelleted by centrifugation (6000 rpm),
and then resuspended in 20 ml of water with
6 nM of the designated 16S rRNA label (sample
specific) and another 6 nM of a shared refer-
ence 16S rRNA probe (table S3). Each sample
was then mixed with 30 ml of prewarmed pri-
mary hybridization buffer [50% formamide,
10% dextran sulfate, and 2× saline-sodium
citrate (SSC)] by gentle pipetting, incubated at
37°C for >16 hours, washed twice with 100 ml of
wash buffer (55% formamide and 0.1% Triton
X-100 in 2× SSC; 5 min at 8000 rpm for the
viscous hybridization buffer), and then incu-
bated at 37°C in 100 ml of wash buffer for 30 min
to remove nonspecific probe binding. Samples
were then washed twice with 100 ml of 2× SSC
and pooled together into a new microcentrifuge
Dar et al., Science 373, eabi4882 (2021)
13 August 2021
12 of 16
RES EARCH | R E S E A R C H A R T I C L E
in equal volumes. The mixture was pelleted and
resuspended in 40 ml of water, and 10 ml of the
mixture was added to 10 ml of the gene probe
library mixture and mixed well with 30 ml of
prewarmed primary hybridization buffer. The
hybridizations were incubated for >16 hours
at 37°C and then washed and prepared as de-
scribed above. The final mixture was resus-
pended in 20 to 25 ml of 1× PBS, and 5 to 10 ml
was gently spotted at the center of the cover-
slip and incubated at room temperature for
10 min to allow the cells to sediment and bind
the surface. The coverslips were centrifuged
for 5 min at 1000 rpm to create a smooth, dense
cell monolayer. The cells were immobilized
using a hydrogel as previously described (36)
and stained with 10 ml ml−1 DAPI (Sigma-
Aldrich, #D8417) for 5 min before imaging so
that cells could be visualized.
In biofilm experiments, the fixed and per-
meabilized surface-attached microaggregates
were air dried, covered with a hydrogel, and
hybridized with the gene library and rRNA
probes in one single reaction, as described
above.
seqFISH imaging
All seqFISH experiments were performed using
a combined imaging and automated fluidics
delivery system as previously described (36).
DAPI-stained samples mounted on coverslips
were connected to the fluidic system. The re-
gions of interest were registered using the
DAPI fluorescence, and a set of sequential sec-
ondary hybridizations, washes, and imaging
was performed.
Each hybridization round contained three
unique 15-nt readouts probes, each conju-
gated to Alexa Fluor 647 (A647), Cy3B, or
Alexa Fluor 488 (A488). All readout probes
were ordered from Integrated DNA Technol-
ogies and prepared as 500 nM stock solutions.
Each serial probe mixture was prepared in EC
buffer [10% ethylene carbonate (Sigma-Aldrich,
#E26258), 10% dextran sulfate (Sigma-Aldrich,
#D4911), and 4× SSC]. Hybridizations were
incubated with the sample for 20 min to allow
for secondary probe binding. The samples
were then washed to remove excess readout
probes and to limit nonspecific binding using
~300 ml of 10% formamide wash buffer (10%
formamide and 0.1% Triton X-100 in 2× SSC).
Samples were then rinsed with ~200 ml of 4×
SSC and stained with DAPI solution (10 mg ml−1
DAPI and 4× SSC). Last, an antibleaching buf-
fer solution [10% (w/v) glucose, 1:100 diluted
catalase (Sigma-Aldrich, #C3155), 0.5 mg ml−1
glucose oxidase (Sigma-Aldrich, #G2133), and
50 mM, pH 8 Tris-HCl in 4× SSC] was flowed
through the samples. Imaging was performed
with a Leica DMi8 microscope equipped with a
confocal scanner unit (Yokogawa, #CSU-W1), a
sCMOS camera (Andor Zyla 4.2 Plus), a 63× oil
objective lens (Leica, 1.40 numerical aperture),
and a motorized stage (ASI, #MS2000). Lasers
from CNI and filter sets from Semrock were
used. Snapshots were acquired using 647-, 561-,
488-, and 405-nm fluorescent channels with
0.5-mm z-steps for all experiments, with the
exception of the 35h biofilm experiment, in
which 1.0-mm z-steps were collected. After
imaging, readout probes were stripped using
55% wash buffer (55% formamide, 0.1% Triton-
X 100, and 2× SSC) that was flowed through
for 1 min, followed by incubation for 15 min
before rinsing with 4× SSC solution. For this
protocol, serial hybridizations, imaging, and
signal quenching steps were repeated for ~40
rounds to capture 16S rRNA for multiplexing,
mRNA expression, and background signal.
The integration of the automated fluidics
delivery system and imaging was controlled
using mManager (96).
Image analysis demultiplexing and gene
expression measurement
Maximal projection images were generated
using ImageJ (97) for DAPI and 16S rRNA, and
hybridization rounds were registered using
DAPI fluorescence. Aberrations between fluo-
rophores were corrected by alignment of 16S
rRNA signals across all channels. Cells were seg-
mented using the DAPI signal with SuperSegger
using the 60XPa configuration (98) and filtered
using custom scripts to eliminate odd shapes or
autofluorescent or low-signal components.
For par-seqFISH demultiplexing, the back-
ground (no readouts) and 16S rRNA fluorescence
intensity for each relevant secondary readout
probe was measured within segmented cell
boundaries to provide a signal-to-background
score for each readout. The cells were classified
according to the positive readout combinations
(table S3). The number of false-positives was
estimated by counting the number of cells
classified into combinations left out of the
experiment.
The mRNA-FISH data were analyzed using
Spätzcells (42). Briefly, spots were detected as
regional maxima with intensity greater than a
threshold value that was set using the negative
control genes and fit with a two-dimensional
(2D) Gaussian model. The integrated inten-
sity of the spot and the position of its esti-
mated maxima were determined (42). Spots
were assigned to cells using cell segmentation
masks (42). In biofilm experiments, spots were
assigned to cells in a z-section–sensitive man-
ner. Deviating spot maxima positions that did
not overlap a cell boundary were tested against
the flanking z-sections to identify their cell of
origin. If no cell was detected, then the spots
were discarded. All predicted low-expression
genes (defined as genes with spots in <30% of
all cells) were identified, and the distribution
of their spot intensities was fit with a Gaussian
mixture model to identify the characteristic
intensity of a single mRNA normalized to the
number of probes used for the specific gene.
The median characteristic single-mRNA signal
was then calculated using all low-expression
genes for each fluorophore (A647, A488, and
cy3B). The variation between different genes
labeled with the same fluorophore was low,
with a coefficient of variation of 18 to 21%.
This median characteristic value was used to
transform fluorescence intensity into discrete
mRNA counts per gene within each cell. The
A488 characteristic signal was corrected by a
factor of 1.5 to account for its lower intensity
in our system. In each cell, the total intensity
of each gene was calculated by summing the
intensities of all spots. The total value was nor-
malized by the characteristic value for a single
mRNA in the corresponding fluorophore.
Single-cell expression analysis and cell biological
parameter calculations
Single-cell UMAP analysis was performed using
Scanpy v1.7.0 (99). Genes detected at consistent-
ly low levels were excluded from the analysis.
These included pilY1, flgK, nasA, algU, purF,
phzH, phzS, and pslG (table S1). The standard
Scanpy normalization and scaling, dimension-
ality reduction, and clustering as described in
the Scanpy tutorial were followed, minus the
high-variance gene selection and without a
library size normalization. Fifteen neighbors
and 15 and 17 PCA components were used for
the LB and merged SCFM analyses, respec-
tively. Clustering was performed using the
Leiden method. Jupyter notebooks with the
chosen parameters, run lines, output files, and
source data are available at Zenodo (see the
Acknowledgments).
Cell nucleoid size was calculated using the
segmentation mask. A chromosome score was
calculated as the median DAPI intensity mul-
tiplied by the nucleoid size. The median chro-
mosome score was calculated for the last time
point in our LB experiment (deep stationary;
OD600 = 3.2). Because most cells in this stage
are in a nondividing state, we set this value as
a reference for a single chromosome copy. We
then normalized the scores of all cells in the
experiment using this value, as seen in Fig. 2.
In addition to using Ribo-Tags to label cells
from different conditions, we also hybridized
another region in the 16S rRNA with a probe
that was shared across all samples (table S3;
described above). We used this reference sig-
nal to compare the 16S rRNA intensity between
cells from different conditions. We measured
the median 16S rRNA signal per cells and mul-
tiplied it by the nucleoid size (which completely
overlaps the 16S signal and estimates cell size).
In E. coli, maximal ribosome numbers appear
at the maximal growth rate and have been
estimated at 72,000 (100). The median rRNA
score was calculated for the maximal growth
(OD600 = 0.2) and normalized to 72,000 as in
E. coli for a rough estimate (Fig. 2).
Dar et al., Science 373, eabi4882 (2021)
13 August 2021
13 of 16
RES EARCH | R E S E A R C H A R T I C L E
Image analysis in surface
colonization experiments
Images were registered as described above,
and segmentation was performed using the
pixel classification workflow in Ilastik (101).
The Ilastik classification model was trained
with background, cell boundaries, and cell
bodies using the 16S rRNA signal and it per-
formed extremely well. However, in high-
density regions, segmentation often resulted
in overconnectivity because of incorrect
3D overlaps and such cell clusters were dis-
connected. The binary masks (segmentation
output) were thinned, and all 3D connected
components (CCs) were recalculated, which
reduced spurious connections. Then, all CCs
traversing >2.5 mm were set aside for reeval-
uation for potential overconnections. For each
such 3D component, each z-slice was examined
individually and all 2D CCs were identified.
Overly large or curved blobs, which represent
segmentation artifacts that often incorrectly
connect distinct cells across z-sections, were
removed. In addition, orientation and overlap
with components in the previous flanking
z-section were calculated for each 2D detected
component. If this component exhibited a sig-
nificant change in its orientation (i.e., the di-
rection it was pointing), it was disconnected
from the component below. The analysis was
then continued using the newly oriented com-
ponent as a seed. Cell clusters that could not
be properly disentangled were removed from
the analysis. At the end of the analysis, the cell
3D masks were re-thickened.
Bulk neighborhood analysis was used to
determine the immediate neighborhoods as-
sociated with high expression of a specific
gene. For the gene of interest, the top 99th per-
centile of cells (99.5th for the pyocin-specific
analysis), denoted as “center cells,” were iden-
tified. Using the 3D centroid coordinates of
center cells, their closest neighbors were iden-
tified within a specified distance (up to 10 mm
for pyocins and 3 mm for the rest). Then, up
to k closest cells (five to 300 neighbors in the
pyocin analysis to view the enrichment decay
and up to five for the rest of the genes) were
collected. All of the neighborhood cells selected
(not including the center cells) were then
analyzed in bulk together and their mean gene
expression was calculated and compared with
the population (minus all center cells not used).
This analysis was conducted across all genes
and a Pearson correlation analysis was per-
formed to identify spatially correlating genes
(fig. S5).
RNA-seq analysis
were analyzed. The raw sequencing data were
aligned to the P. aeruginosa reference genome
using bowtie2 (102), and the reads were
assigned to genes using featureCounts (103).
Reads per kilobase per million values were
calculated for all genes and then compared
with the average seqFISH expression profile
of each LB-grown sample (fig. S2).
RE FERENCES AND NOTES
1.
2.
H.-C. Flemming, S. Wuertz, Bacteria and archaea on Earth
and their abundance in biofilms. Nat. Rev. Microbiol. 17,
247–260 (2019). doi: 10.1038/s41579-019-0158-9;
pmid: 30760902
J. W. Costerton et al., Bacterial biofilms in nature and
disease. Annu. Rev. Microbiol. 41, 435–464 (1987).
doi: 10.1146/annurev.mi.41.100187.002251; pmid: 3318676
3. W. H. DePas et al., Exposing the three-dimensional
7.
5.
8.
4.
6.
biogeography and metabolic states of pathogens in cystic
fibrosis sputum via hydrogel embedding, clearing, and
rRNA labeling. mBio 7, e00796 (2016). doi: 10.1128/
mBio.00796-16; pmid: 27677788
J. L. Mark Welch, B. J. Rossetti, C. W. Rieken, F. E. Dewhirst,
G. G. Borisy, Biogeography of a human oral microbiome
at the micron scale. Proc. Natl. Acad. Sci. U.S.A. 113,
E791–E800 (2016). doi: 10.1073/pnas.1522149113;
pmid: 26811460
J. A. Schaber et al., Pseudomonas aeruginosa forms biofilms
in acute infection independent of cell-to-cell signaling.
Infect. Immun. 75, 3715–3721 (2007). doi: 10.1128/
IAI.00586-07; pmid: 17562773
P. Jorth, M. A. Spero, J. Livingston, D. K. Newman,
Quantitative visualization of gene expression in mucoid and
nonmucoid Pseudomonas aeruginosa aggregates reveals
localized peak expression of alginate in the hypoxic zone.
mBio 10, e02622 (2019). doi: 10.1128/mBio.02622-19;
pmid: 31848278
R. Hatzenpichler, V. Krukenberg, R. L. Spietz, Z. J. Jay,
Next-generation physiology approaches to study microbiome
function at single cell level. Nat. Rev. Microbiol. 18, 241–256
(2020). doi: 10.1038/s41579-020-0323-1; pmid: 32055027
G. L. Chadwick, F. Jiménez Otero, J. A. Gralnick, D. R. Bond,
V. J. Orphan, NanoSIMS imaging reveals metabolic
stratification within current-producing biofilms. Proc. Natl.
Acad. Sci. U.S.A. 116, 20716–20724 (2019). doi: 10.1073/
pnas.1912498116; pmid: 31548422
M. Ackermann, A functional perspective on phenotypic
heterogeneity in microorganisms. Nat. Rev. Microbiol. 13,
497–508 (2015). doi: 10.1038/nrmicro3491; pmid: 26145732
10. C. R. Evans, C. P. Kempes, A. Price-Whelan, L. E. P. Dietrich,
Metabolic heterogeneity and cross-feeding in bacterial
multicellular systems. Trends Microbiol. 28, 732–743 (2020).
doi: 10.1016/j.tim.2020.03.008; pmid: 32781027
F. Schreiber et al., Phenotypic heterogeneity driven by
nutrient limitation promotes growth in fluctuating
environments. Nat. Microbiol. 1, 16055 (2016). doi: 10.1038/
nmicrobiol.2016.55; pmid: 27572840
F. Schreiber, M. Ackermann, Environmental drivers of
metabolic heterogeneity in clonal microbial populations.
Curr. Opin. Biotechnol. 62, 202–211 (2020). doi: 10.1016/
j.copbio.2019.11.018; pmid: 31874388
S. H. Kopf et al., Heavy water and (15) N labelling with
NanoSIMS analysis reveals growth rate-dependent metabolic
heterogeneity in chemostats. Environ. Microbiol. 17,
2542–2556 (2015). doi: 10.1111/1462-2920.12752;
pmid: 25655651
J.-W. Veening, W. K. Smits, O. P. Kuipers, Bistability,
epigenetics, and bet-hedging in bacteria. Annu. Rev.
Microbiol. 62, 193–210 (2008). doi: 10.1146/annurev.
micro.62.081307.163002; pmid: 18537474
A. Z. Rosenthal et al., Metabolic interactions between
dynamic bacterial subpopulations. eLife 7, e33099 (2018).
doi: 10.7554/eLife.33099; pmid: 29809139
13.
12.
15.
14.
11.
9.
Previously published RNA-seq datasets (51)
in which wild-type P. aeruginosa PA14 was
grown in LB medium to OD600 values similar
to those collected in our experiment (OD600 =
1.1 and 2.0 in the RNA-seq experiment; Fig. 2A)
16. C. R. Armbruster et al., Heterogeneity in surface sensing
suggests a division of labor in Pseudomonas aeruginosa
populations. eLife 8, e45084 (2019). doi: 10.7554/
eLife.45084; pmid: 31180327
17. M. Diard et al., Stabilization of cooperative virulence by the
expression of an avirulent phenotype. Nature 494, 353–356
(2013). doi: 10.1038/nature11913; pmid: 23426324
19.
18.
P. S. Stewart, M. J. Franklin, Physiological heterogeneity in
biofilms. Nat. Rev. Microbiol. 6, 199–210 (2008).
doi: 10.1038/nrmicro1838; pmid: 18264116
P. S. Stewart, Diffusion in biofilms. J. Bacteriol. 185,
1485–1491 (2003). doi: 10.1128/JB.185.5.1485-1491.2003;
pmid: 12591863
E. Wolfsberg, C. P. Long, M. R. Antoniewicz, Metabolism in
dense microbial colonies: 13C metabolic flux analysis of E. coli
grown on agar identifies two distinct cell populations with
acetate cross-feeding. Metab. Eng. 49, 242–247 (2018).
doi: 10.1016/j.ymben.2018.08.013; pmid: 30179665
21. O. Kotte, B. Volkmer, J. L. Radzikowski, M. Heinemann,
20.
22.
Phenotypic bistability in Escherichia coli’s central carbon
metabolism. Mol. Syst. Biol. 10, 736 (2014). doi: 10.15252/
msb.20135022; pmid: 24987115
F. Bocci, Y. Suzuki, M. Lu, J. N. Onuchic, Role of metabolic
spatiotemporal dynamics in regulating biofilm colony
expansion. Proc. Natl. Acad. Sci. U.S.A. 115, 4288–4293
(2018). doi: 10.1073/pnas.1706920115; pmid: 29610314
23. C. D. Nadell, K. Drescher, K. R. Foster, Spatial structure,
cooperation and competition in biofilms. Nat. Rev. Microbiol.
14, 589–600 (2016). doi: 10.1038/nrmicro.2016.84;
pmid: 27452230
24. O. X. Cordero, M. S. Datta, Microbial interactions and
community assembly at microscales. Curr. Opin. Microbiol.
31, 227–234 (2016). doi: 10.1016/j.mib.2016.03.015;
pmid: 27232202
25. H. Shi et al., Highly multiplexed spatial mapping of microbial
communities. Nature 588, 676–681 (2020). doi: 10.1038/
s41586-020-2983-4; pmid: 33268897
26. C. Tropini, K. A. Earle, K. C. Huang, J. L. Sonnenburg, The gut
microbiome: Connecting spatial organization to function.
Cell Host Microbe 21, 433–442 (2017). doi: 10.1016/
j.chom.2017.03.010; pmid: 28407481
27. S. A. Wilbert, J. L. Mark Welch, G. G. Borisy, Spatial ecology
of the human tongue dorsum microbiome. Cell Rep. 30,
4003–4015.e3 (2020). doi: 10.1016/j.celrep.2020.02.097;
pmid: 32209464
F. Imdahl, E. Vafadarnejad, C. Homberger, A.-E. Saliba,
J. Vogel, Single-cell RNA-sequencing reports growth-
condition-specific global transcriptomes of individual bacteria.
Nat. Microbiol. 5, 1202–1206 (2020). doi: 10.1038/
s41564-020-0774-1; pmid: 32807892
28.
29. A. Kuchina et al., Microbial single-cell RNA sequencing by
split-pool barcoding. Science 371, eaba5257 (2021).
doi: 10.1126/science.aba5257; pmid: 33335020
30. S. B. Blattman, W. Jiang, P. Oikonomou, S. Tavazoie,
Prokaryotic single-cell RNA sequencing by in situ
combinatorial indexing. Nat. Microbiol. 5, 1192–1201 (2020).
doi: 10.1038/s41564-020-0729-6; pmid: 32451472
A. Raj, P. van den Bogaard, S. A. Rifkin, A. van Oudenaarden,
S. Tyagi, Imaging individual mRNA molecules using multiple
singly labeled probes. Nat. Methods 5, 877–879 (2008).
doi: 10.1038/nmeth.1253; pmid: 18806792
31.
32. A. M. Femino, F. S. Fay, K. Fogarty, R. H. Singer, Visualization
of single RNA transcripts in situ. Science 280, 585–590
(1998). doi: 10.1126/science.280.5363.585; pmid: 9554849
L.-H. So et al., General properties of transcriptional time
series in Escherichia coli. Nat. Genet. 43, 554–560 (2011).
doi: 10.1038/ng.821; pmid: 21532574
33.
34. H. M. T. Choi, V. A. Beck, N. A. Pierce, Next-generation in situ
hybridization chain reaction: Higher gain, lower cost, greater
durability. ACS Nano 8, 4284–4294 (2014). doi: 10.1021/
nn405717p; pmid: 24712299
35. K. H. Chen, A. N. Boettiger, J. R. Moffitt, S. Wang, X. Zhuang,
RNA imaging. Spatially resolved, highly multiplexed RNA
profiling in single cells. Science 348, aaa6090 (2015).
doi: 10.1126/science.aaa6090; pmid: 25858977
37.
36. C. L. Eng et al., Transcriptome-scale super-resolved imaging
in tissues by RNA seqFISH. Nature 568, 235–239 (2019).
doi: 10.1038/s41586-019-1049-y; pmid: 30911168
E. Lubeck, A. F. Coskun, T. Zhiyentayev, M. Ahmad, L. Cai,
Single-cell in situ RNA profiling by sequential hybridization.
Nat. Methods 11, 360–361 (2014). doi: 10.1038/nmeth.2892;
pmid: 24681720
J. R. Moffitt et al., Molecular, spatial, and functional single-cell
profiling of the hypothalamic preoptic region. Science 362,
eaau5324 (2018). doi: 10.1126/science.aau5324;
pmid: 30385464
38.
39. S. Shah, E. Lubeck, W. Zhou, L. Cai, In situ transcription
profiling of single cells reveals spatial organization of cells in
the mouse hippocampus. Neuron 92, 342–357 (2016).
doi: 10.1016/j.neuron.2016.10.001; pmid: 27764670
Dar et al., Science 373, eabi4882 (2021)
13 August 2021
14 of 16
RES EARCH | R E S E A R C H A R T I C L E
40. A. Lignell, L. Kerosuo, S. J. Streichan, L. Cai, M. E. Bronner,
41.
Identification of a neural crest stem cell niche by Spatial
Genomic Analysis. Nat. Commun. 8, 1830 (2017).
doi: 10.1038/s41467-017-01561-w; pmid: 29184067
S. Shah et al., Dynamics and spatial genomics of the nascent
transcriptome by intron seqFISH. Cell 174, 363–376.e16
(2018). doi: 10.1016/j.cell.2018.05.035; pmid: 29887381
42. S. O. Skinner, L. A. Sepúlveda, H. Xu, I. Golding, Measuring
mRNA copy number in individual Escherichia coli cells
using single-molecule fluorescent in situ hybridization.
Nat. Protoc. 8, 1100–1113 (2013). doi: 10.1038/nprot.2013.066;
pmid: 23680982
43. A. Y. Bhagirath et al., Cystic fibrosis lung environment and
Pseudomonas aeruginosa infection. BMC Pulm. Med. 16, 174
(2016). doi: 10.1186/s12890-016-0339-5; pmid: 27919253
44. S. Malhotra, D. Hayes Jr.., D. J. Wozniak, Cystic Fibrosis and
Pseudomonas aeruginosa: The Host-Microbe Interface.
Clin. Microbiol. Rev. 32, e00138-18 (2019). doi: 10.1128/
CMR.00138-18; pmid: 31142499
45. Y. Takei et al., Integrated spatial genomics reveals global
46.
47.
architecture of single nuclei. Nature 590, 344–350 (2021).
doi: 10.1038/s41586-020-03126-2; pmid: 33505024
I. Vallet-Gely, F. Boccard, Chromosomal organization and
segregation in Pseudomonas aeruginosa. PLOS Genet. 9,
e1003492 (2013). doi: 10.1371/journal.pgen.1003492;
pmid: 23658532
J. Lee, L. Zhang, The hierarchy quorum sensing network in
Pseudomonas aeruginosa. Protein Cell 6, 26–41 (2015).
doi: 10.1007/s13238-014-0100-x; pmid: 25249263
61. C. Lombardi et al., Structural and functional characterization
of the type three secretion system (T3SS) needle of
Pseudomonas aeruginosa. Front. Microbiol. 10, 573 (2019).
doi: 10.3389/fmicb.2019.00573; pmid: 31001211
62. T. Bjarnsholt et al., Pseudomonas aeruginosa biofilms in the
respiratory tract of cystic fibrosis patients. Pediatr. Pulmonol.
44, 547–558 (2009). doi: 10.1002/ppul.21011; pmid: 19418571
63. N. Høiby et al., The clinical impact of bacterial biofilms.
Int. J. Oral Sci. 3, 55–65 (2011). doi: 10.4248/IJOS11026;
pmid: 21485309
64. K. L. Palmer, L. M. Aye, M. Whiteley, Nutritional cues control
Pseudomonas aeruginosa multicellular behavior in cystic
fibrosis sputum. J. Bacteriol. 189, 8079–8087 (2007).
doi: 10.1128/JB.01138-07; pmid: 17873029
65. G. Gicquel et al., The extra-cytoplasmic function sigma factor
66.
sigX modulates biofilm and virulence-related properties in
Pseudomonas aeruginosa. PLOS ONE 8, e80407 (2013).
doi: 10.1371/journal.pone.0080407; pmid: 24260387
L. L. Burrows, Pseudomonas aeruginosa twitching motility:
Type IV pili in action. Annu. Rev. Microbiol. 66, 493–520
(2012). doi: 10.1146/annurev-micro-092611-150055;
pmid: 22746331
67. G. A. O’Toole, R. Kolter, Flagellar and twitching motility are
necessary for Pseudomonas aeruginosa biofilm development.
Mol. Microbiol. 30, 295–304 (1998). doi: 10.1046/
j.1365-2958.1998.01062.x; pmid: 9791175
68. R. Belas, Biofilms, flagella, and mechanosensing of surfaces
by bacteria. Trends Microbiol. 22, 517–527 (2014).
doi: 10.1016/j.tim.2014.05.002; pmid: 24894628
48. A. Price-Whelan, L. E. P. Dietrich, D. K. Newman, Pyocyanin
69. B.-J. Laventie et al., A surface-induced asymmetric program
49.
alters redox homeostasis and carbon flux through central
metabolic pathways in Pseudomonas aeruginosa PA14. J.
Bacteriol. 189, 6372–6381 (2007). doi: 10.1128/JB.00505-07;
pmid: 17526704
J. Jo, K. L. Cortez, W. C. Cornell, A. Price-Whelan,
L. E. Dietrich, An orphan cbb3-type cytochrome oxidase
subunit supports Pseudomonas aeruginosa biofilm growth
and virulence. eLife 6, e30205 (2017). doi: 10.7554/
eLife.30205; pmid: 29160206
51.
50. H. Arai, Regulation and function of versatile aerobic and
anaerobic respiratory metabolism in Pseudomonas
aeruginosa. Front. Microbiol. 2, 103 (2011). doi: 10.3389/
fmicb.2011.00103; pmid: 21833336
S. Schulz et al., Elucidation of sigma factor-associated
networks in Pseudomonas aeruginosa reveals a modular
architecture with limited and function-specific crosstalk.
PLOS Pathog. 11, e1004744 (2015). doi: 10.1371/journal.
ppat.1004744; pmid: 25780925
52. S. E. Finkel, Long-term survival during stationary phase:
53.
54.
Evolution and the GASP phenotype. Nat. Rev. Microbiol. 4,
113–120 (2006). doi: 10.1038/nrmicro1340; pmid: 16415927
L. McInnes, J. Healy, J. Melville, UMAP: Uniform Manifold
Approximation and Projection for Dimension Reduction.
arXiv:1802.03426 [stat.ML] (9 February 2018).
E. Bosdriesz, D. Molenaar, B. Teusink, F. J. Bruggeman,
How fast-growing bacteria robustly tune their ribosome
concentration to approximate growth-rate maximization.
FEBS J. 282, 2029–2044 (2015). doi: 10.1111/febs.13258;
pmid: 25754869
55. B. R. Borlee et al., Pseudomonas aeruginosa uses a
cyclic-di-GMP-regulated adhesin to reinforce the biofilm
extracellular matrix. Mol. Microbiol. 75, 827–842 (2010).
doi: 10.1111/j.1365-2958.2009.06991.x; pmid: 20088866
56. C. Reichhardt, C. Wong, D. Passos da Silva, D. J. Wozniak,
M. R. Parsek, CdrA: interactions within the Pseudomonas
aeruginosa biofilm matrix safeguard it from proteolysis and
promote cellular packing. mBio 9, e01376 (2018).
doi: 10.1128/mBio.01376-18; pmid: 30254118
57. O. Zaborina et al., Structure-function aspects of PstS in
multi-drug-resistant Pseudomonas aeruginosa. PLOS Pathog.
4, e43 (2008). doi: 10.1371/journal.ppat.0040043;
pmid: 18282104
58. M. L. Gil-Marqués et al., Role of PstS in the pathogenesis of
Acinetobacter baumannii under microaerobiosis and
normoxia. J. Infect. Dis. 222, 1204–1212 (2020).
doi: 10.1093/infdis/jiaa201; pmid: 32324853
59. A. R. Hauser, The type III secretion system of Pseudomonas
aeruginosa: Infection by injection. Nat. Rev. Microbiol. 7,
654–665 (2009). doi: 10.1038/nrmicro2199; pmid: 19680249
60. O. Wurtzel et al., The single-nucleotide resolution
transcriptome of Pseudomonas aeruginosa grown in body
temperature. PLOS Pathog. 8, e1002945 (2012).
doi: 10.1371/journal.ppat.1002945; pmid: 23028334
promotes tissue colonization by Pseudomonas aeruginosa.
Cell Host Microbe 25, 140–152.e6 (2019). doi: 10.1016/
j.chom.2018.11.008; pmid: 30581112
71.
72.
70. M. G. K. Ghequire, R. De Mot, The tailocin tale: Peeling off
phage tails. Trends Microbiol. 23, 587–590 (2015).
doi: 10.1016/j.tim.2015.07.011; pmid: 26433692
L.-M. Bobay, M. Touchon, E. P. C. Rocha, Pervasive
domestication of defective prophages by bacteria. Proc. Natl.
Acad. Sci. U.S.A. 111, 12127–12132 (2014). doi: 10.1073/
pnas.1405336111; pmid: 25092302
L. Turnbull et al., Explosive cell lysis as a mechanism for the
biogenesis of bacterial membrane vesicles and biofilms.
Nat. Commun. 7, 11220 (2016). doi: 10.1038/ncomms11220;
pmid: 27075392
J. Vacheron, C. M. Heiman, C. Keel, Live cell dynamics
of production, explosive release and killing activity of phage
tail-like weapons for Pseudomonas kin exclusion. Commun.
Biol. 4, 87 (2021). doi: 10.1038/s42003-020-01581-1;
pmid: 33469108
73.
74. C. B. Whitchurch, T. Tolker-Nielsen, P. C. Ragas, J. S. Mattick,
Extracellular DNA required for bacterial biofilm formation.
Science 295, 1487 (2002). doi: 10.1126/
science.295.5559.1487; pmid: 11859186
78.
76.
75. M. D. Brazas, R. E. W. Hancock, Ciprofloxacin induction of a
susceptibility determinant in Pseudomonas aeruginosa.
Antimicrob. Agents Chemother. 49, 3222–3227 (2005).
doi: 10.1128/AAC.49.8.3222-3227.2005; pmid: 16048929
T. L. Povolotsky, A. Keren-Paz, I. Kolodkin-Gal, Metabolic
microenvironments drive microbial differentiation and
antibiotic resistance. Trends Genet. 37, 4–8 (2021).
doi: 10.1016/j.tig.2020.10.007; pmid: 33203570
77. A. K. Wessel et al., Oxygen limitation within a bacterial
aggregate. mBio 5, e00992 (2014). doi: 10.1128/
mBio.00992-14; pmid: 24736225
L. E. P. Dietrich et al., Bacterial community morphogenesis is
intimately linked to the intracellular redox state. J. Bacteriol.
195, 1371–1380 (2013). doi: 10.1128/JB.02273-12;
pmid: 23292774
E. S. Cowley, S. H. Kopf, A. LaRiviere, W. Ziebis,
D. K. Newman, Pediatric cystic fibrosis sputum can be
chemically dynamic, anoxic, and extremely reduced due to
hydrogen sulfide formation. mBio 6, e00767 (2015).
doi: 10.1128/mBio.00767-15; pmid: 26220964
80. S. S. Yoon et al., Pseudomonas aeruginosa anaerobic
respiration in biofilms: Relationships to cystic fibrosis
pathogenesis. Dev. Cell 3, 593–603 (2002). doi: 10.1016/
S1534-5807(02)00295-2; pmid: 12408810
79.
81. M. Eschbach et al., Long-term anaerobic survival of the
opportunistic pathogen Pseudomonas aeruginosa via
pyruvate fermentation. J. Bacteriol. 186, 4596–4604 (2004).
doi: 10.1128/JB.186.14.4596-4604.2004; pmid: 15231792
82. D. J. Hassett, L. Charniga, K. Bean, D. E. Ohman, M. S. Cohen,
Response of Pseudomonas aeruginosa to pyocyanin:
Mechanisms of resistance, antioxidant defenses, and
demonstration of a manganese-cofactored superoxide
dismutase. Infect. Immun. 60, 328–336 (1992).
doi: 10.1128/iai.60.2.328-336.1992; pmid: 1730464
83. S. M. Brown, M. L. Howell, M. L. Vasil, A. J. Anderson,
D. J. Hassett, Cloning and characterization of the katB gene
of Pseudomonas aeruginosa encoding a hydrogen
peroxide-inducible catalase: Purification of KatB, cellular
localization, and demonstration that it is essential for optimal
resistance to hydrogen peroxide. J. Bacteriol. 177,
6536–6544 (1995). doi: 10.1128/jb.177.22.6536-6544.1995;
pmid: 7592431
84. S. Su et al., Catalase (KatA) plays a role in protection against
anaerobic nitric oxide in Pseudomonas aeruginosa. PLOS ONE
9, e91813 (2014). doi: 10.1371/journal.pone.0091813;
pmid: 24663218
F. Cutruzzolà, N. Frankenberg-Dinkel, Origin and impact of
nitric oxide in Pseudomonas aeruginosa biofilms. J. Bacteriol.
198, 55–65 (2016). doi: 10.1128/JB.00371-15;
pmid: 26260455
85.
86. D. W. Basta, M. Bergkessel, D. K. Newman, Identification of
fitness determinants during energy-limited growth arrest in
Pseudomonas aeruginosa. mBio 8, e01170 (2017).
doi: 10.1128/mBio.01170-17; pmid: 29184024
87. U. A. Ochsner, M. L. Vasil, E. Alsabbagh, K. Parvatiyar,
D. J. Hassett, Role of the Pseudomonas aeruginosa oxyR-recG
operon in oxidative stress defense and DNA repair:
OxyR-dependent regulation of katB-ankB, ahpB, and ahpC-ahpF.
J. Bacteriol. 182, 4533–4544 (2000). doi: 10.1128/
JB.182.16.4533-4544.2000; pmid: 10913087
88. M. Toyofuku et al., Quorum sensing regulates denitrification
in Pseudomonas aeruginosa PAO1. J. Bacteriol. 189,
4969–4972 (2007). doi: 10.1128/JB.00289-07;
pmid: 17449629
89. N. E. Van Alst, L. A. Sherrill, B. H. Iglewski, C. G. Haidaris,
Compensatory periplasmic nitrate reductase activity
supports anaerobic growth of Pseudomonas aeruginosa
PAO1 in the absence of membrane nitrate reductase.
Can. J. Microbiol. 55, 1133–1144 (2009). doi: 10.1139/W09-065;
pmid: 19935885
90. K. Schreiber et al., Anaerobic survival of Pseudomonas
aeruginosa by pyruvate fermentation requires an Usp-type
stress protein. J. Bacteriol. 188, 659–668 (2006).
doi: 10.1128/JB.188.2.659-668.2006; pmid: 16385055
S. K. Dolan, M. Welch, The glyoxylate shunt, 60 years on.
Annu. Rev. Microbiol. 72, 309–330 (2018). doi: 10.1146/
annurev-micro-090817-062257; pmid: 30200852
91.
92. A. Crousilles, S. K. Dolan, P. Brear, D. Y. Chirgadze, M. Welch,
Gluconeogenic precursor availability regulates flux
through the glyoxylate shunt in Pseudomonas aeruginosa.
J. Biol. Chem. 293, 14260–14269 (2018). doi: 10.1074/
jbc.RA118.004514; pmid: 30030382
93. S. K. Dolan, G. Pereira, R. Silva-Rocha, M. Welch,
Transcriptional regulation of central carbon metabolism
in Pseudomonas aeruginosa. Microb. Biotechnol. 13,
285–289 (2020). doi: 10.1111/1751-7915.13423;
pmid: 31187593
94. T. Kawakami, M. Kuroki, M. Ishii, Y. Igarashi, H. Arai,
Differential expression of multiple terminal oxidases
for aerobic respiration in Pseudomonas aeruginosa. Environ.
Microbiol. 12, 1399–1412 (2010). pmid: 19930444
95. A. M. Stern, J. Zhu, An introduction to nitric oxide sensing
and response in bacteria. Adv. Appl. Microbiol. 87, 187–220
(2014). doi: 10.1016/B978-0-12-800261-2.00005-0;
pmid: 24581392
96. A. Edelstein, N. Amodaj, K. Hoover, R. Vale, N. Stuurman,
Computer control of microscopes using µManager. Curr. Protoc.
Mol. Biol. Chapter 14, 20 (2010). pmid: 20890901
97. C. A. Schneider, W. S. Rasband, K. W. Eliceiri, NIH Image to
ImageJ: 25 years of image analysis. Nat. Methods 9, 671–675
(2012). doi: 10.1038/nmeth.2089; pmid: 22930834
98. S. Stylianidou, C. Brennan, S. B. Nissen, N. J. Kuwada,
P. A. Wiggins, SuperSegger: Robust image segmentation,
analysis and lineage tracking of bacterial cells. Mol. Microbiol.
102, 690–700 (2016). doi: 10.1111/mmi.13486;
pmid: 27569113
F. A. Wolf, P. Angerer, F. J. Theis, SCANPY: Large-scale
single-cell gene expression data analysis. Genome Biol. 19, 15
(2018). doi: 10.1186/s13059-017-1382-0; pmid: 29409532
99.
100. R. Milo, P. Jorgensen, U. Moran, G. Weber, M. Springer,
BioNumbers—The database of key numbers in molecular and
cell biology. Nucleic Acids Res. 38 (suppl_1), D750–D753
(2010). doi: 10.1093/nar/gkp889; pmid: 19854939
Dar et al., Science 373, eabi4882 (2021)
13 August 2021
15 of 16
RES EARCH | R E S E A R C H A R T I C L E
101. S. Berg et al., ilastik: Interactive machine learning for (bio)
image analysis. Nat. Methods 16, 1226–1232 (2019).
doi: 10.1038/s41592-019-0582-9; pmid: 31570887
102. B. Langmead, S. L. Salzberg, Fast gapped-read alignment
with Bowtie 2. Nat. Methods 9, 357–359 (2012).
doi: 10.1038/nmeth.1923; pmid: 22388286
103. Y. Liao, G. K. Smyth, W. Shi, featureCounts: An efficient
general purpose program for assigning sequence reads to
genomic features. Bioinformatics 30, 923–930 (2014).
doi: 10.1093/bioinformatics/btt656; pmid: 24227677
104. Y. Zhang, Z. Hu, Combined treatment of Pseudomonas
aeruginosa biofilms with bacteriophages and chlorine.
Biotechnol. Bioeng. 110, 286–295 (2013). doi: 10.1002/
bit.24630; pmid: 22886888
105. D. Dar, N. Dar, L. Cai, D. K. Newman, Data for: Spatial
transcriptomics of planktonic and sessile bacterial
populations at single-cell resolution, Zenodo (2021);
doi: 10.5281/zenodo.4767568
106. D. Dar, N. Dar, L. Cai, D. K. Newman, Imaging data for: Spatial
transcriptomics of planktonic and sessile bacterial
populations at single-cell resolution, Zenodo (2021);
doi: 10.5281/zenodo.4771778
ACKN OWLED GMEN TS
We thank G. A. O’Toole and M. Whiteley for help with designing the
gene set, M. Bergkessel and R. Sorek for critically reading the
manuscript, and members of the Newman laboratory for critically
reading the manuscript and for fruitful discussions and comments,
particularly M. Bergkessel for assistance with RNA-Seq analysis.
Funding: This work was supported by the National Institutes
of Health (grants 1R01AI127850-01A1 and 1R01HL152190-01 to D.K.N.)
and the Army Research Office (grant W911NF-17-1-0024 to D.K.N.).
L.C. was supported by the Allen Frontier group. D.D. was
supported by the Rothschild foundation, EMBO Long-Term, and
Helen Hay Whitney postdoctoral fellowships, as well as a
Geobiology Postdoctoral Fellowship from the Division of Geological
and Planetary Sciences, Caltech. Author contributions: D.D., N.D.,
L.C., and D.K.N. designed the study. D.D. led the study, designed the
experiments, and performed the experiments with N.D. D.D.
analyzed the data. D.K.N. and L.C. supervised the study. All authors
contributed to writing the manuscript. Competing interests:
L.C. is a cofounder of Spatial Genomics, Inc. A provisional patent
(No. 63/153,234) has been filed by California Institute of Technology
with inventors Daniel Dar, Dianne K. Newman, Kirsten Frieda, and
Long Cai entitled “Multiplexing of experimental conditions and
samples in spatial genomics.” Data and materials availability:
Custom MATLAB scripts and single-cell source data from this study
are available at Zenodo (105). Imaging data obtained during this
study have also been deposited at Zenodo (106). All other data are
presented in the main text or the supplementary materials.
SUPPLEMENTARY MATERIALS
science.sciencemag.org/content/373/6556/eabi4882/suppl/DC1
Figs. S1 to S8
Tables S1 to S4
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
12 March 2021; accepted 25 June 2021
10.1126/science.abi4882
Dar et al., Science 373, eabi4882 (2021)
13 August 2021
16 of 16
|
10.1126_science.abi8794
|
RES EARCH
TOPOLOGICAL MATTER
Probing topological spin liquids on a programmable
quantum simulator
G. Semeghini1, H. Levine1, A. Keesling1,2, S. Ebadi1, T. T. Wang1, D. Bluvstein1, R. Verresen1,
H. Pichler3,4, M. Kalinowski1, R. Samajdar1, A. Omran1,2, S. Sachdev1,5, A. Vishwanath1*,
M. Greiner1*, V. Vuletic´ 6*, M. D. Lukin1*
Quantum spin liquids, exotic phases of matter with topological order, have been a major focus in physics
for the past several decades. Such phases feature long-range quantum entanglement that can
potentially be exploited to realize robust quantum computation. We used a 219-atom programmable
quantum simulator to probe quantum spin liquid states. In our approach, arrays of atoms were placed on
the links of a kagome lattice, and evolution under Rydberg blockade created frustrated quantum states
with no local order. The onset of a quantum spin liquid phase of the paradigmatic toric code type
was detected by using topological string operators that provide direct signatures of topological order
and quantum correlations. Our observations enable the controlled experimental exploration of
topological matter and protected quantum information processing.
M otivated by theoretical work carried
out over the past five decades, a broad
search has been underway to identify
signatures of quantum spin liquids
(QSLs) in correlated materials (1, 2).
Moreover, inspired by the intriguing predic-
tions of quantum information theory (3),
approaches to engineer such systems for topo-
logical protection of quantum information are
being actively explored (4). Systems with frus-
tration (5) caused by the lattice geometry or
long-range interactions constitute a promising
avenue in the search for QSLs. In particular,
such systems can be used to implement a class
of so-called dimer models (6–10), which are
among the most promising candidates to host
QSL states. However, realizing and probing
such states is challenging because they are
often surrounded by other competing phases.
Moreover, in contrast to topological systems
that involve time-reversal symmetry breaking,
such as in the fractional quantum Hall effect
(11), these states cannot be easily probed by
means of, for example, quantized conductance
or edge states. Instead, to diagnose spin liquid
phases, it is essential to access nonlocal ob-
servables, such as topological string operators
(1, 2). Although some indications of QSL phases
in correlated materials have been previously
reported (12, 13), thus far, these exotic states
of matter have evaded direct experimental
detection.
1Department of Physics, Harvard University, Cambridge,
MA 02138, USA. 2QuEra Computing, Boston, MA 02135, USA.
3Institute for Theoretical Physics, University of Innsbruck,
Innsbruck A-6020, Austria. 4Institute for Quantum Optics
and Quantum Information, Austrian Academy of Sciences,
Innsbruck A-6020, Austria. 5School of Natural Sciences,
Institute for Advanced Study, Princeton, NJ 08540, USA.
6Department of Physics and Research Laboratory of
Electronics, Massachusetts Institute of Technology,
Cambridge, MA 02139, USA.
*Corresponding author. Email: avishwanath@g.harvard.edu
(A.V.); greiner@physics.harvard.edu (M.G.); vuletic@mit.edu
(V.V.); lukin@physics.harvard.edu (M.D.L.)
Programmable quantum simulators are well
suited for the controlled exploration of these
strongly correlated quantum phases (14–21).
In particular, recent work showed that various
phases of quantum dimer models can be effi-
ciently implemented by using Rydberg atom
arrays (22) and that a dimer spin liquid state of
the toric code type could be potentially created
in a specific frustrated lattice (23). Toric code
states have been dynamically created in small
systems by using quantum circuits (24, 25).
However, some of the key properties, such
as topological robustness, are challenging to
realize in such systems. Spin liquids have also
been explored by using quantum annealers,
but the lack of coherence in these systems has
precluded the observation of quantum fea-
tures (26).
Dimer models in Rydberg atom arrays
The key idea of our approach is based on a
correspondence (23) between Rydberg atoms
placed on the links of a kagome lattice (or
equivalently, the sites of a ruby lattice) (Fig. 1A)
and dimer models on the kagome lattice (8, 10).
The Rydberg excitations can be viewed as
“dimer bonds” that connect the two adjacent
vertices of the lattice (Fig. 1B). Because of the
Rydberg blockade (27), strong and properly
tuned interactions constrain the density of
excitations so that each vertex is touched by
a maximum of one dimer. At 1/4 filling, each
vertex is touched by exactly one dimer, result-
ing in a perfect dimer covering of the lattice.
Smaller filling fractions result in a finite den-
sity of vertices with no proximal dimers, which
are referred to as monomers. A QSL can emerge
within this dimer-monomer model close to
1/4 filling (23) and can be viewed as a co-
herent superposition of exponentially many
degenerate dimer coverings with a small ad-
mixture of monomers (Fig. 1C) (10). This cor-
responds to the resonating valence bond (RVB)
state (6, 28), which was predicted long ago
but is so far still unobserved in any experi-
mental system.
To create and study such states experimental-
ly, we used two-dimensional arrays of 219 87Rb
atoms individually trapped in optical tweez-
ers (29, 30) and positioned on the links of a
kagome lattice (Fig. 1A). The atoms were ini-
tialized in an electronic ground state gj i and
coupled to a Rydberg state rj i by means of a
two-photon optical transition with Rabi fre-
quency W. The atoms in the Rydberg state rj i
interact with one another through a strong
van der Waals potential V = V0/d6, where d is
the interatomic distance. This strong inter-
action prevents the simultaneous excitation
of two atoms within a blockade radius Rb =
(V0/W)1/6 (27). We adjusted the lattice spacing
a and the Rabi frequency W so that for each
atom in rj i, its six nearest neighbors are all
within the blockade radius (Fig. 1B), result-
ing in a maximum filling fraction of 1/4. The
resulting dynamics correspond to unitary evo-
lution U(t) governed by the Hamiltonian
H
ℏ
¼
X
W tð Þ
2
X
þ
i
Vijninj
sx
i
(cid:2) D tð Þ
X
ni
i
ð1Þ
i<j
¼ gij
where ℏ is Planck’s constant h divided by 2p,
ni ¼ rij i rih j is the Rydberg state occupation at
j, and D(t) is the
i rih j þ rij i gih
site i, sx
i
time-dependent two-photon detuning. After
the evolution, the state was analyzed by means
of projective readout of ground-state atoms
(Fig. 1A, right) (29).
To explore many-body phases in this system,
we used quasi-adiabatic evolution, in which
we slowly turned on the Rydberg coupling W
and subsequently changed the detuning D
from negative to positive values by using a
cubic frequency sweep over ~2 ms (Fig. 1D). We
stopped the cubic sweep at different endpoints
and first measured the density of Rydberg ex-
citations nh i. Away from the array boundaries
(which result in edge effects permeating just
two layers into the bulk), we observed that the
average density of Rydberg atoms was uniform
across the array (fig. S4) (31). Focusing on the
bulk density, we found that for D=W ≳ 3, the
system reaches the desired filling fraction
nh i e 1=4 (Fig. 1E, top). The resulting state
does not have any obvious spatial order (Fig.
1A) and appears as a different configuration of
Rydberg atoms in each experimental repeti-
tion (fig. S5) (31). From the single-shot images,
we evaluated the probability for each vertex of
the kagome lattice to be attached to one dimer
(as in a perfect dimer covering), zero dimers (a
monomer), or two dimers (representing weak
blockade violations). Around D/W ~ 4, we ob-
served an approximate plateau at which ~80%
of the vertices were connected to a single dimer
Semeghini et al., Science 374, 1242–1247 (2021)
3 December 2021
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
(Fig. 1E), indicating an approximate dimer
covering.
Measuring topological string operators
A defining property of a phase with topolog-
ical order is that it cannot be probed locally.
Hence, to investigate the possible presence of
a QSL state, it is essential to measure nonlocal
observables. In the case of dimer models, a
particularly convenient set of nonlocal varia-
bles is defined in terms of topological string
operators, which are analogous to those used
in the toric code model (3). For the present
model, there are two such string operators,
the first of which characterizes the effective
dimer description; the second probes quan-
tum coherence between dimer states (23).
We first focused on the diagonal operator
Y
i
i∈s
i , with sz
sz
¼ 1 (cid:2) 2ni , which mea-
Z ¼
sures the parity of Rydberg atoms along a
string S perpendicular to the bonds of the
kagome lattice (Fig. 2A). For the smallest
closed Z loop, which encloses a single ver-
tex of the kagome lattice, Zh i ¼ (cid:2)1 for any
perfect dimer covering. Larger loops can be
decomposed into a product of small loops
around all the enclosed vertices, resulting in
Fig. 1. Dimer model in Rydberg atoms arrays. (A) Fluorescence image of
219 atoms arranged on the links of a kagome lattice. The atoms, initially in the
ground state gj i, evolve according to the many-body dynamics U(t). The final
state of the atoms was determined by means of fluorescence imaging of ground-
state atoms. Rydberg atoms are indicated with red dimers on the bonds of
the kagome lattice. (B) We adjusted the blockade radius to Rb/a = 2.4 by
choosing W = 2p × 1.4 MHz and a = 3.9 mm, so that all six nearest neighbors of
an atom in rj i are within the blockade radius Rb. A state consistent with the
Rydberg blockade at maximal filling can then be viewed as a dimer covering of
the kagome lattice, where each vertex is touched by exactly one dimer. (C) In
the idealized limit, the QSL state corresponds to a coherent superposition of
exponentially many dimer coverings. (D) Detuning D(t) and Rabi frequency W(t)
used for quasi-adiabatic state preparation. (E) (Top) Average density of Rydberg
excitations nh i in the bulk of the system, excluding the outer three layers
(31). (Bottom) Probabilities of empty vertices in the bulk (monomers; blue
symbols), vertices attached to a single dimer (red symbols), or to double dimers
(weakly violating blockade; green symbols). After D/W ~ 3, the system reaches
~1/4 filling, where most vertices are attached to a single dimer, which is
consistent with an approximate dimer phase. The average density of defects
per vertex in the approximate dimer phase is ~0.2.
Fig. 2. Detecting a dimer phase by means of diagonal string operator.
(A) The Z string operator measures the parity of dimers along a string. (B) A
perfect dimer covering always has exactly one dimer touching each vertex of the
Þ#enclosed vertices for
array, so that Zh i ¼ (cid:2)1
larger loops. (C) Z parity measurements following the quasi-adiabatic sweep of Fig. 1D,
Þ around a single vertex and Zh i ¼ (cid:2)1ð
ð
with the addition of a 200-ns ramp-down of W at the end to optimize preparation. At
different endpoints of the sweep and for (D) different loop sizes, we measured a
finite Zh i, which is consistent with an approximate dimer phase. The sign of Zh i
properly matches the parity of the number of enclosed vertices: 6 (red), 11 (green),
15 (blue), and 19 (orange). (E) The measured Zh i for the two largest loops (fig. S9).
Semeghini et al., Science 374, 1242–1247 (2021)
3 December 2021
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Probing coherence between dimer states by means of off-diagonal
string operator. (A) Definition of X string operator on a single triangle of
the kagome lattice. (B) On any closed loop, the X operator maps any dimer
covering into another valid dimer covering, so that Xh i measures the coherence
between pairs of dimer configurations. (C) The X operator is measured by
evolving the initial state under the Hamiltonian (Eq. 1) with D = 0 and reduced
blockade radius to encompass only atoms within each individual triangle,
implementing a basis rotation that maps X into Z. (D) In the experiment, after
the state preparation, we set the laser detuning to Dq = 0 and increased W to
2p × 20 MHz to reach Rb/a = 1.53. (E) By measuring the Z parity on the dual
string (red) of a target X loop (blue) after a variable quench time, we identified
the time t for which the mapping in (C) is implemented. (F) We measured Xh i
for different final detunings of the cubic sweep and (inset) for different loop
sizes and found that the prepared state has long-range coherence that extends
over a large fraction of the array (31). The dual Z loops corresponding to the
X loops shown in the inset are defined in fig. S3 and (31).
Þ#enclosed vertices (Fig. 2B). The pres-
Zh i ¼ (cid:2)1ð
ence of monomers or double-dimers reduces
the effective contribution of each vertex, re-
sulting in a reduced Zh i.
To measure Zh i for different loop shapes
(Figs. 2, C and D), we evaluated the string
observables directly from single-shot images,
averaging over many experimental repetitions
and over all loops of the same shape in the bulk
of the lattice (31). In the range of detunings
where nh i e 1=4, we clearly observed the emer-
gence of a finite Zh i for all loop shapes, with
the sign matching the parity of enclosed ver-
tices, as expected for dimer states (Fig. 2B).
j < 1
The measured values were generally Zh i
and decreased with increasing loop size, sug-
gesting the presence of a finite density of de-
fects. Nevertheless, these observations indicate
that the state we prepared was consistent with
an approximate dimer phase.
j
We next explored quantum coherence prop-
erties of the prepared state. To this end, we con-
sidered the off-diagonal X operator, which acts
on strings along the bonds of the kagome lat-
tice. It is defined in Fig. 3A by its action on a
single triangle (23). Applying X on any closed
string maps a dimer covering to another valid
dimer covering (for example, a loop around a
single hexagon in Fig. 3B). A finite expectation
value for X therefore implies that the state con-
tains a coherent superposition of one or more
pairs of dimer states coupled by that specific
loop, which is a prerequisite for a QSL. The
measurement of X can be implemented by per-
forming a collective basis rotation, illustrated
in Fig. 3C (23). This rotation was implemented
through time evolution under the Rydberg
Hamiltonian (Eq. 1) with D = 0 and reduced
blockade radius Rb/a = 1.53, so that only the
atoms within the same triangle were subject
to the Rydberg blockade constraint. Under
these conditions, it was sufficient to consider
the evolution of individual triangles separate-
ly, where each triangle can be described as a
). Within this
four-level system (
p(cid:3)
ffiffiffi
subspace, after a time t ¼ 4p= 3W
3
, the
collective three-atom dynamics realizes a
unitary Uq that implements the basis rota-
tion that transforms an X string into a dual
Z string (31).
(cid:4)
Experimentally, the basis rotation was imple-
mented after the state preparation by quench-
ing the laser detuning to Dq = 0 and increasing
the laser intensity by a factor of ~200 to reduce
the blockade radius to Rb/a = 1.53 (Fig. 3D)
(31). We calibrated t by preparing the state
at D/W = 4 and evolving under the quench
Hamiltonian for a variable time. We measured
the parity of a Z string that was dual to a target
X loop and observed a sharp revival of the
parity signal at t ~ 30 ns (Fig. 3E) (23). Fixing
the quench time t, we measured Xh i for dif-
ferent values of the detuning D at the end of
the cubic sweep (Fig. 3F) and observed a finite
X parity signal for loops that extend over a
large fraction of the array. These observations
clearly indicate the presence of long-range
coherence in the prepared state.
Probing spin liquid properties
The study of closed string operators showed
that we prepared an approximate dimer phase
with quantum coherence between dimer cov-
erings. Although these closed loops are in-
dicative of topological order, we needed to
compare their properties with those of open
strings to distinguish topological effects from
trivial ordering—the former being sensitive to
the topology of the loop (32–34). This compar-
ison is shown in Fig. 4, D and E, and indicates
several distinct regimes. For small D, we found
that both Z and X loop parities factorize into
the product of the parities on the half-loop
open strings; in particular, the finite Zh i is a
trivial result of the low density of Rydberg
excitations. By contrast, loop parities no longer
factorize in the dimer phase (3 ≲ D=W ≲ 5). In-
stead, the expectation values for both open
string operators vanish in the dimer phase,
indicating the nontrivial nature of the corre-
lations measured with the closed loops (31).
More specifically, topological ordering in the
dimer-monomer model can break down either
because of a high density of monomers, cor-
responding to the trivial disordered phase at
small D/W, or owing to the lack of long-range
resonances, corresponding to a valence bond
solid (VBS) (23). Open Z and X strings distin-
guish the target QSL phase from these proxi-
mal phases: When normalized according to
the definition from Fredenhagen and Marcu
(FM) (Fig. 4, B and C) (32, 33), vanishing
expectation values for these open strings can
be considered to be key signatures for the QSL.
In particular, open Z strings have a finite ex-
pectation value when the dimers form an or-
dered spatial arrangement, as in the VBS phase.
At the same time, open X strings create pairs
Semeghini et al., Science 374, 1242–1247 (2021)
3 December 2021
3 of 6
magnetic (m) anyon. Analogous to e-anyons,
which live at the endpoints of open X strings
(Fig. 4A), m-anyons are created by open Z
strings and correspond to phase errors be-
tween dimer coverings (fig. S11) (31). These
excitations cannot be directly identified from
individual snapshots but are detected with
the measurement of closed X loop operators.
The perimeter law scaling observed in Fig. 4I
indicates that m-anyons only appear in pairs
with short correlation lengths (31). These ob-
servations highlight the prospects for using
topological string operators to detect and probe
quasiparticle excitations in the system.
Toward a topological qubit
To further explore the topological properties
of the spin liquid state, we created an atom ar-
ray with a small hole by removing three atoms
on a central triangle (Fig. 5), which creates
RES EARCH | R E S E A R C H A R T I C L E
of monomers at their endpoints (Fig. 4A), so a
finite Xh i can be achieved in the trivial phase,
where there is a high density of monomers.
Therefore, the QSL can be identified as the
only phase where both FM string order pa-
rameters vanish for long strings (23).
The measured values of the FM order param-
eters are shown in Fig. 4, F and G. We found
that Zh i
FM is compatible with zero over the
entire range of D/W, where we observed a fi-
nite Z parity on closed loops, indicating the
absence of a VBS phase (Fig. 4F), which is
consistent with our analysis of density-density
correlations (fig. S6) (31). At the same time,
Xh i
FM converges toward zero on the longest
strings for D=W ≳ 3:3 (Fig. 4G), indicating a
transition out of the disordered phase. By
combining these two measurements with the
regions of nonvanishing parity for the closed
Z and X loops (Figs. 2 and 3), we conclude that
for 3:3 ≲ D=W ≲ 4:5, our results constitute a
direct detection of the onset of a QSL phase
(Fig. 4, F and G, shaded area).
j
j
j; Xh i
The measurements of the closed-loop oper-
ators in Figs. 2 and 3 show that Zh i
j < 1
and that the amplitude of the signal decreases
with increasing loop size, which results from
a finite density of quasiparticle excitations.
Specifically, defects in the dimer covering such
as monomers and double-dimers can be inter-
preted as electric (e) anyons in the language of
lattice gauge theory (23). Because the presence
of a defect inside a closed loop changes the
sign of Z, the parity on the loop is reduced
according to the number of enclosed e-anyons
Þ#enclosed e(cid:2)anyons
(cid:5)
E
(cid:5)
(cid:5). The average
j
as Zh i
(cid:5)
D
(cid:5)
(cid:5)
(cid:2)1ð
j ¼
j
number of defects inside a loop is expected to
scale with the number of enclosed vertices—
with the area of the loop—and we observed an
approximate area-law scaling of Zh i
j for small
loop sizes (Fig. 4H). However, for larger loops
we observed a deviation from area-law scaling,
closer to a perimeter law. This can emerge if
pairs of anyons are correlated over a char-
acteristic length scale smaller than the loop
size [a discussion of the expected scaling is
provided in (31)]. Pairs of correlated anyons
that are both inside the loop do not change
its parity because their contributions cancel
out; they only affect Zh i when they sit across
the loop, leading to a scaling with the length
of the perimeter. These pairs can be viewed
as resulting from the application of X string
operators to a dimer covering (Fig. 4A), orig-
inating, for example, from virtual excitations
in the dimer-monomer model (31) or from
errors caused by state preparation and detec-
tion. State preparation with larger Rabi fre-
quency (improved adiabaticity) results in
larger Z parity signals and reduced e-anyon
density (fig. S9).
A second type of quasiparticle excitation
that could arise in this model is the so-called
FM. (D) Comparison between Zclosed
FM and Xh i
(cid:7)
2 measured on the strings shown in the inset. The expectation value shown for the open string is
Fig. 4. String order parameters and quasi-particle excitations. (A) An open string operator Xopen acting
on a dimer state Dj i creates two monomers (e-anyons) at its endpoints (m-anyons are shown in fig. S11).
(B and C) Definition of the string order parameters Zh i
(cid:6)
and Zopen
squared to account for a factor of two in the string lengths. (E) Analogous comparison for X. (F and
G) Zooming in on the range with finite closed loop parities, we measured the FM order parameters for
FM is consistent with zero over the entire range of D, whereas
different open strings (insets). We found that Zh i
FM vanishes for D=W ≳ 3:3, which allowed us to identify a range of detunings consistent with the onset
Xh i
of a QSL phase (shaded area). (H) Rescaled parities Zh i1=area and Zh i1=perim evaluated for D/W = 3.6, where
area and perimeter are defined as the number of vertices enclosed by the loop and the number of atoms
on the loop, respectively. For small loops, Z scales with an area law but deviates from this behavior for larger
loops, converging toward a perimeter law. (I) Xh i1=area (the area, in this case, is the number of enclosed
hexagons) and Xh i1=perim evaluated for D/W = 3.5, indicating an excellent agreement with a perimeter-law scaling.
i
h
Semeghini et al., Science 374, 1242–1247 (2021)
3 December 2021
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
A
B
C
Fig. 5. Topological properties in array with a hole. (A) A lattice with nontrivial topology is obtained by
removing three atoms at the center to create a small hole. The dimer states can be divided into two distinct
topological sectors 0 and 1. Z strings connecting the hole to the boundary always have a well-defined
expectation value within each sector and opposite sign between the two sectors; the correlations between
two such strings Z1Z2 are identical for both sectors. (B) Measured expectation values for the operators ZL
and XL, defined in the inset, indicate that in the QSL region (shaded area), we prepared a superposition
state of the two topological sectors ( ZLh
expectation values for the correlations between pairs of hole-to-boundary Z strings (inset), which is
consistent with (A).
i ¼ 0) with a finite overlap with the þj i state ( XLh
i > 0). (C) Finite
i and 1Lj
an effective inner boundary [both inner and
outer boundaries here correspond to the so-
called m-type boundaries (31)]. This resulted
in two distinct topological sectors for the dimer
coverings, where states belonging to different
sectors can be transformed into each other only
through large X loops that enclose the hole,
constituting a highly nonlocal process (involv-
ing at least a 16-atom resonance) (fig. S13). We
i as the
define the logical states 0Lj
superpositions of all dimer coverings from the
topological sectors 0 and 1, respectively. One
can define (23) the logical operator sz
L as being
proportional to any ZL string operator that
connects the hole with the outer boundary,
given that these have a well-defined eigen-
value ±1 for all dimer states in the same sector
but opposite for the two sectors. The logical
sx
L is instead proportional to XL, which is any
X loop around the hole. This operator anti-
i e
commutes with ZL and has eigenstates þj
p
ffiffiffi
Þ=
and (cid:2)j i e 0Lj
ð
0Lj
.
2
We measured ZL and XL on the strings de-
fined in Fig. 5B, inset, following the same
quasi-adiabatic preparation as in Fig. 1D. We
found that in the range of D/W associated
i ¼ 0, and
with the onset of a QSL phase, ZLh
i > 0, indicating that the system is in a
XLh
superposition of the two topological sectors,
with a finite overlap with the þj
i state (Fig.
5B), which is consistent with the symmetric
i (cid:2) 1Lj
i þ 1Lj
ffiffiffi
2
Þ=
p
i
ð
i
h
initial state and the quasi-adiabatic prepara-
tion procedure (31). To further support this
i
conclusion, we evaluated correlations Z1Z2
between hole-to-boundary strings, which are
expected to have the same expectation val-
ues for both topological sectors (Fig. 5A). In
agreement with this prediction, we found that
the correlations between different pairs of
strings have finite expectation values, with
amplitudes decreasing with the distance be-
tween the strings (Fig. 5C) owing to imperfect
state preparation. These measurements rep-
resent the first steps toward initialization
and measurement of a topological qubit in
our system.
Discussion and outlook
It is not possible to classically simulate quan-
tum dynamics for the full experimental sys-
tem, so we compare our results with several
theoretical approaches. First, our observations
qualitatively disagree with the ground-state
phase diagram obtained from density-matrix-
renormalization-group (DMRG) (35, 36) sim-
ulations on infinitely long cylinders. For the
largest accessible system sizes, including van
der Waals interactions only up to intermediate
distances (~4a), we found a ℤ2 spin liquid in
the ground state (fig. S15). However, unlike in
deformed lattices (23), longer-range couplings
destabilize the spin liquid in the ground state
of the Hamiltonian (Eq. 1) on the specific ruby
lattice used in the experiment, leading to a
direct first-order transition from the disordered
phase to the VBS phase (figs. S15 and S16). By
contrast, we experimentally observed the onset
of the QSL phase in a relatively large parameter
range, and no signatures of a VBS phase were
detected.
To develop additional insight, we performed
time-dependent DMRG calculations (35–37)
that simulated the same state preparation pro-
tocol as in the experiment on an infinitely long
cylinder with a seven-atom-long circumference
(31). The results of these simulations are in
good qualitative agreement with our exper-
imental observations (fig. S19). Specifically,
similar to the results in Fig. 4, in the region
D
W ∼ 3:5 to 4:5 we found nonzero signals for
closed Z and X loops, which cannot be fac-
torized into open strings (fig. S19). Consistent
with experimental observations, these indi-
cate the onset of spin liquid correlations. In
addition, exact diagonalization studies of
a simplified blockade model reveal how the
dynamical state preparation creates an ap-
proximate equal-weight and equal-phase super-
position of many dimer states, instead of the
VBS ground state (31). We conclude that quasi-
adiabatic state preparation occurring over a
few microseconds is insensitive to longer-
range couplings and generates states that
retain the QSL character (31). Although this
phenomenon deserves further theoretical
studies, these considerations indicate the
creation of a metastable state with key char-
acteristic properties of a QSL.
Our experiments offer detailed insights into
elusive topological quantum matter. These
studies can be extended along a number of
directions, including improvement of the
robustness of the QSL by using modified lat-
tice geometries and boundaries (22, 23) as
well as optimization of the state preparation
to minimize quasiparticle excitations; under-
standing and mitigation of environmental
effects associated, for example, with dephasing
and spontaneous emission (31); and optimiza-
tion of string operator measurements by using
quasi-local transformations (38), potentially
with the help of quantum algorithms (39). At
the same time, hardware-efficient techniques
for robust manipulation and braiding of topo-
logical qubits can be explored. Furthermore,
methods for anyon trapping and annealing
can be investigated, with eventual applications
toward fault-tolerant quantum information
processing (40). With improved program-
mability and control, a broader class of topo-
logical quantum matter and lattice gauge
theories can be efficiently implemented (41, 42),
opening the door to their detailed explora-
tion under controlled experimental conditions
and providing a route for the design of quan-
tum materials that can supplement exactly
Semeghini et al., Science 374, 1242–1247 (2021)
3 December 2021
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
solvable models (3, 43) and classical numer-
ical methods (35, 36).
Note added in proof: During the completion
of this manuscript, we became aware of re-
lated work demonstrating the preparation of
toric code states by using quantum circuits on
a superconducting processor (44).
RE FE RENCES AND N OT ES
1. X.-G. Wen, Rev. Mod. Phys. 89, 041004 (2017).
2. S. Sachdev, Rep. Prog. Phys. 82, 014001 (2019).
3. A. Kitaev, Ann. Phys. 303, 2–30 (2003).
4. C. Nayak, S. H. Simon, A. Stern, M. Freedman, S. Das Sarma,
Rev. Mod. Phys. 80, 1083–1159 (2008).
5. L. Savary, L. Balents, Rep. Prog. Phys. 80, 016502 (2017).
6. D. S. Rokhsar, S. A. Kivelson, Phys. Rev. Lett. 61, 2376–2379 (1988).
7. N. Read, S. Sachdev, Phys. Rev. Lett. 66, 1773–1776 (1991).
8. S. Sachdev, Phys. Rev. B Condens. Matter 45, 12377–12396 (1992).
9. R. Moessner, S. L. Sondhi, Phys. Rev. Lett. 86, 1881–1884 (2001).
10. G. Misguich, D. Serban, V. Pasquier, Phys. Rev. Lett. 89,
137202 (2002).
11. B. I. Halperin, J. K. Jain, Fractional Quantum Hall Effects
(World Scientific, 2020).
12. T.-H. Han et al., Nature 492, 406–410 (2012).
13. M. Fu, T. Imai, T.-H. Han, Y. S. Lee, Science 350, 655–658 (2015).
14. C. Gross, I. Bloch, Science 357, 995–1001 (2017).
15. H. Weimer, M. Müller, I. Lesanovsky, P. Zoller, H. P. Büchler,
Nat. Phys. 6, 382–388 (2010).
16. M. Hermele, V. Gurarie, A. M. Rey, Phys. Rev. Lett. 103, 135301
(2009).
17. N. Y. Yao et al., Phys. Rev. Lett. 110, 185302 (2013).
18. A. W. Glaetzle et al., Phys. Rev. X 4, 041037 (2014).
19. A. Celi et al., Phys. Rev. X 10, 021057 (2020).
20. S. de Léséleuc et al., Science 365, 775–780 (2019).
21. T. Ozawa et al., Rev. Mod. Phys. 91, 015006 (2019).
22. R. Samajdar, W. W. Ho, H. Pichler, M. D. Lukin, S. Sachdev,
Proc. Natl. Acad. Sci. U.S.A. 118, e2015785118 (2021).
23. R. Verresen, M. D. Lukin, A. Vishwanath, Phys. Rev. X 11,
031005 (2021).
24. C. Song et al., Phys. Rev. Lett. 121, 030502 (2018).
25. C. K. Andersen et al., Nat. Phys. 16, 875–880 (2020).
26. S. Zhou, D. Green, E. D. Dahl, C. Chamon, Phys. Rev. B 104,
L081107 (2021).
27. M. Saffman, T. G. Walker, K. Mølmer, Rev. Mod. Phys. 82,
2313–2363 (2010).
28. P. Anderson, Mater. Res. Bull. 8, 153–160 (1973).
29. S. Ebadi et al., Nature 595, 227–232 (2021).
30. P. Scholl et al., Nature 595, 233–238 (2021).
31. Materials and methods are available as supplementary materials.
32. K. Fredenhagen, M. Marcu, Commun. Math. Phys. 92, 81–119
(1983).
33. J. Bricmont, J. Frölich, Phys. Lett. B 122, 73–77 (1983).
34. K. Gregor, D. A. Huse, R. Moessner, S. L. Sondhi, New J. Phys.
13, 025009 (2011).
35. S. R. White, Phys. Rev. Lett. 69, 2863–2866 (1992).
36. J. Hauschild, F. Pollmann, SciPost Phys. Lect. Notes p. 5 (2018).
37. M. P. Zaletel, R. S. K. Mong, C. Karrasch, J. E. Moore,
F. Pollmann, Phys. Rev. B Condens. Matter Mater. Phys. 91,
165112 (2015).
38. M. B. Hastings, X.-G. Wen, Phys. Rev. B Condens. Matter Mater.
Phys. 72, 045141 (2005).
39. I. Cong, S. Choi, M. D. Lukin, Nat. Phys. 15, 1273–1278 (2019).
40. E. Dennis, A. Kitaev, A. Landahl, J. Preskill, J. Math. Phys. 43,
4452–4505 (2002).
41. M. C. Bañuls et al., Eur. Phys. J. D 74, 165 (2020).
42. L. Tagliacozzo, A. Celi, A. Zamora, M. Lewenstein, Ann. Phys.
330, 160–191 (2013).
43. A. Kitaev, Ann. Phys. 321, 2–111 (2006).
44. K. J. Satzinger et al., Science 374, 1237–1241 (2021).
45. G. Semeghini et al., Probing topological spin liquids on a
programmable quantum simulator. Harvard Dataverse (2021);
doi:10.7910/DVN/BDCTRX.
46. R. Verresen, rubenverr/DMRG-QSL-Rydberg-experiment.
Zenodo (2021); doi:10.5281/zenodo.5553679.
ACKN OWLED GMEN TS
We thank many members of the Harvard AMO community,
particularly E. Urbach, S. Dakoulas, and J. Doyle for their efforts
enabling operation of our laboratories during 2020-2021. We thank
S. Choi, I. Cong, E. Demler, G. Giudici, W. W. Ho, N. Maskara,
K. Najafi, N. Yao, and S. Yelin for stimulating discussions. Funding:
We acknowledge financial support from the Center for Ultracold
Atoms, the National Science Foundation, the U.S. Department
of Energy (DE-SC0021013 and LBNL QSA Center), the Army
Research Office MURI, the DARPA ONISQ program, QuEra
Computing, and Amazon Web Services. We further acknowledge
support from the Max Planck/Harvard Research Center for
Quantum Optics fellowship (to G.S.), the National Defense Science
and Engineering Graduate (NDSEG) fellowship (to H.L.), Gordon
College (to T.T.W,), the NSF Graduate Research Fellowship
Program (grant DGE1745303) and The Fannie and John Hertz
Foundation (to D.B.), the Harvard Quantum Initiative Postdoctoral
Fellowship in Science and Engineering (to R.V.), the Simons
Collaboration on Ultra-Quantum Matter (Simons Foundation grant
651440 to R.V., A.V., and S.S.). R.S. and S.S. were supported
by the U.S. Department of Energy under grant DE-SC0019030.
The DMRG simulations were performed by using the Tensor
Network Python (TeNPy) package developed by J. Hauschild and
F. Pollmann (36) and were run on the FASRC Cannon and Odyssey
clusters supported by the FAS Division of Science Research
Computing Group at Harvard University. Author contributions:
G.S., H.L., A.K., S.E., T.T.W., D.B., and A.O. contributed to building
the experimental setup, performed the measurements, and
analyzed the data. R.V., H.P., and A.V. contributed to developing
methods for detecting QSL correlations, performed numerical
simulations, and contributed to the theoretical interpretation of the
results. M.K. and R.S. contributed to the theoretical interpretation
of the results. All work was supervised by S.S., M.G., V.V., and
M.D.L. All authors discussed the results and contributed to the
manuscript. Competing interests: M.G., V.V., and M.D.L. are
cofounders and shareholders of QuEra Computing. A.K. is CEO and
shareholder of QuEra Computing. A.O. is shareholder of QuEra
Computing. Some of the techniques and methods used in this work
are included in provisional and pending patent applications filed by
Harvard University (US patent application nos. 16/630719, 63/116,321,
and 63/166,165). Data and materials availability: The data and
code are available on Harvard Dataverse (45) and Zenodo (46).
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abi8794
Materials and Methods
Figs. S1 to S19
References (47–56)
8 April 2021; accepted 28 October 2021
10.1126/science.abi8794
Semeghini et al., Science 374, 1242–1247 (2021)
3 December 2021
6 of 6
|
10.1126_science.abj2096
|
RES EARCH
R E S E A R C H A R T I C L E
◥
ATTOSECOND SCIENCE
Attosecond coherent electron motion in
Auger-Meitner decay
Siqi Li1,2†, Taran Driver1,3,4†, Philipp Rosenberger1,3,5,6, Elio G. Champenois3, Joseph Duris1,
Andre Al-Haddad7, Vitali Averbukh4, Jonathan C. T. Barnard4, Nora Berrah8, Christoph Bostedt7,9,
Philip H. Bucksbaum2,3,10, Ryan N. Coffee1,3, Louis F. DiMauro11, Li Fang11,12, Douglas Garratt4,
Averell Gatton1, Zhaoheng Guo1,10, Gregor Hartmann13, Daniel Haxton14, Wolfram Helml15,
Zhirong Huang1,2, Aaron C. LaForge8, Andrei Kamalov1,2,3, Jonas Knurr3, Ming-Fu Lin1,
Alberto A. Lutman1, James P. MacArthur1,2, Jon P. Marangos4, Megan Nantel1,2, Adi Natan3,
Razib Obaid1,8, Jordan T. O’Neal2,3, Niranjan H. Shivaram1,16, Aviad Schori3, Peter Walter1,
Anna Li Wang3,10, Thomas J. A. Wolf1,3, Zhen Zhang1, Matthias F. Kling1,3,5,6,
Agostino Marinelli1,3*, James P. Cryan1,3*
In quantum systems, coherent superpositions of electronic states evolve on ultrafast time scales (few
femtoseconds to attoseconds; 1 attosecond = 0.001 femtoseconds = 10−18 seconds), leading to a time-
dependent charge density. Here we performed time-resolved measurements using attosecond soft
x-ray pulses produced by a free-electron laser, to track the evolution of a coherent core-hole excitation in
nitric oxide. Using an additional circularly polarized infrared laser pulse, we created a clock to time-
resolve the electron dynamics and demonstrated control of the coherent electron motion by tuning
the photon energy of the x-ray pulse. Core-excited states offer a fundamental test bed for studying
coherent electron dynamics in highly excited and strongly correlated matter.
I nterference is a pillar of quantum physics
and a manifestation of one of its most
profound consequences: the wavelike nature
of matter. A quantum system can exist in a
superposition of energy states whose relative
quantum phases progress in time. This behavior
can cause the states to interfere constructively
or destructively as the system evolves, causing
physical observables (e.g., charge density) to
oscillate in time. Such oscillations are known
as quantum beats and have a period of TQB ¼
h=DE, where h is Planck’s constant and DE is
the energetic separation between the states
(1–5). To display a quantum beat, two con-
ditions must be satisfied: First, the quantum
system must be prepared in a superposition
of two or more different energy states that
have a well-defined (or coherent) relationship
between their individual quantum phases, which
remains stable over the beat period between the
relevant phases. Second, the physical observable
must be sensitive to the difference between the
quantum phases of the energy states forming the
coherent superposition.
In this work, we demonstrated the creation
and observation of coherent superpositions of
core-excited states in molecules using atto-
second x-ray pulses. These molecules decayed
nonradiatively via the Auger-Meitner (AM)
mechanism—a multielectron process in which
the core vacancy created by an x-ray pulse is
filled by one electron from a valence orbital,
and another valence electron is emitted to
conserve energy. The AM process is the domi-
nant mechanism for relaxation following x-ray
absorption in most biologically relevant mole-
cules, and in any molecules composed of light
atoms such as carbon, oxygen, and nitrogen.
We demonstrated how coherence in short
x-ray pulses is imprinted on excited electronic
states in x-ray–matter interaction and how
this coherence affects the attosecond evolution
of the excited electronic wave packet. To this
end, we measured the time-dependent AM
yield and found that it was sensitive to the
quantum coherence of the electronic wave
packet, as well as the differences in the excited
state populations. The coherence of the wave
packet was manifested as femtosecond modu-
lations (or quantum beats) in the time-dependent
electron yield. The effect of the wave packet
coherence on the relaxation process could have
implications for a broad class of other ultrafast
experiments in which the need for high tem-
poral resolution necessitates the use of broad-
bandwidth x-ray pulses.
Time-resolved measurements of any correlated
electron interaction (including AM decay) are
challenging because of the extreme time scale
(few to subfemtosecond) on which electron-
electron interactions occur. Previous time-
resolved measurements have extracted a single
parameter (G ) to characterize the decay of a
core-excited system (6–9). In the case of short
excitation or ionization pulses, G corresponds
to the lifetime of the core-excited state, but for
long pulses the extracted decay constant is
altered by interferences with the excitation
process (9–11). Our distinct combination of
short excitation pulses and a sufficiently long
observation window allowed for a direct time-
resolved measurement of the AM emission
process. We measured a quantum beat, dem-
onstrating the creation and observation of
electronic coherence in a core-excited molecu-
lar system. Our technique of mapping coherent
electronic motion to the AM decay profile
offered a distinctive test-bed for studies of
electronic coherence in highly excited and
strongly correlated systems.
Measurement
Our experimental setup is shown in Fig. 1A.
Isolated soft x-ray attosecond pulses from a
free-electron laser (12), tuned near the oxygen
1s → p resonance in nitric oxide (NO) (∼530 to
540 eV), irradiated a gas target in the presence
of a circularly polarized, 2.3 mm, 5 (cid:2) 1012 W/cm2
laser field. The momentum distribution of the
resultant photoelectrons was recorded by a
coaxial velocity map imaging spectrometer
(c-VMI) (13). Interaction with the x-ray pulse
produced electrons from several different
photoionization channels: direct ionization of
nitrogen K-shell electrons, KLL AM emission
resulting from the nitrogen K-shell vacancy,
and resonant oxygen AM emission following
O1s → p excitation. These channels are labeled
in Fig. 1B, which shows the electron momen-
tum distribution recorded without the 2.3-mm
laser field. The 1s → p excitation in nitric oxide
corresponds to the promotion of an oxygen 1s
electron to the degenerate 2p molecular orbital,
which is already partially occupied by an un-
paired valence electron. The resonant AM emis-
sion following this excitation has a dominant
feature corresponding to channels where one
1SLAC National Accelerator Laboratory, Menlo Park, CA, USA. 2Department of Physics, Stanford University, Stanford, CA, USA. 3Stanford PULSE Institute, SLAC National Accelerator Laboratory,
Menlo Park, CA, USA. 4The Blackett Laboratory, Department of Physics, Imperial College London, London, UK. 5Max Planck Institute of Quantum Optics, Garching, Germany. 6Physics
Department, Ludwig-Maximilians-Universität Munich, Garching, Germany. 7Paul Scherrer Institute, Villigen, Switzerland. 8Physics Department, University of Connecticut, Storrs, CT, USA. 9LUXS
Laboratory for Ultrafast X-ray Sciences, École Polytechnique Fédérale de Lausanne, Lausanne, Switzerland. 10Department of Applied Physics, Stanford University, Stanford, CA, USA.
11Department of Physics, The Ohio State University, Columbus, OH, USA. 12Department of Physics, University of Central Florida, Orlando, FL, USA. 13Institut für Physik und CINSaT, Universität
Kassel, Kassel, Germany. 14KLA Corporation, Milpitas, CA, USA. 15Department of Physics, TU Dortmund University, Dortmund, Germany. 16Department of Physics and Astronomy and Purdue
Quantum Science and Engineering Institute, Purdue University, West Lafayette, IN, USA.
*Corresponding author: Email: marinelli@slac.stanford.edu (A.M.); jcryan@slac.stanford.edu (J.P.C.)
†These authors contributed equally to this work.
Li et al., Science 375, 285–290 (2022)
21 January 2022
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
of the degenerate 2p electrons participates in
the decay, leading to excited cationic states.
There is a small contribution from the channel
where both 2p electrons participate, resulting
in a2p0 ground configuration of the cation (14).
The circularly polarized laser field maped the
temporal profile of the electron emission on
to the momentum measured at the detector.
When electrons were released from the mole-
cule following interaction with the x-ray pulse,
their trajectory was altered by the presence of
the infrared (IR) laser field, similar to the prin-
ciple of a time-resolving streak camera (15, 16).
This interaction altered (or “streaked”) the final
electron momentum, which was measured at
the detector. In a semi-classical approxima-
tion, the final momentum of an ionized elec-
tron is given by
p→ t → ∞
ð
Þ ¼ p→
→
0 þ eA
t0ð Þ
ð1Þ
→
(cid:3)∞E
t0ð Þ ¼ (cid:3)∫t0
→
where A
L t′ð Þdt′ is the vector po-
tential of the circularly polarized laser field,
EL tð Þ, at the time of ionization t0, e is the charge
of an electron (1 atomic unit), and p→
0 is the
momentum of the electron in the absence of
the IR laser field. All the quantities are ex-
pressed in atomic units.
In our measurement, the temporal duration
of the circularly polarized “streaking” laser
field (∼100 fs) was much longer than the laser
period ( TL ¼ 7:7 fs). This fact implies that
over the time scale of a single laser period, the
(cid:1)
vector potential had nearly constant ampli-
(cid:1)
→
tude ( A
(cid:1)) but a direction that rotated with
constant angular velocity 2p=TL . Thus, Eq. 1
describes how the streaking technique encodes
the temporal evolution of the electron emission
rate onto the electron momentum spectrum: An
(cid:1)
(cid:1)
(cid:1)
Fig. 1. Experimental observation of Auger-Meitner emission. (A) NO gas is
ionized by an attosecond x-ray free-electron laser (XFEL) pulse (∼530 to
540 eV, central photon energy) in the presence of a 2.3-mm circularly
polarized streaking field. The resultant photoelectron momentum distribution
is measured by a coaxial velocity map imaging spectrometer (c-VMI) (13). The
streaking field maps the instantaneous ionization rate onto the measured
photoelectron momentum distribution. (B) Single-color electron momentum
spectrum projected along the axis of the c-VMI in the absence of the
streaking field. Atomic units are denoted here and throughout as “a.u.” We
define px to lie along the x-ray polarization axis. (C) Applying an inverse Abel
transform to this image, we retrieve the electron kinetic energy distribution
(“arb.” denotes arbitrary units). (D) Change in the projected momentum
→
distribution as the direction of the streaking laser vector potential (A
0,
light-gray line pointing along 0 fs of the stopwatch face) is varied. The projected
momentum distribution is presented as a difference image where the
electron momentum shown in (B) is used as a background. To observe the
temporal evolution of AM emission, we monitor the AM yield in a small (15°)
region of the detector [black box shown in the panels of (D); energetic
position also shown in pale red in (C)]. The time dependence of this yield
is shown in black dots in (E) (dashed red line shows trace with high-frequency
noise filter applied; see supplementary materials for further details).
The AM yield in (E) is plotted as a function of angle between the streaking
laser vector potential at the time of ionization and the angle of the detection
→
box, which is shown as a gray shaded area in (D). E
shows the direction of
rotation of the electric field. The red error bars have a total length of four
times the SEM of the measured electron yield, T2s(cid:1)x.
Li et al., Science 375, 285–290 (2022)
21 January 2022
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
electron emitted at ti will experience a mo-
→
tið Þ. Because
mentum shift in the direction of A
the period TL of the circularly polarized laser is
well known, if two photoemission features are
found to have momentum shifts that differ by
an amount Dq, this difference implies that
the photoemission events were separated by
a time Dt:
Dt ¼
Dq
2p
(cid:2) TL
ð2Þ
This mapping of angle-to-time resembles
the face of a clock, which has led to the term
“attoclock” being used to describe this type of
time-resolved measurement (17–19).
Our method for extracting the temporal pro-
file of the AM electron yield is illustrated in
Fig. 1, D and E. Figure 1D shows the var-
iation in the differential electron yield for
measurements with three different x-ray ar-
rival times, or directions of the streaking laser
→
vector potential, A
0 . The differential images
show the difference between the averaged
electron image when the vector potential of
the IR laser was chosen to lie along the line
→
0 on the figure (and labeled as “0 fs”
labeled A
on the clock face) and the averaged electron
image where the IR laser was intentionally
mistimed with the x-rays to ensure that there
was no effect from the streaking field. To
extract the time-dependent emission rate of
resonant AM electrons, we monitored a small
angular region on the detector (black wedge
in Fig. 1D) and plot this yield as a function
of streak angle, or the angle between the
observation bin and the streaking laser vector
potential, in Fig. 1E. The observation region
was chosen to be slightly higher in momen-
tum than the center of the field-free resonant
emission spectrum shown in Fig. 1B. The elec-
tron yield in this radial bin therefore mapped
to the number of electrons released into the
continuum at the time the vector potential
→
0 tð Þ is pointed in the angular direction of
A
the observation region. The lower momentum
Fig. 2. Model for Auger-Meitner emission. (A) Schematic representation of
the model used for AM emission. Subfemtosecond x-ray pulses coherently
excite four resonances (labeled 2Sþ;2S(cid:3), and a doubly degenerate 2D).
In addition to the resonant pathway, electrons can be directly ionized by the
x-ray pulse, leading to interfering paths from the ground state to the field-
dressed continuum [although the direct ionization pathway is a minor channel
(14)]. (B) Calculated photoelectron momentum spectrum for 0:5-fs x-ray
pulses centered at 533-eV photon energy in the presence of a 2:3-mm laser
field. The blue arrow shows the direction of IR laser vector potential, A(t), at
the x-ray arrival time, t0. (C) Kinetic energy distribution of the continuum
electron as a function of time in the absence of the streaking laser field.
(D) Time-dependent ionization rate for this wave function, summed over electron
kinetic energy, for a central photon energy of 533 eV (red), 534:5 eV (green),
and 536 eV (blue). (E) Total population of the core-excited states as a
function of time delay for the same photon energies as in (D). (F) Time
evolution of the electron density of the bound electronic states. The three-
dimensional contour is drawn at 20% of the maximum electron density,
and its transparency represents the overall bound-state population, which
decays via AM emission. The blue and red dots in the rightmost panel show
the positions of the nitrogen and oxygen atoms, respectively.
Li et al., Science 375, 285–290 (2022)
21 January 2022
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
limit of this region is ∼6.4 atomic units, and
the upper limit extends to include all electrons
at higher momenta. Equation 2 can be used to
convert the streak angle into a time delay, and
this value is used to label the clock face in Fig.
1D and the lower horizontal axis in Fig. 1E.
At the Linac Coherent Light Source (LCLS),
the synchronization of the streaking laser and
x-ray pulse has a jitter of roughly ∼500 fs (20),
which is orders of magnitude below the re-
quired precision for directly timing the AM
process. Thus, to produce the images shown in
Fig. 1D, we must use a single-shot diagnostic of
the relative arrival time between the x-rays
and laser pulse. As described above, in addi-
tion to driving resonant excitation near the
oxygen K-edge, the attosecond x-ray pulse
ionized electrons from the nitrogen K-shell
of the NO molecule (see Fig. 1, B and C). This
direct photoionization process produced high
energy (∼120 eV) electrons. The photoionization
delay between the arrival of the x-ray pulse and
the appearance of these fast photoelectrons in
the continuum was negligibly small (≲ 5 as)
compared to the streaking laser period TL of
7.7 fs (21–23). Therefore, the momentum shift
observed for the nitrogen K-shell photoemission
feature provided an accurate, single-shot mea-
surement of the direction of the streaking laser
→
vector potential A
0 at the time of arrival of
the x-ray pulse.
We monitor the AM yield in a small angular
region of the detector to avoid introducing
artifacts in the extracted time-dependent trace
due to angular anisotropy in the AM emission
(24). The period of the streaking field was
chosen to be longer than the dominant time
scale of the AM process. This fact simplifies
interpretation of the streaking measurement
by limiting the effect of “wrapping,” where
electrons released into the continuum at time
t and t þ TL experience a similar momentum
kick from the streaking field.
The time-dependent electron yield shown
in Fig. 1E shows a maximum at t = 0, when
→
A
0 was directed along the detection direction
and the the core-excited population (and AM
emission rate) was at a maximum. In addition
to an exponentially decaying electron emis-
sion rate, we observed a revival in the time-
dependent emission rate at t = 3.5 fs.
Model
We modeled our measurement according to
the theory of attosecond streaking of multiple
Fano resonances described by Wickenhauser et al.
(25, 26). Our model, illustrated in Fig. 2A, in-
cluded a ground state that is doubly degenerate
and was resonantly coupled to four bound
states, one of which (2D) is also doubly de-
generate, and thus is labeled as a single state
in the figure. These bound states were also
coupled to a single, structure-less continuum,
which was dressed by the circularly polarized,
2:3(cid:3)mm streaking laser field. The coupling
between the bound and continuum states was
the result of electron correlation interactions
and drove the AM decay process. The bound
states had excitation energies of 531:5 eV (2S(cid:3)),
532:6 eV (2D), and 533:5 eV (2Sþ ), which re-
presented the core-excitation spectrum of nitric
oxide (27). The continuum coupling constant
Fig. 3. Comparison between model and experimental results for resonant
Auger-Meitner emission. (A) Measurement of time-resolved Auger-Meitner
emission from core-excited NO. The left panel shows the experimentally
measured time-dependent AM yield for the various central XFEL photon energies
(black dots). Colored error bars have a total length of four times the SEM
of the measured electron yield, T2s(cid:1)x. This measurement is compared with the
results of the model shown in Fig. 2 (solid colored lines). The right panel shows
total electron yield, which decreases as the central photon energy moves away
from the center of the 1 s → p resonance (bars). The time-dependent yields
change by a factor of 2 between the minimum (normalized to 0) and maximum
(normalized to 1) values. The coherent bandwidth of the attosecond XFEL pulse
spans ∼5 eV, as illustrated by a Gaussian curve of equivalent full width at half
maximum at each central photon energy. The black line shows the O1s → p
feature reported in (27), comprising the 2S(cid:3), 2D, and 2Sþ electronic states. The
revival at t ∼ 3.5 fs, marked by the black vertical arrow, is due to the rephasing
(constructive interference) of the AM emission from the core-excited states
(2S and 2D) populated by the x-ray pulse. The coherent revival is suppressed as
the photon energy moves above the 1 s → p resonance and the contribution
from the direct photoionization channel increases. The photon energy–
dependence of the quantum beat is shown in the magnified image in (B), for
experiment (left) and simulation (right). The shaded area represents the streak-
angle–dependent yield with corresponding error bar T2s(cid:1)x, and the solid line
shows this electron yield after application of a high-frequency filter along the
time axis (see supplementary materials for further details). The color of the
curves corresponds to the central photon energies shown on the left side of (A).
(C) Comparison between two different models where core-excited states are
populated coherently (deep red) and incoherently (pale red) at 533-eV central
photon energy. The experimental measurement is shown in black dots with error
bars T2s(cid:1)x. Coherent interaction between the core-excited states is required to
account for the measured data.
Li et al., Science 375, 285–290 (2022)
21 January 2022
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
(G ¼ 170 meV) was consistent with previous
x-ray absorption measurements (27). The rela-
tive amplitude between transitions to the
bound, core-excited states and direct photo-
ionization of valence electrons to the continuum
was represented by the Fano parameter, qi (see
supplementary materials) (28). We choose the
value for qi according to the measured absorp-
tion spectrum of NO (27).
The coherent bandwidth of the exciting x-ray
source was ∼5 eV (12), which was sufficient to
span all core-excited bound states in the model.
Symmetry constraints did not allow for the
coherent population of the 2S(cid:3) and 2Sþ states
because the 2S(cid:3) and 2Sþ states each coupled to
a different component of the doubly degenerate
ground state (14). Moreover, each component of
the ground state coupled to a different compo-
nent of the degenerate 2D state (14). Thus, the
model only included coherence between the
2S(cid:3) and one of the 2D states and the 2Sþ and
the other 2D state, but not the 2S(cid:3) and 2Sþ
states. In both simulation and experiment, we
tune the central wavelength of the x-ray source
across the 1s → 2pp resonance (red, green,
and blue shaded curves in Fig. 3, bandwidth
drawn to scale with energetic separation of
core-excited states).
In the simulation, we could calculate the
energy-resolved continuum wave function in
the absence of the streaking-field, shown in
Fig. 2D, demonstrating the build-up of reso-
nant features. The rate of electron emission
(integrated over electron kinetic energy) is
shown in Fig. 2E), and we clearly observe an
oscillatory emission rate. Finally, in Fig. 2F
we show the population of each core-excited
state as a function of time, which again shows
oscillatory behavior. The periodic modulation
of the electron emission rate resulted from
the coherent population of the two pairs of
excited states 2S(cid:3) and 2D, and 2D and 2Sþ .
Electronic coherence between the pairs of
excited states resulted in consecutive minima
or maxima in the time-dependent ionization
rate, owing to destructive or constructive in-
terference between emission from the core-
excited states. Because the core-excited wave
packet consisted of states with different an-
gular momentum projections along the mo-
lecular axis, the excited state wave packet
produced an excited electron density that
rotated around the molecular axis, as shown
in Fig. 2C.
Results
We directly modeled our experimental observ-
able by computing the asymptotic (t→∞) mo-
mentum distribution of ionized electrons within
the strong-field approximation (SFA) (26) (Fig.
2B) and performing the same analysis routine
as the one we applied to the experimental
data. The asymmetry parameters describing
emission from the oxygen 1s → 2S(cid:3), O 1s → 2D,
and O 1s → 2Sþ excitations were expected to be
different for each of the electronic states (24)
and have not previously been measured, mean-
ing the contribution of each channel to emis-
sion in the direction of our observation window
was not well defined. We fit the simulation to
the experimental data using the lower kine-
tic energy limits of the small detector region
defined in Fig. 1E, and the relative contribu-
tion from each decay channel at the precise
region on the detector, as free parameters.
With a separate measurement taken concur-
rent with the presented data, we determined
an error distribution of s = 30° for single-shot
vector potential determination. We accounted
for this experimental error by convolution
of the time-dependent electron yield with a
Gaussian kernel of s = 30°. Further details are
provided in the supplementary materials. We
also accounted for the possibility of a small
systematic error in t0 determination between
experiment and theory, resulting from the
finite temporal profile of the reference nitro-
gen K-shell photoline produced by the atto-
second x-ray pulse (12). We identified an offset
of ∼1.7% of the full detector angle.
Figure 3A shows the vector-potential direction–
dependent electron yield measured at different
x-ray excitation energies (black dots) compared
with the simulated yield (solid line). The tran-
sient revival at t∼3:5 fs resulting from electronic
coherence in the core-excited state is indicated
by the black arrow and is observed in both ex-
periment and simulation. This feature is a
quantum beat, occurring at the moment when
the quantum phases of the coherently excited
2S(cid:3) and 2D excitations realigned. This align-
ment caused constructive interference between
the two core-excited states, and an increase in
AM emission rate. The feature at ∼1.3 fs mea-
sured at central photon energy 536 eV was
possibly due to the temporal build-up of the
Fano interference between the resonant and
direct excitation channels and has been qual-
itatively reproduced in further simulation.
Analysis of the energetic positions of the
Rydberg series converging to the oxygen K-edge
(27) was not consistent with the interpretation
that this modulation was due to further coherent
excitation involving Rydberg states.
Figure 3B shows a magnified image of the
revival feature. By tuning the central x-ray
photon energy away from the center of the
1s → 2pp resonance, we could suppress the
quantum beat in both experiment (left) and
simulation (right), demonstrating control over
the coherent evolution of the core-excited states.
The beat was suppressed at higher photon en-
ergy because of an increased relative contri-
bution from the direct channel versus the
coherently excited resonant decay pathways.
In Fig. 3C, we compare our measurement, for
a central x-ray excitation energy of 533 eV, to
our simulation, including (deep red) and ex-
cluding (pale red) the coherence between the
core excited states. As expected, the revival
feature could be reproduced only by including
the coherence between the different core-excited
electronic states. The result from incoherent
summation of the AM emission from the dif-
ferent core-excited states failed to reproduce
the feature at 150° streaking angle.
Conclusion
This work reports the real-time measurement
of electronic coherence in the temporal evolution
of a core-excited molecule. Electronic coherence
imparted a modulation in the time-dependent
emission rate of AM electrons, driven by an
isolated attosecond soft x-ray pulse from a free-
electron laser. The AM emission occurred on a
few-femtosecond time scale, and we time-
resolved it using angular streaking. Our mea-
surement provides a testbed for exploring the
effect of electronic coherence in the photo-
excitation dynamics and subsequent photo-
chemical behavior of molecular systems. The
existence of this electronic coherence provides
the opportunity to explore interatomic site
electronic wave packet coupling, which can
reveal interactions between different parts of
an extended system (29–31). Measuring this
coupling can reveal important information on
the system’s fundamental physical properties
(32, 33). For example, the spectral makeup of
the observed modulations provides rich infor-
mation on the composition of the excited super-
position state. This information opens the
possibility for resolving in time the evolution
and decay of coherent electronic states, as they
evolve and couple to subsequent nuclear mo-
tion in the first stages of a photochemical
reaction (34–37).
REFERENCES AND NOTES
1. A. T. Forrester, R. A. Gudmundsen, P. O. Johnson, Phys. Rev.
99, 1691–1700 (1955).
2. E. Alexandrov, Optika i Spektrosk. 17, 957 (1964).
3. T. Hadeishi, W. A. Nierenberg, Phys. Rev. Lett. 14, 891–892
(1965).
J. Mauritsson et al., Phys. Rev. Lett. 105, 053001 (2010).
4.
5. E. Goulielmakis et al., Nature 466, 739–743 (2010).
6. M. Drescher et al., Nature 419, 803–807 (2002).
7. M. Uiberacker et al., Nature 446, 627–632 (2007).
8. A. J. Verhoef et al., Laser Phys. 21, 1270–1274 (2011).
9. D. Haynes et al., Nat. Phys. 17, 512–518 (2021).
10. O. Smirnova, V. S. Yakovlev, A. Scrinzi, Phys. Rev. Lett. 91,
253001 (2003).
11. A. K. Kazansky, I. P. Sazhina, N. M. Kabachnik, J. Phys. At. Mol.
Opt. Phys. 42, 245601 (2009) Publisher: IOP Publishing.
12. J. Duris et al., Nat. Photonics 14, 30–36 (2020).
13. S. Li et al., AIP Adv. 8, 115308 (2018) Publisher: American
Institute of Physics.
14. H. Wang et al., Chem. Phys. 289, 31–44 (2003).
15. D. Bradley, B. Liddy, W. Sleat, Opt. Commun. 2, 391–395
(1971).
16. Y. Tsuchiya, IEEE J. Quantum Electron. 20, 1516–1528
(1984) Conference Name: IEEE Journal of Quantum
Electronics.
17. J. Itatani et al., Phys. Rev. Lett. 88, 173903 (2002).
18. N. Hartmann et al., Nat. Photonics 12, 215–220 (2018).
19. P. Eckle et al., Science 322, 1525–1529 (2008).
20. J. M. Glownia et al., Opt. Express 18, 17620–17630 (2010).
Li et al., Science 375, 285–290 (2022)
21 January 2022
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
21. J. M. Dahlström, A. L’Huillier, A. Maquet, J. Phys. At. Mol. Opt.
Phys. 45, 183001 (2012).
22. J. M. Dahlström et al., Chem. Phys. 414, 53–64 (2013).
23. V. V. Serov, V. L. Derbov, T. A. Sergeeva, Phys. Rev. A 87,
37. A. Marciniak et al., Nat. Commun. 10, 337 (2019).
38. T. Driver, S. Li, A. Marinelli, J. Cryan, Attosecond coherent
electron motion in Auger-Meitner decay, Zenodo (2022);
https://doi.org/10.5281/zenodo.5711217.
063414 (2013).
24. P. V. Demekhin et al., J. Phys. At. Mol. Opt. Phys. 43, 165103
(2010).
25. M. Wickenhauser, J. Burgdörfer, F. Krausz, M. Drescher,
J. Mod. Opt. 53, 247–257 (2006).
26. M. Wickenhauser, J. Burgdörfer, F. Krausz, M. Drescher,
Phys. Rev. Lett. 94, 023002 (2005).
27. R. Püttner et al., Phys. Rev. A 59, 3415–3423 (1999).
28. U. Fano, Phys. Rev. 124, 1866–1878 (1961).
29. L. S. Cederbaum, J. Zobeley, Chem. Phys. Lett. 307, 205–210
(1999).
30. F. Calegari et al., Science 346, 336–339 (2014).
31. P. M. Kraus et al., Science 350, 790–795 (2015).
32. L. S. Cederbaum, W. Domcke, J. Schirmer, W. V. Niessen, in
Advances in Chemical Physics, I. Prigogine, S. A. Rice, Eds.
(Wiley, 1986), pp. 115–159; https://doi.org/10.1002/
9780470142899.ch3.
33. A. I. Kuleff, L. S. Cederbaum, J. Phys. At. Mol. Opt. Phys. 47,
124002 (2014).
34. F. Lépine, M. Y. Ivanov, M. J. J. Vrakking, Nat. Photonics 8,
195–204 (2014).
35. M. Vacher, M. J. Bearpark, M. A. Robb, J. P. Malhado,
Phys. Rev. Lett. 118, 083001 (2017).
36. V. Despré, N. V. Golubev, A. I. Kuleff, Phys. Rev. Lett. 121,
203002 (2018).
ACKN OWLED GMEN TS
Funding: S.L., Z. Z., and A.M. acknowledge support from US
Department of Energy (DOE), BES Scientific User Facilities Division
Field Work Proposal 100317; J.D. and A. M. were supported by
the Laboratory Directed Research and Development Program in
support of the Panofsky fellowship. The contributions from T.D.,
P.H.B., A.K., A.N., J.T.O., T.J.A.W., A.L.W., and J.P.C. were
supported by the US DOE, Office of Science, Office of Basic Energy
Sciences (BES), Chemical Sciences, Geosciences, and Biosciences
Division (CSGB); E.G.C. was supported by the DOE Laboratory
Directed Research and Development program at SLAC National
Accelerator Laboratory, under contract DE-AC02-76SF00515.
P.R. and M.F.K. acknowledge support by the German Research
Foundation via KL-1439/10, and the Fellow program of the Max
Planck Society. V.A, J.C.T.B., D.G., and J.P.Mar. gratefully acknowledge
funding support from UK EPSRC grants EP/R019509/1, EP/
T006943/1, and EP/I032517/1. N.B., R.O., and A.C.L. acknowledge the
Chemical Sciences, Geosciences and Biosciences Division, US DOE,
Office of Science, BES, grant DE-SC0012376. C.B. acknowledges the
Swiss National Science Foundation and the National Center of
Competence in Research–Molecular Ultrafast Science and Technology
NCCR–MUST. L.F.D. and L.F. acknowledge support from NSF grant
1605042 and DOE DE-FG02-04ER15614. W.H. thanks the German
BMBF for funding of the project “SpeAR_XFEL” under contract
05K19PE1. Use of the Linac Coherent Light Source (LCLS), SLAC
National Accelerator Laboratory, is supported by the US DOE, Office
of Science, BES, under Contract DE-AC02-76SF00515. Author
contributions: S.L., A.M. and J.P.C. devised the experimental scheme.
S.L., A.M., and J.P.C. developed the experimental apparatus. S.L.,
J.D., J.P.Mac., Z.Z., and A.M. prepared the attosecond x-ray pulses.
M.-F.L., N.H.S., and P.W. prepared the experimental beam-line. All
Authors participated in the collection and interpretation of the
experimental data. T.D. led the data analysis. S.L., T.D., P.R., and
E.G.C. worked on the single-shot “streaking” diagnostic. S.L.,
T.D., A.M., and J.P.C. prepared an initial version of the manuscript.
All authors provided critical feedback in preparing the submitted
manuscript. Competing interests: None declared. Data and
materials availability: The partially analyzed raw data and the
raw data from the calculations is available on the Zenodo repository
(38). All (other) data needed to evaluate the conclusions in the paper
are present in the paper or the supplementary materials.
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abj2096
Materials and Methods
Supplementary Text
Figs. S1 to S8
References (39–42)
28 April 2021; accepted 29 November 2021
Published online 6 January 2022
10.1126/science.abj2096
Li et al., Science 375, 285–290 (2022)
21 January 2022
6 of 6
|
10.1126_science.abi7281
|
RES EARCH
R E S E A R C H A R T I C L E
◥
CARBON CAPTURE
A scalable metal-organic framework as a durable
physisorbent for carbon dioxide capture
Jian-Bin Lin1†, Tai T. T. Nguyen2, Ramanathan Vaidhyanathan1,3, Jake Burner4, Jared M. Taylor1,5,
Hana Durekova4, Farid Akhtar6, Roger K. Mah1,5, Omid Ghaffari-Nik7, Stefan Marx8, Nicholas Fylstra1,
Simon S. Iremonger1 , Karl W. Dawson1 , Partha Sarkar2, Pierre Hovington7*, Arvind Rajendran2*,
Tom K. Woo4*, George K. H. Shimizu1,5*
Metal-organic frameworks (MOFs) as solid sorbents for carbon dioxide (CO2) capture face the challenge
of merging efficient capture with economical regeneration in a durable, scalable material. Zinc-based
Calgary Framework 20 (CALF-20) physisorbs CO2 with high capacity but is also selective over water.
Competitive separations on structured CALF-20 show not just preferential CO2 physisorption below
40% relative humidity but also suppression of water sorption by CO2, which was corroborated by
computational modeling. CALF-20 has a low enthalpic regeneration penalty and shows durability to
steam (>450,000 cycles) and wet acid gases. It can be prepared in one step, formed as composite
materials, and its synthesis can be scaled to multikilogram batches.
C apture of CO2 after fossil fuel combustion
requires CO2 removal from a localized
emission source but also regeneration
and recycling of the capture system.
Major challenges for the capture stage
span materials design and development
through to process engineering (1, 2). Flue
gas has a low concentration of CO2 diluted in
mostly N2 along with water and acid gases (3).
Amine and solvent systems (4, 5) rely on con-
tacting flue gas with a liquid that absorbs the
CO2 through a combination of chemical and
physical absorption. Although CO2 removal is
effective, regeneration is energy intensive and
can lead to chemical decomposition (6).
Solid sorbents represent a step-change tech-
nology for carbon capture (7–10) and have been
demonstrated at smaller scales (11). Solids can
bind CO2 through either chemical or physical
sorption (3, 7–10). In most cases, chemisorp-
tive materials have higher capacity and selec-
tivity for CO2 (12). However, factors that enhance
CO2 binding often proportionally increase the
energy needed to regenerate the sorbent and
can enhance binding of competing gases. For
the absolute CO2 uptake, the relevant parameter
1Department of Chemistry, University of Calgary, Calgary,
Alberta, Canada. 2Department of Chemical and Materials
Engineering, University of Alberta, Edmonton, Alberta,
Canada. 3Indian Institute of Science Education and Research,
Dr. Homi Bhabha Road, Pashan, Pune, Maharashtra, 411008,
India. 4Department of Chemistry and Biomolecular Science,
University of Ottawa, Ottawa, Ontario, Canada. 5ZoraMat
Solutions Inc., Calgary, Alberta, Canada. 6Department of
Materials Engineering, Luleå University of Technology, Luleå,
Sweden. 7Svante Inc., Vancouver, British Columbia, Canada.
8BASF SE, Ludwigshafen am Rhein, Germany.
*Corresponding author. Email: phovington@svanteinc.com
(P.H.); arvind.rajendran@ualberta.ca (A.R.); twoo@uottawa.ca
(T.K.W.); gshimizu@ucalgary.ca (G.K.H.S.)
†Present address: C-CART, CREAIT Network, Memorial University of
Newfoundland, St. John’s, Newfoundland and Labrador, Canada.
is working capacity under the operational
cycling conditions to regenerate the solid
sorbent (3). Selectivity over N2 is typically
reported, but sorption of CO2 in the presence
of water vapor is much less reported, espe-
cially for physisorptive capture systems (12–14).
A physisorptive CO2 capture solid would offer
much lower regeneration costs, but it must
have sufficient working capacity and selectiv-
ity in an actual flue stream in which gases are
present with stronger intermolecular attractive
forces than those of CO2. Moreover, to translate
to process productivity, the kinetics of sorption
and release are as important as capacity.
Nearly all classes of porous solids have
potential as solid sorbents for CO2 capture
(1–3, 7–10), including metal-organic frame-
works (MOFs) (2, 3, 9, 12–14), in which chemical
building blocks, pore sizes and shapes, surface
functionalities, and even degrees of order can
be varied to optimize CO2 capture ability. More
robust MOFs (15, 16), including ones that are
stable in the presence of water (17–19) and
steam (20), have been reported, although stabil-
ity to wet acid gases is less common (21–23).
For sorbent powder to be a usable material, it
must be capable of formation in macroscopic
shape for rapid mass transfer and thermal
management, be durable in that form, and be
available at scale (hundreds of thousands of
tonnes) and reasonable cost (24, 25).
Solid sorbents optimized in an adsorption
process have the potential to substantially
decrease the CO2 capture cost compared with
traditional amine absorption processes because
of lower regeneration energy, less chemical de-
composition versus the solvent capture system,
extensive use of stainless-steel owing to the
corrosivity of amine solvents, and large plant
footprint (4–6). Optimization of the solid sorbent
process must include high volumetric product-
ivity in the presence of water (present in the
flue gas) and the lowest regeneration energy.
For regeneration, several processes are under
evaluation, including vacuum swing, pressure
swing, and temperature swing (26). Although
cycling performance per sorbent volume or
productivity is one of the main drivers of final
CO2 capture cost, there are several other pa-
rameters that affect operating and capital
expenses of CO2 capture. For solid sorbents, un-
like solvent-based absorption, it is not feasible
to continuously replace deactivated sorbents
with fresh ones.
Here we present Calgary Framework 20
(CALF-20), a MOF with high capacity and
selectivity for CO2 despite a physisorptive
mechanism and modest heat of adsorption.
Its selectivity extends beyond N2 to capture
CO2 in a wet gas. CALF-20 is exceptionally robust
and stable to steam, wet acid gases, and even
prolonged exposure to direct flue gas from
natural gas combustion. Its single-step syn-
thesis from commercially available compo-
nents is highly scalable. The origin of the CO2
philicity, despite CALF-20 being highly water
resistant, was studied by simulation. Structur-
ing of CALF-20 was performed, as well as
competitive breakthrough experiments in wet
gas streams that aligned with pure-component
isotherms, heats of adsorption, and molecular
modeling. In particular, not only can CALF-20
physisorb CO2 up to and beyond 40% RH, but
the presence of CO2 actually suppresses water
sorption. Finally, we present durability and
CO2 capture data on the MOF that are based
on industrial testing.
Synthesis, structure, and gas sorption
CALF-20, [Zn2(1,2,4-triazolate)2(oxalate)], was
initially prepared solvothermally and single
crystals obtained through the in situ degra-
dation of a dihydroxybenzoquinone derivative
(see supplementary materials). CALF-20 is com-
posed of layers of 1,2,4-triazolate-bridged zinc(II)
ions pillared by oxalate ions to form a three-
dimensional (3D) lattice and 3D pore structure
(Fig. 1, A to C). Channels of 2.73 Å by 2.91 Å,
1.94 Å by 3.11 Å, and 2.74 Å by 3.04 Å along
[100], [011], and [0(cid:1)1 1], respectively (factoring
van der Waals radii), that permeate the MOF
result in a ~38% void volume. The one crystal-
lographically unique Zn center is five-coordinate
with a distorted trigonal bipyramidal geometry
[Zn-O = 2.022(2), 2.189(3) Å; Zn-N = 2.007(2),
2.016(3), 2.091 (3) Å]. The N atoms in the 1,2
positions of the triazolate bridge Zn dimers are
linked to the next dimer by the N atom in the
4-position. The Zn coordination is completed
by two oxygen atoms of a chelating oxalate
group, and there are no open coordination
sites. The bulk powder shows the same phase
(Fig. 1D). Detailed structural analyses on pillared
zinc triazolates have shown that layers can exist
Lin et al., Science 374, 1464–1469 (2021)
17 December 2021
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Single-crystal structure of CALF-20. (A) View of the two-dimensional zinc triazolate grid. (B) View
orthogonal to (A) showing the pillaring of the zinc triazolate layers by oxalate anions. (C) View of the zinc
coordination sphere (H atoms removed). (D) Powder x-ray pattern simulated from the single-crystal
structure (top) and obtained experimentally.
in different manifestations with varying de-
grees of buckling (27, 28). Indeed, since a
provisional patent application was filed in 2014,
a hydrated form of [Zn2(1,2,4-triazolate)2(oxalate)]
has been reported (29). This structure has the
same connectivity but slightly different unit
cell and pore dimensions. The specific pore
structure affects sorption properties, and
modeling was carried out with our obtained
crystal data.
Gas adsorption experiments were performed
for CO2 and N2 (Fig. 2A). The Langmuir surface
area calculated from the N2 isotherm at 77 K
was 528 m2 g−1, and the uptake for CO2 was
4.07 mmol g−1 at 1.2 bar and 293 K. The zero-
loading heat of adsorption for CO2 was −39 kJ
mol−1 (fig. S6), and the calculated selectivity
for CO2/N2 by ideal adsorbed solution theory
was 230 for a 10:90 CO2/N2 mixture. CALF-20
structured readily as a 20% polysulfone com-
posite and retained the expected porosity (Fig.
2, B and C, and fig. S5). For CO2 capacity and
selectivity over N2 as metrics, there are numerous
other materials with noteworthy performance
(30–36). The water sorption profile of CALF-
20 was unusual in that, for a solid with good
physisorptive capacity for CO2, it exhibited
poor water uptake at low partial pressures
(Fig. 2, D and E). Comparisons to zeolite 13X
(37), as well as two other water-resistant MOFs,
CAU-10 (38) and Al fumarate (39), are included in
Fig. 2 and fig. S16. Moreover, higher-temperature
water isotherms showed that water uptake
decreased more readily at higher temperatures
than did the corresponding CO2 isotherms.
Binding-site modeling
To gain insights into the nature of CO2 bind-
ing in CALF-20 and its unusual water sorption
behavior, we performed atomistic grand canon-
ical Monte Carlo (GCMC) simulations (see sup-
plementary materials). The experimental and
simulated CO2 and N2 isotherms were in excel-
lent agreement (see supplementary materials).
Probability distributions of the guest molecules
within the MOF allowed us to identify binding
sites. The most probable CO2 binding, which lies
in the middle of the CALF-20 pore (Fig. 3A), had
a binding energy of −34.5 kJ mol−1 based on the
GCMC force field; the density functional theory
(DFT) value with dispersion corrections was
−36.5 kJ mol−1. The interatomic distances dis-
played were consistent with physisorption; the
shortest distance was 3.03 Å between the CO2
oxygen and a hydrogen of the triazole (fig. S8).
Analysis of the binding energy revealed that
the CO2-MOF interaction was dominated by
attractive dispersion interactions (85%), with
electrostatics contributing the balance.
Water adsorption isotherms are more chal-
lenging to simulate given the polar nature of
water, which enables potentially strong inter-
actions with the framework and with itself.
The experimental water isotherm had a general
S-shape, where the water uptake was initially
low until ~10% relative humidity (RH), at which
point there was a steep rise until ~30% RH
(Fig. 3B). These features indicated that water
condensed in the pores, and they were re-
produced in the simulated isotherm. After
the initial steep rise in water uptake beyond
30% RH, the experimental isotherm showed
a more gradual increase in adsorption until
reaching a saturation limit at ~11 mmol g−1.
However, for the simulated isotherm, the steep
rise continued until full saturation at 40% RH
and then flattened. The general S-shape and
the saturation capacity of ~11 mmol g−1 were
reproduced by the simulation.
A snapshot from the pure water simulation
at 20% RH, where the water uptake was roughly
half the saturation limit (Fig. 3C), revealed that
the pores were either full of water molecules,
forming a hydrogen-bonded network, or com-
pletely empty. In comparison, at 60% RH, where
uptake had fully saturated, all the pores were
full of hydrogen-bonded water molecules (fig.
S9). The equilibrium distribution at 20% RH,
where partially filled pores were not observed,
suggested rapid condensation or evaporation
of water. We extracted the water binding sites
at 20% RH with the highest probability from
the GCMC simulations, and the top three bind-
ing sites, in order, are labeled i, ii, and, iii in
Fig. 3D. The binding energies with the frame-
work of the sites, −17.5, −8.9, and −29.1 kJ mol−1,
respectively, were calculated by placing a single
water molecule in the site with no other guest
molecule present. The two most probable bind-
ing sites had a relatively low binding energy
and were oriented away from the framework
such that there were no hydrogen-bonding
interactions with the oxalate linkers. Water
molecules in these sites were poised to form
hydrogen-bonding interactions with other water
molecules, which suggested that the main driver
for the initial water uptake was the interac-
tion with other water molecules. This result was
consistent with the experimentally observed
water-uptake properties of CALF-20 at low RH.
Breakthrough studies
The intriguing CO2 and water isotherms
prompted a series of dynamic breakthrough
studies (Fig. 4) on the CALF-20–polysulfone
composite (see supplementary materials). Com-
petitive CO2/N2 studies, with CO2/N2 mixtures
of 5/95, 15/85, and 30/70 , respectively (Fig. 4,
A and B), confirmed the selectivity suggested
by the pure-component isotherms. In the N2
profiles, a sharp front, indicating complete
breakthrough of N2, was observed at dimen-
sionless time (ratio of experimental time to
the time taken by a nonadsorbing tracer to
travel through the column) (cid:1)t∼4 in all three
cases. The “roll-up” effect of N2, whereby the
outlet composition of N2 was higher than its
inlet value, until CO2 broke through is clearly
Lin et al., Science 374, 1464–1469 (2021)
17 December 2021
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Equilibrium gas uptake data on pure CALF-20. (A) CO2 and N2 isotherms from 273 to 373 K on
pure CALF-20. (B to D) Structured CALF-20 (80% MOF:20% polysulfone). (B) CO2 isotherms from 303
to 373 K. (C) N2 isotherms from 303 to 353 K. (D) H2O isotherms from 295 K to 373 K. (E) A comparison of
H2O isotherm on zeolite 13X (37), CAU-10 (38), Al fumarate (39) and structured CALF-20 at 295 K. The
isotherms of CO2 and N2 were measured by volumetry, and that of H2O was measured by gravimetry.
visible. The CO2 concentration profiles showed
different breakthrough times for various
CO2/N2 compositions (Fig. 4A). Higher CO2
composition in the feed led to shorter break-
through times
We measured the competitive adsorption of
CO2 and H2O using a combination of gravim-
etry, a study that measured the loading of
CO2 + H2O by subjecting the sample to a moist
stream of CO2 whose RH was controlled, and a
breakthrough experiment that provided the
competitive loading of H2O in the presence of
CO2 (Fig. 4, B and C). The difference between
the total loading from the gravimetry and the
H2O loading from the breakthrough provided the
competitive CO2 loading. Up to a value of 30%
RH, the CO2 loading was nearly unaffected (Fig.
4F), which was unexpected for a physisorptive
material but corroborated by the atomistic sim-
ulations. The CO2 loading gradually decreased
until it became negligible at RH > 80%. Ad-
ditionally, the distinct shift in the H2O isotherm
in the presence of CO2, compared to its pure-
component isotherm, also confirmed the sup-
pression of water sorption by CO2.
To further demonstrate the physisorption
of CO2 by CALF-20 in wet environments, we
measured the water breakthrough curves in
air or CO2 at two different RHs (Fig. 4, D and
E). With the air experiment as a background,
the water breakthrough was actually accelerated
in CO2, providing definitive support for the
physisorptive preference of CALF-20 for CO2
over water below 40% RH. The difference in
water loading, exemplified by the area behind
the breakthrough curve, between the two curves
was pronounced. A comparison of both CO2 and
H2O loading in competitive experiments (Fig. 4F
and figs. S15 and S16) corroborated not only the
sustained CO2 capacity in wet gas but also the
ability to suppress water sorption.
The nature of the water and CO2 binding
from the single-component water simulations
and dry CO2/N2 simulations presented in Fig.
3 was consistent with the preferential binding
of CO2 over water observed at low RHs. Namely,
CO2 has strong binding sites in the center of the
CALF-20 pores that precluded the formation of a
hydrogen-bonded network that was responsible
for the large uptake of water at high RHs. To
corroborate this model, we performed multi-
component simulations of CO2, N2, and water
at varying RH. Figure S10 shows the compar-
ison of simulated water uptake at various RHs
from a single-component water simulation to
that of multicomponent simulations with 0.20
bar of CO2 and 0.80 bar of N2. The results were
in good agreement with the experimental
competitive isotherms shown in Fig. 4F. The
simulations showed that without CO2, water
uptake at 20% RH is 6 mmol/g, whereas in
the presence of CO2, it was negligible. Only at
40% RH did water uptake reach 6 mmol/g when
CO2 was present. Calculated binding energies
of most probable CO2 and H2O binding sites
taken from the multicomponent CO2/N2/H2O
simulations give −17.5 kJ mol−1 for H2O and
−33.5 kJ mol−1 for CO2 (table S6). Calculated
heats of adsorption, at zero loading and high
loading (table S7), suggest that partially water-
loaded pores were more attractive for subse-
quent water sorption than empty pores, and CO2
had a stronger zero-loading heat of adsorption
than water. A binding site analysis of the mixed
CO2/N2/H2O simulations is presented in the
supplementary materials.
The low water-affinity yet CO2-phillic behav-
ior of CALF-20 was enabled by its pore structure.
Although a pore that is ideal for CO2 is, of course,
targeted in carbon capture, it is much less a
focus that a pore be nonideal for water. Notably,
a key feature of CALF-20 was the absence of any
Lin et al., Science 374, 1464–1469 (2021)
17 December 2021
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
strongly interacting functionality with CO2.
Although this property would be expected to
moderate the affinity of the MOF for CO2, the
less specific dispersion interactions cumula-
tively compensate. As previously mentioned,
dispersion interactions account for >85% of
the binding energy in the most favorable
CO2 binding site. The boiling point of H2O is
157°C higher than that of CO2, so it is not
expected that CO2 would preferentially phys-
isorb, but we can connect the competitive
sorption in Fig. 4F with modeling in fig. S10.
Interactions between guest molecules (40, 41),
or in this case, the lack thereof as CO2 blocked
cooperative H2O binding, could tip subtle
balances in binding enthalpies. The pore itself
is the critical element in performing a sorptive
function (42). Other MOFs with low water
affinity such as CAU-10 (38) and Al fumarate
(39) have been reported and studied for CO2
capture from wet gas. These MOFs have good
CO2/N2 selectivity and reasonably low water
affinity, as indicated by stepped water isotherms.
However, CAU-10 loses CO2 capacity above a
RH value of 20%, and aluminum fumarate loses
17% CO2 capacity at 14% RH (fig. S16). However,
CALF-20 has a higher CO2 capacity and retains
it up to and beyond 40% RH. Also, neither
shows the suppression of water sorption by
CO2 that is observed with CALF-20.
Flue gas sorption and scaling
Industrially, materials must absorb CO2 from
postcombustion flue gases at 100°C contain-
ing water vapor and acid gases, and endure
stresses during regeneration as the sorbent
goes through a temperature swing, pressure
swing, or vacuum swing process. CALF-20 has
been run through stability assessments from
multiple academic, government, and industry
partners and shows robust performance, as
confirmed by retention of structure and gas
adsorption properties. The retention of CO2
capacity after being repeatedly heated to dry
air at 150°C in the thermal gravimetric analysis
(Fig. 5A) showed excellent stability (6, 43).
This feature is key to high sorbent lifetime, as
there is residual O2 in the flue gas and during
conditioning of the bed where air can oxidize
reactive groups.
Powder x-ray diffraction (Fig. 5B) and N2
sorption isotherms (Fig. 5C) are shown after a
week of exposure to 150°C steam. CALF-20
was also tested for retention of structure and
porosity (figs. S12 and S13) after treatment with
20 parts per million (ppm) SO2 and 100 ppm
NOx at 20°C in separate experiments. We sub-
jected CALF-20 to a real flue gas stream (50°C,
flow of 100 cm3 min−1) from natural gas com-
bustion containing 7.3% H2O, 7.1% O2, 147 ppm
CO, 78 ppm NO, and 13 ppm NO2 (see sup-
plementary materials, fig. S3, and table S4).
Under these flowing flue gas conditions,
powdered CALF-20 lost only 1.3% of its capac-
Fig. 3. Most probable CO2 binding site determined from the single-component CO2 GCMC simulation
at 0.15 atm. (A) Select distances between heavy atoms of the CALF-20 framework and the CO2 atoms
are highlighted. These are the shortest distance between atoms of the framework and any atom of the
CO2 molecule. (B) Experimental and simulated single-component water isotherms at 293 K. Simulated
adsorption results refer to the values obtained starting from empty pores, whereas the desorption results
refer to the values obtained by starting the simulation with the pores saturated with water. (C) A snapshot
from a 20% RH simulation of water in CALF-20. (D) The three most probable H2O binding sites
determined from the single-component water simulation at 20% RH. For (A), (C), and (D), ball-and-stick
representations are used for the guest, whereas a tube representation is used for the framework. Atom
colors are the same as shown in Fig. 1B.
ity after 6 days, as confirmed by gravimetric
CO2 uptake in a 15/85 mixture of CO2/N2. A
process demonstration unit using the Svante
VeloxoTherm process was built on the basis
of rotating beds and fast cycles (~1 min) at
0.1 tonne per day CO2 capacity, and it was
deployed to test the CALF-20 lifetime with
simulated cement flue gas. This simulated
flue gas was generated by enriching real flue
gas from a natural gas boiler with pure CO2
and air to bring CO2, water, and O2 concen-
trations to cement kiln flue gas composition
(17% CO2, 10% O2, 5% H2O, balance N2, at 45°C).
The gas analyzer recorded around 60 ppm NO
and 12 ppm NO2 in the generated flue gas that
was fed to the CALF-20 beds. The process was
continuously tested for over 2000 hours with
expected key performance indicators and no
appreciable performance loss, as can be seen
in fig. S17 (44). Furthermore, the process was
able to achieve US Department of Energy target
CO2 purity of 95%.
For large-scale applications, it is important
that the scale-up be feasible from an economic
and technical viewpoint (45). Analysis of the
cost driver for different MOF syntheses reveals
that the costs of the raw materials, especially
for the linker and less commonly for the metal,
are often prohibitive. In addition, synthetic
process conditions can have a substantial
impact on the economics—for example, the
necessity of high-pressure equipment is not
only expensive but also results in costly safety
precautions to protect employees and the envi-
ronment. For CALF-20, none of these dis-
advantageous conditions apply (46). The raw
materials are commercially available on a large
scale from qualified vendors. Both linkers
are low-cost bulk chemicals with large global
production capacity (47): 200,000 metric tonnes
per annum (MTPA) for oxalic acid, found mainly
in pharmaceutical, textile and mining industries;
10,000 MTPA for triazoles, used mainly in the
agricultural sector as a building block for azole-
based fungicides. In addition, the reaction could
be carried out in a water/methanol mixture, where
the organic solvent represents <25 wt % with
ongoing improvements. These conditions are
particularly advantageous from a safety and
environmental aspect. Further, in large-scale
Lin et al., Science 374, 1464–1469 (2021)
17 December 2021
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Competitive dynamic column breakthrough (DCB) and equilibrium measurements on
structured CALF-20 at 295 K and 97 kPa. (A) Competitive DCB of CO2 and N2 at different
compositions. (B) Competitive CO2 breakthrough curves measured at various RH values. (C) Competitive
H2O breakthrough curves at various RH values corresponding to the curves shown in (B). (D) A
comparison of breakthrough curves obtained from experiments with Air + H2O and that with
CO2 + H2O at 13% RH. (E) A comparison of breakthrough curves obtained from experiments with
Air + H2O and that with CO2 + H2O at 47% RH. (F) Competitive CO2 loadings (red triangle) and
competitive H2O loadings (blue circle) at various RH values. The loading of pure H2O isotherm
(green square) is shown as a reference. The breakthrough curves are plotted in dimensionless time,
which is the ratio of the actual time to the average retention time taken of a nonadsorbed component.
Also, there is a break in the abscissa of breakthrough curves.
batch synthesis, CALF-20 can obtain an
unusually high solid content (total amount
of dried MOF per total amount of solvents
used) of >35%. The high yield of >90%, the
reasonable reaction time, and the very high
solid content result in an exceptional space-
time yield (STY) for the precipitation step of
550 kg/m3 day. In comparison, the STYs for
zeolites are in the range of 50 to 150 kg/m3
day (48). Critically, the CO2 uptake of CALF-20
was retained through a wide range of scaling
and structuring. Figure 5D shows a 3 million–
fold difference in scale with matching CO2
isotherms.
Outlook
An ideal adsorbent for the postcombustion
CO2 capture should exhibit several properties,
including (i) high CO2 adsorption capacity;
(ii) fast adsorption/desorption kinetics; (iii)
high CO2 selectivity over N2, O2, and ability
to function in wet gas; (iv) mild regeneration
conditions; (v) the ability to be formed into
structures, e.g., beads, laminates, or mono-
liths; (vi) chemical, mechanical, and thermal
stability during adsorption-desorption cycling;
and (vii) low cost and scalability of production.
We have shown that CALF-20 can meet all of
these criteria and help make industrial-scale
CO2 capture cost effective and reliable (44).
Other MOFs have better reported properties
in one or more of the aforementioned criteria,
but not in all of them. For example, most
reported MOFs cannot tolerate even ambi-
ent moisture or steam despite having very
high CO2 capacity or high CO2/N2 selectivity.
The other important factor to consider is cost
and scalability of synthesis. Most MOFs need
aprotic solvents (such as dimethyl formamide
or diethyl formamide) or contain expensive
and noncommercial-grade organic linkers.
With CALF-20, the components are commer-
cially available at low cost and large volume,
and water and methanol are the solvents used
to synthesize this MOF.
In terms of gas separations, there is an
increasing body of evidence showing that
simple metrics such as selectivity and working
capacity correlate poorly with ultimate process
performance (49–53). A recent study that
screened >5000 MOFs (54) showed that
sorbent screening should include detailed
process modeling and optimization. The Svante
VeloxoTherm capture process used direct
steam to rapidly desorb all the captured CO2.
In comparison to a traditional temperature
swing process, the steam regeneration step
of the VeloxoTherm process provided con-
centration swing in addition to heat, which
allowed the extraction of the entire quantity
of physisorbed CO2, through its cyclic working
capacity. Beyond steam stability, key aspects of
the CALF-20 adsorbent synergizing with this
process are its low water affinity and its ability
to rapidly physisorb CO2 in a wet gas, facili-
tating faster cycling, higher productivity, and
ultimately resulting in a smaller plant foot-
print. In the Svante process with CALF-20, less
energy is required to remove moisture in the
drying cycle, and it also bears a higher moisture
tolerance, which allows the capture cycle to
recommence more rapidly.
Although materials can have one or more
exceptional features, the key point is to merge
those properties with process engineering
conditions that best exploit them, such as cap-
ture conditions and available waste energy
or heat for regeneration. The high uptakes at
lower partial pressures of CO2 make CALF-20
Lin et al., Science 374, 1464–1469 (2021)
17 December 2021
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. CALF-20 scalability and stability. (A) Cycling of heating and introduction of CO2 showing
30 cycles heated to 150°C. The left y axis is truncated to show the CO2 mass gain on each cycle. CALF-20
survived more than 450,000 steam treatments in another test, but CO2 uptake was only measured on the
terminal sample. (B) Powder x-ray diffractograms and treatment with steam and running gas sorption shown
in (C) N2 isotherms at 77 K run on steam-treated samples and compared to pristine CALF-20. (D) CO2
isotherms on 3 million–fold different scale batch preparations of CALF-20, showing retention of the CO2
capacity. Comparisons with simulated uptake from the crystal structure and the structured CALF-20 scaled
by a factor of 0.2 to account for 20% polysulfone are also shown.
a more suitable sorbent for a temperature or
concentration swing process or potentially a
pressure or vacuum swing at elevated tem-
peratures. Independent of the regeneration
process, the competitive nature of CO2 and
H2O will enhance sorbent efficiency as coad-
sorption of water is reduced. Factoring its
scalable preparation and durability, CALF-20
should derisk the use of MOFs for large-scale
gas separation in industrial settings, and in par-
ticular, the challenge of postcombustion CO2
capture (55, 56). In terms of carbon capture
and climate change, efficient capture is only a
step, albeit a very important one, in reducing
greenhouse gases. With the integration of better
materials into advanced processes, this derisking
should lead to additional and larger demon-
stration projects for critical testing of MOFs in
CO2 capture and other strategic challenges.
RE FE RENCES AND N OT ES
7. M. Pardakhti et al., ACS Appl. Mater. Interfaces 11,
34533–34559 (2019).
8. L. A. Darunte, K. S. Walton, D. S. Sholl, C. W. Jones, Curr. Opin.
Chem. Eng. 12, 82–90 (2016).
9. C. A. Trickett et al., Nat. Rev. Mater. 2, 17045 (2017).
10. A. E. Creamer, B. Gao, Environ. Sci. Technol. 50, 7276–7289 (2016).
11. J. C. Abanades et al., Int. J. Greenh. Gas Control 40, 126–166
(2015).
12. D. G. Madden et al., Philos. Trans. A, Math. Phys. Eng. Sci. 375,
20160025 (2017).
13. P. G. Boyd et al., Nature 576, 253–256 (2019).
14. E. González-Zamora, I. A. Ibarra, Mater. Chem. Front. 1,
1471–1484 (2017).
15. J. H. Cavka et al., J. Am. Chem. Soc. 130, 13850–13851 (2008).
16. A. J. Howarth et al., Nat. Rev. Mater. 1, 15018 (2016).
17. J. M. Kolle, M. Fayaz, A. Sayari, Chem. Rev. 121, 7280–7345
(2021).
18. S. Yuan et al., Adv. Mater. 30, e1704303 (2018).
19. J. Duan, W. Jin, S. Kitagawa, Coord. Chem. Rev. 332, 48–74 (2017).
20. E. J. Kim et al., Science 369, 392–396 (2020).
21. D. F. Sava Gallis, D. J. Vogel, G. A. Vincent, J. M. Rimsza,
T. M. Nenoff, ACS Appl. Mater. Interfaces 11, 43270–43277 (2019).
22. S. Bhattacharyya et al., Chem. Mater. 30, 4089–4101 (2018).
23. G. W. Peterson, J. J. Mahle, J. B. DeCoste, W. O. Gordon,
J. A. Rossin, Angew. Chem. Int. Ed. 55, 6235–6238 (2016).
24. P. A. Julien, C. Mottillo, T. Friscic, Green Chem. 19, 2729–2747
(2017).
1.
“Accelerating Breakthrough Innovation in Carbon Capture,
Utilization, and Storage” (U.S. Department of Energy, 2017).
2. Z. Hu, Y. Wang, B. B. Shah, D. Zhao, Adv. Sustainable Syst. 3,
25. S. Wang, C. Serre, ACS Sustain. Chem.& Eng. 7, 11911–11927 (2019).
26. M. Bui et al., Energy Environ. Sci. 11, 1062–1176 (2018).
27. R. B. Lin, D. Chen, Y. Y. Lin, J. P. Zhang, X. M. Chen,
1800080 (2018).
3. K. Sumida et al., Chem. Rev. 112, 724–781 (2012).
4. B. Dutcher, M. Fan, A. G. Russell, ACS Appl. Mater. Interfaces 7,
2137–2148 (2015).
5. M. Wang, A. S. Joel, C. Ramshaw, D. Eimer, N. M. Musa,
Appl. Energy 158, 275–291 (2015).
Inorg. Chem. 51, 9950–9955 (2012).
28. Y. Y. Lin, Y. B. Zhang, J. P. Zhang, X. M. Chen, Cryst. Growth Des.
8, 3673–3679 (2008).
29. X.-F. Wei, J. Miao, L.-L. Shi, Synth. React. Inorg. Met.-Org.
Nano-Met. Chem. 46, 365–369 (2016).
30. S. R. Caskey, A. G. Wong-Foy, A. J. Matzger, J. Am. Chem. Soc.
32. P. Nugent et al., Nature 495, 80–84 (2013).
33. T. M. McDonald et al., Nature 519, 303–308 (2015).
34. P. M. Bhatt et al., J. Am. Chem. Soc. 138, 9301–9307 (2016).
35. S. Nandi et al., J. Am. Chem. Soc. 139, 1734–1737 (2017).
36. S. Xiang et al., Nat. Commun. 3, 954 (2012).
37. N. S. Wilkins, J. A. Sawada, A. Rajendran, Adsorption 26,
765–779 (2020).
38. V. B. López-Cervantes et al., Polyhedron 155, 163–169 (2018).
39. J. A. Coelho et al., Ind. Eng. Chem. Res. 55, 2134–2143 (2016).
40. K. S. Walton et al., J. Am. Chem. Soc. 130, 406–407 (2008).
41. R. Vaidhyanathan et al., Science 330, 650–653 (2010).
42. S. Kitagawa, R. Matsuda, Coord. Chem. Rev. 251, 2490–2509
(2007).
43. M. Jahandar Lashaki, S. Khiavi, A. Sayari, Chem. Soc. Rev. 48,
3320–3405 (2019).
44. P. Hovington, O. Ghaffari-Nik, L. Mariac, A. Liu, B. Henkel,
S. Marx, Rapid Cycle Temperature Swing Adsorption Process
Using Solid Structured Sorbent for CO2 capture from Cement
Flue Gas, Proceedings of the 15th Greenhouse Gas Control
Technologies Conference, 15–18 March 2021.
45. M. Rubio-Martinez et al., Chem. Soc. Rev. 46, 3453–3480 (2017).
46. J. M. Taylor, R. K. Mah, G. K. H. Shimizu, Synthesis of Zinc
MOF Materials, PCT/CA2019/050530, filed 24 April 2019.
47. S. N. Bizzari, M. Blagoev, CEH Marketing Research Report,
Chemical Economics Handbook, SRI Consulting, April 2010.
48. A. U. Czaja, N. Trukhan, U. Müller, Chem. Soc. Rev. 38,
1284–1293 (2009).
49. A. K. Rajagopalan, A. M. Avila, A. Rajendran, Int. J. Greenh.
Gas Control 46, 76–85 (2016).
50. J. Park et al., Ind. Eng. Chem. Res. 59, 7097–7108 (2020).
51. K. T. Leperi, Y. G. Chung, F. You, R. Q. Snurr, ACS Sustain.
Chem.& Eng. 7, 11529–11539 (2019).
52. M. Khurana, S. Farooq, Ind. Eng. Chem. Res. 55, 2447–2460
(2016).
53. A. H. Farmahini, S. Krishnamurthy, D. Friedrich, S. Brandani,
L. Sarkisov, Chem. Rev. 121, 10666–10741 (2021).
54. T. D. Burns et al., Environ. Sci. Technol. 54, 4536–4544 (2020).
55. A. Samanta, A. Zhao, G. K. H. Shimizu, P. Sarkar, R. Gupta,
Ind. Eng. Chem. Res. 51, 1438–1463 (2012).
56. R. L. Siegelman, E. J. Kim, J. R. Long, Nat. Mater. 20,
1060–1072 (2021).
AC KNOWLED GME NTS
Funding: This research was undertaken thanks in part to funding
from Alberta Innovates Technology Futures (Strategic Research
Grant), the Natural Sciences and Engineering Research Council
(NSERC) of Canada (CREATE Grant), the US Department of
Energy’s (DOE) office of Fossil Energy (FE) DE- FOA-0001792,
GreenSTEM from Alberta Jobs, Economy, and Innovation, Carbon
Management Canada’s Carbon Capture and Conversion Institute,
MITACS, Innovate Calgary, the Canada First Research Excellence
Fund (Global Research Initiative in Sustainable Low Carbon
Unconventional Resources), and a Parex Innovation Fellowship to
GKHS. We also thank Compute Canada for computing resources.
Author contributions: Methodology and Investigation: J.-B.L.,
T.T.T.N., R.V., J.M.T., N.J.F., R.K.M., O.G.-N., S.S.I., K.W.D., P.S.,
S.M.; Formal Analysis: J.-B.L, T.T.T.N., J.B., H.D., O.G.-N., A.R.,
T.K.W., G.K.H.S.; Funding acquisition/Supervision/Project
administration: A.R., T.K.W., P.H., G.K.H.S.; Writing: J.-B.L, T.T.T.N.,
A.R., T.K.W., and G.K.H.S. wrote the first draft. All authors
contributed to the final draft. Competing interests: Two patents
(CA2904546A1 and EP3784824A1) related to CALF-20 are licensed
to Svante Inc. and ZoraMat Solutions Inc. for different fields of
use. J.-B.L., R.V., R.K.M., J.M.T., S.S.I., K.W.D., and G.K.H.S. receive
royalties from the license. T.T.T.N., H.D., J.B., F.A., S.M., N.J.F.,
P.S., A.R., T.K.W. have no competing interests. Data and materials
availability: The CIF file for CALF-20 is available at the Cambridge
Crystallographic Data Centre with deposition number CCDC
2084733. All other data, excepting the large-scale synthesis of
CALF-20, which is patent-pending, are available in the manuscript
or the supplementary materials. Samples of CALF-20 are available
for data reproduction purposes from BASF/Svante under a
material transfer agreement via P.H. (phovington@svanteinc.com).
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abi7281
Materials and Methods
Figs. S1 to S17
Tables S1 to S7
References (57–74)
6. C. Gouedard, D. Picqa, F. Launay, P.-L. Carrette, Int. J. Greenh.
130, 10870–10871 (2008).
Gas Control 10, 244–270 (2012).
31. J. An, S. J. Geib, N. L. Rosi, J. Am. Chem. Soc. 132, 38–39 (2010).
26 March 2021; accepted 19 October 2021
10.1126/science.abi7281
Lin et al., Science 374, 1464–1469 (2021)
17 December 2021
6 of 6
|
10.1126_science.abo7651
|
RES EARCH
PEPTIDE SEQUENCING
Real-time dynamic single-molecule protein
sequencing on an integrated semiconductor device
Brian D. Reed1*, Michael J. Meyer1, Valentin Abramzon1, Omer Ad1, Pat Adcock1, Faisal R. Ahmad1,
Gün Alppay1, James A. Ball1, James Beach1, Dominique Belhachemi1, Anthony Bellofiore1,
Michael Bellos1, Juan Felipe Beltrán1, Andrew Betts1, Mohammad Wadud Bhuiya1,
Kristin Blacklock1, Robert Boer1, David Boisvert1, Norman D. Brault1, Aaron Buxbaum1,
Steve Caprio1, Changhoon Choi1, Thomas D. Christian1, Robert Clancy1, Joseph Clark1,
Thomas Connolly1, Kathren Fink Croce1, Richard Cullen1, Mel Davey1, Jack Davidson1,
Mohamed M. Elshenawy1, Michael Ferrigno1, Daniel Frier1, Saketh Gudipati1, Stephanie Hamill1,
Zhaoyu He1, Sharath Hosali1, Haidong Huang1, Le Huang1, Ali Kabiri1, Gennadiy Kriger1,
Brittany Lathrop1, An Li1, Peter Lim1, Stephen Liu1, Feixiang Luo1, Caixia Lv1, Xiaoxiao Ma1,
Evan McCormack1, Michele Millham1, Roger Nani1, Manjula Pandey1, John Parillo1, Gayatri Patel1,
Douglas H. Pike1, Kyle Preston1, Adeline Pichard-Kostuch2, Kyle Rearick1, Todd Rearick1,
Marco Ribezzi-Crivellari2, Gerard Schmid1, Jonathan Schultz1, Xinghua Shi1, Badri Singh1,
Nikita Srivastava1, Shannon F. Stewman1, T. R. Thurston1, Philip Trioli1, Jennifer Tullman1,
Xin Wang1, Yen-Chih Wang1, Eric A. G. Webster1, Zhizhuo Zhang1, Jorge Zuniga1, Smita S. Patel3,
Andrew D. Griffiths2, Antoine M. van Oijen4, Michael McKenna1,
Matthew D. Dyer1, Jonathan M. Rothberg1
Studies of the proteome would benefit greatly from methods to directly sequence and digitally
quantify proteins and detect posttranslational modifications with single-molecule sensitivity.
Here, we demonstrate single-molecule protein sequencing using a dynamic approach in which single
peptides are probed in real time by a mixture of dye-labeled N-terminal amino acid recognizers
and simultaneously cleaved by aminopeptidases. We annotate amino acids and identify the peptide
sequence by measuring fluorescence intensity, lifetime, and binding kinetics on an integrated
semiconductor chip. Our results demonstrate the kinetic principles that allow recognizers to identify
multiple amino acids in an information-rich manner that enables discrimination of single amino
acid substitutions and posttranslational modifications. With further development, we anticipate
that this approach will offer a sensitive, scalable, and accessible platform for single-molecule
proteomic studies and applications.
M easurements of the proteome provide
deep and valuable insight into bio-
logical processes. However, methods
with higher sensitivity are needed
to fully understand the complex and
dynamic states of the proteome in cells and
changes to the proteome that occur in disease
states and to make this information more
accessible. The complex nature of the pro-
teome and the chemical properties of proteins
present several fundamental challenges to
achieving sensitivity, throughput, cost, and
adoption on par with DNA sequencing tech-
nologies (1, 2). These challenges include the
large number of different proteins per cell
(>10,000) and yet larger number of proteo-
forms (3); the very wide dynamic range of
protein abundance in cells and biological fluids
(4, 5) and lack of correlation with transcript
levels (6); the costs and high detection limits
of current methods of mass spectrometry (2);
1Quantum-Si, Inc., Guilford, CT 06437, USA. 2Laboratoire de
Biochimie, ESPCI Paris, Université PSL, CNRS UMR 8231,
Paris, France. 3Department of Biochemistry and Molecular
Biology, Rutgers University, Piscataway, NJ 08854, USA.
4Molecular Horizons, University of Wollongong, Wollongong,
NSW 2522, Australia.
*Corresponding author. Email: breed@quantum-si.com
and the inability to copy or amplify proteins.
Methods to directly sequence single protein
molecules offer the maximum possible detec-
tion sensitivity, with the potential to enable
single-cell inputs, digital quantification based
on read counts, detection of posttranslational
modifications (PTMs) and low-abundance or
aberrant proteoforms, and cost and throughput
levels that favor broad adoption.
Here, we present a single-molecule protein
sequencing approach and integrated system
for proteomic studies. We immobilize peptides
in nanoscale reaction chambers on a semi-
conductor chip and detect N-terminal amino
acids (NAAs) with dye-labeled NAA recogniz-
ers in real time. Aminopeptidases sequentially
remove individual NAAs to expose subsequent
amino acids for recognition, eliminating the
need for complex chemistry and fluidics (Fig. 1).
We built a benchtop device with a 532-nm
pulsed laser source for fluorescence excitation
and electronics for signal processing (fig. S1A).
Our semiconductor chip uses fluorescence
intensity and lifetime, rather than emission
wavelength, for discrimination of dye labels.
Our recognizers detect one or more types of
NAAs and provide information for peptide
identification based on the temporal order
of NAA recognition and the kinetics of on-off
binding.
A complementary metal-oxide
semiconductor chip and integrated
system for single-molecule measurements
We used complementary metal-oxide semi-
conductor fabrication technology to build a
custom time domain–sensitive semiconductor
chip with nanosecond precision, containing
fully integrated components for single-molecule
detection, including photosensors, optical wave-
guide circuitry, and reaction chambers for
biomolecule immobilization (fig. S1, B and C).
We achieve observation volumes of <5 attoliters
through evanescent illumination at reaction-
chamber bottoms from the nearby waveguide,
enabling sensitive single-molecule detection
in the context of high concentrations (>1 mM)
of freely diffusing dye.
The semiconductor chip uses a filterless sys-
tem that excludes excitation light on the basis
of photon arrival time, achieving >10,000-
fold attenuation of incident excitation light.
Elimination of an integrated optical filter layer
increases the efficiency of fluorescence collec-
tion and enables scalable manufacturing of the
chip. To discriminate fluorescent dye labels
attached to NAA recognizers by fluorescence
lifetime and intensity, the chip rapidly alter-
nates between early and late signal collection
windows associated with each laser pulse,
thereby collecting different portions of the
exponential fluorescence lifetime decay curve.
The relative signal in these collection windows
(termed “bin ratio”) provides a reliable indica-
tion of fluorescence lifetime (fig. S1, D to H,
and materials and methods).
Ordered recognition and cleavage of NAAs on
single peptide molecules in real time
For NAA binding proteins to function as re-
cognizers, the recognizer-peptide complex must
remain bound long enough (typically >120 ms
on average) to generate detectable single-
molecule binding events. We first focused on
proteins from the N-end rule adapter family
ClpS that naturally bind to N-terminal phenyl-
alanine, tyrosine, and tryptophan (7–9). Using
PS610, a recognizer we derived from ClpS2
from Agrobacterium tumefaciens (table S1),
we established that this recognizer binds detect-
ably to immobilized peptides with these NAAs.
We also determined that the kinetics of bind-
ing differ for each NAA. To demonstrate these
properties, we incubated immobilized peptides
containing the initial N-terminal sequences
FAA, YAA, or WAA (A, alanine; F, phenylala-
nine; W, tryptophan; Y, tyrosine) on separate
chips with PS610 and collected data for 10 hours
(see materials and methods). We observed NAA
recognition by PS610, characterized by contin-
uous on-off binding during the incubation pe-
riod, with a distinct pulse duration (PD) for each
Reed et al., Science 378, 186–192 (2022)
14 October 2022
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
Peptides
On-off binding
Proteins
R
L
I
F
A
Reaction
chamber
Semiconductor chip
R
L
I
F
A
R
L
I
F
R
A
L
L
I
A
F
I
I
F
A
F
Recognizers
L I V
F Y W
R
Aminopeptidases
Fig. 1. Overview of real-time dynamic protein sequencing. Protein samples
are digested into peptide fragments, immobilized in nanoscale reaction
chambers, and incubated with a mixture of freely diffusing NAA recognizers
and aminopeptidases that carry out the sequencing process. The labeled
recognizers bind on and off to the peptide when one of their cognate NAAs is
exposed at the N terminus, thereby producing characteristic pulsing
patterns. The NAA is cleaved by an aminopeptidase, exposing the next amino
acid for recognition. The temporal order of NAA recognition and the kinetics
of binding enable peptide identification and are sensitive to features that
modulate binding kinetics, such as PTMs.
Fig. 2. NAA recognition and dynamic sequencing. (A to C) Example traces
demonstrating single-molecule N-terminal recognition by PS610 (A), PS961
(B), and PS691 (C). Scatterplots of the number of pulses per RS versus RS
mean PD are displayed for each peptide in (A) to (C), with median PD
indicated. (D) Example traces from dynamic sequencing of the synthetic
peptide FAAWAAYAADDD. Median PD is indicated above each RS.
(E to G) Dynamic sequencing of the synthetic peptide LAQFASIAAYASDDD
using PS610 and PS961. Example traces are shown in (E). A scatterplot
of RS mean PD versus bin ratio illustrating discrimination of recognizers by
bin ratio and NAAs by PD is shown in (F). A scatterplot of the number
of pulses per RS versus RS mean PD, grouped by the amino acid label
assigned to the RS, is shown in (G).
peptide (Fig. 2A). Median PDs were 2.49, 0.73,
and 0.31 s for FAA, YAA, and WAA, respective-
ly. These values reflect differences in binding
affinity driven by different dissociation rates
for each type of protein-NAA interaction (7)
(fig. S2, A and B).
To expand the set of recognizable NAAs, we
further investigated N-end rule pathway pro-
teins as a source of additional recognizers. In a
comprehensive screen of diverse ClpS family
proteins, we discovered a group of ClpS proteins
from the bacterial phylum Planctomycetes with
native binding to N-terminal leucine, isoleucine,
and valine. We applied directed evolution tech-
niques to generate a Planctomycetes ClpS
variant—PS961 (table S1)—with submicromolar
affinity to N-terminal leucine, isoleucine, and
valine, and demonstrated recognition of these
Reed et al., Science 378, 186–192 (2022)
14 October 2022
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Dynamic sequencing of diverse peptides with high-precision
kinetic outputs. (A to E) Dynamic sequencing of the peptide DQQRLIFAG. An
example trace is shown in (A). A scatterplot of RS mean PD versus bin ratio is
shown in (B). Shown in (C) are additional example traces of dynamic
sequencing of DQQRLIFAG peptide. Shown in (D) are distributions of the
duration of each RS and NRS acquired during sequencing, with mean
durations indicated. Kinetic signature plots summarizing the characteristic
sequencing behavior of DQQRLIFAG peptide are shown in (E). (F to G)
Dynamic sequencing of the synthetic peptides DQQIASSRLAASFAAQQYPDDD
(top), RLAFSALGAADDD (middle), and EFIAWLV (bottom). Example
traces for each peptide are shown in (F). Corresponding kinetic signature
plots are shown in (G).
NAAs (Fig. 2B). The median PD of binding to
peptides with N-terminal LAA, IAA, and VAA
(I, isoleucine; L, leucine; V, valine) was 1.21,
0.28, and 0.21 s, respectively, in agreement
with bulk characterization (fig. S2C).
In a separate screen, we investigated a diverse
set of UBR-box domains from the UBR family of
ubiquitin ligases that natively bind N-terminal
arginine, lysine, and histidine (10). The UBR-
box domain from the yeast Kluyveromyces
marxianus UBR1 protein (table S1) exhibited
the highest affinity for N-terminal arginine, and
we used this protein to generate an arginine
recognizer, PS691. PS691 recognized arginine
in a peptide with N-terminal RLA (R, arginine)
with a median PD of 0.23 s (Fig. 2C). Lower-
affinity binding to N-terminal lysine and
histidine (fig. S2, D and E) was insufficient for
single-molecule detection.
To demonstrate that amino acids in a single
peptide molecule can be sequentially exposed
by aminopeptidases and recognized in real time
with distinguishable kinetics, we incubated
an immobilized peptide containing the initial
sequence FAAWAAYAA with PS610 for 15 min,
followed by the addition of PhTET3, an amino-
peptidase from Pyrococcus horikoshii (11). The
collected traces consisted of regions of distinct
pulsing, which we refer to as recognition seg-
ments (RSs), separated by regions lacking re-
cognition pulsing [nonrecognition segments
(NRSs)]. We developed analysis software to
automatically identify pulsing regions and
transition points within traces on the basis of
fluorescence properties and pulsing kinetics
(see materials and methods). Traces began
with the recognition of phenylalanine with a
median PD of 2.36 s (Fig. 2D), in agreement
with the PD observed for FAA in recognition-
only assays. This pattern terminated after amino-
peptidase addition (on average, 11 min after
addition) and was followed by the ordered
appearance of two RSs with median PDs of
0.25 and 0.49 s (Fig. 2D), corresponding to
the short and medium PDs obtained in our
YAA and WAA recognition-only assays. Thus,
the introduction of aminopeptidase activity to
the reaction resulted in the sequential appear-
ance of discrete RSs with the expected kinetic
properties in the correct order.
To demonstrate dynamic sequencing with
two NAA recognizers, we labeled PS610 and
PS961 with the distinguishable dyes atto-Rho6G
and Cy3, respectively, and exposed an immobi-
lized peptide of sequence LAQFASIAAYASDDD
(D, aspartate; Q, glutamine; S, serine) to a solu-
tion containing both recognizers. After 15 min,
we added two P. horikoshii aminopeptidases
with combined activity covering all 20 amino
acids—PhTET2 and PhTET3 (11, 12). The col-
lected traces displayed discrete segments of
pulsing alternating between PS961 and PS610
according to the order of recognizable amino
acids in the peptide sequence (Fig. 2E). The
average bin ratio and average PD associated
with each RS readily distinguished the two
dye labels and four types of recognized NAAs
(Fig. 2F). Median PDs were 2.71, 1.40, 0.25, and
0.64 s for N-terminal LAQ, FAS, IAA, and YAS,
respectively (Fig. 2G).
Previous studies have shown that NAA-bound
ClpS and UBR proteins also make contacts with
the residues at position 2 (P2) and position 3
(P3) from the N terminus that influence binding
affinity (9, 13, 14). These influences are reflected
in the modulation of PD depending on the
downstream P2 and P3 residues, as we ob-
served above for LAA (1.21 s) compared with
Reed et al., Science 378, 186–192 (2022)
14 October 2022
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Detection of single amino acid changes and PTMs. (A and B) Dynamic
sequencing of synthetic peptides that differ by a single amino acid: RLAFAYPDDD
(top), RLIFAYPDDD (middle), and RLVFAYPDDD (bottom). Example traces are
shown in (A). Scatterplots of RS mean PD versus bin ratio are shown in (B).
(C and D) Detection of oxidized methionine in the peptide RLMFAYPDDD.
Distributions of mean PD for leucine are shown in (C); labels indicate populations
with leucine followed by methionine (LM) or methionine sulfoxide (LMo). Shown
in (D) are example traces in which methionine is recognized by PS961 and
leucine exhibits a long PD (top) or in which methionine is not recognized, owing
to oxidation, and in which leucine exhibits a short PD (bottom). (E) Scatterplots
of RS mean PD versus bin ratio for runs in which oxidation was not controlled
(left) or in which methionine was fully oxidized (right).
LAQ (2.70 s). We find that these influences
on PD vary within informatically advantageous
ranges and can be determined empirically or
approximated in silico to model peptide se-
quencing behavior a priori (fig. S2, F to H). A
powerful feature of this recognition behavior
with respect to peptide identification is that
each RS contains information about potential
downstream P2 and P3 residues or PTMs,
regardless of whether these positions are the
targets of an NAA recognizer.
Principles of dynamic protein sequencing
illustrated with model peptides
To evaluate the kinetic principles of our dy-
namic sequencing method when applied to
diverse sequences, we first characterized the
synthetic peptide DQQRLIFAG (G, glycine),
corresponding to a segment of human ubi-
quitin (Fig. 3, A to E). We performed sequencing
reactions through a combination of three dif-
ferentially labeled recognizers—PS610, PS961,
and PS691—and two aminopeptidases—PhTET2
and PhTET3 (see materials and methods). The
example trace in Fig. 3A starts with an NRS that
corresponds to the time interval during which
residues in the initial DQQ motif are present at
the N terminus. The first RS starts at 120 min,
upon exposure of N-terminal arginine to recog-
nition by PS691. Subsequent cleavage events
sequentially expose N-terminal leucine, iso-
leucine, and phenylalanine to their corresponding
recognizers, with fast transitions (average <10 s)
from one RS to the next. The transition from
leucine to isoleucine recognition by PS961 is
readily identified as a sharp change in average
PD. This overall pattern is replicated across
many instances of sequencing of the same
peptide, with similar PD statistics across traces,
as each peptide molecule follows the same re-
action pathway over the course of the sequenc-
ing run (Fig. 3, B and C). Owing to the stochastic
timing of cleavage events, each trace displays
distinct start times and durations for each RS
(Fig. 3C).
This approach reports the binding kinetics
at each recognizable amino acid position and
the kinetics of aminopeptidase cleavage along
the peptide sequence. High-precision kinetic
information on binding is obtained from a
single trace because each RS typically contains
tens to hundreds of on-off binding events, re-
sulting in a distribution of PD and interpulse
duration (IPD) measurements that can be
analyzed statistically. The repetitive probing
of each NAA also provides accurate recognizer
calling because calls are not based on the error-
prone detection of a single event associated with
one fluorophore molecule (fig. S1F). Recognizer
on-rate and concentration govern IPD for each
RS; higher recognizer concentrations result in
shorter average IPDs and faster rates of pulsing
(fig. S3, A and B). Higher recognizer concen-
trations, however, increase the fluorescence
background from freely diffusing recognizers,
resulting in lower pulse signal-to-noise ratios,
and can compete with aminopeptidases for
N-terminal access. In practice, IPDs in the
range of ~2 to 10 s provide a favorable balance
among these factors.
The distribution of RS durations across an
ensemble of replicate traces defines the rate
of cleavage of each recognizable NAA. For
DQQRLIFAG peptide, we observed average
cleavage times of 30, 55, 40, and 88 min for
N-terminal arginine, leucine, isoleucine, and
phenylalanine, respectively, with approximate
single-exponential decay statistics for each
position (Fig. 3D and fig. S3C). The distribution
of NRS durations reports the cleavage rate of a
run of one or more nonrecognized NAAs. The
average NRS duration for the initial DQQ motif
was 155 min (Fig. 3D). Average cleavage rates
are a key parameter and are controlled by the
aminopeptidase concentration in the assay (fig.
S3, D and E). Given the exponential behavior,
we target average RS durations of 10 to 40 min
Reed et al., Science 378, 186–192 (2022)
14 October 2022
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. Discrimination of peptides in mixtures and mapping peptides to
the human proteome. (A) Example traces from sequencing a mixture of the
peptides DQQRLIFAG and RLAFSALGAADDD on the same chip; the chip window
indicates the location of reaction chambers producing a sequencing readout
for each peptide. (B) Example traces from the dynamic sequencing of
two peptides, DQQRLIFAGK (top) and EFIAWLVK (bottom), isolated from the
recombinant human proteins ubiquitin and GLP-1, respectively. (C) Diagram
illustrating the identification of the protein ubiquitin as a match to the kinetic
signature from DQQRLIFAGK peptide in an in silico digest of the human
proteome based on kinetic information.
to provide sufficient time for pulsing data col-
lection, avoid missed RSs due to rapid cleavage,
and minimize excessively long RS durations.
We found it helpful to visualize the sequenc-
ing profiles of peptides as kinetic signature
plots—simplified trace-like representations of
the time course of complete peptide sequenc-
ing containing the median PD for each RS and
the average duration of each RS and NRS (Fig.
3E). These highly characteristic features provide
a wealth of sequence-dependent information
for mapping traces from peptides to their pro-
teins of origin.
To demonstrate that this core methodology
and its kinetic principles apply to a wide range
of peptide sequences, we sequenced the syn-
thetic peptides DQQIASSRLAASFAAQQYPDDD,
RLAFSALGAADDD, and EFIAWLV (E, glu-
tamate; P, proline)—a segment of human
glucagon-like peptide–1 (GLP-1)—under the same
sequencing conditions used for DQQRLIFAG
(Fig. 3F). Each peptide generated a characteristic
kinetic signature in accordance with its se-
quence (Fig. 3G). We obtained readouts as far
as position 18 (the furthest recognizable amino
acid) in the peptide DQQIASSRLAASFAAQ-
QYPDDD, illustrating that the method is com-
patible with long peptides and capable of deep
access to sequence information in peptides.
Distinctive kinetic signatures from single
amino acid changes and PTMs
To illustrate how the kinetic parameters acquired
from sequencing are sensitive to changes in
sequence composition, we performed sequenc-
ing with a set of three peptides—RLAFAYPDDD,
RLIFAYPDDD, and RLVFAYPDDD—that differ
only at a single position, located immediately
downstream from the PS961 N-terminal target
leucine (Fig. 4A). Each type of amino acid at this
position had a distinct effect on the PD acquired
during recognition of N-terminal leucine by
PS961. We observed median PDs of 1.29, 2.22,
and 4.21 s for LAF, LIF, and LVF, respectively
(Fig. 4B). In addition to differences in PD for
leucine, each peptide displayed a characteristic
RS or NRS in the interval between leucine and
phenylalanine recognition (Fig. 4A and fig. S4A).
These results demonstrate the sensitivity of the
sequencing readout to variation at a single posi-
tion and illustrate that both directly recognized
NAAs and adjacent residues can influence the
full kinetic signature obtained from sequencing.
Because the aminoacyl-proline bond of the
YP motif in peptides such as RLIFAYPDDD
cannot be cleaved by the PhTET aminopepti-
dases (11, 12), observation of YP pulsing at
the end of a trace ensures that cleavage has
progressed completely from the first to last
recognizable amino acid. The sequencing
output from RLIFAYPDDD, therefore, provided
a convenient dataset for examining biochem-
ical sources of nonideal behavior that could
lead to errors in peptide identification. The
main sources of incomplete information in
traces were deletions of expected RSs due to
the stochastic occurrence of rapid sequential
cleavage events (fig. S4B) and early termi-
nation of reads resulting from photodamage
or surface detachment (fig. S4C).
In addition to changes in amino acid se-
quence composition, sequencing readouts
are sensitive to changes due to PTMs. As an
example, we examined methionine oxidation.
The thioether moiety of the methionine side
chain is susceptible to oxidation during peptide
synthesis and sequencing. We determined
that PS961 binds a peptide with N-terminal
methionine with a dissociation constant (Kd)
of 947 ± 47 nM (fig. S4D) and hypothesized
that oxidation, resulting in a polar methionine
sulfoxide side chain, would eliminate binding
and reduce NAA binding affinity when located
at P2. We determined computationally that
methionine sulfoxide is highly unfavorable in
the PS961 NAA binding pocket and that non-
polar residues are preferred at P2 (fig. S4E and
fig. S2H). We sequenced the synthetic peptide
RLMFAYPDDD (M, methionine) and observed
two populations of traces with distinct kinetic
signatures—a first population containing leucine
recognition with a median PD of 0.86 s and a
second population with a median PD of 0.35 s
(Fig. 4C). Traces from the first population also
displayed methionine recognition with a short
PD in the time interval between leucine and
phenylalanine recognition (Fig. 4D). Methio-
nine recognition was absent in traces from the
second population (Fig. 4D), indicating that
the methionine side chain in these peptides
was not capable of recognition by PS961.
When we fully oxidized methionine by prein-
cubation with hydrogen peroxide (see mate-
rials and methods), we observed elimination of
both methionine recognition and the leucine
Reed et al., Science 378, 186–192 (2022)
14 October 2022
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
recognition cluster with a long median PD, as
expected (Fig. 4E). These results demonstrate
the capability for extremely sensitive detec-
tion of PTMs owing to their kinetic effects on
recognition.
Sequencing peptide mixtures and mapping
peptides derived from human proteins
Proteomics applications require identification
of peptides in mixtures derived from biological
sources. To extend our results to peptide
mixtures and biologically derived peptides,
we performed two experiments. First, we
mixed DQQRLIFAG and RLAFSALGAADDD
peptides, immobilized them on the same chip,
and performed a sequencing run. Data analy-
sis (see materials and methods) identified two
populations of traces corresponding to each
peptide, with kinetic signatures in close agree-
ment with those identified in runs with indi-
vidual peptides (Fig. 5A and fig. S4F). Second,
to demonstrate that our method extends to
biologically derived peptides, we performed
sequencing runs with peptide libraries gen-
erated using a simple workflow from recom-
binant human ubiquitin (76 amino acids) and
GLP-1 (37 amino acids) proteins digested with
AspN/LysC and trypsin, respectively (see mate-
rials and methods). For both libraries, data
analysis readily identified traces matching
the expected recognition pattern for the pro-
tease cleavage products DQQRLIFAGK and
EFIAWLVK (K, lysine) for ubiquitin and GLP-1,
respectively, and produced kinetic signatures
in agreement with synthetic versions of these
peptides (Fig. 5B and fig. S4G). We identified
matches to the kinetic signature of the ubi-
quitin peptide DQQRLIFAGK across the human
proteome, taking advantage of simple sequence
constraints provided by kinetic information
(see materials and methods). We found only
one protein other than ubiquitin that con-
tained a peptide that could potentially match
this signature (Fig. 5C); thus, even short sig-
natures can exhibit proteome abundance of
<1 in 104 proteins. These results illustrate the
potential of the full kinetic output from se-
quencing to enable digital mapping of peptides
to their proteins of origin.
Conclusions
Our simple, real-time dynamic approach differs
markedly from other recently described single-
molecule approaches that rely on complex,
iterative methods involving stepwise Edman
chemistry or hundreds of cycles of epitope
probing (15–17). Nanopore approaches offer
the potential for real-time readouts and sim-
plicity but face substantial challenges related
to the size and biophysical complexity of poly-
peptides (18–20). Our sequencing technology
is readily expanded in its capabilities, and there
are multiple areas for improvement. Expansion
of proteome coverage can be achieved through
directed evolution and engineering of recog-
nizers. The NAA targets demonstrated here
make up ~35.6% of the human proteome, but
lower-affinity NAA targets require longer PDs
to enable detection in all sequence contexts.
Recognizers for new amino acids or PTMs
can be evolved from current recognizers or
identified in screens of other scaffolds, such
as other types of NAA- or PTM-binding pro-
teins or aptamers. Extension to detection
of all 20 natural amino acids and multiple
PTMs is feasible for de novo sequencing; how-
ever, partial sequences are sufficient for most
proteomics applications, which rely on map-
ping to predefined sets of candidate proteins
(21). Aminopeptidases can be engineered to
optimize cleavage rates and minimize RS
deletions from rapid sequential cleavage. We
envision that the dynamic range of samples
and the applications most suitable for the
system will tend to scale with the number of
reaction chambers on the chip and that com-
pression of dynamic range will be necessary
for certain applications. We anticipate that
future developments of the platform will in-
crease the accessibility of proteomics studies
and enable discoveries in biological and clinical
research.
RE FERENCES AND NOTES
1. S. Goodwin, J. D. McPherson, W. R. McCombie, Nat. Rev. Genet.
17, 333–351 (2016).
2. W. Timp, G. Timp, Sci. Adv. 6, eaax8978 (2020).
3. R. Aebersold et al., Nat. Chem. Biol. 14, 206–214 (2018).
4. M. Beck et al., Mol. Syst. Biol. 7, 549 (2011).
5. N. L. Anderson, N. G. Anderson, Mol. Cell. Proteomics 1,
845–867 (2002).
6. Y. Liu, A. Beyer, R. Aebersold, Cell 165, 535–550
7.
(2016).
J. Tullman, N. Callahan, B. Ellington, Z. Kelman, J. P. Marino,
Appl. Microbiol. Biotechnol. 103, 2621–2633 (2019).
8. D. A. Dougan, B. G. Reid, A. L. Horwich, B. Bukau, Mol. Cell 9,
673–683 (2002).
9. B. J. Stein, R. A. Grant, R. T. Sauer, T. A. Baker, Structure 24,
232–242 (2016).
10. T. Tasaki, Y. T. Kwon, Trends Biochem. Sci. 32, 520–528
(2007).
11. M. A. Durá et al., Mol. Microbiol. 72, 26–40 (2009).
12. M. A. Durá et al., Biochemistry 44, 3477–3486 (2005).
13. W. S. Choi et al., Nat. Struct. Mol. Biol. 17, 1175–1181
(2010).
15. J. Swaminathan et al., Nat. Biotechnol. 36, 1076–1082 (2018).
16. J. D. Egertson et al., bioRxiv 2021.10.11.463967 [Preprint]
(2021); https://doi.org/10.1101/2021.10.11.463967.
17. M. Chee, K. Gunderson, M. P. Weiner, Macromolecule analysis
employing nucleic acid encoding (2019); https://pdfaiw.uspto.
gov/.aiw?Docid=20190145982.
18. S. Zhang et al., Nat. Chem. 13, 1192–1199 (2021).
19. H. Brinkerhoff, A. S. W. Kang, J. Liu, A. Aksimentiev, C. Dekker,
Science 374, 1509–1513 (2021).
20. J. A. Alfaro et al., Nat. Methods 18, 604–617 (2021).
21. J. Swaminathan, A. A. Boulgakov, E. M. Marcotte, PLOS
Comput. Biol. 11, e1004080 (2015).
22. B. D. Reed, M. J. Meyer, J. F. Beltrán, Code and data for
“Real-time dynamic single-molecule protein sequencing
on an integrated semiconductor device.” Zenodo (2022);
https://doi.org/10.5281/zenodo.7017750.
AC KNOWLED GME NTS
We thank E. Chen for helpful discussions in preparing the
manuscript. Funding: This work was funded by Quantum-Si, Inc.
Author contributions: A.D.G., A.L., A.M.v.O., A.P.-K., B.D.R.,
B.S., C.L., D.H.P., E.M., F.L., G.P., H.H., J.A.B., J.P., J.T., J.Z., K.B.,
K.F.C., M.Mi., M.M.E., M.P., M.R.-C., M.W.B., N.D.B., N.S., O.A., R.B.,
R.N., S.Ha., S.G., S.S.P., T.D.C., X.S., and Y.-C.W. developed
methods and reagents for protein sequencing; A. Bu., D. Be., D.F.,
G.A., G.K., J.F.B., M.D., M.J.M., S.F.S., S.L., V.A., and Z.Z. developed
software; and A. Bel., A. Bet., A.K., B.L., C.C., D.Bo., E.A.G.W.,
F.R.A., G.S., J.B., J.C., J.D., J.S., K.P., K.R., L.H., M.B., M.F., P.A.,
P.L., P.T., R.Cl., R.Cu., S.Ho., S.C., T.C., T.R., T.R.T., X.M., X.W., and
Z.H. developed semiconductor chips, nanophotonics, lasers, and
instruments. J.M.R., M.Mc., T.R., B.D.R., M.D., G.S., M.F., P.L.,
and M.D.D. supervised and/or acquired resources and funding. B.D.R.,
M.J.M., M.P., and B.S. designed experiments and/or generated
the single-molecule recognition and sequencing data presented
in the figures. B.D.R., M.J.M., and J.F.B. analyzed sequencing
data and prepared the figures. M.P. and G.P. characterized
recognizer ensemble kinetic properties. O.A. generated libraries
from recombinant proteins. H.H. and Y.-C.W. generated model
peptides. D.H.P. performed computational modeling. B.D.R. led the
study and wrote the manuscript with review and commentary from
coauthors. All authors met the criteria for authorship and
contributed critically to the development of the sequencing method
and platform by conducting experiments, developing methods and
concepts, analyzing and interpreting data, developing software,
supervising research, or acquiring funding. Competing interests:
All authors affiliated with Quantum-Si, Inc., along with A.D.G. and
A.M.v.O., are shareholders and/or are listed as inventors on
patents owned by Quantum-Si, Inc.; A.D.G. and A.M.v.O. are on the
scientific advisory board of Quantum-Si, Inc., and are paid
consultants. The technology presented in this paper is the subject
of numerous pending or awarded patents filed by Quantum-Si,
Inc., with the US Patent and Trademark Office and international
offices. Data and materials availability: Data and custom code used
in this paper are available for download online at Zenodo (22).
Sequencing reagents and instruments are available from Quantum-Si,
Inc., under a material transfer agreement. License information:
Copyright © 2022 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science. No
claim to original US government works. https://www.science.org/
about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abo7651
Materials and Methods
Figs. S1 to S4
Table S1
References (23–27)
View/request a protocol for this paper from Bio-protocol.
14. J. Muñoz-Escobar, E. Matta-Camacho, C. Cho, G. Kozlov,
K. Gehring, Structure 25, 719–729.e3 (2017).
Submitted 28 February 2022; accepted 13 September 2022
10.1126/science.abo7651
Reed et al., Science 378, 186–192 (2022)
14 October 2022
6 of 6
|
10.1126_science.abo4940
|
RES EARCH
METALLURGY
Machine learning–enabled high-entropy
alloy discovery
Ziyuan Rao1, Po-Yen Tung1,2, Ruiwen Xie3, Ye Wei1*, Hongbin Zhang3, Alberto Ferrari4, T.P.C. Klaver4,
Fritz Körmann1,4, Prithiv Thoudden Sukumar1, Alisson Kwiatkowski da Silva1, Yao Chen1,5,
Zhiming Li1,6, Dirk Ponge1, Jörg Neugebauer1, Oliver Gutfleisch1,3, Stefan Bauer7, Dierk Raabe1*
High-entropy alloys are solid solutions of multiple principal elements that are capable of reaching
composition and property regimes inaccessible for dilute materials. Discovering those with valuable
properties, however, too often relies on serendipity, because thermodynamic alloy design rules alone often
fail in high-dimensional composition spaces. We propose an active learning strategy to accelerate the
design of high-entropy Invar alloys in a practically infinite compositional space based on very sparse data.
Our approach works as a closed-loop, integrating machine learning with density-functional theory,
thermodynamic calculations, and experiments. After processing and characterizing 17 new alloys out of
millions of possible compositions, we identified two high-entropy Invar alloys with extremely low thermal
expansion coefficients around 2 × 10−6 per degree kelvin at 300 kelvin. We believe this to be a suitable
pathway for the fast and automated discovery of high-entropy alloys with optimal thermal, magnetic,
and electrical properties.
A lloy design refers to a knowledge-guided
approach to the development of high-
performance materials. The strategy was
established in the Bronze Age and has
undergone further developments since
that time. Alloy design is the basis for the
development of different materials that en-
able technological progress. Several thousand
metallic alloys have been developed so far
that serve in engineering applications. The
first essential alloy groups developed, such
as bronze and steel, are all based on one main
element that forms the matrix of the material.
Over time, alloys with a higher number of al-
loying elements in larger fractions, such as
austenitic stainless steels, have been devel-
oped. Today, with the development of high-
entropy alloys (HEAs), we have reached a
stage where multiple elements are used in
similar fractions (1, 2). Considering only the
most used elements of the periodic table, this
spans a composition space of at least 1050 alloy
variants, a space so large that it cannot be
managed by conventional alloy design methods
(3). These conventional methods for designing
alloys, which have been applied to small sub-
spaces of the HEA composition realm, include
calculation of phase diagrams (CALPHAD) and
density-functional theory (DFT) (4–6). However,
CALPHAD provides equilibrium-phase diag-
1Max-Planck-Institut für Eisenforschung GmbH, Düsseldorf,
Germany. 2Department of Earth Sciences, University of
Cambridge, Cambridge, UK. 3Institut für Materialwissenschaft,
Technische Universität Darmstadt, Darmstadt, Germany.
4Materials Science and Engineering, Delft University of
Technology, Delft, Netherlands. 5School of Civil Engineering,
Southeast University, Nanjing, China. 6School of Materials
Science and Engineering, Central South University, Changsha,
China. 7KTH Royal Institute of Technology, Stockholm, Sweden.
*Corresponding author. Email: y.wei@mpie.de (Y.W.);
d.raabe@mpie.de (D.R.)
rams only, and DFT is computationally costly
and cannot be readily applied to higher tem-
peratures and disordered alloys (5, 7). Like-
wise, combinatorial experiments (8) are very
labor intensive and only cover the limited com-
position space of HEAs.
Because of these methodological limitations
to finding materials with promising functional
and mechanical features, we present a differ-
ent approach to accelerating the discovery of
HEAs. We based our approach on the use of
machine learning (ML) techniques, with a
focus on probabilistic models and artificial
neural networks. Limited by the amount of
available composition-property data, conven-
tional ML approaches in alloy design have to
predominantly rely on simulation data, often
with only limited experimental validation
(9, 10). As the experimental microstructure
database continues to expand, ML obtains
higher accuracy in predicting the phase or
microstructure of materials (11). However,
the direct composition-property prediction is
still elusive because of the comparably small
databases and the human bias in feature se-
lection. Recently, active learning has emerged
as an alternative choice for functional mate-
rials discovery (12). Active learning is a subfield
of ML in which surrogate models iteratively
select unseen data points that are most in-
formative to improve the predictive power of
the models (13). In this approach, the next set
of experiments is guided by the previous model
trained based upon the results seen so far,
yielding data points that will again be used
iteratively for updating the model. Active learning
has the potential to reduce the computational
costs of alloy design and to both incorporate
and guide experimental data and routines.
However, active learning approaches to guid-
ing the experimental discovery of materials
have relied on simple surrogate models and
Bayesian optimization methods, which are
limited to low-dimensional data, thus showing
property improvements only after many iter-
ations (14, 15).
To overcome these obstacles, we propose an
active learning framework for the composition
discovery of HEAs that is efficient for very
sparse experimental datasets. The approach
comprises ML-based techniques, DFT, mean-
field thermodynamic calculations, and experi-
ments. We focused on the design of high-entropy
Invar alloys with a low thermal expansion
coefficient (TEC) for several reasons: (i) a high
demand exists for different types of Invar al-
loys to serve emerging markets for the transport
of liquid hydrogen, ammonia, and natural gas;
(ii) the mechanical properties of the original
Fe63.5Ni36.5 (wt %) alloy for which Charles
Edouard Guillaume received the 1920 physics
Nobel Prize leave room for improvement; (iii)
alternative Invar alloys (e.g., intermetallic,
amorphous, or antiferromagnetic Invar com-
pounds) come at forbiddingly high alloy costs
and/or poor ductility (16, 17); (iv) although a
few HEAs have the potential to fill this gap
(18–20), the lowest TEC (~10 × 10−6 K−1) of
HEAs reported in the literature exceeds the
value of the original Fe63.5Ni36.5 (wt %) alloy
(~1.6 × 10−6 K−1) (19); and (v) our active learn-
ing framework mainly considers compositional
information instead of the alloy manufacturing
process, which makes the Invar effect an ideal
target because these alloys are mostly determined
by composition and less by processing (6, 19)
(see fig. S1 and table S1 for more background).
Results and discussion
Generative alloy design
The active learning framework includes three
main steps: targeted composition generation,
physics-informed screening, and experimental
feedback (Fig. 1). Considering the large num-
ber of possible composition combinations of
HEAs and the small experimental datasets
(699 compositions; fig. S2), the challenge is to
directly sample new compositions with the de-
sired properties. Therefore, we developed an
HEA generative alloy design (HEA-GAD) ap-
proach that is based on a generative model
(GM) (21). First, the HEA-GAD uses GM,
mathematical modeling, and sampling to per-
form a large-scale search of potential Invar
alloys. GM learns an efficient and effective re-
presentation of the high-dimension data, which
not only provides direct data visual represen-
tation, but also converts the search in high-
dimensional design spaces to those of lower
dimensionality (22). Different GMs are compared
and analyzed on the basis of the evaluation
metrics. The results show that the Wasserstein
autoencoder (WAE) architecture performs bet-
ter than other models with similar architec-
tures (21) (figs. S3 and S4). The encoder takes
Rao et al., Science 378, 78–85 (2022)
7 October 2022
1 of 7
RES EARCH | R E S E A R C H A R T I C L E
Latent space sampling
Alloy
Encoder Decoder
q (z|x)
p (x|z)
Autoencoder
training
Targeted
generation
Physical properties
~1,000
candidates
DFT and Thermodynamic
calculation
Active learning loop
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...
+
...
Neural network
Boosting Tree
Two-step ensemble
regression model
Invar database
Fe Ni Co Cr Cu
0.9 0.1 0.0 0.0 0.0
0.8 0.2 0.0 0.0 0.0
1.0 0.0 0.0 0.0 0.0
.
.
.
+
0.5 0.2 0.3 0.0 0.0
0.3 0.2 0.2 0.2 0.1
0.4 0.4 0.0 0.2 0.0
Feedback to
the database
PPMS
Fe Ni Co Cr Cu
Rank
score
0.5 0.2 0.3 0.0 0.0
1.2
Top-3
0.3 0.2 0.2 0.2 0.1
0.4 0.4 0.0 0.2 0.0
.
.
.
1.4
1.8
.
.
.
10-30
candidates
Experimental validation
Ranking policy
Fig. 1. Approach overview. We developed an active learning framework for
the targeted composition design and discovery of HEAs, which combines ML
models, DFT calculations, thermodynamic simulations, and experimental
feedback. First, the promising candidates are generated under the HEA-GAD
framework consisting of two primary steps: (i) an autoencoder for composition
generation and (ii) stochastic sampling for composition selection. Second,
the selected candidates from the HEA-GAD are further processed by the TERM
framework, which includes two ensemble models composed of multilayer
perceptrons and gradient-boosting decision trees. In the last step, the most
promising compositions are selected by a ranking-based policy. The top three
candidates are experimentally measured and fed back to the database. The
iteration is repeated until the discovery of Invar alloys.
compositions of alloys as the input and learns
to compress them down to low-dimensional
representations, and the decoder can then act
as a generator for producing alloy compositions
given the learned continuous latent z repre-
sentation. Although WAE is trained with only
compositional information of alloys, it may im-
plicitly include information on composition-
related properties, which makes the latent space
physically meaningful and informative. In our
case, Invar alloys show extremely low TEC
(hereafter used to refer to the TEC around
room temperature unless otherwise specified)
values, and the composition-TEC relation obeys
specific physical laws. Subsequently, HEA-GAD
uses the Gaussian mixture model (GMM) and
Markov chain Monte Carlo (MCMC) sampling
(23, 24) to perform a large-scale search for the
Invar compositions generated from WAE latent
representation.
Two-stage ensemble regression
Next, we use the two-stage ensemble regres-
sion model (TERM) to further investigate the
TEC of the HEA-GAD–generated alloy compo-
sitions. The first stage concerns composition-
based regression models aiming at fast and
large-scale composition inference. Then, the
top ~1000 results with potentially low TEC
from the HEA-GAD model are screened and
enter the second-stage model, where DFT and
thermodynamic calculations are included as
part of the input, making it a physics-informed
model (table S4). In the following section, we
demonstrate that incorporating the physical
inputs does increase the model accuracy. To
increase the robustness of TERM without
sacrificing the prediction accuracy, TERM
taps into the advantages of the multilayer
perceptron (25–27) and gradient-boosting
decision tree approaches (28–30) by combin-
ing both into a single ensemble (31). Based on
prediction and uncertainty, the exploration
and exploitation strategy is used to adaptively
guide the discovery of desirable compositions
(31). Exploration prefers the composition with
higher uncertainty (curiosity), whereas ex-
ploitation favors the composition with lower
predicted TEC (perceived usefulness). Such a
baseline strategy is premised on the model’s
ability to generalize beyond the known data,
which is, however, often hampered by the highly
nonlinear nature of the composition-property
relation and sparsity of the available dataset.
To overcome this issue, we designed a rank-
order strategy that allows predictions to be
rearranged and ranked in a specific order
(32, 33). This strategy is particularly advanta-
geous when the underlying distribution of
the data is largely unknown. The rank-based
strategy ensures that the candidate selection
is less affected by model inaccuracy and pro-
vides a systematic way to combine model
prediction and uncertainty (31). Finally, the
TEC values of the top three selected candidate
materials are experimentally determined by
the physical properties measurement system.
These experimental results then augment the
training database for the next active learning
iteration.
Compositional latent space distribution
We produced a large benchmark dataset with
699 data points of Invar alloys mainly from
former publications (fig. S2 and table S3) (34–39).
Then, on the basis of the HEA-GAD-TERM
framework proposed above, we performed six
iterations and cast 18 alloys including 17 new
Rao et al., Science 378, 78–85 (2022)
7 October 2022
2 of 7
RES EARCH | R E S E A R C H A R T I C L E
alloys and one Fe63.5Ni36.5 (wt %) classic Invar
alloy as a reference alloy. Because of the data
imbalance (figs. S5 and S6), the discovery of
FeNiCoCrCu HEAs is much more difficult
than the discovery of FeNiCoCr HEAs. For this
reason, we focused on the design of FeNiCoCr
HEAs for the first three iterations and on
FeNiCoCrCu HEAs for the last three iterations.
We show the WAE latent space and GMM-
modeled two-dimensional probability density
of the first iteration in Fig. 2, A and B. The la-
tent space yields certain islands that indicate
the compositional differences. For example,
the HEAs tend to stay in the middle, whereas the
binary and ternary alloys tend to stay in the
edges of the latent space. Also, a smooth tran-
sition among the Fe–Ni, Ni–Co binary alloys
and the Fe–Ni–Co ternary alloys can be ob-
served. FeNiCoCrCu forms a single island, in-
dicating that features of compositions with
nonzero Cu content are indeed captured by
HEA-GAD. The new FeCoNiCr HEAs candidates
are cross-marked, whereas the best-ranked HEAs
are illustrated with white dots in Fig. 2, A and
B. We also show the last iteration result of
FeCoNiCrCu HEAs discovered by HEA-GAD-
TERM in Fig. 2, C and D (in red). The entire
latent space is slightly rotated because of the
addition of new data into the training data-
set from previous iterations. The augmented
dataset also leads to a modified GMM-modeled
probability density shown in Fig. 2D, in which
the left Gaussian ellipse extends more to the left
region compared with Fig. 2B. Such pheno-
mena suggest that the HEA-GAD-TERM frame-
work is interpretable and sensitive to the dataset.
Physics-descriptor–informed model
So far, the Masumoto empirical rule (34, 35)
has played an important role in the discov-
ery of several Invar alloys. As exemplified in
Fig. 3D for the Fe60Ni35Co5 (wt %) Super Invar
alloy, according to this rule, the TEC is related
to the ratio ws/Tc (magnetostriction/Curie tem-
perature): Because of the Invar effect, Invar
alloys have lower TEC in the ferromagnetic
state (below Curie temperature Q) than in the
paramagnetic state (above Q). The TEC in the
ferromagnetic state can thus be estimated as
ws
Tc
QA (cid:2) SA
Tc
≈ tan d (cid:2)
QA
Tc
TEC ¼
SA
Tc
QS
RS
¼
(cid:2)
¼
CALPHAD for FeCoNi alloys. The alloys from
our experimental dataset were slowly cooled
from high-temperature homogenization, so
an equilibrium temperature to calculate the
phase fractions in our samples cannot be de-
termined unambiguously. We nevertheless cal-
culated ws/Tc for the annealing temperatures
Tann = 1273 K, 1073 K, and 873 K (Fig. 3, A to C)
and observed a good correlation with the ex-
perimentally observed TEC values, especially
for the values taken at Tann = 873 K. ws and Tc
are thus useful descriptors that can be exploited
to increase the accuracy of TERM. We show
the comparison of the model training history
with and without the use of the descriptor Tc
(Fig. 3E). This history reflects the performance
evolution with time (epoch) as more data were
fed to the model. The final testing error was
notably reduced from 0.19 to 0.14 upon in-
clusion of DFT and CALPHAD data, a piece of
strong evidence that the physics-descriptor–
informed model can achieve better accuracy
than that based only on compositions.
Learning curve and thermal expansion behavior
We demonstrated the correlation between ws/
Tc and the experimental TEC with DFT and
We show the measured and predicted TEC
values of the 17 alloys experimentally measured
WAE latent space
GMM-MCMC sampling
Fig. 2. First and last (sixth) itera-
tions of the HEA-GAD generation.
(A and B) WAE latent space
and GMM-modeled density of the
first iteration. (C and D) WAE
latent space and GMM-modeled
density of the last iteration. The
WAE latent space distribution
of the different compositions is
marked with different symbols.
The colors of the data points in
the latent space denote their
corresponding TEC. The GMM
shows the probabilistic density
in the latent space. The new
candidates proposed by the first
stage of the TERM are marked
by crosses, and the new
compositions proposed by the
second stage of the TERM
are marked by circles. The
learned latent spaces are
informative of the TEC
of the HEAs.
n
o
i
t
a
r
e
t
i
t
s
r
i
F
n
o
i
t
a
r
e
t
i
h
t
x
S
i
With V
Fe-Ni-Co-Cr
A
Fe-Ni
2
B
)
K
2
i
n
o
s
n
e
m
D
i
6
-
/
0
1
(
Fe-Ni-Co
C
E
T
With Cu
Fe-Co
Fe-Co-Cr
i
n
o
s
n
e
m
D
i
C
2
i
n
o
s
n
e
m
D
i
Dimension 1
D
Dimension 1
With V
Fe-Ni-Co-Cr
Fe-Ni
Fe-Ni-Co
FeCoNiCrCu
HEAs
)
K
6
-
/
0
1
(
C
E
T
With Cu
Fe-Co
Fe-Co-Cr
i
2
n
o
s
n
e
m
D
i
d
o
o
h
i
l
e
k
i
l
g
o
L
d
o
o
h
i
l
e
k
i
l
g
o
L
Rao et al., Science 378, 78–85 (2022)
7 October 2022
3 of 7
Dimension 1
Dimension 1
RES EARCH | R E S E A R C H A R T I C L E
A
)
K
/
6
-
0
1
(
C
E
T
T = 873 K
ann
Invar alloys
B
)
K
/
6
-
0
1
(
C
E
T
T = 1073 K
ann
Invar alloys
C
)
K
/
6
-
0
1
(
C
E
T
-4
(10 /K)
s
T c
R
s
-4
(10 /K)
s
T c
E
s
s
o
L
Super Invar
T c
Q
S
A
TEC =
QS =
RS
QA - SA
RS
≈ tan -
s
T c
D
)
m
µ
(
h
t
g
n
e
l
n
i
e
g
n
a
h
C
Training history
T = 1273 K
ann
Invar alloys
-4
(10 /K)
s
T c
Test (without DFT)
Test (with DFT)
Train (without DFT)
Train (with DFT)
Temperature (K)
Epoch
Fig. 3. Importance of the physics-informed descriptors. (A to C) Correlation between the proposed descriptor ws/Tc and the experimental TEC. (D) Schematic
model of the Masumoto empirical rule for discovering Invar alloys. (E) Comparison of training and testing history with and without use of the descriptor ws/Tc. Both
the final training and testing errors decrease after considering the physics-informed descriptors; for example, the testing error decreases from 19 to 14%.
Table 1. Compositions and TEC of the HEAs designed in this work.*
Alloys
Iteration
Fe
(wt %)
Ni
(wt %)
Co
(wt %)
Cr
(wt %)
Cu
(wt %)
Predicted TEC
(×10−6/K)
Predicted uncertainty
(×10−6/K)
Experimental TEC
(×10−6/K)
1st
1st
1st
2nd
2nd
3rd
3rd
3rd
4th
4th
4th
5th
5th
5th
6th
6th
6th
A1
............................................................................................................................................................................................................................................................................................................................................
A2
............................................................................................................................................................................................................................................................................................................................................
A3
............................................................................................................................................................................................................................................................................................................................................
A4
............................................................................................................................................................................................................................................................................................................................................
A5
............................................................................................................................................................................................................................................................................................................................................
A7
............................................................................................................................................................................................................................................................................................................................................
A8
............................................................................................................................................................................................................................................................................................................................................
A9
............................................................................................................................................................................................................................................................................................................................................
B1
............................................................................................................................................................................................................................................................................................................................................
B2
............................................................................................................................................................................................................................................................................................................................................
B3
............................................................................................................................................................................................................................................................................................................................................
B4
............................................................................................................................................................................................................................................................................................................................................
B5
............................................................................................................................................................................................................................................................................................................................................
B6
............................................................................................................................................................................................................................................................................................................................................
B7
............................................................................................................................................................................................................................................................................................................................................
B8
............................................................................................................................................................................................................................................................................................................................................
B9
............................................................................................................................................................................................................................................................................................................................................
*The original Fe63.5Ni36.5 Invar (A6) is a reference alloy and is not listed here.
7.54
10.52
1.41
7.97
3.24
4.09
4.83
2.02
5.84
4.38
8.56
4.94
5.31
9.68
5.60
5.13
6.29
3.41
3.13
4.39
3.91
4.20
4.58
5.88
5.16
7.57
5.48
4.43
8.41
4.50
9.32
5.49
5.65
5.56
16.7
27.1
40.9
20.8
34.6
37.7
33.2
17.2
39.5
22.2
14.6
38.3
8.3
27.5
20.9
18.3
15.8
55.2
49.2
41.8
52.5
44
42.4
44.2
54.1
40
48.8
57
40.6
57.7
51.6
48.3
50
50.7
1.29
0.75
0.79
0.53
0.96
1.40
2.17
1.43
1.45
1.01
1.33
1.70
1.00
3.49
0.92
1.16
1.05
23.9
17.2
9.4
22
13.8
12.6
15.8
22.8
6.9
17.8
16.4
6.9
22.9
6.8
17.8
18.3
19.9
4.2
6.5
8
4.7
7.6
7.3
6.8
5.9
7.9
6.2
5.1
9.2
5.2
7.8
7.9
8
7.9
0
0
0
0
0
0
0
0
5.7
5
6.9
5
5.9
6.3
5.1
5.4
5.7
Rao et al., Science 378, 78–85 (2022)
7 October 2022
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
in the six iterations in Table 1. A3 and A9 HEAs
with four principal elements show extremely
low TECs that are comparable to the classical
Fe63.5Ni36.5 (wt %) binary Invar alloy. B2 and
B4 HEAs with five principal elements show
TECs that are comparable to the commercially
used Fe54Co17Ni29 (wt %) ternary Kovar alloys.
In addition, a tabular comparison between
HEA-GAD-TERM and trial and error can be
found in table S2, where our method shows a
fivefold higher discovery rate than that achieved
by the trial and error approach alone.
We illustrate the alloy discovery process in
two scenarios (Fig. 4, A and B). In the ideal
case, the composition-TEC curve is simple and
‘Ideal’
Path I
Path II
Invar
B
C
E
T
-
t
s
e
w
o
L
‘Real-world’
Path II
Path I
Local
minima
Composition
Composition
Observed
lowest TEC
D
A2
A1
A3
Cr
B3
B1
B2
Unknown
Cu
Starting point
Unknown
Known
Exploration
s
y
o
l
l
a
f
o
r
e
b
m
u
N
FeCoNiCr HEAs
FeCoNiCrCu HEAs
E
)
K
6
-
/
0
1
(
C
E
T
F
E
P
A
M
1st
2nd
3rd
4th
5th
6th
FeCoNiCr HEAs
FeCoNiCrCu HEAs
Active learning
Composition (wt.%)
Composition (wt.%)
1st
2nd
3rd
4th
5th
6th
Iteration
I
IPF map
A2
BCC
HAGB
Phase map
100 µm
100 µm
J
IPF map
A3
FCC
HAGB
Phase map
A2
ModelBCC
ModelFCC
Temperature (K)
A3
K
)
Å
(
a
L
)
Å
(
a
=0.00
=0.04
=0.08
=0.15
=0.24
=0.00
=0.03
=0.07
=0.12
=0.17
=0.25
=0.50
=0.00
=0.03
=0.07
=0.12
=0.17
=0.24
=0.50
=0.40
100 µm
100 µm
ModelFCC
Temperature (K)
A
C
E
T
-
t
s
e
w
o
L
C
)
K
/
6
-
0
1
(
C
E
T
-
t
s
e
w
o
L
G
A2
111
001
101
H
A3
111
001
101
Fig. 4. Analysis of the results after six iterations in the active learning
loop. (A and B) Representation of the alloy discovery process in the ideal
scenario and the real world. (C and D) Cr and Cu distribution histogram.
The Cr histogram has various concentrations (from 0 to 20%). By contrast,
the vast majority (>95%) of the compositions have zero Cu concentration. The
lowest known TEC as a change of composition is plotted as a solid line, and the
unknowns are shown as a dashed line. Gray arrows illustrate the discovery paths
of HEA-GAD-TERM. (E) Experimental and predicted TEC of the FeNiCoCr and
FeNiCoCrCu HEAs. (F) MAPE of active learning. The dots represent the MAPE
between experiment and predictions. Rapid decrease of the MAPE is akin to a
natural learning process. (G and I) Electron backscatter diffraction (EBSD)
phase and boundary maps of the A2 alloy. (H and J) EBSD phase and
boundary maps of the A3 alloy. (K and L) Change of lattice constants with
↑
↑
↑
↓
↓
↓
temperature in the A2 [(Fe
h)26.1(Cr
1(cid:2)hCo
h)16.7(Co
1(cid:2)hNi
h)50.1(Ni
1(cid:2)hFe
↑
↑
↑
↓
↓
h)8.7] alloys
h)39.5(Cr
1(cid:2)hCo
h)9.1(Co
1(cid:2)hNi
and A3 [(Fe
for different values of h, where h denotes the pseudo-alloy concentration
(0 ≤ h ≤ 0.50).
↓
h)42.7(Ni
↑
1(cid:2)hFe
↓
1(cid:2)hCr
↓
1(cid:2)hCr
↑
h)7.1]
Rao et al., Science 378, 78–85 (2022)
7 October 2022
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
convex, which means that this specific relation
is readily learned and “never forgotten.” Even
with a small dataset present, the global maxima
can be easily found regardless of their initial
starting points: Both path 1 and path 2 can lead
to the Invar point. However, in the reality, the
lowest TEC curve is highly nonlinear because
of the complex underlying composition-property
relations, and the composition landscape re-
mains largely unknown. Both experts with ap-
propriate domain knowledge and algorithms
will have to explore the unknown territory and
accumulate knowledge about the system by
making mistakes. Furthermore, the composi-
tion axis is multidimensional and therefore the
design space is huge. Therefore, the chosen
paths, available data, and starting points will
notably influence the final results. Path 1 may
lead to local minima, whereas path 2 is rather
difficult initially, and multiple high TEC non-
Invar HEAs can be discovered before the even-
tual Invar discovery.
We provide the concentration histogram of
Cr and Cu in the current dataset in Fig. 4, C
and D. We also plotted the observed lowest
TEC curve to illustrate the discovery path in
two HEAs. The GAD-TERM framework shows
its high efficiency by quickly identifying the
Invar points in the first iteration (A3 and B2).
However, the algorithm is designed for ex-
ploration. The algorithm inevitably discov-
ers some non-Invar alloys along the path (e.g.,
A4 and A8, denoted by gray arrows in Fig. 4, C
and D). As mentioned before, the discovery of
FeNiCoCr HEAs and FeNiCoCrCu HEAs are
different tasks because of the different data
distribution. The distribution of Cu in the
alloys is extremely imbalanced (Fig. 4D); that
is, by far most of the alloys in the dataset do
not contain Cu at all and only a few alloys
have 5% Cu. Such distributional difference
likely accounts for the substantially different
learning behavior observed (Fig. 4, E and F).
We show the measured and predicted TEC
values for FeNiCoCr and FeNiCoCrCu HEAs in
Fig. 4E and the mean absolute percentage er-
ror (MAPE) between experiments and predic-
tions versus experimental iteration in Fig. 4F,
with each exploitation and exploration step
marked by arrows. For FeNiCoCr HEAs, the
average experimental TEC value gradually de-
creases: 6.49 × 10−6 per degree kelvin (/K) in
the first, 5.61 × 10−6/K in the second, and
3.65 × 10−6/K in the third iteration (Table 1).
Exploration and exploitation take place alter-
nately, akin to a natural learning process, and
such a plot represents the “learning curve” of
the HEA-GAD-TERM model. The learning curve
indicates a progressive trend as the MAPE
error decreases notably (from 1.5 to 0.2).
Because of the exploration step, the model
predictions deviate considerably from their
experimental counterparts in the first itera-
tion. Alloy A3 (Table 1) has the highest pre-
dicted TEC value (4.39 ± 0.79 × 10−6/K), but
the experimental TEC value shows exactly the
opposite, namely, the lowest measured TEC
value (1.41 × 10−6/K). In the second and third
iterations, the standard deviation of the ex-
perimental TEC values declines substantially
(3.34 × 10−6/K and 1.46 × 10−6/K, respectively).
This demonstrates excellent exploration prog-
ress in which HEA-GAD-TERM converges
quickly and can predict TEC with high ac-
curacy after only three iterations. Conversely,
FeNiCoCrCu shows a different learning behav-
ior. The discovery path shows no significant
improvements, from experimentally mea-
sured 6.26 × 10−6/K in the first iteration, to
6.64 × 10−6/K in the second, and 5.67 × 10−6/K
in the third (for more numerical details, see
Table 1). We can attribute this trend to the lack
of Cu-containing FeNiCoCrCu data (only three
data points are available at the beginning; Fig.
4D). Despite this shortcoming, the experimen-
tal mean deviation narrows down, from 33.9%
for the first iteration to 10.2% in the last ite-
ration, indicating a gradually improved model
accuracy.
To reveal the physical origin behind the
properties, we show experimental and DFT
analyses of the A2 and A3 alloys (TEC = 10.52 ×
10−6/K and 1.41 × 10−6/K, respectively, in Fig. 4,
G to L). It can be seen in Fig. 4, G to J, that A2
and A3 alloys have a single-phase bcc and fcc
structure, respectively. The partial disordered
local moment (PDLM) model within the co-
herent potential approximation simulations
(40) reveals that the Invar effect is qualitatively
related to such volume reduction at finite-
temperature PDLM phase compared with the
0 K ferromagnetic ground state (41). In con-
trast to the fcc A3 alloy, the bcc A2 alloy, with a
higher Tc around 950 K, exhibits a slight up-
ward trend of the lattice parameter a. Using
DFT simulations, we also validate that if the A2
alloy can be stabilized in its fcc phase state, then
an Invar effect can be realized as well [Fig. 4, K
and L, red dash-dot line; for simulation de-
tails, see (31)]. The TEC value is also affected
by the occurrence of phase transformations in
some HEAs (18, 20). Our measurements show
that the low TEC values of our A3 alloys are
not caused by any phase transformation.
A
)
K
/
6
-
0
1
(
C
E
T
HEAs + MEAs
MnFeNi
Cantor alloy
MnCoNi
CrFeCoNi
CrFeNi
CrCoNi
FeCoNi
FeCoNiMnCu
Compositionally complex alloys
This work
Invar
Amorphous
Invar
Kovar
B
)
K
/
6
-
0
1
(
C
E
T
Kovar
Invar
B4
B2
A9
A3
This work
Conventional
alloys
Anti-Invar
MEAs
HEAs
s
e
l
c
y
c
l
a
m
r
e
h
t
o
t
t
n
a
t
s
i
s
e
R
Temperature (K)
Configurational entropy (R)
Fig. 5. Summary of the properties of the ML-designed HEAs. (A) TEC of the
ML-designed HEAs as a function of the change in temperature. As a comparison,
we plotted the thermal expansion curve of the HEAs and MEAs. A3 and A9
FeNiCoCr HEAs show extremely low TECs around 2 × 10−6/K at 300 K, which can
be used as Invar alloys. B2 and B4 FeNiCoCrCu HEAs show low TECs around
5 × 10−6/K at 300 K, which qualifies them as Kovar alloys. (B) Configurational
entropy plotted against the TEC values for various known alloys and alloys
discovered in this work. ML enables this approach to efficiently discover new
alloys with excellent properties (high resistance to thermal cycles) in an infinite
phase spectrum (compositionally complex alloys).
Rao et al., Science 378, 78–85 (2022)
7 October 2022
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
We show the TEC as a function of tempera-
ture for the two Invar alloys (TEC ≈2 × 10−6/K)
and two Kovar alloys (TEC ≈5 × 10−6/K) that
we developed in Fig. 5A, compared with HEAs
and medium-entropy alloys (MEAs) (19, 42).
The new alloys show abnormally low TEC val-
ues compared with the HEAs, MEAs, and con-
ventional alloys previously reported (Fig. 5B)
(43–45). Most Invar alloys show a low TEC but
also low configurational entropy. The Invar
alloys developed in this work offer a good com-
bination of low TEC and high configurational
entropy. This indicates the high potential of
the HEA concept for the design of Invar alloys,
which, beyond their beneficial thermal expan-
sion response, also offer high strength, ductil-
ity, and corrosion resistance.
Conclusions
Understanding the underlying physics behind
composition-property relations is the key mis-
sion in alloy design, a task particularly chal-
lenging in the case of compositionally complex
materials. In principle, HEAs with interesting
features can hide in practically infinite and
vastly unexplored composition space, a sce-
nario that puts targeted alloy design to its
hardest test. We have therefore developed a
widely applicable active learning framework
that combines a generative model, regression
ensemble, physics-driven learning, and experi-
ments for the compositional design of HEAs.
Our method demonstrates its proficiency in
designing high-entropy Invar alloys using
very sparse experimental data. The entire
workflow required only a few months, in con-
trast to the conventional alloy design approach,
which requires years and many more experi-
ments. We expect that more than one prop-
erty can be optimized simultaneously using
the GAD-TERM framework in the composi-
tional spectrum of HEAs.
RE FE RENCES AND N OT ES
J. W. Yeh et al., Adv. Eng. Mater. 6, 299–303 (2004).
1.
2. B. Cantor, I. Chang, P. Knight, A. Vincent, Mater. Sci. Eng. A
375–377, 213–218 (2004).
3. E. P. George, D. Raabe, R. O. Ritchie, Nat. Rev. Mater. 4,
515–534 (2019).
4. H. Mao, H.-L. Chen, Q. Chen, J. Phase Equilibria Diffus. 38,
353–368 (2017).
7. S. Huang, E. Holmström, O. Eriksson, L. Vitos, Intermetallics
95, 80–84 (2018).
8. Z. Rao, H. Springer, D. Ponge, Z. Li, Materialia (Oxf.) 21, 101326
(2022).
42. G. Laplanche et al., J. Alloys Compd. 746, 244–255 (2018).
43. T. Schneider, M. Acet, B. Rellinghaus, E. F. Wassermann,
W. Pepperhoff, Phys. Rev. B Condens. Matter 51, 8917–8921
(1995).
9. B. O. Mukhamedov, K. V. Karavaev, I. A. Abrikosov, Phys. Rev. Mater.
44. C. Chanmuang, M. Naksata, T. Chairuangsri, H. Jain, C. E. Lyman,
5, 104407 (2021).
Mater. Sci. Eng. A 474, 218–224 (2008).
10. J. Schmidt, L. Pettersson, C. Verdozzi, S. Botti, M. A. L. Marques,
45. K. Fukamichi, M. Kikuchi, S. Arakawa, T. Masumoto, Solid State
Sci. Adv. 7, eabi7948 (2021).
Commun. 23, 955–958 (1977).
11. Z. Pei et al., Adv. Sci. (Weinh.) 8, e2101207 (2021).
12. P. V. Balachandran, B. Kowalski, A. Sehirlioglu, T. Lookman,
Nat. Commun. 9, 1668 (2018).
13. J. Gubernatis, T. Lookman, Phys. Rev. Mater. 2, 120301 (2018).
14. D. Xue et al., Nat. Commun. 7, 11241 (2016).
15. R. Yuan et al., Adv. Mater. 30, 1702884 (2018).
16. E. Wassermann, M. Acet, in Magnetism and Structure
in Functional Materials, A. Planes, L. Mañosa, A. Saxena, Eds.
(Springer, 2005), pp. 177–197.
17. M. Shiga, Curr. Opin. Solid State Mater. Sci. 1, 340–348
(1996).
18. C. L. Lin et al., Appl. Phys. Lett. 119, 171902 (2021).
19. Z. Rao et al., Phys. Rev. Mater. 5, 044406 (2021).
20. C.-L. Lin et al., Mater. Chem. Phys. 271, 124907 (2021).
21.
I. Tolstikhin, O. Bousquet, S. Gelly, B. Schoelkopf, Wasserstein
auto-encoders. arXiv:1711.01558 [stat.ML] (2017).
22. R. Gómez-Bombarelli et al., ACS Cent. Sci. 4, 268–276 (2018).
23. A. Christophe, N. de Freitas, A. Doucet, M. I. Jordan, Mach. Learn.
50, 5–43 (2003).
24. W. K. Hastings, Biometrika 57, 97–109 (1970).
25. K. Hornik, M. Stinchcombe, H. White, Neural Netw. 2, 359–366
(1989).
26. R. Kohavi, D. H. Wolpert, ICML 96, 275–283 (1996).
27. U. Von Luxburg, B. Schölkopf, in Handbook of the History of
Logic, D. M. Gabbay, S. Hartmann, J. Woods, Eds. (Elsevier,
2011), vol. 10, pp. 651–706.
28. G. Ke et al., in Proceedings of the 31st International Conference
on Neural Information Processing Systems, Long Beach, CA,
December 2017, U. von Luxburg, I. Guyon, S. Bengio,
H. Wallach, R. Fergus, Eds. (NIPS, 2017), p. 3149–3157.
29. J. H. Friedman, Comput. Stat. Data Anal. 38, 367–378 (2002).
30. A. H. Li, J. Bradic, J. Am. Stat. Assoc. 113, 660–674 (2018).
31. Materials and methods are available as supplementary materials.
32. F. Wilcoxon, in Breakthroughs in Statistics, S. Kotz, N. L. Johnson,
Eds. (Springer, 1992), pp. 196–202.
33. R. Lowry, “Concepts and applications of inferential statistics”
(DOER, 2014); http://doer.col.org/handle//123456789/4853.
34. H. Masumoto, Science Reports of the Tohoku Imperial
University 20, 101–123 (1931).
35. H. Masumoto, Science Reports of the Tohoku Imperial University
23, 265–275 (1934).
36. M. Hakaru, S. Hideo, K. Tatsuo, Science Reports of the Research
Institutes, Tohoku University. Series A, Physics, Chemistry and
Metallurgy 6, 529–538 (1954).
37. M. Hakaru, S. Hideo, S. Yutaka, Science Reports of the Research
Institutes, Tohoku University. Series A, Physics, Chemistry and
Metallurgy 7, 533–540 (1955).
38. M. Hakaru, S. Hideo, G. Kimiyosi, Science Reports of the Research
Institutes, Tohoku University. Series A, Physics, Chemistry and
Metallurgy 9, 159–169 (1957).
39. M. Hakaru, S. Hideo, S. Yutaka, Science Reports of the Research
Institutes, Tohoku University. Series A, Physics, Chemistry and
Metallurgy 9, 170–175 (1957).
40. I. A. Abrikosov et al., Phys. Rev. B Condens. Matter Mater. Phys.
46. Code for: Z. Rao et al., Machine learning-enabled high-entropy
alloy discovery, GitHub (2022); https://github.com/ziyuanrao11/
Machine-learning-enabled-high-entropy-alloy-discovery.
47. Data for: Z. Rao et al., Machine learning-enabled high-entropy
alloy discovery, Zenodo (2022); https://doi.org/10.5281/
zenodo.7019194.
AC KNOWLED GME NTS
We thank M. Acet from University of Duisburg-Essen and
M. Nellessen, M. Adamek, and F. Schlüter from Max-Planck-Institut
für Eisenforschung GmbH. The staff of TU Darmstadt is
gratefully acknowledged for providing computational resources
with the Lichtenberg high-performance computer for the exact
muffin-tin orbital (EMTO) calculations in the present work.
Funding: Z.R., R.X., O.G., and H.Z. were supported by Deutsche
Forschungsgemeinschaft (DFG, German Research Foundation
project ID 405553726-TRR 270). Y.W. was supported by BiGmax,
the Max Planck Society’s Research Network on Big-Data-Driven
Materials Science, and the ERC-CoG-SHINE-771602. P.T. was
supported by the Electron and X-Ray Microscopy Community for
Structural and Chemical Imaging Techniques for Earth materials
(EXCITE grant no. G106564) and the International Max Planck
Research School for Interface Controlled Materials for Energy
Conservation (IMPRS-SurMat). Author contributions: Z.R., Y.W.,
and D.R. conceived the study. Y.W. and Z.R. designed the active
learning framework. Z.R., Y.W., P.T., and S.B. developed the
algorithm and analyzed the results. Z.R. performed the
experiments. R.X., H.Z., A.F., P.K., and F.K. performed the DFT
calculations. P.T.S., A.K.S., and Z.R. performed the thermodynamic
calculations. Z.R., Y.W., and P.T. wrote the main parts of the
manuscript. P.T. produced the final figures. All authors discussed
the results and commented on the manuscript. Competing
interests: The authors declare no competing interests. Data and
materials availability: The training dataset is curated from the
previous publications (34–39) and can be found in (46). The
necessary data produced by our model in this work can be found in
the supplementary materials. The original code used to perform
this work is available on GitHub (46). We also provide a simplified
version of the code integrated into a single Jupyter Notebook
on GitHub (46), which is easier to perform and understand. In
addition, data are provided in the Zenodo repository (47).
License information: Copyright © 2022 the authors, some rights
reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government
works. https://www.science.org/about/science-licenses-journal-
article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abo4940
Materials and Methods
Figs. S1 to S13
Tables S1 to S16
References (48–63)
Submitted 13 February 2022; resubmitted 30 June 2022
Accepted 30 August 2022
10.1126/science.abo4940
5. F. Körmann et al., Appl. Phys. Lett. 107, 142404 (2015).
6. Z. Rao et al., Intermetallics 111, 106520 (2019).
76, 014434 (2007).
41. A. V. Ruban, Phys. Rev. B 95, 174432 (2017).
Rao et al., Science 378, 78–85 (2022)
7 October 2022
7 of 7
|
10.1126_science.abq1218
|
RES EARCH
EXPATRIATE SCHOLARS
Has China’s Young Thousand Talents program
been successful in recruiting and nurturing
top-caliber scientists?
Dongbo Shi1*, Weichen Liu2, Yanbo Wang3*
In this study, we examined China’s Young Thousand Talents (YTT) program and evaluated its effectiveness
in recruiting elite expatriate scientists and in nurturing the returnee scientists’ productivity. We find
that YTT scientists are generally of high caliber in research but, as a group, fall below the top category
in pre-return productivity. We further find that YTT scientists are associated with a post-return publication
gain across journal-quality tiers. However, this gain mainly takes place in last-authored publications and
for high-caliber (albeit not top-caliber) recruits and can be explained by YTT scientists’ access to greater
funding and larger research teams. This paper has policy implications for the mobility of scientific
talent, especially as early-career scientists face growing challenges in accessing research funding in
the United States and European Union
I mmigrants are playing an increasing role
in US science and engineering (1–3), and
China particularly has been the top sender
of international students to the US’s STEM
programs (4, 5). In recent years, China
launched an ambitious Thousand Talents Pro-
gram (TTP) to recruit elite expatriate scientists
to return to China. This program has received
intense attention both from the US govern-
ment, as reflected in the launch of the China
Initiative, and from the academic community,
especially over the Federal Bureau of Inves-
tigation’s arrest of Massachusetts Institute of
Technology professor Gang Chen.
Despite the attention, there has been little
evidence-based research on the operation, im-
pact, and policy implications of China’s talent
programs. Prior research on scientific retur-
nees has found productivity declines among
those returning to lower-income home coun-
tries, but such research has focused on countries
other than China (6). China-specific studies have
suggested that returnees face difficulties re-
integrating into the country’s research envi-
ronment (7), where administrative intervention
and personal connections hinder scientific in-
quiry (8). This raises the question of whether
China’s talent programs have been effective
in recruiting top-caliber scientists (9) and in
nurturing the returnees’ productivity (10).
Studying talent programs is important for
understanding the evolving landscape of global
knowledge production; it is also policy-relevant
because an increasing number of govern-
ments across both high-income (e.g., Canada
and Singapore) and middle- or lower-income
1School of International and Public Affairs, Shanghai Jiao
Tong University, Shanghai, China. 2School of Public Policy
and Management, Tsinghua University, Beijing, China.
3Faculty of Business and Economics, The University of Hong
Kong, Hong Kong.
*Corresponding author. Email: shidongbo@sjtu.edu.cn (D.S.);
yanbo.wang@hku.hk (Y.W.)
(e.g., Brazil and India) countries are pursuing
means to tap expatriates and migrant networks
for domestic knowledge production and talent
development. Some governments have come
to believe that expats and returnees are the
key to building globally competitive research
institutions and dynamic, knowledge-based
economies.
China’s Young Thousand Talents program
We examined China’s Young Thousand Talents
(YTT) program, the “youth” branch of the TTP.
Among the country’s 200-plus talent recruit-
ment programs, TTP is the most prominent
initiative to bring leading global scientists to
China. In principle, TTP is open to researchers
of any nationality; but in practice, few non-
Chinese have availed themselves of the program.
Established in 2010, the YTT program targets
outstanding young STEM scholars and offers
generous financial support to each awardee,
including a one-off tax-exempt income subsidy
of 500,000 yuan RMB (~$150,200 in 2010 USD
purchasing power parity) and start-up grants
of 1 million to 3 million yuan. This package is
matched by the host institution and even local
governments. All awardees are also provided
with fringe benefits such as housing subsidies
and are prioritized when applying for local and
national grants.
To be eligible, YTT applicants should ideally
(i) work in the STEM field and be 40 years of
age or younger, (ii) have a PhD from a rep-
utable overseas university and three or more
years of overseas research experience, (iii) have
a full-time overseas research position, (iv) be
committed to full-time employment in China,
and (v) be a top talent in their cohort and have
the potential to become a research-field leader.
However, these criteria are not rigid. YTT also
welcomes freshly minted overseas PhDs and
expatriate researchers with Chinese PhDs to ap-
ply if they have outstanding research records.
China’s YTT program made 3576 offers be-
tween 2011 and 2017. Although designed to im-
prove China’s prospect of becoming a global
STEM leader, the program’s effectiveness in
attracting top talents and nurturing their pro-
ductivity is unclear. On one hand, a program
providing substantive research support could
motivate expatriate talents to return and even
help grow their productivity; on the other hand,
returnees may struggle to reintegrate into
China’s academia (7–9) and thus experience a
research output slowdown. YTT scientists may
also be incentivized to focus on publication
quantity rather than quality, because program
officials have motivations to demonstrate YTT’s
impact on publication counts, even at the cost
of quality and originality.
Data and methods
We studied the YTT program’s first four co-
horts, totaling 721 awardees. Our main analy-
ses excluded 309 individuals because they
either returned for nonacademic jobs (27),
received PhDs in China (196), were not of Chi-
nese origin (34), left China within 5 years of
returning (5), or lacked CV information (47).
This left us with 73 scientists who rejected
the YTT offers to remain overseas (hereafter,
“rejectors”) and 339 returnees who received
PhDs abroad, accepted the offers, and spent
at least 5 years conducting research in China
(hereafter, “acceptors”).
We implemented two sets of analyses. First,
we examined YTT returnees’ educational cre-
dentials and pre-return productivity. We specif-
ically used rejector-versus-acceptor comparisons
to estimate the YTT program’s relative attract-
iveness to scientists across different levels of
research caliber and career opportunity. To
further contextualize YTT returnees’ research
capability, we benchmarked them against all
early-career, research-active scientists based
in the US with Chinese surnames.
Next, we evaluated the YTT program’s impact
on returnees’ productivity using a selection-
on-observables approach. We matched each
returnee with comparable “stayers,” that is, sci-
entists who attended college in China and re-
ceived a PhD in the same field from the same
overseas university (11) during the same period
(± 3 years) as the returnees but remained in
overseas academia. To ensure YTT and stayer
comparability, we further collected data on their
publications and citations and used coarsened
exact matching (CEM) to identify matched
pairs (12).
We used the difference-in-differences (DID)
method to estimate the YTT program’s impact
on the productivity of overseas-educated Chi-
nese scientists who returned. We ran Poisson
models because the outcomes are publication
counts. The supplementary materials provide
more details about the data and methods, and,
below, we report the empirical results.
Shi et al., Science 379, 62–65 (2023)
6 January 2023
1 of 4
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Publication trajectories of YTT scientists and their overseas
counterparts. The y axis reports coefficients estimated from Poisson regressions
comparing the knowledge productivity of returnee scientists with that of their
overseas counterparts for the CEM sample. The publication data are lagged by
2 years to take into consideration the necessary delay between knowledge
production and in-print publication. (A) Annual article count without differ-
entiating authorship position in publication. (B and C) First- and last-authored
publications, respectively. The bars represent the 95% confidence intervals of
the estimates. The sample includes (i) 151 returnee scientists who attended
colleges in China, received their PhDs overseas, accepted the YTT offers, and spent
at least 5 years of their professional careers in China and (ii) 340 overseas
counterparts who had similar pre-return knowledge productivity and educational
backgrounds as the YTT returnees (i.e., having attended colleges in China, received
PhDs overseas, and graduated from the same doctoral institutions in the same
fields around the same time period) but have stayed in overseas academia rather
than returning to China.
Table 1. YTT offer receiver comparison. This table compares 339 YTT offer acceptors who have returned to China with 73 YTT offer rejectors who stayed
overseas.
Covariates
Mean
Acceptors
Rejectors
Difference
P value
0.525
Articles per year
First-authored articles per year
First-authored articles in top 10% of journals per year
Last-authored articles per year
Last-authored articles in top 10% of journals per year
PhD from globally top-100 STEM program
0.701
............................................................................................................................................................................................................................................................................................................................................
Research productivity before return
............................................................................................................................................................................................................................................................................................................................................
0.098
............................................................................................................................................................................................................................................................................................................................................
0.658
............................................................................................................................................................................................................................................................................................................................................
0.040
............................................................................................................................................................................................................................................................................................................................................
0.000
............................................................................................................................................................................................................................................................................................................................................
0.001
............................................................................................................................................................................................................................................................................................................................................
0.000
Overseas faculty appointments
............................................................................................................................................................................................................................................................................................................................................
0.006
Research funding per year ($1000 in 2010 USD)
............................................................................................................................................................................................................................................................................................................................................
−0.541
−0.055
0.119
−0.412
−0.156
−0.755
−25.925
2.932
1.058
0.403
0.608
0.202
0.890
30.365
2.390
1.003
0.523
0.196
0.046
0.136
4.439
0.551
−0.026
High- but not top-caliber recruits
China’s YTT program has attracted high-caliber
researchers, more than half of whom received
PhDs from globally top-100 STEM programs (fig.
S3). These recruits were also highly productive,
averaging 2.39 publications annually in the pre-
return period. When benchmarked against all
early-career, research-active scientists based in
the US (with Chinese surnames), these scien-
tists, as a group, would rank in the top-15th
(17th) percentile for productivity (table S10).
The majority (73%) of the YTT recruits worked
overseas as postdocs or research fellows.
The program was less successful in recruit-
ing top-caliber scientists. Among YTT offer
receivers, the rejectors were more produc-
tive (2.93 versus 2.39 publications per year;
top-10th versus top-15th percentile in ranking),
more likely (89% versus 14%) to have over-
seas faculty appointments, and associated with
larger annual research grants ($30,365 versus
$4439 in 2010 USD) than were the acceptors
in the pre-return period (Table 1 and table S10).
Furthermore, the more top-journal, first-authored
publications that an offer-receiver had, the
more likely they were to have accepted the
YTT offer; in contrast, there was a negative
association between top-journal, last-authored
publications and offer acceptance (table S11).
These results jointly show that typical YTT
returnees were of high research caliber but that
their pre-return productivity was ranked right
below the top-10th percentile; they held no
faculty positions, worked in other people’s labs,
and received minimal research grants. Or, to
put it differently, while “the best are yet to come”
(9), China’s YTT program was attractive to young
expatriates who had the capability but not
the funding to run their own labs for indepen-
dent research.
Returnees’ productivity gain and research
independence
Figure 1A shows that despite an initial drop,
YTT scientists’ post-return productivity was
27.4% higher than that of their CEM-matched
overseas peers in terms of total publications
(table S14). As matched-group scientists aver-
aged 3.77 publications in 2016, an additional
1.03 (27.4% of 3.77) articles would have raised
their ranking in Microsoft Academic Graph
publication count from the 88th to the 92nd
percentile. YTT scientists’ performance gain
continued to hold when we looked only at
Shi et al., Science 379, 62–65 (2023)
6 January 2023
2 of 4
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Publication trajectories of YTT scientists and their overseas counterparts, controlling for team size and research funding. We obtained grant
information from the Dimensions database and proxied for a scientist’s team size by the annual number of unique coauthors that were listed on the scientist’s
last-authored publications and affiliated with their research institution. (A) Annual article count without differentiating authorship position in publication. (B and
C) First- and last-authored publications, respectively.
Fig. 3. Effects of the YTT program across academic fields and scientist groups. The bars represent the difference-in-differences (DID) coefficients for the
coarsened exact matching (CEM) sample across academic fields (A), pre-return overseas faculty appointments (B), and pre-return scientific productivity ranking
(C). *P < 0.1, **P < 0.05, and ***P < 0.01.
publications in high-impact journals, that is,
journals ranked among the top 50%, 25%, and
10% in field-specific journal impact factor.
Compared with the stayers, YTT returnees
had slightly fewer first-authored post-return
publications (Fig. 1B); however, returnees over-
performed in last-authored post-return pub-
lications by 144.3%, and this gain held across
journal-quality tiers (Fig. 1C). As the STEM
fields’ norm is to list the principal investiga-
tor as a publication’s last and corresponding
author, these results suggest that YTT retur-
nees were more likely to become independent
researchers pursuing their own scientific agen-
das in the post-return period than were their
overseas peers. In contrast, the stayers were
more likely to work in others’ research groups.
Numus and manus
We further investigated each scientist’s research
funding and team size. As early-career scientists
in the US and EU often lack sufficient funding
(13), YTT scientists may have benefited from
the program’s generous start-up grants and
China’s abundant supply of STEM students.
Figure 2A shows that once funding and team
size were controlled for, YTT scientists barely
outperformed the control-group scientists in
terms of publications. This result held for both
total publications and high-impact journal pub-
lications. While YTT returnees continued to pub-
lish more last-authored articles than did the
stayers, the effect size became much smaller (com-
pare Fig. 2C and Fig. 1C). For example, the incident
rate ratio dropped from 2.443 to 1.371 in overall
publications (compare table S14 and table S15).
These statistics suggest that funding and team
size play a critical role in explaining the publica-
tion gap between the returnees and the stayers.
Heterogeneity across fields and scientists
We conducted subgroup analyses across re-
search fields and scientist profiles in pre-return
employment and productivity. Figure 3A shows
that YTT returnees overperformed in the fields
of chemistry and life sciences, which require
large amounts of physical assets, financial re-
sources, and human power (14). YTT returnees
also outperformed in environmental and earth
science, engineering and material science, and
information science. However, we saw perfor-
mance loss (although it was not statistically
significant) for returnees in the fields of math-
ematics and physics. Figure 3, B and C, further
shows that the post-return publication boost
was confined to returnees who were neither
overseas faculty members nor top-ranked in
pre-return productivity.
Discussion and policy implications
Our empirical results show that China’s YTT
program has been successful in recruiting and
nurturing high-caliber scientists and that YTT
scientists outperform their overseas peers in
post-return publication, mainly owing to their
access to greater funding and larger research
teams. These results show the potential of
talent programs as a policy tool for countries
to attract expatriate scientists and promote
their productivity.
We also find that few top-caliber scien-
tists have availed themselves of this program,
suggesting room for improvement in Chinese
research institutions. With the option to pur-
sue independent research either in the US
Shi et al., Science 379, 62–65 (2023)
6 January 2023
3 of 4
RES EARCH | R E S E A R C H A R T I C L E
or in China, top-caliber expatriates remain
unlikely to return even given the YTT offers,
probably reflecting a social and cultural envi-
ronment conducive to scientific inquiry in the
US (15). The departure of Chenyang Xu—a YTT
recruit and a Breakthrough Prize winner in
mathematics—back to the US has specifically
raised questions about whether a research en-
vironment distinguished by administrative in-
terventions and personal connections could
be conducive to nurturing top-caliber scien-
tists (7–9).
This study’s context is noteworthy. Although
generously funded, the YTT program accounted
for only a small portion (0.36% in 2017) of the
Chinese central government’s academic research
and development (R&D) budget (16). China
has been increasing its higher education ex-
penditure (e.g., by 23.2%, compared with a 5.7%
increase in US higher education expenditure,
between 2018 and 2019); given YTT’s relatively
small budget share and the program’s success
in recruiting high-caliber scientists, it is highly
probable that such talent programs will be
sustained or even scaled up.
This study has important implications for
global academic mobility, because Chinese
citizens not only account for a large share of
the US and EU STEM PhD graduates but also
are among the most productive graduates (5).
As China continues to invest in higher educa-
tion and academic talent, we can expect more
Western-trained Chinese students to return to
China, although findings were mixed; whereas
a National Science Foundation survey showed
that 87% of Chinese STEM PhDs wanted to
stay in the US (15), another study revealed that
70% of them would prefer to return to China if
offered salaries comparable to what they could
expect to receive in the US (17). We can also
expect Chinese universities to become more
attractive locations for Chinese (and interna-
tional) students intending to pursue scientific
research careers—students who would other-
wise study in the US or EU.
If either of the previously mentioned sce-
narios materializes, it may disrupt the current
model of university science in the US, partic-
ularly in certain academic fields. In biomedical
research, for example, the field’s knowledge-
production function critically hinges on a
large supply of postdoctoral fellows that ac-
cept minimal compensation from these tem-
porary positions despite facing dim prospects
of finding long-term tenure-track positions
(1, 18). The success of talent programs in coun-
tries such as China, and possibly elsewhere,
would offer science-oriented international stu-
dents a viable alternative to US universities and
institutions. If this trend persists, the biomed-
ical labs in the US could be facing a shrinking
pool of foreign students, raising doubts about
their current research model’s sustainability.
Our findings also point to the need for pol-
icy adjustments to allocate more support for
young scientists. It has been documented that
a declining share of research grants has been
going to early-career researchers in the US and
EU, such that many talented young scientists
cannot get a healthy start to pursue indepen-
dent research (13, 19). The empirical evidence
from our study underscores this issue, as the
relative success of the recruits of China’s talent
program can largely be attributed to the avail-
ability of better funding and larger research
teams supporting their research. As a major
driver for researchers to stay in academia is to
pursue independent research (20), the dearth
of necessary resources in the US and EU may
not only expedite expatriates’ return decisions
but also motivate young US- and EU-born sci-
entists to seek international research oppor-
tunities (21).
RE FERENCES AND NOTES
1. R. Freeman, E. Weinstein, E. Marincola, J. Rosenbaum,
F. Solomon, Science 294, 2293–2294 (2001).
2. G. C. Black, P. E. Stephan, in American Universities in a Global
Market, C. T. Clotfelter, Ed. (Univ. of Chicago Press, 2010),
pp. 129–161.
J. Bound, S. Turner, P. Walsh, “Internationalization of U.S.
Doctorate Education,” NBER Working Paper No. 14792
(2009); http://www.nber.org/papers/w14792.
3.
4. National Science Board, National Science Foundation, Science
and Engineering Indicators 2022: The State of U.S. Science
and Engineering, NSB-2022-1 (2022); https://ncses.nsf.gov/
pubs/nsb20221.
5. P. Gaulé, M. Piacentini, Rev. Econ. Stat. 95, 698–701 (2013).
6. S. Kahn, M. MacGarvie, Rev. Econ. Stat. 98, 397–414 (2016).
7. C. Cao, D. F. Simon, in Innovation and China’s Global Emergence,
E. Baark, B. Hofman, J. Qian, Eds. (NUS Press, 2021),
pp. 90–112.
8. Y. Xie, C. Zhang, Q. Lai, Proc. Natl. Acad. Sci. U.S.A. 111,
9437–9442 (2014).
9. D. Zweig, S. Kang, H. Wang, J. Contemp. China 29, 776–791
(2020).
10. Prior studies such as (9) have examined China’s other talent
programs (e.g., the TTP and the Changjiang Scholars Program)
and found that the research quality of part-time participants
was higher than that of full-time participants. Although insightful,
these studies have examined neither the participants’ research
quality compared with that of nonparticipants nor these
programs’ impact on returnee productivity.
11. W. W. Ding, A. Ohyama, R. Agarwal, Nat. Biotechnol. 39,
1019–1024 (2021).
14. Chinese scientists in such fields may also have a “regulatory
advantage” over their US and EU peers.
15. R. Zwetsloot, J. Feldgoise, R. Dunham, “Trends in U.S.
intention-to-stay rates of international Ph.D. graduates across
nationality and STEM fields,” Center for Security and Emerging
Technology (CSET) Issue Brief (Georgetown University, 2020);
https://cset.georgetown.edu/publication/trends-in-u-s-
intention-to-stay-rates-of-international-ph-d-graduates-across-
nationality-and-stem-fields/.
16. The central government budget for the 2017 (ninth cohort)
YTT program was 1.26 billion RMB, less than 0.36% of the total
R&D budget (354.6 billion RMB) for universities and state
research institutions (http://www.stats.gov.cn/tjsj/tjgb/
rdpcgb/qgkjjftrtjgb/201908/t20190830_1694754.html).
17. R. Zeithammer, R. Kellogg, J. Mark. Res. 50, 644–663
(2013).
18. J. Miller, M. Feldman, Camb. J. Regions Econ. Soc. 7, 289–305
(2014).
19. Y. Xie, A. A. Killewald, Is American Science in Decline? (Harvard
Univ. Press, 2012).
20. S. Stern, Manage. Sci. 50, 835–853 (2004).
21. S. Williams, Science 354, 644–647 (2016).
22. D. Shi, W. Liu, Y. Wang, Replication Data for: Has China’s Young
Thousand Talents Program been Successful in Recruiting and
Nurturing Top Caliber Scientists?, version 1, Harvard Dataverse
(2022); https://doi.org/10.7910/DVN/8SR0V9.
AC KNOWLED GME NTS
We thank Z. Zhang, N. Liu, M. Li, J. Zhang, H. Zhou, and H. Zeng for
capable research assistance. We also thank W. Ding, T. Stuart,
E. Zuckerman, P. Gaulé, J. Li, W. Ng, I. Png, J. Bian, C. Marquis,
Q. Wang, M. Peng, and the audiences from the National University
of Singapore, the Taiwan Symposium on Innovation Economics
and Entrepreneurship, the University of Hong Kong, Sun Yat-Sen
University, Beijing Normal University, the Leibniz Centre for
European Economic Research (ZEW), the 2022 Academy of
Management Annual Conference, and the National Academies of
Sciences, Engineering, and Medicine for feedback and suggestions.
We are particularly grateful for the insightful input from
M. Macgarvie, an early member of the research team, on the
research design and empirical strategy. We deeply appreciate the
insightful guidance provided by our editor and three anonymous
reviewers. Funding: D.S. was supported by funding from the
National Natural Science Foundation of China (grant #71704107),
Shanghai Chen Guang Project (grant #17CG04), and Shanghai
Research Center for Innovation and Policy Evaluation. Y.W. was
supported by the National University of Singapore’s Humanities and
Social Sciences (HSS) Research Fellowship and the University of
Hong Kong’s Startup Grant. Author contributions: D.S. and Y.W.
collaboratively conceived of and designed the study. Y.W. drafted
the manuscript. D.S. and Y.W. revised and edited the manuscript.
D.S. and W.L. collected and analyzed the data. D.S. implemented
all the regressions and produced all visualizations. Competing
interests: The authors declare that they have no competing
interests. Data and materials availability: Data for replicating the
results in this paper are available in the Harvard Dataverse (22).
License information: Copyright © 2023 the authors, some rights
reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government
works. https://www.science.org/about/science-licenses-journal-
article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abq1218
Materials and Methods
Figs. S1 to S6
Tables S1 to S32
References (23–56)
12. S. Iacus, G. King, G. Porro, Polit. Anal. 20, 1–24 (2012).
13. A. I. Leshner, Science 320, 849–849 (2008).
Submitted 28 March 2022; accepted 30 November 2022
10.1126/science.abq1218
Shi et al., Science 379, 62–65 (2023)
6 January 2023
4 of 4
|
10.1126_science.abo6499
|
RES EARCH
ANIMAL MOVEMENT
Behavioral responses of terrestrial mammals
to COVID-19 lockdowns
Marlee A. Tucker1*, Aafke M. Schipper1, Tempe S. F. Adams2, Nina Attias3,4, Tal Avgar5, Natarsha L. Babic6,
Kristin J. Barker7, Guillaume Bastille-Rousseau8, Dominik M. Behr9,10, Jerrold L. Belant11, Dean E. Beyer Jr.11,
Niels Blaum12, J. David Blount13, Dirk Bockmühl14, Ricardo Luiz Pires Boulhosa15, Michael B. Brown16,17,
Bayarbaatar Buuveibaatar18, Francesca Cagnacci19, Justin M. Calabrese20,21, Rok Cˇerne22,
Simon Chamaillé-Jammes23,24, Aung Nyein Chan25,17, Michael J. Chase2, Yannick Chaval26,27,
Yvette Chenaux-Ibrahim28, Seth G. Cherry29, Duško Ćirović30, Emrah Çoban31, Eric K. Cole32, Laura Conlee33,
Alyson Courtemanch34, Gabriele Cozzi9,10, Sarah C. Davidson35,36,37, Darren DeBloois38, Nandintsetseg Dejid39,
Vickie DeNicola40, Arnaud L. J. Desbiez3,41,42, Iain Douglas-Hamilton43,44, David Drake45, Michael Egan8,27,
Jasper A.J. Eikelboom46, William F. Fagan21, Morgan J. Farmer47, Julian Fennessy16, Shannon P. Finnegan48,
Christen H. Fleming21,49, Bonnie Fournier50, Nicholas L. Fowler48,51, Mariela G. Gantchoff52,53,
Alexandre Garnier26,54, Benedikt Gehr55, Chris Geremia56, Jacob R. Goheen57, Morgan L. Hauptfleisch58,
Mark Hebblewhite59, Morten Heim60, Anne G. Hertel61, Marco Heurich62,63,64, A. J. Mark Hewison26,27,
James Hodson65, Nicholas Hoffman66, J. Grant C. Hopcraft67, Djuro Huber68, Edmund J. Isaac28,
Karolina Janik69, Miloš Ježek70, Örjan Johansson71,72, Neil R. Jordan73,74,10, Petra Kaczensky75,76,
Douglas N. Kamaru57,77, Matthew J. Kauffman78, Todd M. Kautz48, Roland Kays79,80, Allicia P. Kelly81,
Jonas Kindberg82,83, Miha Krofel84,85, Josip Kusak68, Clayton T. Lamb86, Tayler N. LaSharr87,
Peter Leimgruber17, Horst Leitner88, Michael Lierz89, John D.C. Linnell60,90, Purevjav Lkhagvaja91,
Ryan A. Long92, José Vicente López-Bao93, Matthias-Claudio Loretto35,94,95, Pascal Marchand96,
Hans Martin59, Lindsay A. Martinez97, Roy T. McBride Jr.98, Ashley A.D. McLaren99,100, Erling Meisingset101,
Joerg Melzheimer14, Evelyn H. Merrill102, Arthur D. Middleton7, Kevin L. Monteith87, Seth A. Moore28,
Bram Van Moorter60, Nicolas Morellet26,27, Thomas Morrison67, Rebekka Müller14, Atle Mysterud103,
Michael J Noonan104, David O’Connor105,106,107, Daniel Olson38, Kirk A. Olson108, Anna C. Ortega109,110,
Federico Ossi19, Manuela Panzacchi60, Robert Patchett111, Brent R. Patterson112,113, Rogerio Cunha de Paula114,
John Payne115, Wibke Peters116, Tyler R. Petroelje48, Benjamin J. Pitcher74,117, Boštjan Pokorny118,119,120,
Kim Poole121, Hubert Potocˇnik122, Marie-Pier Poulin123, Robert M. Pringle124, Herbert H.T. Prins125,
Nathan Ranc19,126,26, Slaven Reljić68,127, Benjamin Robb109, Ralf Röder14, Christer M. Rolandsen60,
Christian Rutz111, Albert R. Salemgareyev128, Gustaf Samelius72,129, Heather Sayine-Crawford65,
Sarah Schooler48, Çag˘an H. S¸ekerciog˘lu13,130,31, Nuria Selva131,132, Paola Semenzato133,19, Agnieszka Sergiel131,
Koustubh Sharma134,135,136,137, Avery L. Shawler7, Johannes Signer138, Václav Silovský70, João Paulo Silva139,140,
Richard Simon141, Rachel A. Smiley87, Douglas W. Smith56, Erling J. Solberg60, Diego Ellis-Soto142,143,144,
Orr Spiegel145, Jared Stabach17, Jenna Stacy-Dawes146, Daniel R. Stahler56, John Stephenson147,
Cheyenne Stewart148, Olav Strand60, Peter Sunde149, Nathan J. Svoboda150, Jonathan Swart151,
Jeffrey J. Thompson152,153, Katrina L. Toal141, Kenneth Uiseb154, Meredith C. VanAcker155,17,
Marianela Velilla152,153,156, Tana L. Verzuh87, Bettina Wachter14, Brittany L. Wagler87, Jesse Whittington157,
Martin Wikelski35,158, Christopher C. Wilmers159, George Wittemyer160,43, Julie K. Young161,162,
Filip Zie¸ba163, Tomasz Zwijacz-Kozica163, Mark A. J. Huijbregts1, Thomas Mueller39,164,17
COVID-19 lockdowns in early 2020 reduced human mobility, providing an opportunity to disentangle its effects
on animals from those of landscape modifications. Using GPS data, we compared movements and road
avoidance of 2300 terrestrial mammals (43 species) during the lockdowns to the same period in 2019.
Individual responses were variable with no change in average movements or road avoidance behavior, likely due
to variable lockdown conditions. However, under strict lockdowns 10-day 95th percentile displacements
increased by 73%, suggesting increased landscape permeability. Animals’ 1-hour 95th percentile
displacements declined by 12% and animals were 36% closer to roads in areas of high human footprint,
indicating reduced avoidance during lockdowns. Overall, lockdowns rapidly altered some spatial behaviors,
highlighting variable but substantial impacts of human mobility on wildlife worldwide.
I n 2020, governments around the world
introduced lockdown measures in an at-
tempt to curb the spread of the novel severe
acute respiratory syndrome coronavirus 2
(SARS CoV-2) virus. This resulted in a
drastic reduction in human mobility including
human confinement to living quarters, closure
of recreation and protected areas, and reduc-
tions in the movement of vehicles and their
associated by-products (e.g., noise and pol-
lutants) (1). This “anthropause” provides a
unique opportunity to quantify the effects of
human mobility on wildlife by decoupling
these from landscape modification effects
(e.g., roads) (2, 3). It is established that an-
thropogenic landscape modifications affect
how animals use habitats (4) and interact with
each other (5). For example, human infrastruc-
ture may induce various behavioral responses
in animals, including avoidance (6), shifts in
movement speed or habitat selection near
roads (7), and altered diurnal patterns of hab-
itat use (8). In addition to these landscape
modification effects, animals can react directly
to the presence and activity of humans (9).
These often are perceived as a risk (10), which
can lead to changes in habitat use due to the
avoidance of areas heavily used by humans,
increased energetic costs and physiological
stress (11), and altered demography (e.g., re-
duced fecundity) (12). As large-scale, high-
resolution human mobility data are rare, our
ability to decouple the effects of landscape
modification and human mobility has been
limited. In particular, little is known about
the overall impact of human mobility on ter-
restrial mammalian behavior across species
and continents. Here, we make use of the quasi-
experimental alteration of human mobility
during COVID-19 lockdowns in early 2020 to
study the effect of human mobility on ani-
mal behavior, specifically on movement and
road avoidance in terrestrial mammals.
Using animal tracking data to study behavioral
changes during lockdowns
We used global positioning system (GPS) track-
ing data to evaluate how 2300 individual ter-
restrial mammals, representing 43 species
across 76 studies (Fig. 1 and table S1), changed
their spatial behavior during the initial 2020
COVID-19 lockdowns compared with the same
time period a year earlier. For the initial 2020
lockdown period we included the date of the
first government-mandated lockdown in each
study area (between 1 February and 28 April,
2020) until 15 May, 2020. We used matching
time periods from 2019 as a baseline for com-
parison. Individuals were tracked for an av-
erage of 59 days per observation period (range:
10 to 72 days). We focused on two behaviors:
displacement distance (straight-line distance
between two consecutive GPS locations) and
distance to the nearest road. As changes in
displacement might be scale-dependent, we
considered displacements at 1-hour and 10-day
intervals based on Tucker et al. (13). Changes
in 1-hour displacements reflect immediate re-
sponses to altered human mobility (14). We
expected that reduced human mobility during
strict lockdowns would lead to an overall re-
duction in 1-hour displacements due to fewer
avoidance and escape responses, or easier ac-
cess to foraging areas due to reduced distur-
bance as has been previously shown for red
deer (14). For the 10-day displacements, we
expected a different response because previous
analyses of the effects of land-modifications
on mammal movements (13) have shown longer
displacement distances in areas with low human
footprint. Accordingly, displacement distances
Affiliations are listed at the end of this Research Article.
*Corresponding author. Email: marlee.tucker@ru.nl
Tucker et al., Science 380, 1059–1064 (2023)
9 June 2023
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
at the 10-day scale might be longer under
lockdown conditions as animals might be able
to cross barriers linked to human mobility dur-
ing such periods (e.g., roads with lower traffic
volumes). For each time scale, we evaluated
the 50th (median) and 95th percentiles of the
displacements. Median displacements rep-
resent a suite of behaviors including resting
and sleeping (1-hour scale) or residency in the
same area (10-day scale). The 95th percentile
eliminates stationary behaviors and repre-
sents longer and more directed movements
such as avoidance behaviors on the 1-hour
time scale and long-distance displacements
at the 10-day time scale (13). Because longer
displacements generally have a greater prob-
ability of encountering humans or infrastruc-
ture, we expected stronger responses for the
95th-percentile displacements.
Although roads may benefit some species by
providing foraging opportunities or movement
corridors (15), their effects are more often
negative as they not only create barriers but
also increase mortality and facilitate human
access to remote areas (16). We expected that
declines in vehicular traffic during the early
2020 lockdowns (17) would reduce the per-
ceived risk level and mammals would there-
fore be closer to roads.
To evaluate possible changes in displace-
ments or distance to the nearest roads be-
tween the lockdown and baseline periods, we
calculated log response ratios for each mea-
sure (medians and 95th percentiles of the 1-hour
and 10-day displacements, and distance to
roads) and each individual. Our analyses of the
response ratios involved a two-step process
following previous work (18). First, we used
Bayesian mixed-effects models to examine
the overall effect of lockdowns on movement
distance and distance to the nearest road (i.e.,
intercept-only model) (19). Second, we used
Bayesian mixed-effects models to examine pos-
sible relationships between the response ratios
and various covariates indicative of environ-
mental context (i.e., lockdown strictness, hu-
man footprint, and productivity) and species
traits (i.e., body mass, diet, activity, and relative
brain size) (19). For both steps of the analyses,
we included random effects for species-study
combined to account for nonindependence
between effect sizes from the same study and/or
species. For the second step of the analysis,
we included the Oxford COVID-19 government
response tracker stringency index (SI) (20) in
our models to examine country-level variation
in lockdown strictness, ranging from 0 (no
lockdown) to 100 (very strict lockdown; e.g.,
confined to home). We used the human foot-
print index [(HFI) 1-km resolution] (21) as a
proxy of direct and indirect human activities
including roads, agriculture, and human pop-
ulation density. The HFI values range from 0
to 50, where low values represent areas rela-
Stringency
Index
75
50
25
Fig. 1. Distribution of GPS data from 43 terrestrial mammal species. The map represents the mean
Oxford COVID-19 government response tracker stringency index (SI) (20), which measures lockdown
strictness, ranging from 0 (no lockdown) to 100 (very strict lockdown). Values are presented per country
during the 2020 study period (i.e., initial lockdown date to 15 May, 2020), where higher values (red)
represent countries with a stricter lockdown policy. Light gray represents countries with no SI data. SI values
range from 10 to 92. Black points represent the centroids of each study-species combination (n = 90). Map in
Mollweide projection.
2
1
0
A
B
1
2
0
)
R
R
(
e
z
i
S
t
c
e
ff
E
t
n
e
m
e
v
o
M
e
Fig. 2. Changes in 1-hour
animal movement during the
COVID-19 lockdowns. (A) Overall
reduction in the 1-hour 95th-
percentile displacements (inter-
cept-only model). (B) Overall
reduction in the 10-day 95th-
percentile displacements (inter-
cept-only model). Colored points
represent individuals (n = 423 and
1725), with point sizes propor-
tional to the inverse sampling
variance of the response ratio for
each individual. The black points
and error bars indicate the overall
effect with 95% CI. The 1-hour
95% CI do not overlap 0 (−0.25 to
−0.01) but the 10-day CI did
overlap 0 (−0.36 to 0.05). Nega-
tive values indicate reduced
movement distances during the
early 2020 lockdowns whereas
positive values indicate increased movement distances during the lockdowns.
)
R
R
(
e
z
i
S
t
c
e
ff
E
t
n
e
m
e
v
o
M
e
l
i
t
n
e
c
r
e
P
h
t
5
9
y
a
D
-
0
1
l
i
t
n
e
c
r
e
P
h
t
5
9
r
u
o
H
-
1
-3
-2
-1
-2
-4
-6
tively undisturbed by humans and high values
represent areas with high human develop-
ment levels. We expected stronger behavioral
responses to lockdowns in areas with a higher
human footprint and in countries with stricter
lockdowns for both displacement distances
and distance to roads. To account for movement
capacity, differences in movements related to
diet, activity cycle, and behavioral flexibility,
we included body mass (range: 10 to 4000 kg),
diet (carnivore, omnivore, herbivore), activity
(diurnal or nocturnal), and relative brain size as
additional explanatory variables. Finally, we
also included the between-year difference in
normalized difference vegetation index (NDVI)
between 2019 and 2020 to account for potential
differences in seasonality and productivity. We
fit models for the median and 95th percentile
of the 1-hour and 10-day displacements, and
for distance to roads including all covariates
for lockdown strictness, environmental context,
and species traits (19). We report our results as
the percentage increase or decrease in move-
ment distance or distance to roads by back-
transforming the response ratios (19) and
reporting the 95% credible intervals (CI).
Tucker et al., Science 380, 1059–1064 (2023)
9 June 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
4A
)
R
R
(
e
z
i
S
t
c
e
ff
E
t
n
e
m
e
v
o
M
e
l
i
t
n
e
c
r
e
P
h
t
5
9
y
a
D
-
0
1
2
0
-2
-4
-6
B
4
)
R
R
(
e
z
i
S
t
c
e
ff
E
t
n
e
m
e
v
o
M
e
l
i
t
n
e
c
r
e
P
h
t
5
9
y
a
D
-
0
1
2
0
-2
-4
-6
50
60
70
Stringency Index
80
90
100
- 2
-1
0
1
2
NDVI difference (scaled)
Fig. 3. Changes in 10-day animal movement during the COVID-19 lockdowns.
(A) Increasing 10-day 95th-percentile displacements in response to the Stringency
Index and (B) 10-day 95th-percentile displacements were longer during 2020 when
we observed higher NDVI values compared with 2019. Colored points represent
individuals (n = 1725), with point size proportional to the inverse sampling variance
of the response ratio for each individual. The black line is the fitted effect size
(response ratio; RR). The shaded area indicates 95% CI, and the dashed gray line at
zero illustrates no change. Negative values indicate reduced movement distances
during the early 2020 lockdowns whereas positive values indicate increased
movement distances during the lockdowns.
Changes in movement displacements
during lockdowns
We found an average 12% reduction in 1-hour
95th-percentile displacements when evaluat-
ing the impact of only the lockdown itself
(intercept-only model, 95% CI: 1 and 22%,
Fig. 2 and table S2). This may indicate re-
duced avoidance and escape behavior of hu-
mans (e.g., no need to travel longer distances
to avoid humans) (22, 23) as a result of al-
tered human mobility levels during lockdowns.
When exploring potential correlates of this
response, no covariates had an effect that dif-
fered from zero (table S3). For the 1-hour me-
dian displacements, we found no overall effect
(table S2) and again, no effect of the covar-
iates (table S4). Taken together, these results
suggest that responses at the 1-hour scale were
highly variable and not dependent on the se-
lected species traits (body mass, diet, activity,
or relative brain size) or on the variables de-
scribing the local context (lockdown stringency
or HFI).
The overall lockdown response was not dif-
ferent from zero for the 10-day 95th-percentile
or long-distance displacements (15%, 95% CI;
−30 to 5%; Fig. 2B and table S2). However,
when exploring the covariates that might ex-
plain variation in response ratios the 95% CI
of the stringency index did not overlap zero
(table S5), with displacements increasing 73%
on average in areas of stricter lockdown (i.e.,
areas with an SI of 90; Fig. 3A). This may
indicate that tighter restrictions on human
movements, including confinement to living
spaces and reduced human mobility in green
spaces (e.g., Italy and France; Fig. 1) led to
increased landscape permeability for mam-
mals. This effect of human mobility is sim-
ilar in magnitude to previous work that used
the same displacement metric but examined
the effect of permanent landscape alterations
(land conversion and infrastructure) on ter-
restrial mammal movements (13). Although
this work used a spatial comparison rather
than comparing changes over time within the
same individuals, they found a decline of 67%
of the 10-day 95th-percentile displacements in
areas where the human footprint is high (13).
We found no effect of the remaining covariates
(HFI, body mass, diet, activity, or relative brain
size) (table S5).
We found that the 10-day 95th-percentile
displacements in areas with lower lockdown
stringency (SI values 50 to 70) were actually
shorter (on average 22 to 72%) during the
lockdown than in 2019 (Fig. 3A). The re-
duction in movement may reflect increased
human mobility in seminatural areas such
as parks and other green spaces (24, 25). In
fact, green space use by people in some areas
of intermediate lockdown increased up to
350% (25). In addition to the lockdown effects,
seasonality played a role in determining 10-day
movement distances. The 10-day median (fig.
S1) and 95th percentile (Fig. 3B) displacements
were longer during 2020, when we observed
higher NDVI values compared with 2019, which
may have led some individuals to begin their
spring migration or reproduction earlier in
2020. For the 10-day median displacements,
we found no overall lockdown effect (table S2),
no effect of lockdown stringency, and no ef-
fects of the other covariates (HFI, body mass,
diet, activity, or relative brain size) (table S6).
This difference in responses between 95% and
median movements suggests that lockdown
stringency may have affected mainly wide-
ranging behavior such as migratory move-
ments, long-distance dispersal, exploratory
excursions, or long displacements within in-
dividuals’ home ranges.
Mammals were closer to roads during lockdowns
We found no overall lockdown response in the
distance of individuals to roads (−1%, 95% CI;
−5 to 3%, table S2) nor a relationship with the
Stringency Index, NDVI difference, or species
traits (table S7). However, the response ratios
were negatively related to HFI, showing that
animals in areas with a high human footprint
were on average 36% closer to roads during
lockdown (HFI = 36, Fig. 4). In many parts of
the world, traffic volume was substantially
reduced during lockdowns (26, 27), which in
turn lessened the impact of roads on animals,
including reduced barrier effects (15, 28) and
road-kill numbers (17, 29). Our findings add con-
text to these previous results by demonstrating
that not only were road-kill numbers lower dur-
ing lockdown (17, 29), but also animals were
closer on average to roads in human-modified
areas, indicating reduced avoidance.
Overall, we detected three main signals of a
lockdown effect on terrestrial mammal behavior,
Tucker et al., Science 380, 1059–1064 (2023)
9 June 2023
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Changes in animal
distance to roads during
the COVID-19 lockdowns.
Decreasing distance to roads in
response to the human footprint
index (HFI). Colored points
represent individuals (n = 2160),
with point size proportional to
the inverse sampling variance of
the response ratio for each
individual. The black line is the
predicted effect size (response
ratio; RR). The shaded area
indicates 95% CI, and the dashed
gray line at zero illustrates no
change. Negative values indicate
closer proximity to roads during
the early 2020 lockdowns,
whereas positive values indicate
increased distance from roads
during the lockdowns.
)
R
R
(
e
z
i
S
t
c
e
ff
E
d
a
o
R
o
t
e
c
n
a
t
s
i
D
2
0
-2
-4
0
4
1 6
36
Human Footprint Index
although they were heterogeneously distrib-
uted across species and populations. These
were (i) reductions in 1-hour 95th-percentile
displacements suggesting relaxed avoidance
behavior, reduced disturbance, and/or fewer
escape responses, (ii) increased 10-day 95th-
percentile displacements under strict lockdown
conditions, suggesting increased landscape
permeability, and (iii) closer proximity to roads
in areas heavily used by humans, suggesting
a reduction in traffic disturbance. A number
of species-specific case studies are consistent
with these findings. For example, evidence
suggests that during the lockdowns, mountain
lions’ (Puma concolor) usual aversion to urban
edges ceased (9), crested porcupine (Hystrix
cristata) abundance increased in urban areas
(30), diurnal activity of invasive Eastern cot-
tontails (Sylvilagus floridanus) increased (30),
and brown bears (Ursus arctos) exploited novel
connectivity corridors (12).
Despite these three general responses to
the lockdowns considerable variation in re-
sponses existed across species and study re-
gions (Fig. 2). This variability highlights that
lockdown impacts are highly context-dependent.
For example, mountain lions explored more
urban areas during the lockdown whereas
other species including American black bears
(Ursus americanus), bobcats (Lynx rufus), and
coyotes (Canis latrans) in the same areas did
not (31). In addition, in our study lockdown
stringency was only measured at the country
level and did not account for local variability
in restrictions. We also note that our data were
predominantly from Europe and North Amer-
ica so our results should be interpreted with
caution for other regions. Finally, we note that
a given movement metric could capture differ-
ent behaviors in different species, especially at
the 10-day scale, whereas displacements could
capture behaviors ranging from within home
range movements to dispersal.
We show that human mobility is a key driver
of some terrestrial mammal behaviors, with a
magnitude potentially similar to that of land-
scape modifications. Therefore, when evalu-
ating human impacts on animal behavior or
designing mitigation measures both phys-
ical landscape alteration and human mobility
should be taken into consideration [see also
(32)]. Disentangling the effects of human mo-
bility and landscape modification will allow
the implementation of conservation measures
specifically targeted at mitigating the impacts
of human mobility, such as enticements to
adjust timing, frequency, and volume of traffic
in areas important for animal movement. Mam-
mals have been living with human disturbance
for a long time, but we demonstrate that many
wildlife populations retain the capacity to re-
spond to changes in human behavior, providing
a positive outlook for future mitigation strat-
egies designed to maintain animal movement
and the ecosystem functions they provide.
RE FERENCES AND NOTES
1. C. Rutz, Nat. Rev. Earth Environ. 3, 157–159 (2022).
2. C. Rutz et al., Nat. Ecol. Evol. 4, 1156–1159 (2020).
3. A. E. Bates et al., Biol. Conserv. 263, 109175 (2021).
4. M. Ciach, Ł. Pęksa, Curr. Zool. 65, 129–137 (2019).
5. C. A. DeMars, S. Boutin, J. Anim. Ecol. 87, 274–284 (2018).
6. M. Howe, M. M. Okello, J. M. Davis, Afr. Zool. 48, 159–166 (2015).
7. M. A. Scrafford, T. Avgar, R. Heeres, M. S. Boyce, Behav. Ecol.
29, 534–542 (2018).
8. W. Neumann, G. Ericsson, H. Dettki, V. C. Radeloff, Landsc.
Urban Plan. 114, 9–27 (2013).
9. C. C. Wilmers, A. C. Nisi, N. Ranc, Curr. Biol. 31, 3952–3955.e3
(2021).
10. K. M. Gaynor, J. S. Brown, A. D. Middleton, M. E. Power,
J. S. Brashares, Trends Ecol. Evol. 34, 355–368 (2019).
11. B. A. Nickel, J. P. Suraci, A. C. Nisi, C. C. Wilmers, Proc. Natl.
Acad. Sci. U.S.A. 118, e2004592118 (2021).
12. A. Corradini et al., Biol. Conserv. 253, 108818 (2021).
13. M. A. Tucker et al., Science 359, 466–469 (2018).
14. P. Sunde, C. R. Olesen, T. L. Madsen, L. Haugaard, Wildl. Biol.
15, 454–460 (2009).
15. J. E. Hill, T. L. DeVault, J. L. Belant, Mammal Rev. 51, 214–227
(2021).
16. A. W. Coffin, J. Transp. Geogr. 15, 396–406 (2007).
17. M. Bíl et al., Biol. Conserv. 256, 109076 (2021).
18. W. Viechtbauer, J. Stat. Softw. 36, 1–48 https://www.jstatsoft.
org/v36/i03/ (2010).
19. See Supplementary Materials and Methods.
20. T. Hale et al., Nat. Hum. Behav. 5, 529–538 (2021).
21. B. A. Williams et al., One Earth 3, 371–382 (2020).
22. J. K. Rogala et al., Ecol. Soc. 16, art16 (2011).
23. C. M. Prokopenko, M. S. Boyce, T. Avgar, J. Appl. Ecol. 54,
470–479 (2017).
24. H. Burnett, J. R. Olsen, N. Nicholls, R. Mitchell, BMJ Open 11,
e044067 (2021).
25. D. C. Geng, J. Innes, W. Wu, G. Wang, J. For. Res. 32, 553–567 (2021).
26. S. S. Patra, B. R. Chilukuri, L. Vanajakshi, Transp. Lett. 13, 1–9 (2021).
27. Ò. Saladié, E. Bustamante, A. Gutiérrez, Transp. Res. Interdiscip.
Perspect. 8, 100218 (2020).
28. G. Shannon, K. R. Crooks, G. Wittemyer, K. M. Fristrup,
L. M. Angeloni, Behav. Ecol. 27, 1370–1375 (2016).
29. F. Shilling et al., Biol. Conserv. 256, 109013 (2021).
30. R. Manenti et al., Biol. Conserv. 249, 108728 (2020).
31. R. Vardi, O. Berger-Tal, U. Roll, Biol. Conserv. 254, 108953 (2021).
32. B. A. Nickel, J. P. Suraci, M. L. Allen, C. C. Wilmers, Biol. Conserv.
241, 108383 (2020).
33. M. A. Tucker et al., Supplementary data for Behavioral responses of
terrestrial mammals to COVID-19 lockdowns. Dryad (2023);
https://doi.org/10.5061/dryad.c59zw3rbd.
34. M. A. Tucker et al., Supplementary code for Behavioral responses
of terrestrial mammals to COVID-19 lockdowns. Zenodo (2023);
https://doi.org/10.5281/zenodo.6915169.
35. M. A. Tucker et al., Supplementary spatial data for Behavioral
responses of terrestrial mammals to COVID-19 lockdowns.
Zenodo (2023).
AC KNOWLED GME NTS
We thank numerous people helping with fieldwork including the
Ktunaxa Nation for their support of the Elk Valley grizzly bear
collaring project, NP Tara (brown bear research), Tanzania Wildlife
Research Institute and Tanzania National Parks, the leadership of
Gorongosa National Park for allowing and facilitating research
on spiral-horned antelope, the farmers of the Seeis, Hochfeld, and
Auas Oanob Conservancies for collaboration and the Namibian
Ministry of Environment, Forestry, and Tourism. We also thank all
in Madison, WI, USA, who support the UW Urban Canid Project.
Collaring of khulan in Mongolia’s South Gobi Region was conducted
within the framework of the Oyu Tolgoi LLC (OT) Core Biodiversity
Monitoring Program, implemented by the Wildlife Conservation
Society (WCS) through a cooperative agreement with Sustainability
East Asia LLC (SEA). We thank the field team from Tatra National
Park (Poland). We want to thank the Mongolian Ministry of Nature,
Environment, and Tourism and staff from WCS, OT, SEA, and
protected areas for the logistical and practical support during khulan
capture. A.K., J.H., B.F., and H.S.C. wish to thank the Indigenous
governments and organizations across the Northwest Territories’
boreal and barren-ground caribou ranges for their support of
caribou monitoring programs. We wish to thank G. Mowat and
L. Smit for their support on the Elk Valley grizzly bear collaring
project. We thank Save the Elephants and the Welgevonden Game
Reserve for their elephant tracking data. We thank the Eastern
Shoshone and Northern Arapahoe Fish and Game Department and
the Wyoming State Veterinary Laboratory for assistance with the
Wyoming bighorn sheep project. The Afognak and Sitkalidak
islands elk and bear project was implemented through a
cooperative effort of the Alaska Department of Fish and Game,
Kodiak Brown Bear Trust, Rocky Mountain Elk Foundation,
Koniag Native Corporation, Old Harbor Native Corporation,
Afognak Native Corporation, Ouzinkie Native Corporation, Natives
of Kodiak Native Corporation and the State University of New York,
College of Environmental Science and Forestry. We also thank
four anonymous reviewers for comments on the manuscript.
Funding: Supported by the Radboud Excellence Initiative
(to M.T.), the German Federal Ministry of Education and Research
[MORESTEP, 01LC1710A and 01LC1820A (to T.M. and N.D.)],
the National Science Foundation [IIBR 1915347 (to J.M.C., C.H.F.,
and W.F.F.)], Serbian Ministry of Education, Science and
Technological Development [451-03-68/2022-14/ 200178 (to
D.C.)], Dutch Research Council NWO program “Advanced
Instrumentation for Wildlife Protection” (to H.H.T.P. and J.A.J.E.),
Fondation Segré, RZSS, IPE, Greensboro Science Center, Houston
Zoo, Jacksonville Zoo and Gardens, Nashville Zoo, Naples Zoo, Reid
Park Zoo, Miller Park, WWF, ZCOG, Zoo Miami, Zoo Miami
Foundation, Beauval Nature, Greenville Zoo, Riverbanks zoo and
Tucker et al., Science 380, 1059–1064 (2023)
9 June 2023
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
garden, SAC Zoo, La Passarelle Conservation, Parc Animalier
d’Auvergne, Disney Conservation Fund, Fresno Chaffee zoo,
Play for nature, North Florida Wildlife Center, Abilene Zoo, a Liber
Ero Fellowship (to C.T.L.), the Fish and Wildlife Compensation
Program, Habitat Conservation Trust Foundation, Teck Coal (to
K.P.), and the Grand Teton Association. The collection of Norwegian
moose data was funded by the Norwegian Environment Agency,
the German Ministry of Education and Research via the SPACES II
project ORYCS [FKZ:01LL1804A (to N.B.)], the Wyoming Game
and Fish Department, Wyoming Game and Fish Commission,
Bureau of Land Management, Muley Fanatic Foundation (including
Southwest, Kemmerer, Upper Green, and Blue Ridge Chapters),
Boone and Crockett Club, Wyoming Wildlife and Natural Resources
Trust, Knobloch Family Foundation, Wyoming Animal Damage
Management Board, Wyoming Governor’s Big Game License
Coalition, Bowhunters of Wyoming, Wyoming Outfitters and Guides
Association, Pope and Young Club, US Forest Service, US Fish and
Wildlife Service, the Rocky Mountain Elk Foundation, Wyoming Wild
Sheep Foundation, Wild Sheep Foundation, Wyoming Wildlife/
Livestock Disease Research Partnership, the US National Science
Foundation [IOS-1656642 and IOS-1656527 (to R.A.L. and R.M.P.)],
the Spanish Ministry of Economy, Industry and Competitiveness
[RYC-2015-18932; CGL2017-87528-R AEI/FEDER EU (to J.V.L.B.)],
and by a GRUPIN research grant from the Regional Government of
Asturias [IDI/2021/000075 (to J.V.L.B.)], Sigrid Rausing Trust,
Batubay Özkan, Barbara Watkins, NSERC Discovery Grant [RGPIN-
2021-02758 (to M.J.N.)], the Federal Aid in Wildlife Restoration act
under Pittman-Robertson project [AKW-12 (to N.J.S.)], the State
University of New York, College of Environmental Science and
Forestry, the Ministry of Education, Youth and Sport of the Czech
Republic [CZ.02.1.01/0.0/0.0/16_019/0000803; CZ.02.2.69/0.0/
0.0/19_073/0016944 (to M.J., M.S.P., and V.S.)], the Ministry of
Agriculture of the Czech Republic [QK1910462 (to M.J., M.S.P.,
and V.S.)], Rufford Foundation [grant 29681-1(to D.N.K.)], an
American Society of Mammalogists African Graduate Student
Research Fund (to D.N.K.), the German Science Foundation [HE
8857/1-1 (to A.G.H.)], the Israeli Science Foundation [grant
396/20 (to O.S.)], the BSF-NSF [2019822 and IOS2015662 (to O.S.)],
the Ministry of Agriculture, Forestry and Food and Slovenian
Research Agency (CRP V1-1626), the Aage V. Jensen Naturfond
(project: Kronvildt - viden, værdier og værktøjer), the Deutsche
Forschungsgemeinschaft (DFG, German Research Foundation)
under Germany’s Excellence Strategy [EXC 2117 – 422037984 (to
M.W.)], National Centre for Research and Development in Poland
POLNOR/198352/85/2013 (project GLOBE), the Slovenian
Research Agency [P4-0059 and N1-0163 (to M.K.)], the David
Shepherd Wildlife Foundation, Disney Conservation Fund, Whitley
Fund for Nature, Acton Family Giving, Zoo Basel, Columbus,
Bioparc de Doué-la-Fontaine, Zoo Dresden, Zoo Idaho, Kolmården
Zoo, Korkeasaari Zoo, La Passarelle, Zoo New England, Tierpark
Berlin, Tulsa Zoo, the Ministry of Environment and Tourism,
Government of Mongolia, the Mongolian Academy of Sciences, the
Federal Aid in Wildlife Restoration act and the Illinois Department
of Natural Resources, the National Science Foundation [LTREB
1556248 and 2038704 9 (to M.H., E.H.M., and H.M.)], Parks
Canada, Natural Sciences and Engineering Research Council,
Alberta Environment and Parks, Rocky Mountain Elk Foundation,
Safari Club International and Alberta Conservation Association, the
Consejo Nacional de Ciencias y Tecnología (CONACYT) of
Paraguay [14-INV-208 and PRONII], the Norwegian Environment
Agency and the Swedish Environmental Protection Agency, EU
funded Interreg SI-HR 410 Carnivora Dinarica project, Paklenica
and Plitvice Lakes National Parks, UK Wolf Conservation Trust,
EURONATUR and Bernd Thies Foundation, the Messerli Foundation
in Switzerland and WWF Germany, the European Union’s
Horizon 2020 research and innovation program under the
Marie Skłodowska-Curie Actions [grant agreement 798091 (to M.-C.L.)],
NASA Ecological Forecasting Program [80NSSC21K1182], the
Ecotone Telemetry company (to S.C.-J.), the French National
Research Agency (to S.C.-J.), LANDTHIRST [ANR-16-CE02-0001-01
(to S.C.-J.)], grant REPOS awarded by the i-Site MUSE thanks to the
“Investissements d’avenir” program [ANR-16-IDEX-0006 (to S.C.-J.)],
the ANR Mov-It project [ANR-16-CE02-0010 (to A.J.M.H. and N.M.)],
the USDA Hatch Act Formula Funding (MSN201473), the Fondation
Segre and North American and European Zoos listed at http://
www.giantanteater.org/, the Utah Division of Wildlife Resources (to
D.O., T.A., and J.K.Y.), the Yellowstone Forever and the National
Park Service (to D.R.S, D.W.S., and C.G.), Missouri Department of
Conservation, Federal Aid in Wildlife Restoration Grant PROJECT
MO W103-R1, and State University of New York, SNSF [31003A_
182286 and 310030_204478 (to G.C.)], various donors to the
Botswana Predator Conservation Program (to G.C.), data from
collared caribou in the Northwest Territories (to A.P.K., J.H., B.F.,
and H.S.C.) were made available through funds from the Department
of Environment and Natural Resources, Government of the
Northwest Territories. The European Research Council Horizon2020
[AfricanBioServices 641918], the British Ecological Society, the
Paul Jones Family Trust, and the Lord Kelvin Adam Smith fund, the
Tanzania Wildlife Research Institute and Tanzania National Parks
(to J.G.C.H. and T.M.). The Eastern Shoshone and Northern Arapahoe
Fish and Game Department and the Wyoming State Veterinary
Laboratory, the Alaska Department of Fish and Game, Kodiak
Brown Bear Trust, Rocky Mountain Elk Foundation, Koniag Native
Corporation, Old Harbor Native Corporation, Afognak Native
Corporation, Ouzinkie Native Corporation, Natives of Kodiak Native
Corporation and the State University of New York, College of
Environmental Science and Forestry, and the Slovenia Hunters
Association and Slovenia Forest Service. F.C. was partly supported by
the Resident Visiting Researcher Fellowship, IMéRA/Aix-Marseille
Université, Marseille. This work was partially funded by the Center of
Advanced Systems Understanding (CASUS), which is financed by
Germany’s Federal Ministry of Education and Research (BMBF) and
by the Saxon Ministry for Science, Culture and Tourism (SMWK) with
tax funds on the basis of the budget approved by the Saxon State
Parliament. This article is a contribution of the COVID-19 Bio-Logging
Initiative, which is funded in part by the Gordon and Betty Moore
Foundation (GBMF9881) and the National Geographic Society
(NGS-82515R-20) (both grants to C.R.). Author contributions:
M.A.T. and T.M. conceived the manuscript; M.A.T. conducted the
analyses and M.A.T. and T.M. wrote the first manuscript draft.
Co-authors contributed data and assisted with writing the final version
of the manuscript. Competing Interests: H.H.T.P. is a Member of
the Welgevonden Scientific Advisory Committee and A.D.M. is a
Senior Advisor for Wildlife Conservation for the US Department of
Agriculture. C.R. is the President of the International Bio-Logging
Society, a member of an expert group providing advice on
animal culture and social complexity to the Convention on the
Conservation of Migratory Species of Wild Animals (CMS), and
member of the advisory committee of a WILDLABS research
program aimed at identifying research and funding priorities in
movement ecology. Data and materials availability: The full
dataset used in the final analyses (33) and associated code (34)
are available at Dryad. A subset of the spatial coordinate datasets
is available at Zenodo (35). Certain datasets of spatial coordinates
will be available only through requests made to the authors due
to conservation and Indigenous sovereignty concerns (see table S1
for more information on data use restrictions and contact information
for data requests). These sensitive data will be made available
upon request to qualified researchers for research purposes,
provided that the data use will not threaten the study populations,
such as by distribution or publication of the coordinates or detailed
maps. Some datasets, such as those overseen by government
agencies, have additional legal restrictions on data sharing, and
researchers may need to formally apply for data access. Collaborations
with data holders are generally encouraged, and in cases where
data are held by Indigenous groups or institutions from regions
that are under-represented in the global science community,
collaboration may be required to ensure inclusion. License
information: Copyright © 2023 the authors, some rights reserved;
exclusive licensee American Association for the Advancement of
Science. No claim to original US government works. https://www.
sciencemag.org/about/science-licenses-journal-article-reuse
1Department of Environmental Science, Radboud Institute for
Biological and Environmental Sciences, Radboud University, P.O.
Box 9010, 6500, GL Nijmegen, Netherlands. 2Elephants Without
Borders, P.O. Box 682, Kasane, Botswana. 3Instituto de Con-
servação de Animais Silvestres (ICAS), Campo Grande, Mato
Grosso do Sul, Brazil. 4Department of Wildlife Ecology and
Conservation, University of Florida, Gainesville, FL, USA. 5Depart-
ment of Wildland Resources and the Ecology Center, Utah State
University, Logan, UT 84322, USA. 6School of Biological Sciences,
Monash University, Clayton, Victoria 3800, Australia. 7Department of
Environmental Science, Policy, and Management, University of
California, Berkeley, Berkeley, CA 94720, USA. 8Cooperative Wildlife
Research Laboratory, Southern Illinois University, Carbondale, IL,
62901, USA. 9Department of Evolutionary Biology and Environmental
Studies, University of Zurich, Winterthurerstrasse 190, CH - 8057
Zürich, Switzerland. 10Botswana Predator Conservation, Private Bag
13, Maun, Botswana. 11Department of Fisheries and Wildlife, Michigan
State University, 480 Wilson Road, East Lansing, MI 48824, USA.
12University of Potsdam, Plant Ecology and Nature Conservation, Am
Mühlenberg 3, 14476 Potsdam, Germany. 13School of Biological
Sciences, University of Utah, 257 S 1400 E, Salt Lake City, UT 84112,
USA. 14Department of Evolutionary Ecology, Leibniz Institute for
Zoo and Wildlife Research, Alfred-Kowalke-Str. 17, 10315 Berlin,
Germany. 15Instituto Pró-Carnívoros, Atibaia, SP, 12945010 Brazil.
16Giraffe Conservation Foundation, Eros, PO Box 86099, Windhoek,
Namibia. 17Smithsonian National Zoo and Conservation Biology
Institute, Conservation Ecology Center, 1500 Remount Rd, Front Royal,
VA, 22630, USA. 18Wildlife Conservation Society, Mongolia
Program, Ulaanbaatar, Mongolia. 19Animal Ecology Unit, Research
and Innovation Centre, Fondazione Edmund Mach, Via E. Mach 1,
38010 San Michele all’Adige, Italy. 20Center for Advanced Systems
Understanding (CASUS), Goerlitz, Germany. 21Department of
Biology, University of Maryland, College Park, 4094 Campus
Dr, College Park, MA, USA. 22Slovenia Forest service,
Večna pot 2, 1000 Ljubljana, Slovenia. 23CEFE, CNRS, Univ
Montpellier, EPHE, IRD, Montpellier, France. 24Mammal Research
Institute, Department of Zoology and Entomology, University of
Pretoria, South Africa. 25Dept. Fish, Wildlife and Conservation
Biology, Colorado State University, Fort Collins, CO 80525, USA.
26Université de Toulouse, INRAE, CEFS, F-31326 Castanet-Tolosan,
France. 27LTSER ZA PYRénées GARonne, F-31320 Auzeville-Tolosane,
France. 28Department of Biology and Environment, Grand Portage
Band of Lake Superior Chippewa, Grand Portage, MN 55605, USA.
29Parks Canada Agency, Box 220, Radium Hot Springs, BC, V0A
1M0, Canada. 30Faculty of Biology, University of Belgrade,
Studentski trg 16, 11000 Belgrade, Serbia. 31KuzeyDoğa Society,
Ortakapı Mah. Şehit Yusuf Cad. 69, 36100 Kars, Turkey.
32U.S. Fish and Wildlfe Service, National Elk Refuge, PO Box 510,
Jackson, WY 83001, USA. 33Missouri Department of Conservation,
Columbia, MO, 65201, USA. 34Wyoming Game and Fish Depart-
ment, Jackson, WY 83001, USA. 35Department of Migration, Max
Planck Institute of Animal Behavior, 78315 Radolfzell, Germany.
36Department of Biology, University of Konstanz, 78464 Konstanz,
Germany. 37Department of Civil, Environmental and Geodetic
Engineering, The Ohio State University, 43210 Columbus, OH, USA.
38Utah Division of Wildlife Resources. 39Senckenberg Biodiversity
and Climate Research Centre, Senckenberganlage 25, 60325
Frankfurt am Main, Germany. 40White Buffalo Inc., 26 Davison
Road, Moodus, CT 06469, USA. 41Royal Zoological Society of
Scotland (RZSS), Murrayfield, Edinburgh, UK. 42Instituto de
Pesquisas Ecológicas (IPÊ), Nazaré Paulista, São Paulo, Brazil.
43Save the Elephants, Marula Manor, Marula Lane, Karen, Nairobi
00200, Kenya. 44Department of Zoology, Oxford University, Oxford
OX1 3PS, UK. 45Department of Forest and Wildlife Ecology,
University of Wisconsin-Madison, Madison, WI 53706, USA.
46Wildlife Ecology and Conservation Group, Wageningen University
and Research, Droevendaalsesteeg 3a, 6708 PB, Wageningen,
Netherlands. 47Department of Forest and Wildlife Ecology, Univer-
sity of Wisconsin, 1630 Linden Drive, Madison, WI 53706, USA.
48Global Wildlife Conservation Center, State University of New York
College of Environmental Science and Forestry, 1 Forestry Drive,
Syracuse, NY 13210, USA. 49Smithsonian Conservation Biology
Institute, 1500 Remount Rd, Front Royal, VA, USA. 50Wildlife and
Fish Division, Department of Environment and Natural Resources,
Government of the Northwest Territories, P.O. Box 1320,
Yellowknife, NT, Canada. 51Alaska Department of Fish and Game,
43961 Kalifornsky Beach Road, Suite B, Soldotna, AK 99669, USA.
52State University of New York College of Environmental Science
and Forestry, Syracuse, NY 13210, USA. 53Department of Biology,
College of Arts and Sciences, University of Dayton, Dayton, OH
45469, USA. 54Parc National des Pyrénées, 65000 Tarbes, France.
55Department of Evolutionary Biology and Environmental Studies,
University of Zurich, Winterthurerstrasse 190, 8057 Zurich,
Switzerland. 56Yellowstone Center for Resources, P.O. Box 168,
Yellowstone National Park, WY 82190, USA. 57Department of
Zoology and Physiology, University of Wyoming, Laramie, WY
82071, USA. 58Biodiversity Research Centre, Namibia University of
Science and Technnology Pvt bag 13388 Windhoek, Namibia.
59Wildlife Biology Program, Franke College of Forestry and
Conservation, University of Montana, Missoula, MT, 59801, USA.
60Norwegian Institute for Nature Research, Terrestrial Ecology
Department, P.O. Box 5685 Torgarden, 7485 Trondheim, Norway.
61Behavioural Ecology, Department of Biology, Ludwig Maximilian
University of Munich, Großhaderner Str. 2, 82152 Planegg-
Martinsried, Germany. 62Department of Visitor Management and
National Park Monitoring, Bavarian Forest National Park, Freyunger
Straße 2, 94481 Grafenau, Germany. 63Chair of Wildlife Ecology
and Conservation Biology, Faculty of Environment and Natural
Resources, University of Freiburg, Tennenbacher Straße 4, 79106
Freiburg, Germany. 64Institute for forest and wildlife management,
Faculty of Applied Ecology, Agricultural Sciences and Bio-
technology, Campus Evenstad, Inland Norway University of
Applied Science, NO-2480 Koppang, Norway. 65Wildlife and Fish
Division, Department of Environment and Natural Resources,
Government of the Northwest Territories, P.O. Box 1320,
Yellowknife, NT X1A 2L9, Canada. 66Ecological Program,
Tucker et al., Science 380, 1059–1064 (2023)
9 June 2023
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
Pennsylvania Department of Military and Veterans Affairs, Fort
Indiantown Gap National Guard Training Center, Annville,
PA 17003, USA. 67Institute of Biodiversity, Animal Health and
Comparative Medicine, University of Glasgow, Glasgow G12 8QQ,
UK. 68Veterinary Biology Department, Faculty of Veterinary
Medicine, University of Zagreb, Heinzelova 55, HR-10000 Zagreb,
Croatia. 69City of New York Parks and Recreation, Wildlife Unit,
1234 5th Avenue, 5th Floor, NY 10029, USA. 70Faculty of Forestry
and Wood Sciences, Czech University of Life Sciences Prague,
Czech Republic. 71Grimsö Wildlife Research Station, Swedish
University of Agricultural Sciences, 739 93 Riddarhyttan, Sweden.
72Snow Leopard Trust, 4649 Sunnyside Avenue North, Seattle,
WA 98103, USA. 73Centre for Ecosystem Science, School of
Biological, Earth and Environmental Sciences, University of New
South Wales, Sydney, NSW, 2052, Australia. 74Taronga Institute of
Science and Learning, Taronga Conservation Society, Sydney,
NSW, 2088, Australia. 75Inland Norway University of Applied
Sciences, Department of Forestry and Wildlife Management,
Norway. 76University of Veterinary Medicine Vienna, Research
Institute of Wildlife Ecology, Austria. 77Wildlife Department, Ol
Pejeta Conservancy, Private Bag-10400, Nanyuki, Kenya. 78U.S.
Geological Survey, Wyoming Cooperative Fish and Wildlife
Research Unit, Department of Zoology and Physiology, University
of Wyoming, Laramie, WY 82071, USA. 79North Carolina Museum of
Natural Sciences, Raleigh, NC 27601, USA. 80Department of
Forestry and Environmental Resources, North Carolina State
University, Raleigh, NC, 27695, USA. 81Department of Environment
and Natural Resources, Government of the Northwest Territories,
P.O. Box 2668, Yellowknife, NT X1A 2P9, Canada. 82Norwegian
Institute for Nature Research, NO-7484 Trondheim, Norway.
83Department of Wildlife, Fish and Environmental studies, Swedish
University of Agricultural Sciences, SE- 901 83 Umeå, Sweden.
84Department of Forestry, Biotechnical Faculty, University of
Ljubljana, Jamnikarjeva 101, SI-1000 Ljubljana, Slovenia.
85Department of Evolutionary Ecology, Leibniz Institute for Zoo and
Wildlife Research, Alfred- Kowalke- Str. 17, 10315 Berlin, Germany.
86Biological Sciences Centre, University of Alberta, Edmonton,
Alberta T6G 2E9, Canada. 87Haub School of Environment and
Natural Resources, Wyoming Cooperative Fish and Wildlife
Research Unit, Department of Zoology and Physiology, University
of Wyoming, 804 East Fremont, Laramie, WY 82072, USA. 88Büro
für Wildökologie und Forstwirtschaft, Klagenfurth, Austria. 89Clinic
for birds, reptiles, amphibians and fish, Justus-Liebig-University
Giessen, Germany. 90Inland Norway University of Applied Sciences,
Department of Forestry and Wildlife Management, Anne Evenstads
vei 80, 2480 Koppang, Norway. 91Snow Leopard Conservation
Foundation, Ulaanbaatar, Mongolia. 92Department of Fish and
Wildlife Sciences, University of Idaho, Moscow, ID 83844, USA.
93Biodiversity Research Institute (CSIC - Oviedo University -
Principality of Asturias), Oviedo University, E-33600 Mieres, Spain.
94Technical University of Munich, TUM School of Life Sciences,
Ecosystem Dynamics and Forest Management Group, 85354
Freising, Germany. 95Berchtesgaden National Park, 83471
Berchtesgaden, Germany. 96Office Français de la Biodiversité,
Direction de la Recherche et de l’Expertise, Unité Ongulés
Sauvages, Juvignac, France. 97Wyoming Cooperative Fish and
Wildlife Research Unit, Department of Zoology and Physiology,
University of Wyoming, Laramie, WY 82071, USA. 98Faro Moro Eco
Research, Estancia Faro Moro, Departmento de Boquerón,
Paraguay. 99Ontario Ministry of Natural Resources and Forestry,
Wildlife Research and Monitoring Section, Trent University, 2140
East Bank Drive, Peterborough, Ontario, K9J 7B8, Canada.
100Department of Environment and Natural Resources, Govern-
ment of the Northwest Territories, Highway 5, PO Box 900, Fort
Smith, Northwest Territories, X0E 0P0, Canada. 101Department of
Forestry and Forestry resources, Norwegian Institute of Bioecon-
omy Research, Tingvoll gard, NO-6630 Tingvoll, Norway.
102Department of Biological Sciences, University of Alberta,
Edmonton, Alberta T6G 2E9, Canada. 103Centre for Ecological and
Evolutionary Synthesis (CEES), Department of Biosciences,
University of Oslo, P.O. Box 1066 Blindern, NO-0316 Oslo, Norway.
104Department of Biology, University of British Columbia Okanagan,
Kelowna, British Columbia, Canada. 105Save Giraffe Now, 8333
Douglas Avenue, Suite 300, Dallas, Texas 75225, USA.
106The Faculty of Biological Sciences, Goethe University,
Max-von-Laue-Str. 9, 60438 Frankfurt am Main, Germany.
107National Geographic Partners, 1145 17th Street NW, Washington,
DC 20036, USA. 108Wildlife Conservation Society, Mongolia
Program. Post 20A, Box 21, Ulaanbaatar 14200, Mongolia.
109Wyoming Cooperative Fish and Wildlife Research Unit, Depart-
ment of Zoology and Physiology, University of Wyoming, Laramie,
WY 82071, USA. 110Program in Ecology, University of Wyoming,
Laramie, WY 82071, USA. 111Centre for Biological Diversity, School
of Biology, University of St. Andrews, Sir Harold Mitchell Building,
St. Andrews, KY16 9TH, UK. 112Department of Environmental and
Life Sciences, Trent University, 2140 East Bank Drive,
Peterborough, Ontario K9J 7B8, Canada. 113Ontario Ministry of
Natural Resources and Forestry, Wildlife Research and Monitoring
Section, Trent University, 2140 East Bank Drive, Peterborough,
Ontario K9J 7B8, Canada. 114Centro Nacional de Pesquisa e
Conservação de Mamíferos Carnívoros, Instituto Chico Mendes de
Conservação da Biodiversidade, Atibaia, SP, 12952011 Brazil.
115Research Institute of Wildlife Ecology, University of Veterinary
Medicine, Vienna, Austria. 116Department of Biodiversity, Conser-
vation and Wildlife Management, Bavarian State Institute for
Forestry, Hans-Carl-von Carlowitz Platz 1, 85354 Freising, Germany.
117School of Natural Sciences, Faculty of Science and Engineering,
Macquarie University, NSW, 2109, Australia. 118Faculty of
Environmental Protection, Trg mladosti 7, 3320 Velenje, Slovenia.
119Slovenian Forestry Institute, Večna pot 2, 1000 Ljubljana, Slovenia.
120Department of Biodiversity, Faculty of Mathematics, Natural
Sciences and Information Technologies, University of Primorska,
Glagoljaška 8, 6000 Koper, Slovenia. 121Aurora Wildlife Research, 1918
Shannon Point Rd., Nelson, BC, V1L 6K1, Canada. 122Department of
Biology, Biotechnical Faculty, University of Ljubljana, Jamnikarjeva 101,
1000 Ljubljana, Slovenia. 123Department of Zoology and Physiology,
University of Wyoming, Laramie, WY 82071, USA. 124Department of
Ecology and Evolutionary Biology, Princeton University, Princeton, NJ
08544, USA. 125Department of Animal Sciences, Wageningen
University and Research, De Elst 1, 6708 WD, Wageningen,
Netherlands. 126Department of Organismic and Evolutionary Biology,
Harvard University, 26 Oxford Street, Cambridge MA 02138, USA.
127Oikon Ltd, Institute of Applied Ecology, Trg Senjskih uskoka 1-2,
HR-10020 Zagreb, Croatia. 128Association for the Conservation of
Biodiversity of Kazakhstan (ACBK), Nur-Sultan, 010000, Kazakhstan.
129Nordens Ark, 456 93 Hunnebostrand, Sweden. 130Koç University
Department of Molecular Biology and Genetics, Faculty of Sciences,
Rumelifeneri, Istanbul, Sarıyer, Turkey. 131Institute of Nature
Conservation Polish Academy of Sciences, Adama Mickiewicza 33, 31-
120 Kraków, Poland. 132Departamento de Ciencias Integradas,
Facultad de Ciencias Experimentales, Centro de Estudios Avanzados
en Física, Matemáticas y Computación, Universidad de Huelva, 21071
Huelva, Spain. 133Dimension Research, Ecology and Environment
(D.R.E.Am. Italia), Via Garibaldi, 3, 52015 Pratovecchio Stia (AR), Italy.
134Snow Leopard Trust, Seattle, WA 98103, USA. 135Global Snow
Leopard and Ecosystem Protection Program, Bishkek, Kyrgyzstan.
136Snow Leopard Foundation, Kyrgyzstan Bishkek, Kyrgyzstan.
137Nature Conservation Foundation, Mysore 570002, India. 138Wildlife
Sciences, Faculty of Forest Sciences and Forest Ecology, University
of Goettingen, Göttingen, Germany. 139CIBIO, Centro de Investigação
em Biodiversidade e Recursos Genéticos, InBIO Laboratório Asso-
ciado, Campus de Vairão, Universidade do Porto, 4485-661 Vairão,
Portugal. 140BIOPOLIS Program in Genomics, Biodiversity and Land
Planning, CIBIO, Campus de Vairão, 4485-661 Vairão, Portugal.
141City of New York Parks and Recreation, Wildlife Unit, 1234 5th
Avenue, 5th Floor, NY, NY 10029, USA. 142Department of Ecology and
Evolutionary Biology, Yale University, New Haven, CT, USA.
143Center for Biodiversity and Global Change, Yale University, New
Haven, CT, USA. 144Max Planck - Yale Center for Biodiversity Movement
and Global Change, Yale University. 145School of Zoology, Faculty of Life
Sciences, Tel Aviv University, Tel Aviv 69978, Israel. 146San Diego
Zoo Wildlife Alliance, 15600 San Pasqual Valley Road, Escondido, CA
92027, USA. 147Grand Teton National Park, PO Drawer 170, Moose,
Wyoming 83012, USA. 148Wyoming Game and Fish Department,
700 Valley View Dr. Sheridan, WY 82801, USA. 149Aarhus University,
Department of Ecoscience - Wildlife Ecology, C.F. Møllers Allé 4-8, 8000
Aarhus C, Denmark. 150Alaska Department of Fish and Game, Kodiak,
AK 99615, USA. 151Welgevonden Game Reserve, P.O. Box 433,
Vaalwater, South Africa. 152Guyra Paraguay - CONACYT, Asunción,
Paraguay. 153Instituto Saite, Asunción, Paraguay. 154Ministry of
Environment, Forestry and Tourism, Windhoek, Namibia. 155Ecology,
Evolution and Environmental Biology, Columbia University, NY, NY
10027, USA. 156School of Natural Resources, University of Arizona,
1064 E Lowell St, Tucson, AZ 85719, USA. 157Park Canada, Banff
National Park Resource Conservation. PO Box 900, Banff, Alberta T1L
1K2, Canada. 158Centre for the Advanced Study of Collective Behaviour,
University of Konstanz, 78457 Konstanz, Germany. 159Center for
Integrated Spatial Research, Environmental Studies Department,
University of California, Santa Cruz, CA 95064, USA. 160Department
of Fish, Wildlife and Conservation Biology, Colorado State University,
Fort Collins, CO 80523, USA. 161USDA National Wildlife Research
Center, Predator Research Facility, Millville, UT 84326, USA.
162Department of Wildland Resources, Utah State University, Logan, UT
84322, USA. 163Tatra National Park, Kuźnice 1, 34-500, Zakopane,
Poland. 164Department of Biological Sciences, Goethe University,
Max-von-Laue-Strasse 9, 60438 Frankfurt am Main, Germany.
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abo6499
Materials and Methods
Fig. S1
Tables S1 to S15
References (36–59)
MDAR Reproducibility Checklist
Submitted 23 February 2022; accepted 27 April 2023
10.1126/science.abo6499
Tucker et al., Science 380, 1059–1064 (2023)
9 June 2023
6 of 6
|
10.1126_science.abi9810
|
RES EARCH
R E S E A R C H A R T I C L E
◥
BIOPHYSICS
Live-cell micromanipulation of a genomic locus
reveals interphase chromatin mechanics
Veer I. P. Keizer1,2,3†, Simon Grosse-Holz1,2,4, Maxime Woringer1,2, Laura Zambon1,2,3‡,
Koceila Aizel2§, Maud Bongaerts2, Fanny Delille5, Lorena Kolar-Znika1,2, Vittore F. Scolari1,2,
Sebastian Hoffmann3¶, Edward J. Banigan4, Leonid A. Mirny1,4, Maxime Dahan2#,
Daniele Fachinetti3*, Antoine Coulon1,2*
Our understanding of the physical principles organizing the genome in the nucleus is limited by the lack
of tools to directly exert and measure forces on interphase chromosomes in vivo and probe their
material nature. Here, we introduce an approach to actively manipulate a genomic locus using controlled
magnetic forces inside the nucleus of a living human cell. We observed viscoelastic displacements over
micrometers within minutes in response to near-piconewton forces, which are consistent with a Rouse
polymer model. Our results highlight the fluidity of chromatin, with a moderate contribution of the
surrounding material, revealing minor roles for cross-links and topological effects and challenging the
view that interphase chromatin is a gel-like material. Our technology opens avenues for future research
in areas from chromosome mechanics to genome functions.
R ecent progress in observing and perturb-
ing chromosome conformation has led
to an unprecedented understanding of
the physical principles at play in shap-
ing the genome in four dimensions (4D)
(1). From genomic loops and topologically as-
sociating domains to spatially segregated A/B
compartments and chromosome territories,
the different levels of organization of the eu-
karyotic genome are thought to arise from
various physical phenomena, including phase
separation (2–4), ATP-dependent motors (4, 5),
and polymer topological effects (6). Nonethe-
less, the physical nature of chromatin and chro-
mosomes inside the nucleus and its functional
implications for mechanotransduction remain
an active area of investigation (7, 8). Observation-
based studies assessing the mobility of the
genome in living cells, from single loci (9–11)
and small regions (12) to large domains (13),
underline the possible existence of different
1Institut Curie, PSL Research University, Sorbonne Université,
CNRS UMR3664, Laboratoire Dynamique du Noyau, 75005
Paris, France. 2Institut Curie, PSL Research University,
Sorbonne Université, CNRS UMR168, Laboratoire Physico
Chimie Curie, 75005 Paris, France. 3Institut Curie, PSL
Research University, CNRS UMR144, Laboratoire Biologie
Cellulaire et Cancer, 75005 Paris, France. 4Department of
Physics and Institute for Medical Engineering and Science,
Massachusetts Institute of Technology, Cambridge, MA
02139, USA. 5ESPCI Paris, PSL Research University,
Sorbonne Université, CNRS UMR8213, Laboratoire de
Physique et d’Étude des Matériaux, 75005 Paris, France.
*Corresponding author. Email: antoine.coulon@curie.fr (A.C.);
daniele.fachinetti@curie.fr (D.F.)
†Present address: National Cancer Institute, National Institutes of
Health, Bethesda, MD, USA.
‡Present address: Lifeanalytics srl, Oderzo, Veneto, Italy.
§Present address: TreeFrog Therapeutics, Pessac, France.
¶Present address: Boehringer Ingelheim Pharma GmbH & Co. KG,
Biberach an der Riss, Germany.
#Deceased.
material states of chromatin (e.g., liquid, solid,
and gel-like). Extranuclear mechanical pertur-
bations, including whole-nucleus stretching
(14, 15), micropipette aspiration (16), and ap-
plication of local pressures (16, 17) or torques
(18) onto a cell, all affect the overall geometry
of the nucleus and reveal global viscoelastic
properties. Conversely, intranuclear mecha-
nical manipulation of the genome is rare and
technically challenging (8). Viscoelasticity mea-
surements using a microinjected 1-mm bead
suggested that interphase chromatin may be a
cross-linked polymer network (i.e., a gel) (19).
Recently, intranuclear mechanics were ele-
gantly probed by monitoring the fusion of
both artificial (20) and naturally occurring (21)
droplet-like structures. Active mechanical man-
ipulation of an intranuclear structure was
recently achieved using an optical tweezer to
displace a whole nucleolus in oocytes (22) and
using optically induced thermophoretic flows
within prophase (23) or interphase (24) nuclei.
However, these approaches are limited to the
manipulation of large structures or do not ap-
ply forces directly on chromatin. These limi-
tations have made it difficult to disentangle
various effects (e.g., mechanical response of
the nucleus versus chromatin itself or hydro-
dynamics versus polymer viscoelasticity), lead-
ing to contradictory results. Therefore, an
approach for the direct and active mechanical
manipulation of specific genomic loci inside
living cells is needed. To meet this need, we
developed a technique for targeted microman-
ipulation of a specific genomic locus in the
nucleus of a living cell, allowing us to probe
the physical properties of an interphase chro-
mosome by measuring its response to a con-
trolled point force.
Results
Mechanical manipulation of a genomic locus in a
living cell
Our approach relies on tethering magnetic
nanoparticles (MNPs) to a genomic locus and
applying an external magnetic field (Fig. 1A).
We chose ferritin MNPs for their small size
(25, 26): 12 nm in diameter for ferritin (PDB
1GWG) and 28 nm for the full MNP (26). We
produced ferritin MNPs by synthesizing in vitro
recombinant enhanced green fluorescent pro-
tein (eGFP)–labeled ferritin cages and loading
them with a magnetic core (see the materials
and methods). We microinjected MNPs into
the nuclei of living human U-2 OS cells pre-
viously engineered to contain an artificial ar-
ray inserted at a single genomic location in a
subtelomeric region of chromosome 1 (band
1p36) (27). This genomic array contains ~200
copies of a 20-kb genetic construct, each in-
cluding 96 tetO binding sites and a transgene.
It has been extensively used in the past to
study the function of several chromatin mod-
ifications, RNA polymerase II (Pol II) recruit-
ment, and RNA synthesis during induction of
the transgene (27–29). Therefore, although we
used it here uninduced, this array can recapit-
ulate functional chromatin-based processes
such as transcriptional activation. MNPs were
targeted to the array using a constitutively
expressed fusion protein (TetR, mCherry, and
anti-GFP nanobody) serving as a tether (Fig.
1A). Upon injection, MNPs diffused through
the nucleus and accumulated at the array,
forming a fluorescent spot in both eGFP and
mCherry channels (figs. S1A and S2A). Quan-
tification of the fluorescence signals indicated
that MNPs were at nanomolar concentrations
in the nucleus after injection and accumulated
at the genomic locus in the range of hundreds
to thousands of MNPs (median 1500 MNPs;
figs. S1B and S3, movie S1, table S1, and mate-
rials and methods). The locus should be re-
garded as a condensed and heterochromatic
4-Mb region (1.6% of chromosome 1) resid-
ing in a euchromatic genomic context, as pre-
viously reported (27), with small MNPs (each
being ∼2 to 3 times (26) the size of a nucleo-
some) sparsely decorating chromatin (1 MNP
per ∼2.7 kb). Consistently, we observed that
the locus typically resided in low to interme-
diate DNA density regions and was itself rela-
tively condensed (fig. S4, A and B) and that
binding of MNPs to the locus did not sub-
stantially affect its morphology (fig. S2, A and
B). Microinjection and attraction of unbound
MNPs did not substantially alter chromatin
distribution and densities inside the nucleus
[as assessed by silicon rhodamine (SiR)–DNA
staining; fig. S2, C and D]. Cells were imaged
on a coverglass with custom-made microfab-
ricated pillars (30, 31) that behave as local
magnets only when subjected to an external
magnetizing field (fig. S5). Therefore, ON/OFF
Keizer et al., Science 377, 489–495 (2022)
29 July 2022
1 of 7
RES EARCH | R E S E A R C H A R T I C L E
modulation of the local force field could be
achieved while imaging by placing and remov-
ing an external magnet on the microscope
stage. The shape and orientation of the pillars
were chosen to maximize the magnetic field
gradient and hence the force. We performed
magnetic simulations and experimental cal-
ibrations using two independent methods (see
the materials and methods, figs. S6 and S7,
and movie S2) to determine the magnitude
and orientation of the force applied onto the
genomic locus as a function of the number of
MNPs bound to it and its position relative to
the magnetic pillar (Fig. 1B). The typical forces
applied onto the locus were in the sub-
piconewton (pN) range, occasionally reaching
a few piconewtons (table S1; median force =
0.45 pN). These values are in the range of
A
Micro-injection
of ferritin-GFP
anti-GFP
nanobody
mCherry
TetR
Chr1
tetO array
B
Force
on genomic
locus (pN)
ON/OFF
magnetic pillar
1
0.1
10 µm
E
pillar
C
F
4 µm
F
ON/OFF
magnetic pillar
Force
SiR-DNA
Ferritin-GFP
Pull: 30 min
Release: 30 min
10 min
Fx Fy Fz
0
10
20
30
Time (min)
40
50
60
D
)
m
µ
(
n
o
i
t
i
s
o
P
6
4
2
0
)
N
p
(
e
c
r
o
F
0.8
0.4
0
4 µm
F
F
...
F
...
...
SiR-DNA
Ferritin-GFP
Pull 1
Release 1
Pull 3
Pull 8
Release 10
Pulls:
Releases:
100"
100"
F
)
m
µ
(
n
o
i
t
i
s
o
P
)
N
p
(
e
c
r
o
F
6
4
2
0
6
4
2
0
P1
R1
P2
R2
P3
R3
P4
R4
P5
R5
P6
R6
P7
R7
P8
R8
P9
R9
P10
R10
10 min
Fx Fy
0
200
400
600
800 1000
0021
0041
0061
Time (sec)
0081
0002
0022
0042
2600
0082
3000
Fig. 1. Mechanical micromanipulation of a genomic locus in living cells.
(A) MNPs of GFP-labeled ferritin were microinjected into the cell nucleus and
targeted to a genomic array containing ~19,000 tetO binding sites (27) with a linker
protein. Cells were imaged on a coverslide with microfabricated magnetic pillars
that produce a local magnetic field and attract the genomic locus. (B) The force
exerted onto the locus depends on its position relative to the pillar and is characterized
using a precalculated force map (see the materials and methods), here shown for
1000 MNPs at the locus. (C) Example of a pull-release experiment showing the locus
being displaced during the 30 min of force exertion and recoiling during the 30 min
of force release (30′-PR scheme). (See also movie S3.) (D) Kymograph of the same
experiment showing each time frame, along with the force time profile calculated using
the force map. (E) Experiment in which pulls and releases are 100 s and the pull-
release cycle is repeated 10× (100″-PR scheme). Images are time projections; shown
in green are all positions of the center of mass of the locus over the periods
represented on the timeline. The arrows indicate the direction of the motion. (See also
movie S4.) (F) Kymograph of the same experiment showing the displacement of
the locus and the spatial patterns of DNA density in the nucleus, along with the force
time profile. All SiR-DNA images are band-pass filtered (see the materials and
methods). For (D) and (F), dotted lines indicate the nuclear periphery and white arrows
features of interest in the spatial distribution of DNA density.
Keizer et al., Science 377, 489–495 (2022)
29 July 2022
2 of 7
RES EARCH | R E S E A R C H A R T I C L E
forces exerted by molecular motors in the nu-
cleus, e.g., comparable to the stalling force of
∼0.5 pN for the structural maintenance of
chromosomes (SMC) complex condensin (32)
and a few piconewtons for Pol II (33).
Force-induced movement of a genomic locus
reveals viscoelastic properties of chromatin
We first applied the magnetic force for 30 min
and released it for another 30 min while per-
forming low-illumination three-dimensional
imaging with a 2-min interval (30′-PR scheme).
We observed a clear motion of the locus
toward the magnet upon application of the
force and a slow and partial recoil when the
force stopped (Fig. 1, C and D, and movie S3).
This indicates that a sub-piconewton force,
when applied in a sustained and unidirectional
manner on a genomic locus, elicits a displace-
ment of that locus by several micrometers over
minutes. It also shows that the chromosomal
locus can move across the nuclear environ-
ment, which is believed to be crowded and
entangled. We also applied the force period-
ically, pulling for 100 s, releasing for 100 s, and
repeating this cycle 10 times (100′′-PR scheme)
while performing fast two-dimensional imag-
ing with a 5-s interval (Fig. 1, E and F, and
movie S4). Several observations from these
two experiments hinted at the material prop-
erties of chromatin. First, the trajectories
showed recoils during release periods and a
gradual slowdown during both pulls and re-
leases, characteristic of a viscoelastic material.
Second, spatial heterogeneities in the trajecto-
ries were visible and appeared to relate to the
spatial distribution of DNA density (the motion
of the locus was occasionally hindered where
the DNA density varied; Fig. 1, D and F, white
arrows). Third, recoil after force release was
seen even after collision with the nuclear peri-
phery, indicating that the material there (peri-
pheral heterochromatin and the nuclear lamina)
was not sticky enough to fully retain the locus.
Fourth, the spatial distribution of DNA density
in the nucleus did not show large-scale de-
formations, indicating that the locus did not
drag along large amounts of material (movies
S3 and S4). The force-induced displacements
that we observed are consistent with visco-
elastic and nonconfining chromatin and con-
stitute a basis to further develop and test
physical models of interphase chromosomes.
Quantitative force-response and scaling laws of
interphase chromatin mechanics
To quantify the viscoelastic properties of chro-
matin, we analyzed the trajectories of the
locus in 35 cells undergoing the 30′-PR scheme
(corrected for cell motion and force orienta-
tion; see the materials and methods). We ob-
served a range of behaviors in both pulls and
releases regarding initial speed, total distance
traveled, and shape of time profiles (Fig. 2, A
and B, and fig. S8). Most traces showed a dis-
placement that was clearly distinguishable
from diffusion (Fig. 2, A and B, hatched areas;
see the materials and methods). Collision with
the nuclear periphery (Fig. 2, A and B, open
symbols, and fig. S8, open symbols) was seen
in nine of 35 traces, so the total displacement
during the pull is most often not limited by the
nuclear periphery. The initial force applied
onto the locus largely predicted the variability
seen in the initial motion (Fig. 2C and fig.
S9A). The recoil motion after force release
was in part predicted by the total distance over
which the locus had been displaced during the
pull (Fig. 2D and fig. S9B), with a simple linear
relationship highlighting the elastic nature of
chromatin. Deviations from these simple pro-
portionality relationships indicate that the
specific nuclear context or the state of the
genomic locus might influence its response.
In particular, we observed that when the locus
moved more slowly than expected, it was less
DNA dense, and when the locus moved faster
than expected, it was more DNA dense (fig.
S4C), suggesting that the compaction state
of the locus itself affected its response to the
force. The absolute nuclear position of the lo-
cus did not correlate with its response to the
force, but if the locus reached the periphery
during the pull, it often recoiled more slowly
than expected (fig. S10). Despite the varia-
bility between traces, log-log plots of all the
pulls and releases from the 30′-PR and 100′′-PR
trajectories, together with three additional
high-frame-rate (dt = 0.5′′) pull-release trajec-
tories, revealed linear portions in the curves
with a slope of 0.5 over more than three orders
of magnitude in time (Fig. 2E). This behavior
suggests that the different levels of the hier-
archical genome organization are not charac-
terized by vastly distinct mechanical properties.
In addition, displacements that scale with
time as t1/2 can be empirically described by a
“fractional speed,” i.e., a single value in mm/s1/2
capturing how the motion evolves over time
(Fig. 2, C and D, and fig. S9, right axes). The
first pull of the 100′′-PR trace, represented in
this unit, indeed followed the same relation-
ship as the 30′-PR traces (Fig. 2C, dark green
triangle), and the slope of the resulting force-
displacement plot yielded a unique factor of
0.158 (±0.014) mm/s1/2/pN, characterizing the
dynamic response of chromatin to force. These
results indicate that a large part of the re-
sponse of chromatin to force can be described
by simple laws.
The chromatin force response is well described
by a free polymer model (Rouse chain)
We then sought a model of chromatin that
best explains our quantitative measurement
of force-induced locus displacement. Several
features in our data suggested a classical poly-
mer model known as a Rouse polymer (34) as
a first approximation to describe the response
of chromatin to forces. The Rouse model rep-
resents a polymer in which each monomer
diffuses by thermal motion in a viscous me-
dium and is connected to its two neighbors
by elastic bonds. Rouse ignores steric effects
(contact, hindrance), cross-links (affinity, stick-
iness), and topological effects (fibers can pass
through each other). This model is frequently
invoked for chromatin dynamics because it
predicts the characteristic power-law scaling–
that is, a linear relationship on a log-log plot–of
the mean squared displacement (MSD) versus
time with exponent 0.5, as observed here (fig.
S11, A and E) and for other genomic loci in
eukaryotes (35–37). We extended Rouse theory
to study how a polymer responds to a point
force (see the materials and methods and the
supplementary text). Our calculations pre-
dict a power-law behavior with exponent 0.5
for displacements and recoils in response to
force, consistent with our experimental obser-
vations (Fig. 2E). These two power laws have
the same physical origin, so the diffusion co-
efficient obtained independently from the
MSD (1627 ± 19 nm2 s–1/2; fig. S11A) directly
relates to—and predicts—the slope of the
force-displacement plot (Fig. 2C, red line):
1627 nm2 s-1/2/2kBT = 0.190 ± 0.003 mm/s1/2/pN
(see the materials and methods). This agree-
ment between two independent passive and
active measurements (Fig. 2C and fig. S9A, red
and gray lines, representing diffusion and force
response, respectively) supports the Rouse
model to explain our chromatin dynamics
data. Inspected on a cell-by-cell basis, the
force-free MSD of the locus before and after
the pull-release experiments appeared very
moderately reduced in most cases (fig. S11B).
Its natural variability between cells does not
appear to explain the variability of the re-
sponse to force (fig. S11C). After force release,
the Rouse model also predicted a recoil pro-
portional to the total displacement during
the pull. However, in many cases, the locus re-
coiled somewhat more slowly than predicted
by the Rouse theory. Instead, the theoretical
prediction appears to define an upper bound
for the recoils (Fig. 2D and fig. S9B, red lines),
and deviations from Rouse theory were more
pronounced at the nuclear periphery (fig. S10B).
This analysis suggests that the dynamic re-
sponse of the chromosome to the force can be
described by the Rouse polymer model, with
additional effects from the nuclear environment.
Model-based trajectory analysis reveals
moderate hindrance by surrounding chromatin
To further understand the physical nature of
chromatin, we investigated how alternative
polymer models are able to capture the 100′′-PR
trace (Fig. 3, A to D). Our approach was to use
the displacement trajectory and infer, assum-
ing a given polymer model, the time profile of
Keizer et al., Science 377, 489–495 (2022)
29 July 2022
3 of 7
RES EARCH | R E S E A R C H A R T I C L E
A
Pulls
B
Releases
n=35
E
α = 0.5
α = 0.5
α = 0.5
α = 0.5
α = 0.5
α = 0.5
Average
α = 0.5
Average
α = 0.5
C
D
R10
30'-PR traces:
trace 1
.
.
.
trace 35
e
c
r
o
F
100"-PR trace:
P1, R1
.
.
.
P10, R10
e
m
T
i
dt=0.5" traces:
trace 1
trace 2
trace 3
e
c
r
o
F
diffusion
theory (Rouse)
fit
Fig. 2. Quantitative analysis of locus movement in response to force.
(A) Trajectories of the genomic locus in the direction of the applied force for
35 different cells during force exertion with the 30′-PR scheme. A selection of
trajectories, representative of the breadth of observed behaviors, are
highlighted and color-coded by force (see fig. S8 for individual trajectories).
Hatched areas in (A) to (D) correspond to the null model of pure diffusion
based on MSD measurement (see the materials and methods). (B) Recoil
trajectories relative to the time and position at the start of the release are
shown for the same loci as in (A). Curve R10 is the last release of the 100″-PR
trajectory. (C) Displacements measured at Dt = 5 min of force exertion on all the traces
from (A) plotted against the magnitude of the force. Coordinates are interpolated
between the frames before/after Dt. The green line and triangle correspond to the
envelope of the 100″-PR trajectory. Reported forces are the average over Dt.
Displacements are also expressed in mm/s1/2 (right axis), allowing us to place pull P1
from the 100″-PR trace, measured at Dt = 100 s (dark green triangle). The red line
indicates the expected relationship from Rouse theory, solely based on an MSD
measurement. (D) Recoil after Dt = 5 min of force release on all the traces from (B)
and the last release of the 100″-PR trajectory (R10, green triangle), plotted against the
total displacement during the pull. The red line indicates the expected relationship
from Rouse theory [see fig. S9 for versions of (C) and (D) at different Dt and in linear
scales]. Open symbols in (A) to (D) indicate when loci are within 1.5 mm of the nuclear
periphery [at the moment of measurement in (A) to (C) and at the moment of force
release in (D)]. (E) Displacement and recoil trajectories, aligned on the time and
position at the moment of force switching, are represented as log-log plots for pull-
release experiments imaged with different frame intervals: dt = 0.5 s (see the materials
and methods), dt = 5 s (from the 100″-PR trace; Fig. 1F), and dt = 2 min [all 30′-PR
traces; (A) and (B)]. For the latter, average trajectories (right plots) were calculated
over all displacements where the applied force remains ≤2 pN (28 of 35 traces)
and over all the recoils after a displacement of ≤5 mm (21 of 35 traces). Red dotted
lines indicate the power-law behavior, with exponent 0.5, predicted by Rouse theory.
the force that produced the measured trajec-
tory (see the materials and methods, supple-
mentary text, and fig. S12, A to C). Disagreement
between inferred and actual force profiles
indicates when models are incorrect or in-
complete, allowing one to select and refine the
best model(s). With this approach, we com-
pared a series of models (fig. S12C). First, a
simple Rouse model without any adjustable
parameters (i.e., calibrated using the MSD
versus time plot; fig. S11A) predicted well the
first pull and all of the release periods (Fig.
3B). However, the prediction leaves some of
the applied force unexplained (Fig. 3B, gray
area between curves), suggesting a missing
component in the model that would addi-
tionally slow down or hinder the progres-
sion of the locus. This residual unexplained
force did not scale with speed and thus could
not be explained as an additional viscous drag
on the locus (fig. S12D). Instead, it increased
progressively across successive pulls, suggest-
ing an accumulation of hindrance as the locus
moved through the nucleus. To represent this,
we added a capacity for the locus to interact
with the surrounding chromatin, represented
as extra Rouse chains that are either attached
to or pushed by the locus along its path (Fig.
3C and fig. S12C). These models better pre-
dicted the force profile throughout the trajec-
tory compared with a pure Rouse model. The
only free parameter used for the inference shown
in Fig. 3C is the frequency at which the locus
interacts with other polymers, which we found
to be very low (fig. S12C), indicating that the
interaction with the surrounding chromatin
was moderate. This is also consistent with the
small but detectable reduction in mobility of the
locus before and after pull-release experiments
(fig. S11B) and the subtle redistribution of
DNA densities around the pulled locus (fig.
S4D). These modeling results suggest that,
upon force application and release on our geno-
mic locus, chromatin is well described as a
Rouse polymer (i.e., a free polymer in a viscous
environment), with moderate interactions from
the surrounding chromatin, indicating that
hindrance, cross-links, and topological effects
play a minor role.
Interphase chromatin does not behave as a gel
in force-response experiments
Interphase chromatin has been proposed to
be a gel-like material (11, 12, 19). A gel is a
highly cross-linked polymer, which means that
unlike a linear polymer, in which monomers
are linked to two neighbors, extra links be-
tween nonadjacent monomers form an inter-
connected mesh, giving the gel solid-like
properties. For chromatin, this could in prin-
ciple arise from affinity between nucleosomes,
as well as loops or bridges formed by proteins,
complexes, and condensates and topological
entanglement between chromatin fibers. How-
ever, in such an interconnected mesh structure,
short paths effectively linking the pulled locus
Keizer et al., Science 377, 489–495 (2022)
29 July 2022
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
A
Position (µm)
F
g
n
o
A
l
Lateral
F
F
F
F
8
6
4
2
0
6
4
2
0
6
4
2
0
6
4
2
0
F
)
N
p
(
e
c
r
o
F
)
N
p
(
e
c
r
o
F
)
N
p
(
e
c
r
o
F
Free
polymer
(Rouse)
B
C
Obstruction
by surrounding
chromatin
D
Gel-like
material
(crosslinked)
#
G
Lateral mobility
increase? or decrease?
*
F
F
Actual force
Model inference
*
#
Velocity along F
0.04–0.08 µm/s
0.02–0.04 µm/s
0.01–0.02 µm/s
0–0.01 µm/s
H
)
2
m
µ
(
D
S
M
l
a
r
e
t
a
L
0.12
0.10
0.08
0.06
0.04
0.02
0
500
1000
1500
Time (sec)
2000
2500
3000
E
4
2
0
)
N
p
(
e
c
r
o
f
l
i
a
u
d
s
e
R
# Nuclear periphery, elasticity = 4.81 pN/µm F
* Energy barrier = 46.3 kBT
2
3
4
5
Position (µm)
6
7
8
#
*
4 µm
#
*
P4
P8
I
)
2
m
µ
(
D
S
M
l
a
r
e
t
a
L
0.05
0
0
0
10
20
30
40
50
Delay τ (sec)
τ = 10 sec
τ = 20 sec
τ = 35 sec
0.04
0
Velocity along F (µm/s)
0
0.04
0
0.04
0.08
( forward, backward)
Fig. 3. Model-based analysis and hypothesis testing. (A) Trajectory of the
locus shown in the direction of the force (green curve) and orthogonal to the
force in the imaging plane (blue curve) for the 100″-PR experiment. Arrows
indicate apparent obstacle [see asterisks and hash marks in (C), (E), and (F)].
(B to D) Evaluation of different models through their capacity to reproduce
the experimentally measured force time profile (orange curve) by inferring it
from the trajectory (gray curve). Models shown here are a simple Rouse
polymer (34) (B), the same model with extra polymer chains being pushed by the
locus to represent the surrounding chromatin (C), and a gel-like material
represented as a Rouse polymer in a viscoelastic environment (D). (See full list in
fig. S12.) (E) Residual unexplained force from the second model [area between
curves in (C)] plotted along the trajectory of the locus, highlighting an obstacle
(asterisk) and an elastic region near the nuclear periphery (hash mark) for which
physical parameters are measured (see the materials and methods) and which
are visible in (A), (C), and (F). (F) Time projection images before the first pull and
during pulls P4 and P8 showing how the spatial distribution of DNA density in
the nucleus relates to the identified obstacles. SiR-DNA images were band-pass
filtered (see the materials and methods). (G) Hypotheses on how the lateral
mobility of the locus may change depending on its force-induced displacement.
(H and I) MSD of the lateral movement of the locus calculated as a function
of both time delay and velocity in the direction of the force. Solid lines on both
the MSD-delay (H) and the MSD-velocity (I) representations correspond to a
single-parameter fit describing how lateral mobility increases with velocity in the
direction of the force.
to all other loci in the nucleus would result in
long-range deformation of the spatial pattern
of DNA density, which we did not observe
(Fig. 1, D and F, and movies S3 and S4). Fur-
ther, if the chromatin surrounding the locus
were gel-like, then it would effectively act as
a viscoelastic medium. This assumption does
not recapitulate well the experimental data
(even with two free parameters; Fig. 3D and
fig. S12C) and is inconsistent with the ob-
served scaling of 0.5 in the MSD (fig. S11, A
and E), which argues for a simply viscous and
nonelastic medium. Finally, if the locus were
part of an interconnected mesh, then short
series of links would tether it to large struc-
tures (e.g., the periphery and nucleoli). A
Rouse model that includes a finite tether does
not recapitulate the experimental data (fig.
S12C) and is inconsistent with the linear be-
havior observed in Fig. 2E up to several mic-
rometers. These results suggest again minor
effects of cross-links and topological constraints
and argue against the view that interphase
chromatin behaves like a gel at the spatial and
temporal scale of our observations.
Heterogeneities in the trajectory reveal obstacles
in the nuclear interior and a soft elastic material at
the nuclear periphery
Even the models that best capture the data
leave part of the force unexplained (Fig. 3C,
gray area). We plotted this residual unex-
plained force as a function of spatial position
(Fig. 3E). This revealed an accumulation of
non-null residual forces at specific locations,
matching visible features in the trajectory and
in the spatial distribution of DNA density in
the nucleus. First, the residual force in pulls P3
to P5 corresponds to an apparent obstacle in
the trajectory (Fig. 3, A and C, asterisks) oc-
curring at a high-to-low transition of DNA den-
sity (Figs. 3F and 1F). It appears as a spatially
defined barrier of residual force (Fig. 3E), re-
quiring an energy of ≈ 46 kBT to overcome.
This suggests that, whereas DNA dense regions
are not obstacles per se, the interface between
high- and low-density regions may constitute
a barrier. The energy that we estimated sug-
gests that such barriers may be overcome by
ATP-dependent molecular motors (32, 33) but
not likely by spontaneous thermal fluctuations.
Second, the residual force in pulls P8 to P10
(Fig. 3, A and C, hash marks) corresponds to
the collision with structures near the nuclear
periphery (Figs. 1F and 3F, hash marks). The
observed linear force-distance relationship
(Fig. 3E) indicates a solid-like elastic behav-
ior for these structures over at least 600 nm
and with a spring constant of 4.81 pN/mm. This
is much softer than what was measured by
whole-nucleus stretching experiments (14, 15),
which could be explained by the small size of
the locus and/or the existence of a soft layer of
Keizer et al., Science 377, 489–495 (2022)
29 July 2022
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
elastic peripheral components (e.g., hetero-
chromatin and nuclear lamina) rather than
the material directly contributing to the struc-
tural rigidity of the nucleus.
Lateral mobility of the locus reflects transient
collisions with obstacles in the nucleoplasm
To further investigate the material encoun-
tered by the locus, we analyzed the lateral mo-
tion of the locus as it was being pulled and
released (Fig. 3A, blue curve). We hypothe-
sized that collisions with obstacles could in-
crease lateral mobility or, conversely, that the
locus being dragged into a more constraining
and entangled environment could result in a
reduction of its mobility (Fig. 3G). After com-
puting the MSD of the lateral motion as a
function of both time delay t and velocity uy
along the direction of the force (Fig. 3, H and
I, and fig. S11, F and G), we observed a clear
increase of lateral mobility when the locus
moved (for both forward and backward move-
ments; Fig. 3I), suggesting the existence of
obstacles that deflected the motion. This ad-
ditional mobility in the MSD is captured by a
term proportional to uy, as would be expected
for collisions, and proportional to t (not t0.5),
as would be expected if the force due to the
collision with obstacles persists in the same
direction across several frames, indicating the
existence of large obstacles. Indeed, in P3 for
instance, the lateral motion clearly shows a
directional behavior (Figs. 1E and 3A). How-
ever, the relationships that we observed (Fig. 3,
H and I) held even when excluding all of the
time points before P4 (fig. S11, H and I), indi-
cating that the collision with obstacles was
widespread throughout the nucleus. These re-
sults, together with our observation that very
few chromatin fibers appeared to be carried
along with the locus, indicate that obstacles
are frequently encountered by the locus, but
most interactions are weak and transient.
Discussion
Our measurements of how a genomic locus
inside the nucleus of a living cell responds to a
point force indicates that interphase chroma-
tin has fluid-like properties and behaves as a
free polymer. This contrasts with previous
studies depicting chromatin as a stiff, cross-
linked polymer gel with solid-like properties
(11, 12, 19). Our observation that near-piconewton
forces can easily move a genomic locus across
the nucleus over a few minutes (Fig. 1, D and
F) also contrasts with a previous study report-
ing confined submicrometer displacements
over seconds upon application of 65 to 110 pN
forces to a 1-mm bead (19). We propose that our
results may be reconciled with previous ex-
periments in several ways. First, unlike a
micrometer-sized bead, the locus used in
our experiments is small and may be deform-
able enough to pass through the surrounding
chromatin. Second, chromatin may contain
many small, gel-like patches embedded in a
structure with liquid, Rouse-like properties
at a larger scale. This is also consistent with
our observation that the transiting locus fre-
quently encounters obstacles. Third, chroma-
tin may be a weak gel, i.e., a gel with short-lived
cross-links (11). Such a gel could continuously
maintain a stiff, globally connected network
that resists stresses over large length scales
while permitting fluid-like motions at smaller
scales. Future experiments perturbing chro-
matin state and chromatin associated proteins
will be important to reconcile observed micro-
and mesoscale mechanics.
Organization of chromosomes that allows
movement of genomic loci across large dis-
tances by weak forces could have implications
for a range of genome functions. For example,
large-scale movements of chromosomes occur
during nuclear inversion in rod cell differ-
entiation for nocturnal mammals (38). Specific
genes undergo long-range directional motion
upon transcriptional activation (39, 40). Long
and highly transcribed genes can form ∼5-mm
giant loops, believed to be due to chromosome
fiber stiffening (41). Certain double-strand break
sites undergo large-scale, nuclear F-actin de-
pendent relocation to the nuclear periphery
(42). These DNA-based biological processes
require a nuclear organization in which such
movements are possible. Our results reveal the
mechanical properties of chromatin in which
such large-scale movements would only re-
quire weak (near piconewton) forces. Although
sustained unidirectional forces are unlikely to
occur naturally in the nucleus, the magnitude
of the forces and the time scale of force ex-
ertion in our experiments are comparable to
those of molecular motors such as SMC com-
plexes and Pol II; that is, in the sub-piconewton
(32) or low-piconewton (33) ranges and ap-
plied over minutes (e.g., 10 min for Pol II to
elongate through a 25-kb gene, 5–30 min for
SMC complexes). Therefore, some molecular
motors in the nucleus operate in a force range
that is sufficient to substantially reorganize
the genome in space.
Future work will be important to expand
and complement our results. Although the ge-
nomic array that we used here is known to be
chromatinized and has been used extensively
to recapitulate and study functional chromatin-
based processes (27–29), we cannot exclude
that its repetitive and artificial nature might
prevent some of our measurements from being
applicable to nonrepetitive and native regions.
Manipulating loci other than a subtelomeric
locus on the longest chromosome (chromo-
some 1) in other genomic contexts (e.g., hetero-
chromatin or euchromatin) and in different
cell types will be important to assess the gen-
eralizability of our findings in various biolog-
ical contexts.
Our approach to mechanically manipulate
and relocate genomic loci in the nuclear space
opens many avenues for future research, from
the study of interphase chromosome mechan-
ics to the perturbation of genome functions,
including transcription, replication, DNA dam-
age repair, and chromosome segregation. By
giving access to physical parameters and re-
vealing fundamental scaling laws to describe
chromatin mechanics, our work provides a
foundation for future theories of genome
organization.
REFERENCES AND NOTES
J. Dekker et al., Nature 549, 219–226 (2017).
1.
2. D. Jost, P. Carrivain, G. Cavalli, C. Vaillant, Nucleic Acids Res.
42, 9553–9561 (2014).
3. D. Hnisz, K. Shrinivas, R. A. Young, A. K. Chakraborty,
P. A. Sharp, Cell 169, 13–23 (2017).
4. L. A. Mirny, M. Imakaev, N. Abdennur, Curr. Opin. Cell Biol. 58,
142–152 (2019).
5. E. J. Banigan, L. A. Mirny, Curr. Opin. Cell Biol. 64, 124–138 (2020).
J. D. Halverson, J. Smrek, K. Kremer, A. Y. Grosberg, Rep. Prog.
6.
Phys. 77, 022601 (2014).
7. C. Uhler, G. V. Shivashankar, Nat. Rev. Mol. Cell Biol. 18,
717–727 (2017).
8. A. A. Agbleke et al., Mol. Cell 79, 881–901 (2020).
9. T. Nozaki et al., Mol. Cell 67, 282–293.e7 (2017).
10. B. Chen et al., Cell 155, 1479–1491 (2013).
11. N. Khanna, Y. Zhang, J. S. Lucas, O. K. Dudko, C. Murre,
Nat. Commun. 10, 2771 (2019).
12. H. Strickfaden et al., Cell 183, 1772–1784.e13 (2020).
13. A. Zidovska, D. A. Weitz, T. J. Mitchison, Proc. Natl. Acad.
Sci. U.S.A. 110, 15555–15560 (2013).
14. A. D. Stephens, E. J. Banigan, S. A. Adam, R. D. Goldman,
J. F. Marko, Mol. Biol. Cell 28, 1984–1996 (2017).
15. Y. Shimamoto, S. Tamura, H. Masumoto, K. Maeshima,
Mol. Biol. Cell 28, 1580–1589 (2017).
16. K. N. Dahl, A. J. Engler, J. D. Pajerowski, D. E. Discher, Biophys. J.
89, 2855–2864 (2005).
17. C. M. Hobson et al., Mol. Biol. Cell 31, 1788–1801 (2020).
18. A. Tajik et al., Nat. Mater. 15, 1287–1296 (2016).
19. A. H. B. de Vries, B. E. Krenn, R. van Driel, V. Subramaniam,
J. S. Kanger, Nano Lett. 7, 1424–1427 (2007).
20. Y. Shin et al., Cell 175, 1481–1491.e13 (2018).
21. C. M. Caragine, S. C. Haley, A. Zidovska, eLife 8, e47533 (2019).
22. M. S. Syrchina, A. M. Shakhov, A. V. Aybush, V. A. Nadtochenko,
Optical trapping of nucleolus reveals viscoelastic properties of
nucleoplasm inside mouse germinal vesicle oocytes. bioRxiv
999342 [Preprint] (2020); https://doi.org/10.1101/2020.03.19.
999342.
23. M. Mittasch et al., Active gelation breaks time-reversal-
symmetry of mitotic chromosome mechanics. bioRxiv
296566 [Preprint] (2018); https://doi.org/10.1101/296566.
24. B. Seelbinder, M. Jain, E. Erben, S. Klykov, I. D. Stoev,
M. Kreysing, Non-invasive chromatin deformation and
measurement of differential mechanical properties in the
nucleus. bioRxiv 472786 [Preprint] (2021); https://doi.org/10.
1101/2021.12.15.472786.
25. C. Monzel, C. Vicario, J. Piehler, M. Coppey, M. Dahan, Chem. Sci.
8, 7330–7338 (2017).
26. D. Liße et al., Adv. Mater. 29, 1700189 (2017).
27. S. M. Janicki et al., Cell 116, 683–698 (2004).
28. T. Tsukamoto et al., Nat. Cell Biol. 2, 871–878 (2000).
29. X. Darzacq et al., Nat. Struct. Mol. Biol. 14, 796–806
(2007).
30. L. Toraille et al., Nano Lett. 18, 7635–7641 (2018).
31. M. Bongaerts et al., Int. J. Mol. Sci. 21, 6560 (2020).
32. M. Ganji et al., Science 360, 102–105 (2018).
33. E. A. Galburt, S. W. Grill, C. Bustamante, Methods 48, 323–332
(2009).
34. P. E. Rouse Jr., J. Chem. Phys. 21, 1272–1280 (1953).
35. B. Gu et al., Science 359, 1050–1055 (2018).
36. J. Li et al., Cell 178, 491–506.e28 (2019).
37. H. Hajjoul et al., Genome Res. 23, 1829–1838 (2013).
38. I. Solovei et al., Cell 152, 584–598 (2013).
39. C.-H. Chuang et al., Curr. Biol. 16, 825–831 (2006).
40. N. Khanna, Y. Hu, A. S. Belmont, Curr. Biol. 24, 1138–1144 (2014).
41. S. Leidescher et al., Nat. Cell Biol. 24, 327–339 (2022).
42. C. P. Caridi et al., Nature 559, 54–60 (2018).
Keizer et al., Science 377, 489–495 (2022)
29 July 2022
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
43. V. I. P. Keizer et al., Data, software, and documentation for:
Live-cell micromanipulation of a genomic locus reveals
interphase chromatin mechanics, Zenodo (2022); https://doi.
org/10.5281/zenodo.4626942.
ACKN OW LEDG MEN TS
We dedicate this work to the memory of our co-author Maxime
Dahan, who died in July 2018. We acknowledge scientific inputs
from C. Giovannangeli and J.-P. Concordet (MNHN), T. Pons and
N. Lequeux (ESPCI), M. Coppey (Institut Curie), M. Cosentino
Lagomarsino (IFOM), and the members of the Coulon and
Fachinetti teams. We also acknowledge the current and past
members of the Dahan/Coppey/Hajj team and UMR168, including
C. Monzel, D. Libe, J. Manzi, E. Baloul, C. Vicario, and D. Normanno,
for their earlier work and technical help on nanoparticles and
microinjection, and E. Secret and A. Michel-Tourgis (UMR8234
CNRS) for help with magnetic nanoparticle characterization. We
also acknowledge the Flow Cytometry Core Facility of the Institut
Curie, the BMBC platform of UMR168, the machine shop of
UMR168, the technological platform of the Institut Pierre-Gilles
de Gennes (IPGG) for access to microfabrication facilities, the
PICT-IBiSA@Pasteur Imaging Facility of the Institut Curie (member
of the France Bioimaging National Infrastructure; ANR-10-INBS-04),
and the MPBT platform (Physical Measurements at Low
Temperatures) of Sorbonne Université. Funding: This work
received funding from the LabEx CELL(N)SCALE (ANR-11-LABX-
0038 and ANR-10-IDEX-0001-02 to M.D., D.F., and A.C.); the
Agence Nationale de la Recherche (project CHROMAG, ANR-18-
CE12-0023-01 to M.D. and A.C.); the PRESTIGE program of
Campus France (PRESTIGE-2018-1-0023 to V.K.); the ATIP-Avenir
program of CNRS and INERM, the Plan Cancer of the French
ministry for research and health (A.C. and D.F.); the European
Research Council (ERC) under the European Union’s Horizon 2020
research and innovation program (grant agreement 757956 to
A.C.); the LabEx DEEP (ANR-11-LABX-0044 and ANR-10-IDEX-
0001-02 to A.C.); the program Fondation ARC (grant agreement
PJA 20161204869 to A.C.); the Institut Curie (D.F. and A.C.); the
Centre National de la Recherche Scientifique (CNRS) (A.C. and
D.F.); the European Union’s Horizon 2020 research and innovation
program (Marie Skłodowska-Curie grant agreement no. 666003
to S.H.); the National Institutes of Health (grant GM114190
to L.A.M.); and the MIT-France Seed Fund (L.A.M. and M.D.). L.A.M.
is a recipient of Chaire Blaise Pascal by Île-de-France.
Administration. Author contributions: M.D., D.F., and A.C.
conceptualized the project. M.D., D.F., V.I.P.K., and A.C. obtained
funding. V.I.P.K., M.D., D.F., and A.C. designed the experiments.
V.I.P.K., L.Z., F.D., and A.C. obtained the data. V.I.P.K., L.Z., K.A.,
M.B., F.D., L.K.Z., and S.H. produced reagents and provided key
technical expertise. M.W., V.I.P.K., and A.C. preformed image
analysis. M.W., V.I.P.K., S.G.H., V.F.S., and A.C. analyzed
trajectories. S.G.H., E.J.B., and L.A.M. preformed polymer
modeling. V.I.P.K., S.G.H., M.W., V.F.S., E.J.B., L.A.M., D.F., and
A.C. interpreted the results. D.F. and A.C. supervised the
experimental work. A.C. supervised the image analysis work. E.J.B.
and L.A.M. supervised the modeling work. A.C. drafted the
manuscript with input from V.I.P.K., M.W., D.F., S.G.H., E.J.B.,
and L.A.M. All authors participated in reviewing and editing the
manuscript. Competing interests: The authors declare no
competing interests. K.A. is currently employed by TreeFrog
Therapeutic as a research and development engineer in
encapsulation technologies. Data and materials availability:
Cell lines and plasmids are available upon request. All data
and codes are available through GitHub and Zenodo (43), as
summarized in table S2. License information: Copyright © 2022
the authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.science.org/about/science-
licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abi9810
Materials and Methods
Supplementary Text
Figs. S1 to S12
Tables S1 and S2
References (44–49)
Movies S1 to S4
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
Submitted 13 April 2021; resubmitted 2 May 2022
Accepted 21 June 2022
10.1126/science.abi9810
Keizer et al., Science 377, 489–495 (2022)
29 July 2022
7 of 7
|
10.1126_science.abq0595
|
RES EARCH
EVOLUTIONARY ECOLOGY
Rapid plant trait evolution can alter coastal wetland
resilience to sea level rise
M. L. Vahsen1*, M. J. Blum2, J. P. Megonigal3, S. J. Emrich2,4, J. R. Holmquist3, B. Stiller1,
K. E. O. Todd-Brown5, J. S. McLachlan1*
Rapid evolution remains a largely unrecognized factor in models that forecast the fate of ecosystems
under scenarios of global change. In this work, we quantified the roles of heritable variation in plant
traits and of trait evolution in explaining variability in forecasts of the state of coastal wetland
ecosystems. A common garden study of genotypes of the dominant sedge Schoenoplectus americanus,
“resurrected” from time-stratified seed banks, revealed that heritable variation and evolution
explained key ecosystem attributes such as the allocation and distribution of belowground biomass.
Incorporating heritable trait variation and evolution into an ecosystem model altered predictions
of carbon accumulation and soil surface accretion (a determinant of marsh resilience to sea level
rise), demonstrating the importance of accounting for evolutionary processes when forecasting
ecosystem dynamics.
O rganismal traits have long been under-
stood to drive ecosystem functions such
as elemental cycling (1, 2). There is
mounting evidence that heritable trait
variation within species can mediate
ecosystem processes at a magnitude compa-
rable with that of trait variation between spe-
cies (Fig. 1B) (3, 4), and that traits can evolve at
a fast enough pace to generate feedbacks that
alter ecosystem dynamics on the timescale of
current anthropogenic environmental change
(Fig. 1C) (5). Together, this suggests that evo-
lutionary processes play a larger role in the
regulation of ecosystem function than previ-
ously imagined (6–8) (Fig. 1). Despite grow-
ing appreciation for this possibility, efforts
to predict ecosystem function that account
for genetic variation and evolutionary pro-
cesses remain limited in number and scope
(4, 9–12), in part because empirical studies are
still needed to explicitly demonstrate wheth-
er heritable variation and rapid evolution are
important drivers of ecosystem change. Study-
ing the heritable trait variation of organisms
is a necessary step toward understanding
whether organismal evolution can influence
ecosystem dynamics (6). Examining herita-
ble trait variation over historical time might
further reveal how organismal evolution elicits
substantial ecosystem-level change (13, 14).
In coastal marshes, dominant plants act as
ecosystem engineers by contributing to soil
surface accretion, a process that has allowed
1Department of Biological Sciences, University of Notre
Dame, Notre Dame, IN, USA. 2Department of Ecology &
Evolutionary Biology, University of Tennessee, Knoxville, TN,
USA. 3Smithsonian Environmental Research Center,
Edgewater, MD, USA. 4Department of Electrical Engineering
and Computer Science, University of Tennessee, Knoxville,
TN, USA. 5Department of Environmental Engineering
Sciences, University of Florida, Gainesville, FL, USA.
*Corresponding author. Email: mvahsen@nd.edu (M.L.V.);
jmclachl@nd.edu (J.S.M.)
marshes to keep pace with sea level rise for
millennia and is critical to their resilience (15).
Further, the combination of high plant pro-
ductivity and low decomposition rates from
anoxic conditions in coastal marsh soils re-
sults in disproportionately high carbon ac-
cumulation rates per area relative to the soils
of other ecosystems (16, 17). Models and em-
pirical syntheses have demonstrated how
traits and the growth of dominant marsh
plants contribute to these and other eco-
system processes (18, 19). Because coastal
marshes typically have low plant species
diversity, intraspecific trait variation may
play an important role in ecosystem pro-
cesses (12, 20) (Fig. 1, B and C).
Belowground plant traits exert a strong in-
fluence on marsh ecosystem processes. For
example, marsh accretion responds strongly
to annual root turnover, which expands marsh
soils, and plant-mediated decomposition,
which reduces soil volume (21, 22). Below-
ground structures are consequently major
contributors to carbon pools, and below-
ground productivity is tightly linked to carbon
accumulation. Empirical estimates of below-
ground trait variation, heritable or otherwise,
are sparse (23, 24), resulting in the common
simplifying assumption that belowground
traits vary following a fixed proportion to
aboveground traits (Fig. 1D). This is a poten-
tially unrealistic assumption—especially con-
sidering work that suggests that root-to-shoot
ratios can rapidly evolve and exhibit substan-
tial plasticity in response to stress (25, 26)—
that can bias predictions (27, 28).
In this work, we paired a common garden
experiment with an ecosystem model (29) to
quantify the role of heritable variation in plant
traits and of trait evolution in explaining var-
iability in forecasts of carbon accumulation
and soil surface accretion (Fig. 1). We character-
ized heritable trait variation using 16 genotypes
of Schoenoplectus americanus—a dominant
sedge in North American coastal marshes and
the subject of extensive global change research
related to coastal wetlands—and focused on
belowground traits that are known to influ-
ence carbon sequestration and accretion. We
characterized trait variation and evolution
(Fig. 1, B and C) by applying a resurrection
ecology approach (14, 30–32) in which we
“resurrected” genotypes from time-stratified
seed banks from four nearby marshes in the
Chesapeake Bay (figs. S1 to S3 and table S1).
For genotypes from two of the four marshes,
we assessed the role of genotype provenance
(marsh of origin; Corn Island or Sellman Creek),
and age cohort [ancestral (1931 to 1973) or
descendant (1994 to 2016)] in driving trait
variation. To assess potential nonadditive in-
teractions that can be important when scaling
up from genotype to ecosystem (Fig. 1B), we
compared traits of the 16 genotypes grown
in monoculture (four propagules of one geno-
type; n = 3 monocultures per genotype, total-
ing 48 monocultures) with those grown in
polyculture (one propagule each of four geno-
types; n = 48 total polycultures). We quanti-
fied the potential impact of eco-evolutionary
dynamics on ecosystem processes using esti-
mates of heritable variation and evolution
from the common garden experiment to pa-
rameterize a marsh ecosystem model. Togeth-
er, these approaches provide a framework for
integrating data from common garden experi-
ments typical in the field of evolutionary bio-
logy to predictions at the ecosystem scale
appropriate for forecasting ecosystem responses
to global change.
Characterizing heritable variation in
S. americanus traits
S. americanus exhibited considerable heritable
variation in all traits measured in the common
garden experiment (fig. S4). Comparably more
heritable variation was observed in below-
ground traits—such as the magnitude and
distribution of belowground biomass (fig. S4,
B, D, and F)—than in aboveground traits (fig.
S4, A, C, E, and G). For example, heritable
variation explained on average 49.5% of var-
iation in the shape of the root depth dis-
tribution [intraclass correlation coefficient
(ICC), 49.5; 95% confidence interval (CI) (15.7,
77.3)] and 69.1% of variation in root-to-shoot
ratio [ICC, 69.1; (43.7, 87.7)]. These findings
demonstrate the importance of explicitly char-
acterizing variation in belowground traits,
as their variation does not align with that of
traits aboveground.
Heritable variation in S. americanus traits
was structured by evolution captured across
space (across the two provenances for which
there were multiple genotypes; ~2 km of dis-
tance) and time (~50 years between ancestral
and descendant cohorts) (Fig. 2). Differences
Vahsen et al., Science 379, 393–398 (2023)
27 January 2023
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Potential consequences
of heritable variation in
S. americanus on marsh accre-
tion and carbon sequestration.
(A) Current marsh accretion
models do not account for herita-
ble variation in traits and only
account for plastic responses of
aboveground biomass to flooding.
Increased flooding from light
to dark blue environments
increases aboveground biomass.
(B) Heritable variation may
have nonadditive, within-
generation consequences. For
example, interactions between
genotypes can lead to facilitation
(+) or inhibition (–) shifting the
mean trait values of polycultures
and thus shifting the mean
prediction for ecosystem pro-
cesses. (C) Selection may shift
plant trait means, inducing evolu-
tionary change. (D) Variation
in belowground traits may not
scale with variation in above-
ground traits; thus capturing
belowground trait variation
is important for accurately pre-
dicting variation in ecosystem
processes. (E) Within-generation
diversity effects [from (B)]
can evolve [from (C)]. In (B) to
(E), different colored plants
represent different genotypes
of the same species.
in heritable trait variation were reflected in
patterns of genetic variation elucidated with
single-nucleotide polymorphism (SNP) geno-
typing (fig. S3 and table S3). Heritable varia-
tion attributable to genotype, provenance, and
age cohort explained roughly 15 to 50% of
observed variation across all traits (Fig. 2A,
blue shading, and fig. S5) and, for most traits,
exceeded the variation explained by experi-
mental covariates (Fig. 2A, light gray shading;
initial propagule weight, variation in flooding
due to peat levels, and spatial blocking). Dif-
ferences in the shape of root depth distribu-
tions were strongly consistent within age
cohort and provenance (Fig. 2, A and B) [re-
gression coefficient of age (bage) = –0.015
(–0.022, –0.007); regression coefficient of pro-
venance (bprov) = –0.023 (–0.031, –0.015)]. Root-
ing depth became shallower over time within
both provenances, with more root biomass
proportionally allocated near the marsh sur-
face in descendant genotypes (Fig. 2B). Root-
to-shoot ratios exhibited strong signatures of
provenance, with genotypes from Corn Island
having root-to-shoot ratios that were 17.2%
(1.9%, 34.4%) higher than those of genotypes
from Sellman Creek (fig. S6). Comparisons of
ancestral and descendant genotypes also re-
vealed that root-to-shoot ratios have declined
by 8.3% (–5.2%, 19.6%) since the mid-20th cen-
tury (fig. S6), indicating that over time, plants
have allocated fewer resources toward below-
ground biomass. We hypothesize that below-
ground traits may have evolved in response
to anthropogenic nitrogen loading, which has
increased throughout the Chesapeake Bay
over the most recent century (33). Shifts in root
depth distribution in coastal marsh vegeta-
tion have previously been posited to represent
differences in how plants access nutrients
belowground (34). Excess nitrogen may have
alleviated nutrient limitation, reducing the
need for plants to invest in traits that improve
access to belowground resources (35).
Aboveground traits also evolved, but less so
than belowground traits, as evidenced by
smaller effect sizes of age cohort and prove-
nance. For example, on average, stems became
thinner over time, with stem widths declining
5.6% (–3.3%, 14.4%) between ancestral and
descendant cohorts (fig. S7). This pattern mir-
rors changes in stem morphology exhibited by
S. americanus subjected to 30 years of ele-
vated CO2 exposure—a change that can affect
Vahsen et al., Science 379, 393–398 (2023)
27 January 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
A
ag biomass
stem density
stem height
t
i
a
r
t
stem width
bg biomass
root:shoot
root parameter
genotype
provenance
cohort
other
residual
B
0
)
m
c
(
e
c
a
f
r
u
s
l
i
o
s
w
o
e
b
l
h
t
p
e
d
10
20
30
40
50
)
β
(
r
e
t
e
m
a
r
a
p
n
o
i
t
u
b
i
r
t
s
d
i
t
o
o
r
0.900
0.875
0.850
Corn Island
Sellman Creek
ancestral descendant
ancestral descendant
age cohort
0.00
0.25
0.50
proportion of variance
0.75
1.00
1.00
0.75
0.50
cumulative probability
0.25
0.00
Fig. 2. Provenance and age cohort explain considerable variation in traits,
particularly for root depth distribution. (A) Using (generalized) linear mixed
models with provenance and age cohort as fixed effects and genotype as a
random effect, for each trait we decomposed observed trait variation into five
categories: genotype, provenance, cohort, other (covariates in the model
that accounted for exogenous variation because of experimental setup: initial
propagule weight, variation in flooding because of peat level, and spatial
blocking), and residual. Labels “ag biomass” and “bg biomass” represent
aboveground and belowground biomass, respectively. (B) Differences in
belowground biomass distribution with depth according to provenance and age
cohort. Differences in the parameter b (root distribution parameter) are shown in
the inset and were applied to the equation 1 – bdepth to predict the cumulative
proportion of belowground biomass with depth shown in the main figure (34, 45).
The vertical line at 95% cumulative probability indicates the depth at which 95% of
belowground biomass is contained, which is a parameter in the Cohort Marsh
Equilibrium Model (CMEM) (29).
the ability of marshes to withstand storm
surges (36). Like belowground traits, on aver-
age, more variation in aboveground traits
was attributable to provenance than age co-
hort. For example, mean stem heights differed
according to provenance, with plants from
Sellman Creek being 3.0% (–1.5%, 7.5%) taller
than those from Corn Island (fig. S8), whereas
only a 0.3% (–3.8%, 4.7%) difference in stem
height was found between ancestral and de-
scendant plants.
Assessing the strength of nonadditive
interactions between genotypes
Given that there are high levels of standing
genetic diversity within populations of
S. americanus, even at fine spatial scales of a
few meters (30), it is possible that interactions
among genotypes result in nonadditive effects,
in which trait values for a mixture of geno-
types are not equal to the sum of trait values
for individual genotypes (37–39) (Fig. 1B). Con-
sequently, characterizing the direction and
strength of nonadditive effects can be im-
portant for scaling trait variation from geno-
type to ecosystem. Mechanisms that give rise
to nonadditivity can include facilitation (a
positive nonadditive interaction), inhibition
(a negative nonadditive interaction), and se-
lection effects (a positive or negative nonaddi-
tive interaction) (37). Overall, comparisons of
S. americanus genotypes grown in monocul-
ture with those grown in polyculture did not
reveal strong evidence of nonadditive inter-
actions (Fig. 3A). However, for two below-
ground traits—root-to-shoot ratio and root
depth distribution—variation in the strength
of the nonadditive interactions depended on
whether the polycultures were composed of
ancestral genotypes, modern genotypes, or a
mix of both (Fig. 3, B and C and fig S9). Root-
to-shoot ratios were substantially lower than
additive expectations for polycultures com-
posed of descendant genotypes but not for those
composed of ancestral genotypes [banc vs desc =
–0.06 (–0.13, 0.00)] (Fig. 3B), suggesting that
the strength of within-species interactions can
rapidly evolve. Root depth distributions in
mixed polycultures were shallower than addi-
tive expectations, but those composed of only
ancestral or descendant genotypes were sim-
ilar to additive expectations [bmix vs anc = 0.01
(0.00, 0.02); bmix vs desc = 0.01 (0.00, 0.02)] (Fig.
3C). Mixing genotypes from different age co-
horts in our experiment may have increased the
functional diversity of rooting behavior, allow-
ing for better resource partitioning without
the need for deeper rooting (Fig. 3C).
Scaling heritable variation in plant traits to
ecosystem outcomes
Observed heritable trait variation and rapid
evolution drove downstream effects on soil
surface accretion and carbon accumulation
(Fig. 4). We ran simulations of a marsh ac-
cretion model (29) based on conditions at the
Global Change Research Wetland (Edgewater,
MD, USA), which are typical of large areas of
the Chesapeake Bay. We accounted for herita-
ble variation in peak aboveground biomass,
root-to-shoot ratio, and depth of the 95%
cumulative root distribution (while account-
ing for between-trait covariances) (figs. S10
and S11 and table S2). We projected that heri-
table trait variation could result in differences
in marsh elevation gain of up to 5 cm by the
year 2100 [mean elevation = 34.2 cm NAVD88
(North American Vertical Datum, 1988), (32.1,
37.1)]; (Fig. 4A), which is approximately one-
third of the elevation differential between
mean and high tides and thus is consequential
Vahsen et al., Science 379, 393–398 (2023)
27 January 2023
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Genotypes do not exhibit strong
nonadditive interactions in polyculture
overall, but there is evidence that
within-species interactions have evolved
since the mid-20th century. (A to
C) The “scaled difference” on the y axes
indicates the difference between a trait
exhibited by genotypes grown in
polyculture versus what would be
expected of genotypes grown in
monoculture (scaled by the mean
value of the trait for easier comparison
across traits). (A) Scaled differences
across traits overall indicate no significant
nonadditive interactions. Points indicate the
mean difference, and bars indicate 95%
confidence intervals. The strength of within-
species interactions for (B) root-to-shoot
ratio and (C) root depth distribution
varies systematically based on the
composition of the polyculture. For
(B) and (C), points with error
bars represent marginal means with
95% confidence intervals.
A
0.2
observed > predicted
e
c
n
e
r
e
f
f
i
d
l
d
e
a
c
s
0.1
0.0
−0.1
−0.2
predicted > observed
ag biomass
stem density
stem height
stem width
bg biomass
root:shoot
root parameter
B
)
f
f
i
d
l
d
e
a
c
s
(
t
o
o
h
s
:
t
o
o
r
0.1
0.0
−0.1
−0.2
trait
C
)
f
f
i
d
l
d
e
a
c
s
(
r
e
t
e
m
a
r
a
p
t
o
o
r
0.00
−0.02
−0.04
−0.06
ancestral
mix
age cohort
descendant
ancestral
mix
age cohort
descendant
to how much flooding plants will experience.
These differences also result in average ver-
tical accretion rates that could vary more
than 1.5-fold [mean vertical accretion rate =
1.45 mm/year (1.18, 1.81)] (Fig. 4B, width of
histogram). Predicted rates of carbon accu-
mulation at our sites varied up to 0.32 metric
tons C ha−1 year−1 because of heritable varia-
tion [mean C accumulation rate = 0.35 (0.21,
0.53); Fig. 4C, width of histogram], which
would lead to estimates of soil carbon stor-
age through 2100 varying by more than two-
fold in highly organic peat-forming marshes,
such as those in the Chesapeake Bay.
Vertical accretion rates were 8% higher and
carbon accumulation rates were 18% higher
for ancestral cohorts than descendant cohorts.
(Fig. 4, B and C). This suggests that modern
S. americanus marshes may be less resilient to
sea level rise and store less carbon compared
with marshes from the mid-20th century be-
cause of organismal evolution. Across space,
changes in plant traits drove soil accretion at
Corn Island to be 3% higher than at Sellman
Creek and drove soil carbon accumulation rates
to be 6% higher (Fig. 4, A to C, green versus
gold). Additional evidence from a separate ex-
periment suggests that our estimates of the
impact of heritable variation and rapid tem-
poral evolution on accretion and carbon accu-
mulation are robust and possibly conservative
(40) (figs. S12 and S13 and table S4).
The effect of evolution on ecosystem pro-
cesses captured in this work is comparable
with the effects of rapid environmental change.
For example, the magnitude of evolution’s
influence on vertical accretion that we found
is similar to the modeled effect of shifts in salt
marsh mineral accretion rates from changes in
S. americanus stem morphology over 11 years
of exposure to elevated CO2 in a different but
related study (36). Additionally, by running
simulations of the ecosystem model in which
we varied the total amount of sea level rise, we
found that the percent difference in carbon
accumulation rate between ancestral and de-
scendant genotypes was approximately equal
to an additional 4 cm of sea level rise by 2100.
The effects of organismal evolution on carbon
accumulation are particularly notable given
that they would partially offset predicted large
increases in future carbon storage in response
to global change factors such as increasing at-
mospheric CO2 (41, 42).
Much of the projected variation in our pre-
dictions of marsh accretion and carbon ac-
cumulation was attributable to belowground
trait variation (Fig. 4E), which suggests that
further study of the effects of genotypic var-
iation on belowground traits can improve
forecasts of coastal marshes in which surface
accretion is predominantly driven by organic
matter accretion. This was revealed by run-
ning simulations in which we only accounted
for between-genotype variation in aboveground
biomass, keeping belowground traits (root-
to-shoot ratio and depth of the 95% cumula-
tive root distribution) constant, which aligns
with the simplifying assumption that below-
ground traits covary with aboveground traits
(Fig. 1D). Failing to account for heritable var-
iation in belowground traits beyond the var-
iation in aboveground biomass dramatically
decreased the predicted uncertainty for accre-
tion and carbon sequestration (Fig. 4, A, D,
and E). For example, the variance in final
predicted marsh elevation decreased by 68%
when only variance in aboveground biomass
was included in the simulations (Fig. 4E).
Accounting for heritable belowground trait
variation also altered the average predictions
Vahsen et al., Science 379, 393–398 (2023)
27 January 2023
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
A
40
)
8
8
D
V
A
N
m
c
(
n
o
i
t
a
v
e
e
l
h
s
r
a
m
35
30
25
D
40
)
8
8
D
V
A
N
m
c
(
n
o
i
t
a
v
e
e
l
h
s
r
a
m
35
30
25
2020
2040
2060
year
2080
2100
B
200
100
t
n
u
o
c
0
200
100
0
39
37
35
33
C
t
n
u
o
c
E
)
8
8
D
V
A
N
m
c
(
0
0
1
2
r
a
e
y
n
i
n
o
i
t
a
v
e
e
l
1.2
1.5
vertical accretion rate (mm yr
1.8
2.1
−1)
0.2
0.6
carbon accumulation rate (t C ha
0.3
0.5
0.4
−1yr
0.7
−1)
2020
2040
2060
year
2080
2100
ag only
ag + bg
scenario
Fig. 4. Accounting for heritable trait variation and evolution alters
forecasts of marsh ecosystem structure and function. (A) Using
CMEM (29), we simulated marsh elevation gain to the year 2100. Light
gray lines indicate model simulations (n = 1000) that account for
variation in aboveground biomass, root-to-shoot ratio, and 95%
cumulative root distribution depth due to genotype. Green and gold
lines indicate mean predictions for genotypes from Corn Island and
Sellman Creek, respectively, and the shapes at year 2100 indicate
age cohorts (circle, ancestral; triangle, descendant). (B) Average
vertical accretion rate explained by variation in traits due to heritable
variation (histogram) and average provenance and age cohort trait
values (points). (C) Average carbon accumulation rate explained by
variation in traits due to heritable variation (histogram) and average
provenance and age cohort trait values (points). (D) Simulations of
CMEM that account for heritable variation in only aboveground
biomass due to genotype, provenance, and age cohort. (E) Distribution
of final predicted elevation of CMEM simulations for scenarios in
which aboveground and belowground traits were varied [“ag + bg”
from (A)] and for which only aboveground biomass was varied [“ag only”
from (D)].
for each age cohort and provenance. Notably,
differences in ecosystem outcomes between
age cohorts were larger than those between
provenances when belowground trait varia-
tion was considered (Fig. 4A), but the oppo-
site was true when only aboveground trait
variation was considered (Fig. 4D). However,
although belowground trait variation drove
variation in ecosystem processes in a highly
organic marsh in the Chesapeake Bay, heri-
table variation and evolution of aboveground
plant traits may play a larger relative role in
marshes in which accretion rates are driven
by mineral sediment capture aboveground, a
process mediated by stem density and mor-
phology (36).
Observed shifts in plant traits in long-term
studies of coastal marshes have been previ-
ously thought to reflect plastic responses in-
duced by exposure to environmental pressures
(Fig. 1A). For example, there is evidence that
S. americanus morphology has changed in
response to elevated CO2 and increased nitro-
gen deposition over the course of 30 years,
with putatively plastic trait changes having
substantial consequences for model predic-
tions of aboveground sediment capture rate
(36). Our findings, along with additional evi-
dence of the adaptive capacity of S. americanus
(14, 30), offer an updated perspective suggest-
ing that plants can evolve at a pace and mag-
nitude that feeds back on ecosystem-level
processes (6). Failure to account for heritable
variation and rapid evolutionary change in
ecosystem models (43, 44) might therefore
mischaracterize the role that organismal re-
sponse plays in ecosystem resilience to environ-
mental change that could systematically alter
ecosystem-level predictions. For example,
our results suggest that failing to account
for decadal-scale evolutionary change may
overestimate the potential for coastal marshes
Vahsen et al., Science 379, 393–398 (2023)
27 January 2023
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
to build elevation and store carbon. It is in-
creasingly apparent that organismal evolution
and ecosystem development occur on similar
time scales, which can elicit feedbacks (6). In-
tegrative approaches will thus be increasingly
important as anthropogenic change continues
to challenge our ability to forecast the re-
silience of at-risk ecosystems such as coastal
marshes (12).
RE FE RENCES AND N OT ES
1.
2.
J. Chacón-Labella et al., Trends Ecol. Evol. (2022).
J. P. Grime, Plant Strategies, Vegetation Processes, and
Ecosystem Properties (John Wiley & Sons, 2006).
3. S. Díaz, M. Cabido, Trends Ecol. Evol. 16, 646–655
(2001).
4. S. Des Roches et al., Nat. Ecol. Evol. 2, 57–64 (2018).
5. N. G. Hairston Jr., S. P. Ellner, M. A. Geber, T. Yoshida, J. A. Fox,
6.
Ecol. Lett. 8, 1114–1127 (2005).
J. G. Monroe et al., Trends Ecol. Evol. 33, 213–225
(2018).
7. T. G. Whitham et al., Nat. Rev. Genet. 7, 510–523
(2006).
8. L. A. Donovan, C. M. Mason, A. W. Bowsher,
E. W. Goolsby, C. D. A. Ishibashi, J. Ecol. 102,
302–314 (2014).
9. B. Matthews et al., Ecol. Lett. 14, 690–701
(2011).
10. J. Pastor, Ecosystems 20, 245–252 (2017).
11. C. Loehle, J. H. K. Pechmann, Am. Nat. 132, 884–899
(1988).
12. T. J. Mozdzer, M. K. McCormick, I. J. Slette,
M. J. Blum, J. P. Megonigal, Sci. Total Environ. 853,
157846 (2022).
13. L. J. Weider, P. D. Jeyasingh, D. Frisch, Evol. Appl. 11, 3–10
(2017).
14. M. J. Blum et al., Evol. Lett. 5, 422–431 (2021).
15. M. L. Kirwan, J. P. Megonigal, Nature 504, 53–60
(2013).
16. F. Wang et al., Natl. Sci. Rev. 8, nwaa296
(2020).
18. J. T. Morris, P. V. Sundareshwar, C. T. Nietch,
B. Kjerfve, D. R. Cahoon, Ecology 83, 2869–2877 (2002).
19. S. Fagherazzi et al., Rev. Geophys. 50, RG1002
(2012).
20. T. B. H. Reusch, A. R. Hughes, Estuaries Coasts 29, 159–164
(2006).
21. J. T. Morris et al., Earths Future 4, 110–121 (2016).
22. G. L. Noyce, A. J. Smith, M. L. Kirwan, R. L. Rich,
J. P. Megonigal, Nat. Geosci. 10.1038/s41561-022-01070-6
(2023).
23. G. Ottaviani et al., Trends Ecol. Evol. 35, 763–766
(2020).
24. G. T. Freschet et al., New Phytol. 232, 1123–1158 (2021).
25. F. S. Chapin III, K. Autumn, F. Pugnaire, Am. Nat. 142,
S78–S92 (1993).
26. S. Aikio, M. Mari Markkola, Evol. Ecol. 16, 67–76 (2002).
27. Q. Chen, Y. Wang, C. B. Zou, Z.-L. Wang, Wetlands 37,
985–995 (2017).
28. C. P. Carmona et al., Nature 597, 683–687 (2021).
29. J. R. Holmquist et al., Zenodo (2022).
30. J. L. Summers, B. Bernik, C. J. Saunders,
J. S. McLachlan, M. J. Blum, Evol. Appl. 11, 1715–1731
(2018).
31. M. L. Vahsen et al., Evol. Appl. 14, 2831–2847 (2021).
32. C. Y. Lumibao, L. Torres Martínez, J. P. Megonigal,
S. A. Van Bael, M. J. Blum, Mol. Ecol. 31, 4571–4585
(2022).
33. D. C. Fisher, M. Oppenheimer, Ambio 20, 102–108 (1991).
34. T. J. Mozdzer, J. A. Langley, P. Mueller, J. P. Megonigal,
Biol. Invasions 18, 2619–2631 (2016).
35. L. A. Deegan et al., Nature 490, 388–392 (2012).
36. M. Lu, E. R. Herbert, J. A. Langley, M. L. Kirwan, J. P. Megonigal,
Nat. Clim. Chang. 9, 764–768 (2019).
37. A. R. Hughes, B. D. Inouye, M. T. J. Johnson,
N. Underwood, M. Vellend, Ecol. Lett. 11, 609–623
(2008).
38. K. M. Crawford, K. D. Whitney, Mol. Ecol. 19, 1253–1263
(2010).
39. J. A. Schweitzer et al., in Trait-Mediated Indirect Interactions:
Ecological and Evolutionary Perspectives, T. Ohgushi, O.J. Schmitz,
R.D. Holt, Eds. (Cambridge University Press, 2012), pp. 371–390.
40. See the Supplementary Materials.
41. A. J. Rietl, J. P. Megonigal, E. R. Herbert, M. L. Kirwan,
Geophys. Res. Lett. 48, e2020GL092051 (2021).
42. M. A. Pastore, J. P. Megonigal, J. A. Langley, Biogeochemistry
17. J. R. Holmquist et al., Environ. Res. Lett. 13, 115005
133, 73–87 (2017).
(2018).
43. M. Schuerch et al., Nature 561, 231–234 (2018).
44. M. L. Kirwan, S. Temmerman, E. E. Skeehan,
G. R. Guntenspergen, S. Fagherazzi, Nat. Clim. Chang. 6,
253–260 (2016).
45. M. R. Gale, D. F. Grigal, Can. J. For. Res. 17, 829–834 (1987).
46. M. L. Vahsen et al., Zenodo (2022).
AC KNOWLED GME NTS
We are grateful for the support provided by technicians and
volunteers at the Smithsonian Environmental Research Center. In
particular, we thank G. Peresta and A. Peresta for assisting with
the experimental setup and W. Vahsen for assisting with data
collection. We also thank undergraduate students at the University
of Notre Dame (E. Ackerley, A. Appling, B. Brown, E. Nguyen,
L. Onken, and C. Powers) for assisting with seed germination and
root washing, as well as graduate students H. Kleiner and
J. Summers for helping with root washing. Last, we thank
J. T. Morris for his efforts in transitioning the Cohort Marsh
Equilibrium Model (CMEM) into an R package. Funding: This work
was supported by National Science Foundation DEB 1655781,
DEB 1655702 (M.J.B., J.P.M., J.S.M.); National Science Foundation
DEB 09050080, DEB 1457100, DEB 1557009 (J.P.M.); United
States Coastal Research Program W912HZ-2020003 (M.L.V.);
and the Coastal Carbon Research Coordination Network
DEB-1655622 (J.R.H.). Author contributions: Conceptualization:
M.L.V., M.J.B., J.P.M., and J.S.M. Methodology: M.L.V., B.S.,
and J.S.M. Software development: J.R.H. and K.E.O.T-B. Analysis:
M.L.V. and S.J.E. Funding acquisition: M.L.V., M.J.B., J.P.M.,
and J.S.M.; Writing – original draft: M.L.V. and J.S.M. Writing –
review and editing: M.L.V., M.J.B., J.P.M., S.J.E., J.R.H., B.S., K.E.O.T-B.,
and J.S.M. Competing interests: Authors declare that they
have no competing interests. Data and materials availability:
All data and code are available via Zenodo (46). License
information: Copyright © 2023 the authors, some rights reserved;
exclusive licensee American Association for the Advancement of
Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abq0595
Materials and Methods
Figs. S1 to S13
Tables S1 to S4
References (47–70)
Submitted 22 March 2022; resubmitted 4 October 2022
Accepted 22 December 2022
10.1126/science.abq0595
Vahsen et al., Science 379, 393–398 (2023)
27 January 2023
6 of 6
|
10.1126_science.abp8948
|
RES EARCH
SOLID-STATE PHYSICS
Emergent symmetry in a low-dimensional
superconductor on the edge of Mottness
P. Chudzinski1,2†*, M. Berben3,4†, Xiaofeng Xu5, N. Wakeham6, B. Bernáth3,4, C. Duffy3,4, R. D. H. Hinlopen7,
Yu-Te Hsu3,4, S. Wiedmann3,4, P. Tinnemans4, Rongying Jin8, M. Greenblatt9, N. E. Hussey3,4,7*
Upon cooling, condensed-matter systems typically transition into states of lower symmetry. The converse—i.e.,
the emergence of higher symmetry at lower temperatures—is extremely rare. In this work, we show how an
unusually isotropic magnetoresistance in the highly anisotropic, one-dimensional conductor Li0.9Mo6O17 and
its temperature dependence can be interpreted as a renormalization group (RG) flow toward a so-called
separatrix. This approach is equivalent to an emergent symmetry in the system. The existence of two distinct
ground states, Mott insulator and superconductor, can then be traced back to two opposing RG trajectories.
By establishing a direct link between quantum field theory and an experimentally measurable quantity, we
uncover a path through which emergent symmetry might be identified in other candidate materials.
S ymmetry is one of the most inspiring
concepts in mathematics and physics,
with an impact that spans multiple dis-
ciplines from art to applied physics and
chemistry. In the context of quantum
mechanics, symmetry enables us to identify
conservation laws (through Noether’s theo-
rem) and to group states according to their
symmetry classes. Frequently, this is the only
strictly exact information that we have about a
given many-body system. Among the plethora
of symmetry-related phenomena, the notion
of symmetry breaking (i.e., the lowering of sym-
metry) plays a particularly prominent role. It
provides us with a framework to understand
phase transitions in condensed-matter systems
(1) and, on a more fundamental level, helps
explain mass generation through the Higgs
mechanism (2). Recently, a counter idea has
appeared (3): Can the symmetry of the system
become higher upon decreasing the tempera-
ture? The name emergent symmetry reflects
the fact that a more symmetric state emerges
at low energies from an initial high-energy
state that does not explicitly reveal such sym-
metry. The idea that mutual cooperation of
various strong correlations can lead to a low-
energy state of higher symmetry has become
one of the cornerstones for understanding the
exotic ordering in general and indicates that
1School of Mathematics and Physics, Queen’s University Belfast,
Belfast, UK. 2Institute of Fundamental Technological Research,
Polish Academy of Sciences, Warsaw, Poland. 3High Field
Magnet Laboratory (HFML-EMFL), Radboud University,
Nijmegen, Netherlands. 4Institute for Molecules and Materials,
Radboud University, Nijmegen, Netherlands. 5Key Laboratory of
Quantum Precision Measurement of Zhejiang Province,
Department of Applied Physics, Zhejiang University of
Technology, Hangzhou, China. 6Center for Space Sciences and
Technology, University of Maryland Baltimore, Baltimore, MD,
USA. 7H. H. Wills Physics Laboratory, University of Bristol,
Bristol, UK. 8Center for Experimental Nanoscale Physics,
Department of Physics and Astronomy, University of South
Carolina, Columbia, SC, USA. 9Department of Chemistry and
Chemical Biology, Rutgers University, Piscataway, NJ, USA.
*Corresponding author. Email: p.chudzinski@qub.ac.uk (P.C.);
n.e.hussey@bristol.ac.uk (N.E.H.)
†These authors contributed equally to this work.
multiple order parameters may be united (4)
under the umbrella of a single parent state (5).
Despite sustained interest, it has proved
very difficult to obtain such a state experimen-
tally, with the notable exception of interacting
spins on an insulating one-dimensional (1D)
chain tuned by a magnetic field (6, 7). In con-
densed matter, there is always the problem of
additional symmetry-breaking terms becom-
ing relevant—e.g., upon Mott gap or pseudo-
gap opening—and obscuring the detectability
of any emergent symmetry. As a result, the
main experimental focus has shifted to cold-
atom systems that offer full control over the
parameters of the model (8), thereby providing
hope of engineering any quantum simulator,
albeit with the added difficulty of reaching
sufficiently low temperatures.
In any many-body system with a continuous
spectrum, the notion of symmetry can be
rather subtle. Let us imagine a system of spins
freely fluctuating in plane. In the charge sec-
tor, this corresponds to freely flowing charges
described by a free-particle theory. Adding an
Ising-type perturbation g can drive the system
toward localization, but just before it enters the
gapped state, there is a special point in param-
eter space where the symmetry is enlarged—
i.e., where the spins can rotate equally well
in and out of plane. Likewise, the charge can
become equally localized and itinerant, span-
ning the entire manifold of available states.
Checking whether the symmetry is global (i.e.,
holding at all length and timescales) is an im-
mense task. If the corresponding quantum
field theory (QFT) is renormalizable, however,
then we can apply renormalization group (RG)
methods and inspect the corresponding RG
flow (Fig. 1A)—that is, whether the perturba-
tion g increases or decreases as we integrate
out the highest energy degrees of freedom. If g
is unaffected, then the system must be at the
special point (the so-called separatrix in RG
parlance), and thus we have proof of such hid-
den or emergent symmetry.
Physical realization of emergent symmetry
The challenge, however, is to access this flow
experimentally. Our proposed solution is a
model in which g depends on an external (e.g.,
magnetic, electric, or strain) field X, whereas
the hydrodynamic, free theory—with its veloc-
ities vi, compressibilities Ki, etc.—stays unaf-
fected. Because the system is a perfect conductor
in a hydrodynamic description, its resistivity r
can be expressed as g Xð ÞnM vi; Ki
Þ, where
n is a finite number and M is a correlation
function for the free theory.M vi; Ki
Þ is usually
a complicated functional, a result of years of
study aimed at solving a given QFT, but it has
one important feature: It does not depend on g
and so is independent of X. In this circumstance,
the relative change in the resistivity in response
to the external field becomes dependent solely
on g and the dependence of g on any other
variable parameter, such as temperature.
ð
ð
In this work, we consider a correlated metal
in which all of these theoretical conditions are
seemingly met, with magnetic field H play-
ing the role of X. Li0.9Mo6O17 (LMO) is a low-
dimensional system that both lies close to the
Mott transition (9, 10) and can host super-
conductivity (11, 12). The fact that two distinct
ground states appear so close energetically
raises the prospect of bringing the two states to
a degenerate point (the separatrix in Fig. 1A)
and thereby realizing emergent symmetry.
Being electronically 1D (12), LMO is also a
viable candidate for the realization of the
1D QFT solution known as the Tomonaga-
Luttinger liquid (TLL) state. This is promising
for several reasons. First, the exact solution for
the TLL is well established, and thus its man-
ifestation can be robustly verified in LMO if
and only if g[X] behaves the same for re-
sistivities along different geometrical direc-
tions. Second, LMO is nearly ¼ filled, and its
parametrization implies that the relevant TLL
(compressibility) parameter Kr must be very
close to the critical value K ∗
4= , thus plac-
ing LMO close to the RG flow separatrix (9).
Third, as described below, it has already been
shown (13, 14) that the remaining incommen-
surability in LMO (from the noninteger Li)
can be compensated by an auxiliary potential
whose amplitude depends strongly on the mag-
nitude of H. This indicates not only that the
relevant perturbation (labeled hereafter g3 to
reflect the fact that it is umklapp-like) has a
strong H dependence, but it also provides a
mechanism to steer the system toward the sep-
aratrix (Fig. 1B). At this special point, the
charge has equal probability of being localized
(back-scattered) or itinerant (forward-scattered),
as shown schematically in Fig. 1C. Below, we
provide evidence for this approach to the sepa-
ratrix in LMO.
r ¼ 1
Structurally, LMO comprises stacks of con-
ducting chains (Fig. 2A). A pair of 1D dxy bands
cross the Fermi level close to commensurate ¼
Chudzinski et al., Science 382, 792–796 (2023)
17 November 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Emergent symmetry in a
1D metal. (A) Parametric plot of
tentative RG flow of the system
n
(cid:3)
g3 l½ (cid:2); Kr (cid:3) K(cid:4)
r
on the
l½ (cid:2)
o
(cid:4)
r is the
plane, where Kr (cid:3) K(cid:4)
distance from the critical point
for the TLL parameter Kr.
With lowering energy, the stan-
dard (BKT) trajectory (plus
incommensurability) drives the
system away from the separatrix
(green dashed line) and toward
either the gapless, metallic state
or the gapped (Mott) insulating
state (black dotted lines).
(B) The emergence of the highly
relevant, random component of
g3 steers the system back toward
to the separatrix. (C) Scattering
of fermions in the charge channel
(irrespective of spin orientation):
In real space, an electron
(blue) or hole (red) propagating
in the lattice (purple circles)
encounters a slower obstacle
particle (green) and may either
undergo forward scattering
(f-outgoing, dashed arrows) or
umklapp scattering (u-outgoing
solid arrows). In real space,
the arrows correspond to currents
of probability. For the charge-SU(2)
symmetric case, the amplitudes of both processes (f/u) are equal for both electrons and holes. (Inset) Reciprocal
space image of the same processes. Summing over all f-events gives a term proportional to a gradient of the charge
density bosonic field (times Kr), whereas summing over all u-events gives a cosine of this field (times g3).
filling. Only the slight incommensurability
(from the noninteger Li) prevents the Mott
gap from forming, and at high T, LMO exhibits
metallic (T-linear) resistivity. Although prox-
imity to commensurability makes the status quo
fragile, in the canonical Berezinskii-Kosterlitz-
Thouless (BKT) picture of phase transitions in
a TLL with a cosine perturbation (the sine-
Gordon model), the expectation is that g3 will
become increasingly irrelevant with lowering
T and that the metallic state will be preserved,
as indicated by the black dotted line in Fig. 1B.
One of the distinctive elements of the LMO
band structure is the propensity of the other
4d bands to form (dark) excitons (Fig. 2C)
(13). These excitons introduce new scattering
centers below a temperature scale TH ≈ 100 to
150 K and, through these, an auxiliary poten-
tial (Fig. 2D, red crosses) that divides the chains
into segments of randomized lengths, thereby
bringing the mobile carriers closer to or further
away from commensurability. Indirect evidence
for these excitons and their influence on the
mobile carriers was reported in a recent angle-
dependent magneto-resistance study (14). As a
result of this interaction, the amplitude of g3
also becomes randomized. We argue that the
emergence of this highly relevant, random
component of g3 is the key perturbation that
brings the system back toward the separatrix.
In those chain segments closest to the sepa-
ratrix, the amplitudes of backward and for-
ward scattering of fermions (whether they be
electrons or holes) become equal (Fig. 1C) as
will the likelihood of any instability associated
with these scattering amplitudes. Whereas back-
ward scattering favors instability and a gapped
phase with ℤ2 symmetry between occupied
and unoccupied sites, forward scattering fa-
vors a metal with O(2) symmetry with a free
choice of gauge field. Together at the sepa-
ratrix, they form a state with broader SU(2)
symmetry (15). In this way, an approach to the
separatrix can be interpreted as a manifesta-
tion of emergent symmetry—the symmetry in
question being the equal probability to form
a Mott-localized phase or a metallic and ulti-
mately superconducting (SC) phase.
Obtaining experimental evidence for such
cooperative “steering” rests on finding a quan-
tity that will directly track the RG flow of g3
itself. In the present case, the relevant quantity
is the magnetoresistance (MR). The scattering
centers (excitons) randomly modify the ampli-
tude of g3, and the randomness of this effect
increases with H, resulting in a positive nor-
mal state (i.e., non-SC state) MR. This field-
induced effect enables us to detect when the
RG flow of the system becomes steered toward
the separatrix (Fig. 1B). As we will show below,
for the specific circumstances found in LMO, a
—the
linear T dependence of
inverse square root MR—is expected with an
offset that is proportional to its initial distance
from the separatrix.
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
r 0ð Þ=Dr Hð
Þ
p
Experimental evidence for emergent symmetry
Having sketched out the hypothesis, we now
turn to the experimental study. Figure 3, A, C,
and E, shows, respectively, the in-chain resis-
tivity rb(T) for two SC single crystals and one
non-SC crystal of LMO. In all three samples,
rb(T) is T-linear down to T = TH (marked by
vertical arrows in Fig. 3, A, C, and E), below
which rb(T) develops upward curvature. At a
lower temperature scale, Tmin ≈ 25 to 30 K,
rb(T) passes through a minimum [the origin of
which has remained a mystery for several de-
cades (16–19)] before reaching a maximum at
the onset of superconductivity or diverging (in
the case of the insulating sample). The lack of
any evidence for lattice dynamics associated
with a charge density wave below Tmin (18), as
well as the divergent Lorenz ratio (20) (see
below), allows us to exclude electron-phonon
scattering as the origin of the T-linear resistiv-
ity in LMO. According to TLL theory, T-linear
r(T) implies that the TLL parameter Kr = ¼.
The upturn at Tmin then identifies the energy
scale at which the system evolves from a re-
gime dominated by scattering (on excitonic
fluctuations) to one dominated by tunneling
(through excitonic roadblocks) (21). Whereas
the correlation function M and its T depen-
dence change at this point, crucially, the func-
tional dependence of the resistivity on g3 stays
the same. As a result, the relation between the
MR and g3 is unaffected across Tmin.
¼
p
p
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
r 0ð Þ=A
At low fields, the transverse MR (H//c) of
all three samples is positive and varies as H2
over the entire temperature range studied (22).
Taking the initial slopes A of the H2 MR at
each T (normalized to 1 T), we obtain the
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
r 0ð Þ=Dr Hð
resultant plots of
Þ
for the three samples shown, respectively, in
Fig. 3, B, D, and F. Despite the complex and
p
varied form of rb(T),
displays a
simple T-linear dependence from 300 K down
to 2 K with an absolute magnitude that is
comparable in all three samples. Notably, the
magnitude of Dr Hð
Þ=r 0ð Þ is orders of magni-
tude larger than what one would expect from
Boltzmann transport theory (22). Moreover, as
shown in Fig. 3H, the form and magnitude of
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
r 0ð Þ=Dr Hð
Þ
Chudzinski et al., Science 382, 792–796 (2023)
17 November 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
p
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
r 0ð Þ=Dr Hð
Þ
are found to be markedly in-
sensitive to the relative orientation of the current
and/or the applied field [differing by only a
factor of 4, whereas r(0) along the different
crystallographic axes differs by 3 orders of mag-
nitude], which is in marked contrast with ex-
pectations for and observations (23) in a quasi-1D
Fermi-liquid. Such anomalies compel us to seek
an explanation for this MR behavior that lies
beyond standard Boltzmann theory.
In a TLL, the resistivity can be expressed as
an amplitude for back-scattering multiplied by
a correlation function of the bosonic fields
r T ; Hð
Þ ¼ g2
3 T ; Hð
ÞMij Tð Þ
ð1Þ
The correlation function Mij contains infor-
mation about multiple parameters associated
with the TLL, including the velocity of the charge
mode vr+ and the corresponding TLL param-
eter Kr+. Any field dependence in Kr+ would
lead to a nonmonotonous and rapidly changing
Mij and, in turn, a very complicated MR sig-
nal. This is not what we see experimentally. Thus,
we can assume that Kr+ and vr+ are indepen-
dent of H and that the MR depends only on g3
dr=dH ¼ 2g3 dg3=dH
ð
ÞMij Tð Þ
ð2Þ
and there is one dominant mechanism scatter-
ing the 1D carriers out of their chiral trajec-
tory. To obtain the low-field MR, we calculate
Dr Hð
Þ ¼ ðdr=dHÞ (cid:5) DH and then divide by
the zero-field resistivity r(0) to obtain
Þ
Dr Hð
r 0ð Þ
(cid:5)
(cid:6)
¼ 2@H
¼ 2@H ln g3 Hð
f
dg3
g3
½
DH
gDH
(cid:2)
Þ
ð3Þ
Hence, if only g3 depends on H, Mij —the
most complicated term in r(T,H)—drops out
of the expression for Dr Hð
Þ=r 0ð Þ. The Mij func-
tions are different for rxx, ryy, and rzz, but the
g2
3 prefactor remains the same. From this, one
deduces that in the TLL framework, Dr Hð
Þ=
r 0ð Þ exhibits similar behavior for all orienta-
tions of current and field with only the pre-
factor depending on H. Thus, the MR becomes
effectively the same for all orientations, as ob-
served (Fig. 3H), which confirms the validity of
the TLL description.
½
Because the running variable l of the RG can
be linked with temperature l∼ (cid:3)ln T =Lð
(cid:2)
Þ
(where L is the high-energy cutoff in the
system), the T dependence of the MR also
tracks the RG flow of g3[l] and thus allows
direct access to the RG trajectory itself (Fig.
1B). More precisely, the inverse square ratio of
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
∼ @H ln g3 H; l ¼
Þ=Dr H; T
r 0; Tð
the MR
½fð
Þ
ð
Þ(cid:2)gÞ(cid:3)1=2 . This statement is actually
(cid:3)ln T =Lð
quite strong: No matter what the prefactor’s
dependence on g3 (it might even change as the
system goes from one conductivity regime to
another), the MR will always depend on this
logarithmic derivative. Therefore, if there exists
p
Fig. 2. Crystallographic and electronic structure of LMO. (A) 3D crystal structure showing isolated,
conducting, and zigzag chains of MoO6 octahedra along the b axis in dark purple and nonconducting
octahedra and MoO4 tetrahedra in light orange. Li ions are shown as green spheres. (B) Simplified Fermi surface
of LMO showing the weakly dispersive, quasi-1D bands along the a axis caused by the weak interchain hopping
energy. Note that the actual Fermi surface is close to being ½-filled owing to the zigzag nature of the chains.
(C) This small interchain hopping nevertheless causes an energy gap around the Fermi surface (schematic orange
curves). The small energy gap easily allows excitation of electron-hole pairs—i.e., excitons. (D) Schematic structure
of LMO including the 1D chains (boxes) and lattice periodicity (circles). Below TH ≈ 150 K, the system couples to
an auxiliary potential (crosses) associated with these excitons, which tends to fix its small incommensurability
and hence reactivate the umklapp g3 terms. Owing to the random distribution of scattering centers, different regions
fall closer to or further away from incommensurability, as indicated by the intensity of the yellow background.
As a result, the amplitude of g3 also becomes randomized.
½
f
ln g3 lð Þ
any characteristic energy scale l0 below which
the system opens a gap (or alternatively, g3 dis-
appears exponentially), then this must be re-
g(cid:3)1=2.
flected in the T dependence of@l
(cid:2)
The only way this quantity can stay linear is on
approach to the separatrix. Within this picture,
based on a modification of the known BKT flow
(Fig. 3G, inset), the offset in the MR [the inter-
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Þ=Dr H; T
r 0; Tð
cept in
at zero tempera-
Þ
ð
ture] can be related to an initial distance from
the separatrix before g3 randomness emerges,
p
as illustrated in Fig. 3G, where we observe that
the system tends to move closer to the sepa-
ratrix with decreasing temperature—an oppo-
site trend to that usually found in a doped 1D
Mott system (Fig. 1A) (24). This explains one
further notable detail of the experimental data.
The 0 K intercept in the more metallic (i.e., SC)
samples is finite (Fig. 3, B, D, and H, insets),
whereas in the non-SC sample with a divergent
resistivity, it is effectively zero (Fig. 3F, inset), in
agreement with the expectation that the latter
Chudzinski et al., Science 382, 792–796 (2023)
17 November 2023
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
p
p
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
r 0ð Þ=Dr Hð Þ
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
r 0ð Þ=Dr Hð Þ
. Vertical arrows again indicate where rb(T) deviates from linearity [as shown in (A)].
Fig. 3. MR as a probe of emergent symmetry in LMO. (A and C) In-chain resistivity rb(T) for two SC
samples with Tc = 2.15 K (A) and 2.2 K (C). Dashed lines are linear fits to the high-T data, and colored arrows
indicate roughly where rb(T) deviates from linearity. The inset in (C) shows a zoomed-in view near Tc of
rc(T) taken on a piece of the same crystal (12). (E) rb(T) for a non-SC sample. The deviation from the
high-T T-linear behavior (dashed line) sets in at TH ≈ 150 K. (B, D, and F)
extracted from
the inverse square root of the coefficient A of the low-field H2 b axis MR [and divided by the zero-field
resistivity r(0)] for the samples shown in (A), (C), and (E), respectively. Dashed lines highlight the T-linearity
of
(Insets) Zoomed-in views of the low-T region for each sample. Note that for both SC samples [(B) and (D)],
the intercept is finite, whereas for the non-SC sample (F), it is negligible, as expected were it to locate
closer to the separatrix (see text). (G) Connection between the MR, as expressed through the derivative
@H ln g3 Hð Þ
(cid:2)
½
lower the line at Kr (cid:3) K(cid:4)r e 0:1, the further away the system is initially from the separatrix [green dashed
line in (B)]. (Main panel) Resultant T dependence of @H ln g3 Hð Þ
(for the yellow and
(cid:2)
blue trajectories) to be compared with experiment. The closer that the system is initially to the separatrix, the
lower the intercept at zero temperature. [Corresponding curves for the green and red trajectories in the
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
r 0ð Þ=Dr Hð Þ
inset (not shown in this figure) simply have a larger offset.] Arb. Units, arbitrary units. (H)
for various SC samples (Tc = 2.0 ± 0.2 K) with different orientations of current and field, normalized to
their (extrapolated) absolute value at 300 K. The normalization factors are (in parentheses): I//a; H//c (64);
I//c; H//a (119); I//c; H//b (150); I//c; and H//c (160). (Inset) Zoomed-in view of the low-T region showing,
again, the finite intercept for SC samples.
g, and RG flow. (Inset) Parametric plot of g3[l] flow as it gradually approaches the separatrix—the
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
r 0ð Þ=Dr Hð Þ
g(cid:3)1=2 e
plot
p
p
f
f
½
lies closer to the separatrix (i.e., closer to the
Mott state) already at elevated T. This in turn
suggests that SU(2) symmetry in the charge
sector does not appear accidentally (e.g., owing
to a coincidence of parameters in LMO) but
rather emerges during the RG flow itself.
The validity of this interpretation of the MR
rests on two conjectures: (i) that the only term
responsible for finite resistance is g3 and
(ii) that in the expression for r(T,H), only the am-
plitude of g3 is H dependent, whereas the TLL
parameters Ki are not (i is the TLL mode in-
dex, e.g., r+). Two further transport properties
determined for SC LMO—the Hall coefficient
RH(T) (25, 26) and the Lorenz ratio L = k/sT of
the thermal k to the electrical s conductivity
(20, 27)—appear to confirm these conjectures.
First, RH(T) exhibits a marked T dependence
(Fig. 4A), whose dominant contribution is a
single power law that is the same for all fields.
In TLL theory, the value of the power-law ex-
ponent depends on Ki, and the prefactor
depends on g3. The fact that RH(T) fits at dif-
ferent field strengths contain the same ex-
ponent allows us to infer that the only field
dependence is indeed in g3. Because k con-
tains all back-scattering terms present in the
system, the notable coincidence of L(T) and RH
(T), shown in Fig. 4B, informs us that neither
lattice nor neutral bosonic modes contribute
to the resistivity—only g3 contributes (22).
The experimental result for RH(T) not only
confirms that the charge TLL parameters are
H independent, but it also identifies the mech-
anism steering the system closer toward the
separatrix. Specifically, we find that the power-
law exponent in RH(T) below 150 K closely
matches the characteristic value for random
umklapp processes. This drives the emergent
symmetry in LMO; being more relevant than
standard umklapp processes, these random
processes push the system toward the Mott
phase (Fig. 1B). At the same time, however, they
can only be defined within the metallic phase,
when the electrons are able to propagate freely
and explore various regions with differing
strengths of umklapp events. As a result, the
scattering becomes a self-limiting process, whose
asymptote is located right on the separatrix of
the flow.
Discussion and outlook
We make the following minimal statements
on the basis of our experimental findings. The
monotonous, single–power law T dependence
of the MR indicates that there is only one
mechanism of scattering (otherwise it would
have to result from some serendipitous com-
pensation of various terms), and its overall
isotropy confirms that our choice of 1D QFT
is the correct description (coherent motion is
only along one direction, and we measure this
scattering vertex independently of the sample
orientation). This and supporting evidence
Chudzinski et al., Science 382, 792–796 (2023)
17 November 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
from the Hall effect and the Lorenz ratio (that
Kr does not depend on H) imply that LMO is a
good realization of the QFT proposed above,
and the conjecture that the MR reveals the RG
flow of g3[l] seems valid. Because the T de-
pendence of the (inverse square root) MR is
purely linear, the RG flow must stay on the
separatrix down to the SC Tc, the latter pro-
viding sufficient proof of the emergent sym-
metry. What is notable is the fact that LMO
not only hosts these two ground states (Mott
insulating and SC) but that the RG trajectories
extracted from a MR study are capable of dis-
tinguishing between them.
The relation between the emergence of super-
conductivity and symmetry is now more trans-
parent. Mottness is a tendency of charges to
localize—a tendency that suppresses any SC
instability by diminishing the spectral weight
available for condensation. Emergent sym-
metry, by creating a larger manifold of degen-
erate states, provides a linear combination of
states capable of evading localization and thus
enhances the spectral weight for superconduc-
tivity. The experimental realization of emer-
gent symmetry in SC LMO may also have
implications for our understanding of other
unconventional superconductors proximate to
a Mott state, such as the high-Tc cuprates, the
quantum spin liquids, and the two-leg ladders.
Superconductivity at the edge of Mottness is
an unsolved problem, largely because the sim-
plest model to treat strong correlations, the
Hubbard model, is intractable in dimensions
greater than one. In LMO, however, TLL theo-
ry provides a robust theoretical footing from
which to explore the origins of pair condensa-
tion on the border of localization. Although
the degeneracy or near-degeneracy of multi-
ple ground states provides a natural setting for
emergent symmetry to occur (5, 28), the role
played by fluctuations between such states in
promoting pairing is an interesting avenue for
future research. Finally, the role of disorder in
the vicinity of the transition that we have been
able to uncover is a very general aspect that
connects to other areas, including Griffiths
phases (29) in systems with generic slow dy-
namics (30). In this regard, the notion that dis-
order or randomness on the interaction level
can drive a system toward the separatrix (be-
tween metallicity and Mott localization) sheds
light on this complex transition. Notably, the
notion itself is not specific to 1D systems.
14. J. Lu et al., Sci. Adv. 5, eaar8027 (2019).
15. M. Nakamura, Phys. Rev. B 61, 16377–16392 (2000).
16. G. Gweon et al., Phys. Rev. Lett. 85, 3985 (2000).
17. J. Choi et al., Phys. Rev. B 69, 085120 (2004).
18. C. A. M. dos Santos, B. D. White, Y.-K. Yu, J. J. Neumeier,
J. A. Souza, Phys. Rev. Lett. 98, 266405 (2007).
19. X. Xu et al., Phys. Rev. Lett. 102, 206602 (2009).
20. N. Wakeham et al., Nat. Commun. 2, 396–399 (2011).
21. A. Furusaki, N. Nagaosa, Phys. Rev. B 47, 3827–3831
(1993).
22. See the supplementary materials.
23. N. E. Hussey et al., Phys. Rev. Lett. 89, 086601 (2002).
24. T. Giamarchi, Phys. Rev. B 44, 2905–2913 (1991).
25. H. Chen et al., Europhys. Lett. 89, 67010 (2010).
26. J. L. Cohn, B. D. White, C. A. M. dos Santos, J. J. Neumeier,
Phys. Rev. Lett. 108, 056604 (2012).
27. C. L. Kane, M. P. A. Fisher, Phys. Rev. Lett. 76, 3192–3195
(1996).
28. X. Montiel, T. Kloss, C. Pépin, Sci. Rep. 7, 3477 (2017).
29. K. Agarwal, S. Gopalakrishnan, M. Knap, M. Müller, E. Demler,
Phys. Rev. Lett. 114, 160401 (2015).
30. M. Schiulaz, M. Müller, AIP Conf. Proc. 1610, 11–23 (2014).
31. N. Hussey, Emergent symmetry in lithium molybdate (LMO),
dataset, Dryad (2023); https://doi.org/10.5061/dryad.
qfttdz0pj.
32. ElasticScattering, SmallAngleScatteringLMO,
SmallAngleScatteringLMO/SmallAngleScatteringLMO: Small
angle Boltzmann transport scattering in Li0.9Mo6O17, version
v1.0, Zenodo (2023); https://doi.org/10.5281/zenodo.8252644.
AC KNOWLED GME NTS
We acknowledge stimulating discussions with A. Ghosh,
M. Katsnelson, M. Grüning, and M. Rösner and experimental
assistance from A. F. Bangura, J.-F. Mercure, and A. Narduzzo.
We also acknowledge C. Xu for assistance with preparing Fig. 2.
This work was partially carried out at HFML-RU/NWO, a member
of the European Magnetic Field Laboratory (EMFL). Funding: This
study was supported by Netherlands Organisation for Scientific
Research grant 16METL01 (N.E.H. and M.B.), the European
Research Council under the European Union’s Horizon 2020
research and innovation program grant 835279-Catch-22 (N.E.H.,
B.B., C.D., R.D.H.H., and Y.-T.H.), European Union’s Horizon 2020
research and innovation program grant no. 847639 (P.C.), Engineering
and Physical Sciences Research Council (UK) grant EP/V02986X/1
(P.C. and N.E.H.), and National Science Foundation of China grants
12274369 and 11974061 (X.X.). Author contributions:
Conceptualization: N.E.H., P.C., and X.X. Methodology: P.C., M.B., X.X.,
and N.W. Sample synthesis and characterization: R.J., M.G., and P.T.
Investigation: X.X., N.W., M.B., B.B., C.D., Y.-T.H., and S.W. Analysis:
M.B., P.C., and R.D.H.H. Funding acquisition: N.E.H. and P.C. Supervision:
N.E.H. and P.C. Writing – original draft: P.C., M.B., and N.E.H. with
input from all coauthors. Competing interests: The authors declare
that they have no competing interests. Data and materials availability:
All data used to generate the figures are available at Dryad (31),
and the code used to generate fig. S9 is available at Zenodo (32).
License information: Copyright © 2023 the authors, some rights
reserved; exclusive licensee American Association for the Advancement
of Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abp8948
Materials and Methods
Supplementary Text
Figs. S1 to S11
Table S1
References (33–62)
Submitted 7 March 2022; accepted 29 September 2023
10.1126/science.abp8948
H 1 þ aTm
ð
Fig. 4. Hall effect and Lorenz ratio in LMO.
(A) RH(T) of SC LMO (Tc = 2.15 K) at two different
field strengths: 1.5 T (blue circles) and 8.0 T
(red circles) plotted on a log-log scale. Note the
strengthening field dependence of RH below 30 K.
The dashed lines are fits to the expression for a TLL
H ¼ 1:5mm3=C
(22): RH Tð Þ ¼ R0
represents the band value of the Hall coefficient,
m = −1.5, and a = 3600 and 2000 for m0H = 1.5
and 8.0 T, respectively. (B) Comparison of the
low-field RH(T) (above Tmin ≈ 25 K) (circles) and the
zero-field Lorenz ratio L/L0 (squares) measured on
different SC LMO crystals taken from the same growth
batch (where the Lorenz number L0 = 2.44 × 10−8 V2 K−2).
The Lorenz data are taken from Wakeham et al. (20).
Þ, where R0
RE FERENCES AND NOTES
1. P. Coleman, Introduction to Many-Body Physics (Cambridge
Univ. Press, 2015).
2. P. W. Anderson, Phys. Rev. 130, 439–442 (1963).
3. D. González-Cuadra, A. Bermudez, P. R. Grzybowski,
M. Lewenstein, A. Dauphin, Nat. Commun. 10, 2694 (2019).
4. A. J. A. James, R. M. Konik, P. Lecheminant, N. J. Robinson,
A. M. Tsvelik, Rep. Prog. Phys. 81, 046002 (2018).
5. S. C. Zhang, Science 275, 1089–1096 (1997).
6. R. Coldea et al., Science 327, 177–180 (2010).
7. H. Zou et al., Phys. Rev. Lett. 127, 077201 (2021).
8. M. A. Cazalilla, A. M. Rey, Rep. Prog. Phys. 77, 124401
(2014).
9. P. Chudzinski, T. Jarlborg, T. Giamarchi, Phys. Rev. B 86,
075147 (2012).
10. M. Nuss, M. Aichhorn, Phys. Rev. B 89, 045125 (2014).
11. M. Greenblatt, W. H. McCarroll, R. Neifeld, M. Croft,
J. V. Waszczak, Solid State Commun. 51, 671–674 (1984).
12. J.-F. Mercure et al., Phys. Rev. Lett. 108, 187003 (2012).
13. P. Chudziński, Eur. Phys. J. B 90, 148 (2017).
Chudzinski et al., Science 382, 792–796 (2023)
17 November 2023
5 of 5
|
10.1126_science.abq5375
|
RES EARCH
KONDO INSULATORS
Visualizing the atomic-scale origin of metallic
behavior in Kondo insulators
Harris Pirie1,2, Eric Mascot3, Christian E. Matt1, Yu Liu1, Pengcheng Chen1, M. H. Hamidian1,
Shanta Saha4, Xiangfeng Wang4, Johnpierre Paglione4, Graeme Luke5, David Goldhaber-Gordon6,7,
Cyrus F. Hirjibehedin8,9,10†, J. C. Séamus Davis2,11,12,13, Dirk K. Morr3, Jennifer E. Hoffman1*
A Kondo lattice is often electrically insulating at low temperatures. However, several recent experiments
have detected signatures of bulk metallicity within this Kondo insulating phase. In this study, we visualized
the real-space charge landscape within a Kondo lattice with atomic resolution using a scanning tunneling
microscope. We discovered nanometer-scale puddles of metallic conduction electrons centered around
uranium-site substitutions in the heavy-fermion compound uranium ruthenium silicide (URu2Si2) and around
samarium-site defects in the topological Kondo insulator samarium hexaboride (SmB6). These defects
disturbed the Kondo screening cloud, leaving behind a fingerprint of the metallic parent state. Our results
suggest that the three-dimensional quantum oscillations measured in SmB6 arise from Kondo-lattice defects,
although we cannot exclude other explanations. Our imaging technique could enable the development of
atomic-scale charge sensors using heavy-fermion probes.
W hen the electrons in a material inter-
act strongly with one another, they
often produce unexpected behavior.
Above a characteristic temperature
TK, a lattice of local f moments within
a conducting Fermi sea behaves like an ordinary
magnetic metal, with a Curie-Weiss suscepti-
bility. But below TK, the competition between
antiferromagnetic ordering of the local mo-
ments and their screening by conduction elec-
trons leads to a rich phase diagram, exhibiting
quantum criticality (1), unconventional super-
conductivity (2), and heavy fermions (3, 4)—
quasiparticles with f-electron character (Fig.
1A). A Kondo insulator forms if the spectral
gap opened by hybridization between the con-
duction band and the renormalized f band
spans the Fermi level. Mysteriously, some Kondo
insulators seem to “remember” their metallic
parent state long after this gap is fully de-
veloped. For example, the topological Kondo
insulator samarium hexaboride (SmB6) displays
a sizable bulk optical conductivity at terahertz
frequencies (5) and a finite electronic specific
1Department of Physics, Harvard University, Cambridge, MA
02138, USA. 2Clarendon Laboratory, University of Oxford, Oxford
OX1 3PU, UK. 3Department of Physics, University of Illinois at
Chicago, Chicago, IL 60607, USA. 4Maryland Quantum Materials
Center, Department of Physics, University of Maryland, College
Park, MD 20742, USA. 5Department of Physics and Astronomy,
McMaster University, Hamilton, ON L8S 4M1, Canada.
6Department of Physics, Stanford University, Stanford, CA
94305, USA. 7Stanford Institute for Materials and Energy
Sciences, SLAC National Accelerator Laboratory, Menlo Park, CA
94025, USA. 8London Centre for Nanotechnology, University
College London (UCL), London WC1H 0AH, UK. 9Department of
Physics and Astronomy, UCL, London WC1E 6BT, UK
10Department of Chemistry, UCL, London WC1H 0AJ, UK.
11Department of Physics, University College Cork, Cork T12 R5C,
Ireland. 12Laboratory of Atomic and Solid State Physics,
Department of Physics, Cornell University, Ithaca, NY 14850,
USA. 13Max Planck Institute for Chemical Physics of Solids,
D-01187 Dresden, Germany.
*Corresponding author. Email: jhoffman@physics.harvard.edu
†Present address: Lincoln Laboratory, Massachusetts Institute of
Technology, Lexington, MA 02421, USA.
heat at low temperatures (6–9). A complete
three-dimensional (3D) Fermi surface match-
ing its high-temperature metallic state was re-
constructed from quantum oscillation (9, 10)
and Compton scattering (11) measurements
performed in the insulating regime. Notably,
these metallic properties persist even as the
bulk resistivity of SmB6 increases by 10 orders
of magnitude (12). This discrepancy led to sev-
eral theoretical proposals: Some argue that the
metallic behavior is intrinsic, either a conse-
quence of the small hybridization gap in Kondo
insulators (13) or arising from exotic charge-
neutral quasiparticles (14, 15). Others suggest
an extrinsic origin (16–19), which implies the
presence of microscopic metallic pockets.
Charge inhomogeneity is commonplace at
nanometer length scales, especially in mate-
rials with strong electron interactions that
promote competing orders (20). In a Kondo
lattice, defects that substitute or remove the
f-contributing moment, called Kondo holes,
have a widespread impact on the nearby elec-
tronic structure (21–23). First, these defects
locally untangle the hybridized wave function,
leaving puddles of unhybridized conduction
electrons behind (Fig. 1B). In theory, these
charge puddles should have the same itinerant
character as the metallic parent state (22), but
they have not been imaged directly. Addition-
ally, the excess conduction electrons released
from hybridization adjust the strength of their
interactions with the remaining f moments
(22, 24), leading to enhanced local magnetism
(25) (Fig. 1C). For example, Sm1−xLaxB6 sam-
ples with nonmagnetic La dopants are known
to display increased specific heat and mag-
netic susceptibility compared with undoped
samples (26–28). More recently, the existence
of local metallic puddles around Gd dopants in
Sm1−xGdxB6 was inferred from electron spin-
resonance measurements (29). Meanwhile, an
increased concentration of Sm vacancies in
Sm1−xB6 was shown to globally inhibit the
development of the hybridization gap (30),
eventually leading to bulk conduction (12, 31).
All of these findings suggest that Sm-site de-
fects manifest as Kondo holes in SmB6, yet
their key signature—the accompanying charge
oscillations relating to the parent metallic
Fermi surface (22)—remains undetected by any
microscopic probe.
Directly imaging the metallic puddles around
Kondo holes is difficult, because the inherent
screening strongly renormalizes the bare charge
distribution. However, there are a few prom-
ising approaches (32–34). The most common
is to decorate the tip of a Kelvin probe force
microscope with a single atom or molecule
(35, 36). This technique was used to image
the charge variations within an adsorbed mol-
ecule (37). However, it becomes inaccurate for
small tip-sample separations, because of the
influence of short-range forces (38, 39), com-
plicating further improvements to its spatial
resolution (40). Meanwhile, a scanning tun-
neling microscope (STM) routinely achieves
the subnanometer spatial resolution, cryogenic
temperatures, and sub–milli–electron volt (sub-
meV) energy resolution required to access
atomic charge distributions, but existing meth-
ods to extract the electrostatic potential from
the STM vacuum decay length contain substan-
tial artifacts (41). Consequently, simultane-
ously achieving the high charge precision and
high spatial resolution required to measure
the charge environment around a Kondo hole
is not possible using existing techniques.
We have developed a dedicated STM modal-
ity to image the charge environment within a
Kondo lattice with sub-angstrom resolution. At
temperatures below TK, we image charge oscilla-
tions matching the parent Fermi surface centered
around spinless thorium atoms in the Kondo
metal uranium ruthenium silicide (URu2Si2) and
around three separate Sm-site defects in the
Kondo insulator SmB6. The charge puddles we
image in SmB6 exhibit the same metallic wave
vector seen in recent quantum oscillation ex-
periments (9, 10), suggesting that Kondo-lattice
defects are the source of those oscillations.
Measuring local charge density in
a Kondo lattice
To visualize the conduction-electron density
nc(r) in a Kondo lattice [and hence the local
charge −enc(r), where −e is the electron charge],
we first show theoretically that nc(r) determines
the energy position of the Kondo resonance
~ef rð Þ, which forms near the Fermi level as the
magnetic f moments are screened by conduc-
tion electrons. Then, we establish an exper-
imental metric capable of detecting the sub-meV
variations in ~ef rð Þ around a Kondo hole. Our
technique takes advantage of how the many-
body Kondo resonance responds to local doping.
Pirie et al., Science 379, 1214–1218 (2023)
24 March 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
In the Abrikosov fermion representation for
local moments, ~ef is the Lagrange multiplier
that enforces uniform f-electron density, typ-
ically nf = 1 at each site. As additional charge
carriers Dnc enter a uniform Kondo lattice, the
hybridized Fermi surface reshapes to accom-
modate them, leading to a corresponding
change in ~ef in order to maintain nf = 1 (Fig.
1D, black triangles, and fig. S2E). The mag-
nitude and direction of the shift in ~ef depend
on the details of the band structure. But the
relationship between nc and ~ef is linear over
a wide range of band parameters and charge
doping (fig. S2), implying that the charge den-
sity at position r can usually be inferred by
measuring ~ef rð Þ. In fact, the linear dependence
of ~ef rð Þ on nc(r) was recently verified experi-
mentally by micrometer-scale angle-resolved
photoemission spectroscopy (ARPES) measure-
ments in Eu-doped SmB6 (42).
ð
Þ=I r; (cid:2)V
In STM measurements, the Kondo resonance
normally appears as a peak-dip feature in the
tunneling conductance dI/dV (43) (where I is
the sample-to-tip tunneling current at applied
sample bias V ), because of the presence of
multiple tunneling channels (44, 45) (see cal-
culation in Fig. 1D). In simple cases, ~ef can be
estimated by fitting dI/dV to a Fano-like model
(46, 47). However, the exact value of ~ef de-
pends on the model used, so this approach is
not immediately suitable for detecting the
small, sub-meV energy shifts in ~ef rð Þ expected
around a Kondo hole. Instead, we track the
ratio of forward-to-backward tunneling cur-
rent, that is, the local rectification R r;Vð
Þ ¼
I r; þV
Þ
ð
j. This ratio is insensitive
j
to STM setup artifacts, and it was previously
used to track charge inhomogeneity from the
spectral weight transfer at high biases in hole-
doped cuprates (48). We focus on low biases,
typically V ≲ 10 mV, where the small shifts in
~ef rð Þ generate large variations in R(r,V) owing
to the energy asymmetry of dI/dV about the
Fermi level at V = 0 (Fig. 1, D and E, and fig.
S2). To demonstrate this effect locally, we self-
consistently calculated dI(r,V)/dV, nc(r), and
R(r,V) around a Kondo hole in a metallic
Kondo lattice, as shown in Fig. 1, F to H. The
calculated dI(r,V)/dV at V = 0 tracks the local
Fermi-level density of states, so it reveals the
hybridized Fermi surface of heavy fermions
F. In contrast, both nc(r)
with a wave vector 2kh
and R(r,V) are dominated by static oscillations
at the unhybridized wave vector 2kc
F, asso-
ciated with the Friedel-like redistribution of
the Kondo screening cloud. The correlation
between nc(r) and R(r,V) establishes R(r,V) as
a qualitative probe of local charge, except at
very short distances from a Kondo hole rj j∼a
Þ,
likely because nf = 1 is not enforced at that site.
ð
Kondo holes in URu2Si2
To test our technique, we first studied the
Kondo metal URu2Si2 with 1% thorium dop-
F, as shown schematically. (D) In a uniform Kondo lattice, the Kondo resonance
Fig. 1. Expected disruption of the screening cloud around Kondo holes. (A) In a uniform Kondo lattice,
magnetic moments at each site (gray arrows) are coherently screened by itinerant conduction electrons
(blue cloud) to form a spinless ground state of heavy fermions (orange line), characterized by the wave
vector kh
F. (B) If one moment is removed to create a Kondo hole, the conduction electrons previously
screening it can redistribute themselves. (C) The redistributed screening cloud causes oscillations of the local
conduction electron density nc(r), interaction strength n(r), and magnetic susceptibility cm(r) at the
conduction-band wave vector kc
creates a peak-dip feature in the calculated dI/dV, caused by the quantum interference between
tunneling into the conduction band and the f-electron states with respective amplitudes tc and tf. The
energy position of the peak (black triangles) shifts linearly according to the local conduction-electron
density nc. a.u., arbitrary units. (E) The calculated rectification R Vð Þ ¼ I þVð
peak because dI/dV is asymmetric around the Fermi level EF (which occurs at V = 0). The R(V) peak
amplitude depends on the dI/dV peak energy. These changes are almost linear over the small range of
local doping expected around a Kondo hole (inset). (F) The calculated oscillations in dI(r,V)/dV at the
Fermi level around a Kondo hole match the hybridized Fermi surface (kh
the calculated nc(r) varies according to the circular wave vector of the unhybridized Fermi surface (kc
F,
blue line in inset), as it mainly reflects the disturbance to the screening cloud. (H) Calculated R(r,V)
is dominated by unhybridized electrons for biases within the hybridization gap. The calculations in (D) to
(H) are based on a Kondo-Heisenberg model with nearest-neighbor hopping strength t, Kondo coupling
J = 2t, antiferromagnetic exchange I = 0.002t, and tunneling amplitudes tf/tc = −0.025. In (D) and (E), the
hybridization strength is fixed at n = 0.1t, and the antiferromagnetic correlation strength is fixed at c =
0.0003t. The Fermi wavelength is lc
F
¼ 8a0 in (F) to (H), where a0 is the lattice spacing.
F, orange line in inset). (G) In contrast,
Þ
j acquires a strong
Þ=I (cid:2)Vð
j
ants, which are known to induce Kondo-hole
behavior (24, 49). Previous STM measure-
ments mapped a metal-like Fermi surface in
URu2Si2 for temperatures above To = 17.5 K,
consisting of a single conduction band with
≈ 0:3 p=a, where a is the lattice
wave vector kc
F
constant (Fig. 2A) (46). The onset of coherent
heavy fermion bands below To (50) is accom-
panied by the appearance of a peak-dip feature
in dI/dV, that is, the Kondo-Fano resonance
(Fig. 2B). Close to a thorium dopant, this fea-
ture shifts upward in energy, toward the Fermi
level. This energy shift—and even the barely
perceptible shifts 2 nm away from the dopant—
are easily detected in the amplitude of R(r,V)
(Fig. 2C). For biases within the hybridization
gap Vj
j < D=e ≈ 5 mV (where D is the gap mag-
nitude), R(r,V) displays widespread spatial os-
cillations emanating from thorium dopants, as
shown in Fig. 2, D to F. Their wave vector of
0.29 ± 0.01 (2p/a) agrees with the hybridiza-
tion oscillations previously measured around
Kondo holes in this compound (24). It matches
the URu2Si2 parent metallic Fermi surface de-
tected above To from our measured quasipar-
ticle interference patterns in dI(r,V)/dV at
V = 0, but it is distinct from the heavy bands
that we measured below To (Fig. 2G). As a final
check, we independently extracted ~ef rð Þ by fit-
ting dI(r,V)/dV curves to a Fano model (fig. S3).
Pirie et al., Science 379, 1214–1218 (2023)
24 March 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
The excellent agreement between ~ef rð Þ and
R(r,V) corroborates the existence of charge
oscillations at 2kc
F in URu2Si2, indicating that
some electrons retain their itinerant character
around Kondo holes, even below To.
Metallic puddles in SmB6
In our Kondo insulating SmB6 samples, any
atomic defect that replaces a Sm atom to alter
the 4f moment could generate metallic pud-
dles like those seen in URu2Si2. We searched
for these puddles in flux-grown samples lightly
doped with Fe, which contain two clear Sm-
site defects: Sm vacancies and Fe substitutions
(Fig. 3B). We focused on the (2×1) Sm termi-
nation, as its charge environment most closely
represents that of the bulk (51). As in URu2Si2,
we noticed that the dI/dV peak attributed to
the Kondo resonance changes its energy posi-
tion near candidate Kondo holes (Fig. 3C),
strongly affecting the R(r,V) peak amplitude
(Fig. 3D). Similar shifts in dI/dV peak posi-
tion were previously linked to the buildup of
charge around boron clusters on the Sm (1×1)
termination (52). For biases within the hybrid-
ization gap Vj
j < D=e ≈ 10 mV, R(r,V) reveals
prominent oscillations around Sm-site defects
(Fig. 3, E and F). These oscillations create a
sharp ellipse in the Fourier transform of R(r,V),
as shown in Fig. 3G. The wave vectors of the
R(q,V) ellipse are larger than those of the
surface state detected by quasiparticle inter-
ference imaging (53), and they do not disperse
F to kh
F. (B) Experimental measurement
Fig. 2. Thorium dopants induce Kondo-hole
behavior in URu2Si2. (A) Schematic band structure
of URu2Si2 showing the onset of heavy fermion
bands (gray solid lines) at temperatures below To =
17.5 K, as itinerant conduction electrons (blue
dashed line) hybridize with a renormalized 5f level
(gray dashed line), reducing the Fermi wave
vector from kc
of an asymmetric Fano line shape in the tunneling
conductance at temperatures below To on the U
termination (gray curve). This feature shifts toward
the Fermi level near a thorium dopant (black
triangles), consistent with an expected change
in local charge density. (C) For a fixed bias, the
R(V) peak amplitude (black triangle) is highly
sensitive to the dI/dV peak position. The spectra in
(B) and (C) are averaged over the 18 well-isolated
thorium dopants marked in (D). (D) The measured
R(r,V) exhibits clear oscillations that manifest as a
ring in (E) the fourfold-symmetrized Fourier transform. H, high; L, low. (F) These oscillations match the high-temperature Fermi wave vector of 2kc
F
above and below To. (G) In contrast, a conventional dI/dV measurement couples to the temperature-dependent Fermi surface, which changes drastically from
18.6 K to 5.9 K. For clarity, the 18.6 K data have been scaled in (F) and offset in (G).
ð
≈ 0:3 2p=a
Þ, both
Fig. 3. Kondo holes nucleate metallic puddles in
SmB6. (A) Schematic band structure of SmB6
showing the hybridization between conduction
electrons (blue dashed line) and localized
4f moments (gray dashed line), which leads to
an inverted band structure (gray solid line) hosting
emergent heavy Dirac surface states with a
reduced Fermi wave vector (orange). (B) STM
topography of the (2×1)-reconstructed Sm surface of
lightly Fe-doped SmB6. Both the Fe dopant and
Sm vacancy in this image are expected to act as
Kondo holes because they each displace a 4f
moment. (C and D) Near the Fe dopant, the
measured dI/dV peak changes energy position (black
triangles), leading to large variations in the R(r,V)
peak amplitude. The spectra in (C) and (D) have been
offset for clarity. (E) R(r,V) in the same area as
shown in (B) contains clear oscillations around
the two impurities. (F) Linecut of R(r,V) along the
white dashed line in (B). (G) R(r,V) oscillations
appear as a sharp ring in the twofold-symmetrized
Fourier transform (taken from a larger 65 nm by
80 nm area for enhanced q resolution), which
matches the unhybridized 5d Fermi surface inferred
from ARPES experiments (dashed line) (54). The surface reconstruction creates a sharp peak in R(r,V) at QBragg = (0,p/a).
–Q
Q
Pirie et al., Science 379, 1214–1218 (2023)
24 March 2023
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
for biases within the hybridization gap, indi-
cating a different origin (fig. S4). On the other
hand, the size and shape of the ellipse matches
the unhybridized 5d band found by extrapola-
ting ARPES data (54) to the Fermi level (i.e., it
matches the SmB6 metallic parent state), after
accounting for band folding on the (2×1) sur-
face (Fig. 2G and fig. S5). Our observation of
this 5d wave vector within the Kondo insulat-
ing gap is direct evidence of atomic-scale metal-
licity around Kondo holes. This metallicity is
supported by the large residual dI/dV at V =
0 mV that we measured around Kondo holes
(Fig. 3C, green curve), indicating a sizable Fermi-
level density of states even when the metallic
surface states are suppressed (55). We con-
firmed this discovery by checking for R(r,V)
oscillations around a third type of Kondo hole,
Gd dopants, as detailed in fig. S6.
Magnetic fluctuations at Kondo holes in SmB6
Our R(r,V) maps show the real-space structure
of the metallic puddles around Kondo holes in
SmB6. For these puddles to contribute to the
measured de Haas–van Alphen oscillations in
magnetization, they must have a finite mag-
netic susceptibility. Several Sm-site defects
are already suspected to be locally magnet-
ic from their impact on bulk susceptibility
(7, 26, 27, 56) and their influence on the topo-
logically emergent surface states (53, 55). In
general, topological surface states can provide
a test of local magnetism because they are
protected against backscattering from non-
magnetic defects but not from magnetic de-
fects that locally break time-reversal symmetry
(57). This additional magnetic backscattering
was previously imaged around Fe dopants in
two Bi-based topological insulators (58, 59).
We visualized the intensity of magnetic fluc-
tuations at Sm-site defects in SmB6 by iden-
tifying spatial regions where its surface states
backscatter. For biases within the hybridiza-
tion gap, we measured large-area dI(r,V)/dV
maps that contain clear quasiparticle inter-
ference patterns at the backscattering wave
vector q ≡ kf – ki = 2kss (Fig. 4B), consistent
with our previous report (53). We determined
the spatial origin of this signal by Fourier-
filtering dI(r,V)/dV at the wave vector 2kss to
create an image of the local backscattering
strength (Fig. 4C). Most of the peaks in this
image align with the positions of Sm vacancies
or Fe dopants, indicating that these Kondo
holes harbor the necessary magnetic fluctua-
tions to backscatter topological states.
Discussion and outlook
The charge puddles around Kondo holes pre-
sent an alternative yet compelling origin for
many of the strange observations of metallic
behavior in SmB6. First, the detection of de
Haas–van Alphen (magnetic) oscillations with-
out accompanying Shubnikov–de Haas (resistiv-
parable to the R(r) decay length of g = 2.6 nm,
such that a Landau orbit could fit inside a
metallic puddle. Additionally, many of the
metallic properties were detected in floating
zone–grown samples (5, 6, 9–11), which are
known to have higher concentrations of Sm
vacancies than samples grown with an alumi-
num flux (60). Floating-zone samples also con-
tain a higher concentration of dislocations
(31), which may similarly disrupt the Kondo
screening cloud and thus further enhance the
quantum oscillation amplitude beyond that
expected from Sm vacancies alone. In contrast,
the quantum oscillations completely disappear
in flux-grown samples once embedded alumi-
num is removed (8).
Atomic-scale charge inhomogeneity has a
profound impact on many interacting quan-
tum materials, but it has typically not been
possible to measure. In Kondo-lattice systems,
R(r,V) provides a peek at the ground-state
charge landscape, which is strongly perturbed
by Kondo holes. These Kondo holes nucleate
nanometer-scale metallic puddles that could
explain many of the strange phenomena de-
tected by bulk probes. More broadly, the sen-
sitivity to local charge within a Kondo lattice
may enable atomic-scale charge imaging using
STM tips decorated with a Kondo impurity
(61) or fabricated from heavy-fermion mate-
rials (62).
REFERENCES AND NOTES
1. S. Doniach, Physica B+C 91, 231–234 (1977).
2. F. Steglich et al., Phys. Rev. Lett. 43, 1892–1896 (1979).
3. K. Andres, J. E. Graebner, H. R. Ott, Phys. Rev. Lett. 35,
1779–1782 (1975).
4. Z. Fisk, H. R. Ott, T. M. Rice, J. L. Smith, Nature 320, 124–129
(1986).
5. N. J. Laurita et al., Phys. Rev. B 94, 165154 (2016).
6. K. Flachbart et al., Physica B 378–380, 610–611 (2006).
7. W. T. Fuhrman et al., Nat. Commun. 9, 1539 (2018).
8. S. M. Thomas et al., Phys. Rev. Lett. 122, 166401 (2019).
9. M. Hartstein et al., Nat. Phys. 14, 166–172 (2018).
10. B. S. Tan et al., Science 349, 287–290 (2015).
11. T. E. Millichamp et al., arXiv:2111.07727 [cond-mat.str-el]
(2021).
12. Y. S. Eo et al., Proc. Natl. Acad. Sci. U.S.A. 116, 12638–12641
(2019).
13. J. Knolle, N. R. Cooper, Phys. Rev. Lett. 115, 146401 (2015).
14. G. Baskaran, arXiv:1507.03477 [cond-mat.str-el] (2015).
15. O. Erten, P.-Y. Chang, P. Coleman, A. M. Tsvelik, Phys. Rev. Lett.
119, 057603 (2017).
16. H. Shen, L. Fu, Phys. Rev. Lett. 121, 026403 (2018).
17. B. Skinner, Phys. Rev. Mater. 3, 104601 (2019).
18. W. T. Fuhrman, P. Nikolić, Phys. Rev. B 101, 245118 (2020).
19. M. Abele, X. Yuan, P. S. Riseborough, Phys. Rev. B 101, 094101
(2020).
20. E. Dagotto, Science 309, 257–262 (2005).
21. R. Sollie, P. Schlottmann, J. Appl. Phys. 70, 5803–5805
(1991).
22. J. Figgins, D. K. Morr, Phys. Rev. Lett. 107, 066401 (2011).
23. J. Figgins et al., Nat. Commun. 10, 5588 (2019).
24. M. H. Hamidian et al., Proc. Natl. Acad. Sci. U.S.A. 108,
18233–18237 (2011).
25. R. R. Urbano et al., Phys. Rev. Lett. 99, 146402 (2007).
26. E. V. Nefedova et al., J. Exp. Theor. Phys. 88, 565–573
(1999).
27. S. Gabáni et al., Czech. J. Phys. 52 (suppl. 1), A225–A228
(2002).
28. M. Orendáč et al., Phys. Rev. B 96, 115101 (2017).
29. J. C. Souza et al., Phys. Rev. Res. 2, 043181 (2020).
30. M. E. Valentine et al., Phys. Rev. B 94, 075102 (2016).
Fig. 4. Kondo holes backscatter heavy Dirac
fermions. (A) Topography of an SmB6 region that
contains 15 well-isolated Kondo holes (position
indicated by red and green triangles) on several
(2×1)- or (1×2)-reconstructed domains (dotted
lines). (B) For energies within the Kondo insulating
gap, the Fourier-transformed dI/dV along qy
(perpendicular to Sm rows) contains a linearly
dispersing signal (black dashed line) corresponding
to quasiparticle interference from backscattered
heavy Dirac fermions. The Fourier transform from
the (1×2) domains was rotated by 90° before
being averaged with that from the (2×1) domains.
(C) The intensity of backscattering from topological
states, calculated from Fourier-filtering dI/dV at
the y component of the backscattering wave vector
qy ¼ 2kss
hole. This map is computed only for ordered
patches of the sample, as marked in (A), and
excludes step edges. H, high; L, low.
y , is strongly peaked around each Kondo
ity) oscillations (8–10) is expected for electrically
isolated metallic puddles, provided that they
do not meet the percolation threshold [which
could be unreachable (17)]. Second, the large
Fermi surface size and light effective mass ex-
tracted by bulk probes (9–11) is in excellent
agreement with our observation of itinerant
5d electrons. Third, the magnetic length of
the high-frequency (large-kF) quantum oscilla-
tions that onset above 35 T (9, 10) is com-
Pirie et al., Science 379, 1214–1218 (2023)
24 March 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
31. Y. S. Eo et al., Phys. Rev. Mater. 5, 055001 (2021).
32. B. Koslowski, C. Baur, J. Appl. Phys. 77, 28–33 (1995).
33. C. Wagner et al., Phys. Rev. Lett. 115, 026101 (2015).
34. P. Hapala et al., Nat. Commun. 7, 11560 (2016).
35. M. Nonnenmacher, M. P. O’Boyle, H. K. Wickramasinghe,
Appl. Phys. Lett. 58, 2921–2923 (1991).
36. L. Gross, F. Mohn, N. Moll, P. Liljeroth, G. Meyer, Science 325,
1110–1114 (2009).
37. F. Mohn, L. Gross, N. Moll, G. Meyer, Nat. Nanotechnol. 7,
227–231 (2012).
38. U. Zerweck, C. Loppacher, T. Otto, S. Grafström, L. M. Eng,
Phys. Rev. B 71, 125424 (2005).
39. S. Sadewasser et al., Phys. Rev. Lett. 103, 266103
(2009).
40. F. Albrecht et al., Phys. Rev. Lett. 115, 076101 (2015).
41. F. Albrecht, M. Fleischmann, M. Scheer, L. Gross, J. Repp,
Phys. Rev. B 92, 235443 (2015).
42. E. Kotta et al., Research Square [Preprint] (2022).
https://www.researchsquare.com/article/rs-1239795/v1.
43. V. Madhavan, W. Chen, T. Jamneala, M. F. Crommie,
N. S. Wingreen, Science 280, 567–569 (1998).
44. M. Maltseva, M. Dzero, P. Coleman, Phys. Rev. Lett. 103,
206402 (2009).
45. J. Figgins, D. K. Morr, Phys. Rev. Lett. 104, 187202 (2010).
46. A. R. Schmidt et al., Nature 465, 570–576 (2010).
47. I. Giannakis et al., Sci. Adv. 5, eaaw9061 (2019).
48. Y. Kohsaka et al., Science 315, 1380–1385 (2007).
49. A. de la Torre, P. Visani, Y. Dalichaouch, B. Lee, M. Maple,
Physica B 179, 208–214 (1992).
50. A. F. Santander-Syro et al., Nat. Phys. 5, 637–641
(2009).
51. C. E. Matt et al., Phys. Rev. B 101, 085142 (2020).
52. Z. Sun et al., Phys. Rev. B 97, 235107 (2018).
53. H. Pirie et al., Nat. Phys. 16, 52–56 (2020).
54. J. Jiang et al., Nat. Commun. 4, 3010 (2013).
55. L. Jiao et al., Sci. Adv. 4, eaau4886 (2018).
56. K. Akintola et al., Phys. Rev. B 95, 245107 (2017).
57. Q. Liu, C.-X. Liu, C. Xu, X.-L. Qi, S.-C. Zhang, Phys. Rev. Lett.
102, 156603 (2009).
58. Y. Okada et al., Phys. Rev. Lett. 106, 206805 (2011).
59. B. Jäck, Y. Xie, B. Andrei Bernevig, A. Yazdani, Proc. Natl. Acad.
Sci. U.S.A. 117, 16214–16218 (2020).
60. W. A. Phelan et al., Sci. Rep. 6, 20860 (2016).
61. J. Bork et al., Nat. Phys. 7, 901–906 (2011).
62. A. Aishwarya et al., Science 377, 1218–1222 (2022).
63. H. Pirie et al., Dataset for visualizing the atomic-scale origin of
metallic behavior in Kondo insulators, Zenodo (2023);
https://doi.org/10.5281/zenodo.7510910.
ACKN OWLED GMEN TS
We thank A.-P. Li, B. Skinner, C. Wagner, F. Lüpke, S. Ulrich,
Y. S. Eo, and Z. Fisk for helpful conversations. We thank
A. Soumyanarayanan, M. Yee, and Y. He for their help measuring
Gd-doped SmB6. Funding: This project was supported by the
Gordon and Betty Moore Foundation’s EPiQS Initiative through
grants GBMF4536, GBMF9071, and GBMF9457. The experiments
at Harvard were supported by US National Science Foundation
grant DMR-1410480. The data interpretation received support from
AFOSR grant FA9550-21-1-0429. The work of E.M. and D.K.M.
was supported by the US Department of Energy, Office of Science,
Basic Energy Sciences, under award DE-FG02-05ER46225.
C.E.M. is supported by the Swiss National Science Foundation
under fellowship P400P2_183890. Work at the University of
Maryland was supported by AFOSR FA9550-22-1-0023. Research
at McMaster University was supported by the Natural Sciences
and Engineering Research Council. J.C.S.D. acknowledges support
from the Science Foundation of Ireland under award SFI 17/RP/
5445, from the Royal Society under award R64897, and from
the European Research Council under award DLV-788932.
This project received funding from the European Union’s
Horizon 2020 research and innovation program under the
Marie Skłodowska-Curie grant agreement 893097. Author
contributions: H.P., C.E.M., Y.L., P.C., and M.H.H. carried out
the STM experiments. S.S., X.W., J.P., and G.L. synthesized the
samples. E.M. and D.K.M. developed the theoretical model.
D.G.-G., C.F.H., and J.C.S.D. contributed to the understanding of
the results. H.P. and J.E.H. analyzed the data and wrote the
manuscript with contributions from E.M., C.F.H., J.C.S.D., and
D.K.M. Competing interests: The authors have no competing
interests. Data and materials availability: All data and analysis
presented in this paper are deposited in Zenodo (63). License
information: Copyright © 2023 the authors, some rights reserved;
exclusive licensee American Association for the Advancement of
Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abq5375
Materials and Methods
Supplementary Text
Figs. S1 to S7
Table S1
References (64–69)
Submitted 18 April 2022; accepted 28 February 2023
10.1126/science.abq5375
Pirie et al., Science 379, 1214–1218 (2023)
24 March 2023
5 of 5
|
10.1126_science.abo1633
|
LETTERS
Cite as: H.H. Thorp, Science
10.1126/science.adf8367 (2022).
Editorial Expression of Concern
H. Holden Thorp
Editor-in-Chief
On 15 September 2022, Science published the Research Arti-
cle “Structural basis for strychnine activation of human bitter
taste receptor TAS2R46” by Weixiu Xu et al. (1). The editors
have been made aware that the examination of data provided
after publication revealed potential discrepancies with Fig
S10D. This figure was used to support a proposal of pre-
coupling between TAS2R46 and the G protein gustducin. We
are alerting readers to these concerns while the authors’ in-
stitution investigates further.
REFERENCES AND NOTES
1. W. Xu, et al, Science 377, 6612 (2022).
Published online 22 November 2022
10.1126/science.adf8367
First release: 22 November, 2022
(Page numbers not final at time of first release)
1
www.science.org
RES EARCH
STRUCTURAL BIOLOGY
Structural basis for strychnine activation of human
bitter taste receptor TAS2R46
Weixiu Xu1,2, Lijie Wu1, Shenhui Liu1,2, Xiao Liu1,2, Xiaoling Cao1,2, Cui Zhou1,2, Jinyi Zhang1,2,
You Fu1,2, Yu Guo1, Yiran Wu1, Qiwen Tan1, Ling Wang1, Junlin Liu1, Longquan Jiang1,2,
Zhongbo Fan1,2, Yuan Pei1, Jingyi Yu3, Jianjun Cheng1,2, Suwen Zhao1,2, Xiaojiang Hao4,
Zhi-Jie Liu1,2*, Tian Hua1,2*
Taste sensing is a sophisticated chemosensory process, and bitter taste perception is mediated by
type 2 taste receptors (TAS2Rs), or class T G protein–coupled receptors. Understanding the detailed
molecular mechanisms behind taste sensation is hindered by a lack of experimental receptor structures.
Here, we report the cryo–electron microscopy structures of human TAS2R46 complexed with
chimeric mini–G protein gustducin, in both strychnine-bound and apo forms. Several features of TAS2R46
are disclosed, including distinct receptor structures that compare with known GPCRs, a new “toggle switch,”
activation-related motifs, and precoupling with mini–G protein gustducin. Furthermore, the dynamic
extracellular and more-static intracellular parts of TAS2R46 suggest possible diverse ligand-recognition and
activation processes. This study provides a basis for further exploration of other bitter taste receptors and their
therapeutic applications.
T he taste sensory system helps us to avoid
the ingestion of harmful substances (1–3).
Taste perception is initiated by the phys-
ical interaction of tastants with the re-
ceptors located on the surface of taste
receptor cells on the tongue and palate (1). In
humans, tastants evoke five taste sensations:
sweet, bitter, salty, sour, and umami. Among
the five taste modalities, ion channels trans-
duce sour (4) and salty (5) signals, whereas bit-
ter, sweet, and umami tastes are mediated by
G protein–coupled receptors (GPCRs) (6–9). The
type 1 taste receptor (TAS1R) family, classified
as class C GPCRs, includes three members:
TAS1R1, TAS1R2, and TAS1R3, which com-
bine to form heterodimers that sense sweet
(TAS1R2+TAS1R3) and umami (TAS1R1+TAS1R3)
tastes (9–14). However, a distinct group of
type 2 taste receptors (TAS2Rs) is responsible
for bitter taste perception (15–17). Despite
their structural diversity, TAS1Rs and TAS2Rs
share a common signaling G protein pathway,
which activates the heterotrimeric G protein
gustducin (18–20).
TAS2Rs display low sequence identity (<20%)
with other GPCRs and are classified as a
separate class T GPCR subfamily (21). TAS2Rs
recognize thousands of different bitter mole-
cules, which have diverse scaffolds, shapes,
and molecular weights (22). In humans, there
are only ~25 TAS2Rs to cover this broad che-
1iHuman Institute, ShanghaiTech University, Shanghai
201210, China. 2School of Life Science and Technology,
ShanghaiTech University, Shanghai 201210, China.
3School of Information Science and Technology,
ShanghaiTech University, Shanghai 201210, China.
4State Key Laboratory of Phytochemistry and Plant
Resource in West China, Kunming Institute of
Botany, Chinese Academy of Sciences, Kunming
650210, China.
*Corresponding author. Email: liuzhj@shanghaitech.edu.cn
(Z.-J.L.); huatian@shanghaitech.edu.cn (T.H.)
mosensory space. Therefore, some TAS2Rs,
such as TAS2R46, TAS2R14, and TAS2R10
(23), can be activated by multiple tastants
(24). Other receptor subtypes, by contrast, are
more ligand selective. Furthermore, the TAS2Rs
are distributed not only in the oral cavity but
also in extraoral tissues, including the upper
and lower airways, gut, adipose tissue, brain,
heart, and immune cells (25, 26). These ectopic
bitter taste receptors are involved in a variety
of physiological processes and are associated
with different diseases (27, 28). For example,
TAS2R46 and TAS2R38 are expressed on the
motile cilia of human airway epithelial cells and
are implicated as putative targets for asthma
treatment (25, 29).
Owing to their association with many dis-
eases, TAS2Rs represent promising targets
for pharmacological interventions. However,
no experimental structures have yet been de-
termined for any taste receptors, which are
key for understanding signal transduction and
related drug discovery. Here, we report the
cryo–electron microscopy (cryo-EM) structures
of TAS2R46 in complex with chimeric mini–
G protein gustducin, in a potent neurotoxin
strychnine-bound form, and in the apo form.
In combination with previous computational
modeling and mutagenesis studies (22, 30–32),
the structures reveal molecular features of class
T GPCRs, which is the last structureless class of
GPCRs, contributing to a deeper understanding
of the biology behind bitter taste perception
and signaling.
Results
Structures of TAS2R46-miniGas/gust complexes
in different states
Strychnine is a toxic bitter alkaloid that activ-
ates the TAS2R46-mediated G protein gustducin
(Ggust) signaling pathway (33) (Fig. 1A). It is ex-
tracted from the seeds of Strychnos nux-vomica
and is used as a herb in traditional Chinese
medicine to treat ailments such as dyspepsia
and pain (34). We set out to determine the
structure of human TAS2R46 in complex with
strychnine and Ggust using cryo-EM (see mate-
rials and methods). To overcome the low sur-
face expression and tendency to oligomerize of
TAS2R (35), protein endoglucanase H [Protein
Data Bank (PDB) ID 2CIT] was fused into the
N terminus of wild-type (WT) TAS2R46 with
an optimized linker between the fusion pro-
tein and receptor (fig. S1). Strychnine potency
is comparable for the modified construct and
WT receptor, with median effective concentra-
tion (EC50) values of 6.6 and 1.6 mM, respec-
tively (Fig. 1A). To constitute the complex,
we initially attempted to coexpress the mod-
ified TAS2R46 with heterometric G protein
Gagustb1g2 or Gagustb3g13. However, both Ggust
and TAS2R46-Ggust were unstable. We used
the NanoBiT tethering strategy (36) to stabi-
lize the complex and also generated the chime-
ric G protein miniGas/gust, where the a5 helix
of miniGas was replaced with the a5 helix of
Gagust, with the aim being to achieve a stable G
protein that maintains binding to TAS2R46.
This gave a strychnine-TAS2R46-miniGs/gust
complex with improved homogeneity and
stability (fig. S2, A and B), and we determined
the complex structure at a resolution of 3.01 Å
(Fig. 1B; fig. S2, C to G; and table S1).
The resulting high-quality density map al-
lowed accurate model building of the TAS2R46,
strychnine, and miniGs/gust trimer complex (fig.
S3). TAS2R46 adopts a seven-transmembrane
(7TM) bundle topology, like GPCR structures in
other families, yet some of its helixes and
loops have distinct folding features (Figs. 1, C
and D, and 2A).
Structure comparison of TAS2R46
with other GPCRs
We compared the structure of TAS2R46 as a
representative of class T GPCRs with published
human GPCR–G protein complex structures
(fig. S4), which are grouped into six different
classes (table S2): class A (34 structures), class
B1 (14 structures), class B2 (1 structure), class C
(8 structures), and class F (2 structures). We
used the template modeling (TM)–score (37)
and root mean square deviation (RMSD) to
assess the fold and geometrical similarities
between structures. The calculation was car-
ried out to compare 7TM Ca atoms both be-
tween all structures in each class and between
TAS2R46 and all other GPCRs (fig. S5; see
materials and methods). The results show that
TAS2R46 has a distinct three-dimensional struc-
ture. Based on the average TM-score and
RMSD values of TAS2R46 and different classes
of GPCRs, TAS2R46 is most similar to class F
GPCRs with the highest average TM-score of
0.80 or class A GPCRs with smallest average
Xu et al., Science 377, 1298–1303 (2022)
16 September 2022
1 of 6
An Editorial Expression of Concern was posted on 22 November 2022. RES EARCH | R E S E A R C H A R T I C L E
A
Strychnine
B
TAS2R46
WT
Fusion-TAS2R46
100
)
T
W
f
o
%
(
x
a
m
E
50
0
-12
-10
-6
-8
Log[Strychnine(M)]
-4
Nb35
C
TAS2R46
Strychnine
D
45°
Fig. 1. Cryo-EM structure of the strychnine-TAS2R46-miniGs/gust complex. (A) Strychnine is an agonist
of TAS2R46 in the fluorescence imaging plate reader (FLIPR) Ca2+ assay. Data are means ± SEM of
three biological replicates. (B) Cryo-EM density map of the strychnine-TAS2R46-miniGs/gust complex. (C and
D) Cartoon representation of the strychnine-TAS2R46-miniGs/gust complex structure from two viewpoints,
with strychnine shown as magenta sticks.
RMSD value of 3.53 Å. However, TAS2R46’s
most similar receptor in terms of RMSD (3.35 Å)
is the active-state cannabinoid receptor 1 (PDB
ID 6KPG) (38) or, based on TM-score (0.81),
the active-state CXC chemokine receptor 2 (PDB
ID 6LFO) (39).
TAS2R46 exhibits several distinct features
compared with other GPCR–G protein struc-
tures, including a different structural arrange-
ment of TM3-TM4-TM5 (Fig. 2A and figs. S4
and S6, A and B). Here, the extracellular part
of TM5 packs close to TM3, but the extracel-
lular proximal region (after residue N1765.39,
where N is Asn) moves outward and becomes
disordered with weak EM density. We note
that residue N1765.39 was reported to be glyco-
sylated and important for receptor function
(40). Meanwhile, the extracellular part of
TM4 moves away from TM3 and TM5. This
results in a pocket formed by TM3-TM4-TM5
near the membrane bilayer (fig. S6A). Extra-
cellular loop 2 (ECL2) displays consistently
weak density, reflecting its flexible nature.
Notably, TM7 shows a sharp “kink” in the mid-
dle of the helix at P2727.46 (P, proline), where
the intracellular part bends toward the core
of the 7TM bundle and the extracellular part
swings outward, leading to a large ligand
binding cavity that often appears in peptide-
or protein-bound GPCRs (Fig. 2A and fig. S6A).
The EM density is clear for intracellular loop 2
(ICL2) (fig. S3), which forms a short helix and
orients parallel to the membrane surface in
TAS2R46 instead of forming a loop and par-
ticipating in the interaction with the G protein,
as observed in other GPCR–G protein complex
structures (Fig. 2A and figs. S4 and S6B).
Strychnine binding mode in TAS2R46
TAS2R46 is a bitter taste receptor that re-
sponds to a broad spectrum of bitter sub-
stances, but strychnine is the most potent
agonist identified so far (31, 33, 41). The
orthosteric binding pocket resembles a wide-
open funnel that is formed by residues from
TM2, TM3, TM5, and TM7 (Fig. 2, B and C).
We observed a “baseball cap”–shaped EM den-
sity in the orthosteric binding pocket, which is
similar to that in the strychnine–glycine re-
ceptor (GlyR) structure (fig. S3) (42, 43). We
placed strychnine into that piece of EM den-
sity and identified that W883.32 and E2657.39
(W, Trp; E, Glu) are involved in the direct
coordination of strychnine, which is also con-
sistent with previously reported mutagenesis
data (31, 44). In detail, the side chain of
W883.32 from TM3 sticks into the middle of
the pocket at one-third the height of the
orthosteric funnel (Fig. 2B), and its horizon-
tally placed indole ring functions as a landing
platform for the strychnine molecule and es-
tablishes a p-p interaction with strychnine’s
benzene ring (Fig. 2C). Sequence alignment
shows that W3.32 is highly conserved among
TAS2Rs (fig. S7), and a mutagenesis assay
showed that W883.32A (W883.32→Ala) abolished
strychnine activity, confirming its important
role in ligand binding (Fig. 2D). Additionally,
E2657.39 makes a hydrogen bond with the ter-
tiary amine of strychnine (Fig. 2C), which is
largely protonated at physiological pH (pKa =
8.26, where Ka is the acid dissociation con-
stant). Concordantly, the mutation E2657.39A
also diminishes strychnine activity (Fig. 2D).
Structure of TAS2R46-miniGas/gust complex
in the apo state
During the EM data processing of complex
samples with strychnine, two different particle
components were identified. The major one
yielded the strychnine-TAS2R46-miniGs/gust
complex discussed earlier. The other particle
component was solved at a nominal resolution
of 3.08 Å (fig. S2 and table S1). This structure is
similar to strychnine-TAS2R46-miniGs/gust in
terms of overall architecture and miniGs/gust
coupling; however, the EM density for strych-
nine is absent, and this structure is named as
ligand-free TAS2R46-miniGs/gust (figs. S2 and
S6, C and D). To verify that TAS2R46 could
be coupled by miniGs/gust without strychnine,
TAS2R46 and miniGas/gust were constituted
without adding any agonists. Subsequently, the
cryo-EM structure of the apo-T2R46-miniGs/gust
complex was obtained at a nominal resolu-
tion of 3.01 Å (Fig. 3, A and B; fig. S8; and
table S1). Interestingly, this apo TAS2R46 is
indeed coupled with miniGs/gust and is almost
Xu et al., Science 377, 1298–1303 (2022)
16 September 2022
2 of 6
An Editorial Expression of Concern was posted on 22 November 2022.
RES EARCH | R E S E A R C H A R T I C L E
A
Strychnine
B
Strychnine
TM2
TM4
ICL2
TM7
TM3
TM1
H8
TM5
TM6
W883.32
TAS2R46
C
Y853.29
TM3
D
T1805.43
Strychnine
W883.32
N652.60
F252ECL3
TM6
L622.57
TM2
E2657.39
TM7
WT
N65 2.60A
F252ECL3A
E2657.39A
W883.32A
100
)
T
W
f
o
%
(
x
a
m
E
50
0
-12
-10
-8
-6
-4
Log[Strychnine(M)]
Fig. 2. Architecture of TAS2R46 and the strychnine binding pocket. (A) Side view of the 7TM bundle of the
strychnine-TAS2R46 complex. H, helix. (B) Vertical cross section of the strychnine binding pocket in TAS2R46.
The side chain of W883.32 is shown as orange sticks. (C) Binding pocket of strychnine from the extracellular
view. Key residues of TAS2R46 that interact with strychnine are shown as sticks. Hydrogen bonds are shown as
black dashed lines. (D) Alanine substitution of TAS2R46 residues that interact with strychnine reduced its potency.
Data are means ± SEM of three biological replicates. Emax, maximal efficacy.
identical to the ligand-free structure discussed
earlier.
The overall structure of the apo-TAS2R46-
miniGs/gust complex (the separately determined
structure) is, for the most part, similar to the
strychnine-bound TAS2R46-miniGs/gust complex
structure, with a RMSD of 0.89 Å for 7TM Ca
atoms (fig. S9A). When the structures of the
two states are superimposed, the intracellular
parts of 7TMs align well, but the extracellular
portions are more divergent (Fig. 3, C and D).
The biggest structure difference occurs at
ECL2 (Fig. 3C). In the strychnine-TAS2R46-
miniGs/gust structure, most of ECL2 is assumed
to be disordered owing to the missing EM den-
sity. However, in the apo-T2R46-miniGs/gust
structure, the disordered ECL2 makes a sharp
turn at residue N1765.39 and folds into a short
helix that occupies the orthosteric binding
pocket and partially overlaps with the strych-
nine position in the superimposed structures
(Fig. 3C, fig. S9B, and movie S1). Meanwhile,
the extracellular proximal regions of TM1 and
TM5 bend into the 7TM core in the apo state
structure (Fig. 3C). However, ECL1 and ECL3
move outward about 6.7 Å (A74ECL1) and 13.5 Å
(E256ECL3), respectively, leaving a more opened
extracellular region compared with that in
strychnine-TAS2R46-miniGs/gust (Fig. 3C). We
performed intramolecular fluorescent arsen-
ical hairpin bioluminescence energy transfer
(FlAsH-BRET) experiments (45) to monitor
the strychnine-induced conformational changes
in the extracellular region (fig. S9, C to F). The
results showed that the binding of strychnine
indeed caused the outward movement of ECL2
away from the N terminus in a concentration-
dependent manner (fig. S9F).
Activation of TAS2R46
The apo and strychnine-bound structures en-
able us to investigate the activation process of
TAS2R46 (Fig. 4A and movie S1). In addition
to the converged movements of extracellular
helixes and loops, strychnine binding also in-
duces a side chain flip of W883.32, which trans-
forms the side chain from double conformations
in the apo state to a single conformation in the
strychnine-bound state (fig. S9B). The activation-
related conserved toggle switch W6.48 found in
class A GPCRs is absent in TAS2Rs. The 6.48
residue was assigned to C2386.48 (C, Cys) in
GPCRdb (21). However, structural superimpo-
sition with solved active class A GPCRs, such
as the most fold-similar active-state structure
CXCR2 (PDB ID 6LFO), suggests that the cor-
responding residue to position 6.48 is Y2416.51
(hereafter named Y2416.48T; Y, Try) in TAS2R46
(Fig. 4C). Concordantly, Y2416.48T shows a large
conformational change between the apo and
strychnine-bound states, where the side chain
of Y2416.48T rotates about 90° from pointing
outward to pointing into the core of the trans-
membrane bundle (Fig. 4B and movie S1).
Thus Y2416.48T may play the “toggle switch”
role for TAS2R46 activation (Fig. 4, B and C).
Consistently, the Y2416.48TA mutation abolishes
strychnine activity (Fig. 4F). However, in our
structure, the rotation of Y2416.48T does not
induce the large outward movement of TM6
that is observed in class A GPCRs. An inter-
action network around Y2416.48T stabilizes the
active-state conformation in the strychnine-
bound TAS2R46 structure (Fig. 4D). First, TM5,
TM6, and TM7 are tethered by hydrophobic
and p-p stacking interactions mediated by res-
idues F1885.51, Y2416.48T, and Y2717.45 (F, Phe)
(Fig. 4D). In addition, R552.50 (R, Arg) points
into the 7TM core, and the mutation R552.50A
impairs strychnine-induced TAS2R46 activa-
tion (Fig. 4D). Residues forming the “triad core”
(F1885.51, Y2416.48T, and Y2717.45) and R2.50 are
relatively conserved among bitter taste recep-
tors (fig. S9G), suggesting that they may play a
general role in stabilizing the active state of
TAS2Rs. Also, the P2727.46-induced kink shifts
Y2717.45 into the proximity of the “triad core”
and provides the flexibility for TM7 to yield
space for the rotation of Y2416.48T (Fig. 4D),
which may aid ligand binding and receptor
activation. Residue Y6.48T in TAS2R46 is not
highly conserved among TAS2Rs, implying
that diverse activation mechanisms may exist
in other bitter taste receptors or with other
ligands (30).
TAS2Rs lack other conserved activation-
related motifs, such as N7.49P7.50XXY7.53 and
D3.49R3.50Y3.51 motifs (where D is Asp and X is
any residue), in class A GPCRs (15, 46). The
N7.49P7.50XXY7.53 corresponding motif is HP7.50FIL
in TAS2R46, and H2757.49, I2787.52, and L2797.53
form hydrophobic interactions with residues
A993.43, L1023.46, L2376.47, and F2346.44 in the
strychnine-bound TAS2R46 structure (H, His;
I, Ile; L, Leu) (Fig. 4E). This hydrophobic core
mediates the packing of TM3, TM6, and TM7,
which is different from the interaction net-
work observed in class A GPCRs. In the apo
state, these hydrophobic interactions are weak-
ened owing to the side chain shift of residue
F2346.44 (Fig. 4E). In class A GPCR activation,
structural rearrangement of the D3.49R3.50Y3.51
Xu et al., Science 377, 1298–1303 (2022)
16 September 2022
3 of 6
An Editorial Expression of Concern was posted on 22 November 2022.
RES EARCH | R E S E A R C H A R T I C L E
A
B
TAS2R46
TAS2R46
Nb35
C
Extracellular view
D
Intracellular view
TM3
ECL1
TM5
ICL3
TM4
ECL2
TM5
Strychnine
TM6
TM2
180°
TM6
TM3
ICL2
TM1
TM4
TM2
H8
ICL1
TM1
TM7
Strychnine-TAS2R46
Apo-TAS2R46
ECL3
Fig. 3. Cryo-EM structure of the apo-TAS2R46-miniGs/gust complex. (A) Cryo-EM density map of the
apo-TAS2R46-miniGs/gust complex. (B) Cartoon representation of the apo-TAS2R46-miniGs/gust complex
structure. (C and D) Extracellular (C) and intracellular (D) views of the comparison of the strychnine-bound
(green) and apo (blue) TAS2R46 structures. The conformational changes are indicated with red arrows.
motif usually unleashes TM6 from TM3 to ac-
commodate G protein coupling. In strychnine-
bound TAS2R46, the corresponding residues
are F3.49Y3.50L3.51, and this motif is relatively
conserved in TAS2Rs (fig. S9I and movie S2)
(30). Our results are consistent with the in-
volvement of this region in G protein binding—
Y1063.50 forms a hydrogen bond with the car-
bonyl oxygen of residue C351 from the Gagust
a5 helix, and the mutation F1053.49A or Y1063.50A
impairs the activation of TAS2R46 by strychnine
(Fig. 4F). We note that the extent of activation
may be influenced by the chimeric G protein
and stabilization of the receptor. However, the
preassociation of TAS2R46 with gustducin
without agonist binding might be indicative of
flexibility of the receptor that is required for
binding and rapid activation by many ligands
(47). In our BRET assay, addition of strychnine
decreased bioluminescence resonance energy
transfer between labeled TAS2R46 and WT
gustducin heterotrimers (fig. S10D), further
confirming the precoupling state of TAS2R46
with gustducin.
Additionally, W973.41 in TM3 is the sole
strictly conserved residue in all TAS2Rs (fig.
S9G). Its side chain points into the cavity be-
tween TM4 and TM5 and is sandwiched by
P1364.51 and P1875.50 in TAS2R46 (fig. S9H).
The mutation of W973.41A abolishes the activ-
ity of strychnine on TAS2R46 (Fig. 4F), sug-
gesting that W3.41 stabilizes the interface of
TM4-TM3-TM5. Interestingly, the mutation
to tryptophan at position 3.41 has also been
frequently used to stabilize the conformation
of class A GPCRs (48).
TAS2R46 and miniGas/gust interaction interface
in the strychnine-bound state
Gustducin is closely related both to transducins
(20) and to the Gi/o class, based on the phylo-
genetic tree of G protein a subunits. The struc-
ture of the strychnine-TAS2R46-miniGs/gust
complex reveals several distinct features of the
a5 helix–gustducin engagement to TAS2R46’s
cytoplasmic cavity that are mainly contributed
by TM3, TM5, and TM6 (Fig. 5 and fig. S10,
A and B). More specifically, the a5 helix of
miniGas/gust forms hydrophobic and polar in-
teractions, as well as hydrogen bonds, with the
cytoplasmic cavity (Fig. 5C). Whereas the in-
tracellular loops, especially ICL2, are often in-
volved in receptor–G protein interactions in
other GPCR complex structures, the intra-
cellular loops here have almost no direct in-
teraction with the miniGs/gust trimer in the
strychnine-TAS2R46-miniGs/gust complex
(fig. S10F). Rather, ICL2 takes an unusual con-
formation where part of it forms a short helix
that lies parallel to the cytoplasmic cell mem-
brane (Fig. 5A). By contrast, TM5 and TM6 are
more extended into the cytosol, and residues
H2055.68 and H2246.34 establish hydrogen bonds
with D341 of the a5 helix of miniGas/gust (Fig.
5C). The mutation H2055.68A or H2246.34A im-
pairs the activity of strychnine on TAS2R46 in
the functional assay (fig. S10C). The end of the
a5 helix of miniGas/gust inserts into the recep-
tor core, and residues L353 and C351 form
hydrophobic interactions with F1053.49 and
Y1063.50 from the cytoplasmic proximal region
of TM3 (Fig. 5C). Additionally, K1093.53 (K,
Lys) forms a hydrogen bond with D350 of the
a5 helix in miniGas/gust, which stabilizes the
receptor–G protein interaction (Fig. 5C). K3.53
is highly conserved in TAS2Rs, suggesting a
general role in Gagust coupling.
When comparing TAS2R46 with class A
GPCRs in complex with different G protein
subtypes (i.e., CXCR2-Gi1, A1R-Gi2, b2AR-Gs,
and M1-G11) or GPCR–G protein complexes in
other classes (i.e., GPR97-Go, FZD7-miniGs,
and GLP1R-Gs), major differences occur in
the relative positions and orientations of the
a5 and aN helices of the Ga subunits, as well
as the corresponding position shifts of TM6
and intracellular loops (Fig. 5B and fig. S10,
F to I). TAS2R46’s TM6 has a smaller out-
swing angle, a shallower a5 helix insertion of
miniGas/gust, and a more-tilted a5 helix, which
is, interestingly, more similar to that of the
CXCR2-Gi complex structure (Fig. 5B and fig.
S10, E and F). Of note, the potential impact of
the chimeric G protein warrants further in-
vestigation with WT gustducin.
Bitter substance recognition by TAS2R46
The ligand recognition pattern of TAS2R46 was
further explored through molecular docking
studies. A list of 68 bitter molecules was re-
trieved from BitterDB (22, 30, 33) for molec-
ular docking, and the representative binding
poses were analyzed. In agreement with the
strychnine-bound TAS2R46 structure and prev-
ious molecular simulation studies (32), W883.32
and E2657.39 play the most critical roles in the
recognition and coordination of these mole-
cules. Very similar to strychnine, binding poses
of the therapeutic drugs quinine, berberine,
and chlorpheniramine are stabilized by p-p
stacking interactions between an aromatic
moiety in their structures and W883.32, as well
Xu et al., Science 377, 1298–1303 (2022)
16 September 2022
4 of 6
An Editorial Expression of Concern was posted on 22 November 2022. RES EARCH | R E S E A R C H A R T I C L E
A
B
ECL3
Strychnine
C
ECL2
TM2
TM1
TM7
W
3.32
TM3
Y
6.48T
Y
7.45
P
7.46
F
5.51
R
2.50
F
5.58
F
6.44
Y
3.50
H
7.49
L
7.53
TM5
TM6
Y2416.48T
TM6
TM5
Strychnine-TAS2R46
Apo-TAS2R46
TM6
Y2416.48T
D
Strychnine
E
TM2
TM6
TM6
Y2416.48T
F1885.51
Y2717.45
Triad core
TM3
TM7
TM5
W883.32
Kink
P2727.46
R552.50
L2376.47
TM5
F1955.58
TM2
P2727.46
R552.50
Y2717.45
TM7
90°
F1885.51
Y2717.45
R552.50
P2727.46
TM2
TM7
TM3
HS/P7.50FIL
A993.43
H2757.49 P2767.50
F2777.51
L2797.53
IL8
Strychnine
AM841
W6.48
Y2416.48T
TM6
IL8-CXCR2
AM841-CB1
WT
Y2416.48TA
R552.50A
F1885.51A
F1053.49A
Y2717.45A
Y1063.50A
Y2717.45W
W973.41A
F
)
T
W
f
o
%
(
x
a
m
E
100
50
0
F2346.44
TM7
-12
-10
-8
-6
Log[Strychnine(M)]
-4
Fig. 4. The activation features of TAS2R46. (A) Schematic summarizing the
key translational and rotational movements that contribute to the activation of
TAS2R46. (B) The conformational changes of the extracellular region and key
residues between apo-TAS2R46 and strychnine-bound TAS2R46. (C) Super-
imposition of the strychnine-bound TAS2R46, apo-TAS2R46, monomer IL8-
CXCR2-Gi1, and AM841-CB1-Gi1 structures aligned at the W6.48 position in class
A GPCRs. (D) Detailed interactions of residues that stabilize the conformation
of the TM2-TM3-TM5-TM6-TM7 core in the strychnine-bound TAS2R46 structure.
(E) The detailed interactions around the HS/P7.50FIL motif in the strychnine-
bound and apo-TAS2R46 structures. S, Ser. (F) The effects of mutations in
TM helix tethering residues on strychnine-induced Ca2+ mobilization. Data are
means ± SEM of three biological replicates. The side chains of key residues
are shown as orange and blue sticks in the strychnine-bound and apo-TAS2R46
structures, respectively.
as a salt bridge interaction between a positively
charged nitrogen and E2657.39 (fig. S11, A and
B). The combination of an aromatic moiety
and a positively charged nitrogen may allow
many clinical drugs (49) to be recognized by
TAS2R46, which explains the saying “good
medicine tastes bitter” from ancient Chinese
wisdom (27). For the bitter substances that do
not contain an aromatic ring or a positively
charged nitrogen, most feature a lactone sub-
structure (22, 33), in which the carbonyl group
forms a hydrogen bond with the NH group of
W883.32 and a distant hydroxyl group forms
another hydrogen bond with the carboxyl
group of E2657.39 (fig. S11, A, C, and D). To-
gether with previous modeling and bioinfor-
matic works (23, 30, 33, 50), these findings
provide clues for the understanding of broad-
spectrum ligand recognition of TAS2R46,
which may help to identify or design other
chemical entities for bitter taste receptors.
Conclusion
Our structures disclose the roles of W883.32
and E2657.39 in strychnine binding and iden-
tify features involved in the transition from
the apo state to the strychnine-bound state.
We suggest that Y2416.48T acts as the “toggle
switch” for receptor activation. The dynamic
ECL2 may play a role in diverse ligand binding
(32) and receptor activation. Finally, our struc-
ture and functional assay highlight that apo
TAS2R46 is precoupled by miniGs/gust; such
“preassociated” complexes have also been
described in other GPCRs (47). Precoupling
might facilitate a rapid response by the re-
ceptor when potentially toxic bitter molecules
are detected.
REFERENCES AND NOTES
1. D. A. Yarmolinsky, C. S. Zuker, N. J. Ryba, Cell 139, 234–244
(2009).
2. K. Scott, Neuron 48, 455–464 (2005).
3. F. Bermúdez-Rattoni, Nat. Rev. Neurosci. 5, 209–217 (2004).
4. Y. H. Tu et al., Science 359, 1047–1050 (2018).
J. Chandrashekar et al., Nature 464, 297–301 (2010).
5.
6. G. T. Wong, K. S. Gannon, R. F. Margolskee, Nature 381,
796–800 (1996).
7. N. Chaudhari, A. M. Landin, S. D. Roper, Nat. Neurosci. 3,
8.
113–119 (2000).
J. Chandrashekar, M. A. Hoon, N. J. Ryba, C. S. Zuker, Nature
444, 288–294 (2006).
9. G. Nelson et al., Nature 416, 199–202 (2002).
10. G. Nelson et al., Cell 106, 381–390 (2001).
11. G. Q. Zhao et al., Cell 115, 255–266 (2003).
12. M. Max et al., Nat. Genet. 28, 58–63 (2001).
13. J. P. Montmayeur, S. D. Liberles, H. Matsunami, L. B. Buck,
Nat. Neurosci. 4, 492–498 (2001).
Xu et al., Science 377, 1298–1303 (2022)
16 September 2022
5 of 6
An Editorial Expression of Concern was posted on 22 November 2022.
RES EARCH | R E S E A R C H A R T I C L E
A
BB
TM4
TM3
TM6
ICL2
TM5
Strychnine-TAS2R46-
Apo-TAS2R46-
C
TAS2R46
TM4
TM2
TM6
TM3
ICL2
TM6
TAS2R46-miniGs/gust
CXCR2-Gi1
A1R-Gi2
2AR-Gs
M1-G11
TM5
Y1063.50
F1053.49
TM4
F354
L2025.65
TM6
L353
C351
D350
K1093.53
180°
V2236.33
L348
TM3
ICL2
TM5
H2246.34
H2055.68
I344
ICL3
ICL3
D341
M2095.72
D337
H212ICL3
Fig. 5. General features of miniGs/gust coupling to TAS2R46. (A) The coupling mode of miniGs/gust
with TAS2R46 in the strychnine-bound and apo-TAS2R46-miniGs/gust complex structures. (B) Structural
comparison of the strychnine-TAS2R46-miniGs/gust complex with CXCR2-Gi1 (yellow), A1R-Gi2 (cyan), b2AR-
Gs (pink), and M1-G11 (bluish purple) complexes when aligned with the receptor. The differences in TM6 of the
receptor and the a5 and aN helices of the G protein are indicated by red arrows. (C) Detailed interactions of
TAS2R46 with the a5 helix of miniGs/gust in the strychnine-bound TAS2R46-miniGs/gust complex structure.
14. X. Li et al., Proc. Natl. Acad. Sci. U.S.A. 99, 4692–4696 (2002).
15. E. Adler et al., Cell 100, 693–702 (2000).
16. J. Chandrashekar et al., Cell 100, 703–711 (2000).
17. H. Matsunami, J. P. Montmayeur, L. B. Buck, Nature 404,
601–604 (2000).
18. Y. Zhang et al., Cell 112, 293–301 (2003).
19. L. Huang et al., Nat. Neurosci. 2, 1055–1062 (1999).
20. S. K. McLaughlin, P. J. McKinnon, R. F. Margolskee, Nature 357,
563–569 (1992).
21. A. J. Kooistra et al., Nucleic Acids Res. 49, D335–D343 (2021).
22. A. Dagan-Wiener et al., Nucleic Acids Res. 47, D1179–D1185
(2019).
29. A. S. Shah, Y. Ben-Shahar, T. O. Moninger, J. N. Kline,
M. J. Welsh, Science 325, 1131–1134 (2009).
30. J. Topin et al., Cell. Mol. Life Sci. 78, 7605–7615 (2021).
31. A. Brockhoff, M. Behrens, M. Y. Niv, W. Meyerhof, Proc. Natl.
Acad. Sci. U.S.A. 107, 11110–11115 (2010).
32. M. Sandal et al., J. Chem. Theory Comput. 11, 4439–4449
(2015).
33. A. Brockhoff, M. Behrens, A. Massarotti, G. Appendino,
W. Meyerhof, J. Agric. Food Chem. 55, 6236–6243
(2007).
34. R. Guo et al., Am. J. Chin. Med. 46, 1–23 (2018).
35. C. Kuhn, B. Bufe, C. Batram, W. Meyerhof, Chem. Senses 35,
23. M. Behrens, W. Meyerhof, J. Agric. Food Chem. 66, 2204–2213
395–406 (2010).
(2018).
24. W. Meyerhof et al., Chem. Senses 35, 157–170 (2010).
25. S. J. Lee, I. Depoortere, H. Hatt, Nat. Rev. Drug Discov. 18,
116–138 (2019).
26. S. S. Calvo, J. M. Egan, Nat. Rev. Endocrinol. 11, 213–227
(2015).
27. P. Lu, C. H. Zhang, L. M. Lifshitz, R. ZhuGe, J. Gen. Physiol.
149, 181–197 (2017).
28. F. A. Shaik et al., Int. J. Biochem. Cell Biol. 77, 197–204
(2016).
36. J. Duan et al., Nat. Commun. 11, 4121 (2020).
37. Y. Zhang, J. Skolnick, Proteins 57, 702–710 (2004).
38. T. Hua et al., Cell 180, 655–665.e18 (2020).
39. K. Liu et al., Nature 585, 135–140 (2020).
40. C. Reichling, W. Meyerhof, M. Behrens, J. Neurochem. 106,
1138–1148 (2008).
43. J. Du, W. Lü, S. Wu, Y. Cheng, E. Gouaux, Nature 526, 224–229
(2015).
44. A. N. Pronin, H. Tang, J. Connor, W. Keung, Chem. Senses 29,
583–593 (2004).
45. Y. Q. Ping et al., Nature 589, 620–626 (2021).
46. F. Fierro et al., Front. Mol. Biosci. 4, 63 (2017).
47. W. Jang et al., Proc. Natl. Acad. Sci. U.S.A. 117, 30755–30762
(2020).
48. C. B. Roth, M. A. Hanson, R. C. Stevens, J. Mol. Biol. 376,
1305–1319 (2008).
49. E. Vitaku, D. T. Smith, J. T. Njardarson, J. Med. Chem. 57,
10257–10274 (2014).
50. A. Dagan-Wiener et al., Sci. Rep. 7, 12074 (2017).
AC KNOWLED GME NTS
The cryo-EM data were collected at the Bio-Electron Microscopy
Facility, ShanghaiTech University, with the assistance of
Q.-Q. Sun, Z-H. Zhang, and other staff members. We also thank the
staff members of the Assay, Cell Expression, Cloning, and
Purification Core Facilities of the iHuman Institute for their
support. Funding: This work was supported by the CAS Strategic
Priority Research Program XDB37030104 (Z.-J.L.), the National
Science Fund for Distinguished Young Scholars 32022038
(T.H.), the National Natural Science Foundation of China grants
31930060 (Z.-J.L.) and 31870744 (T.H.), the National Key
Research and Development Program of China grant
2018YFA0507000 (T.H.), and the Shanghai Rising-Star Program
20QA1406500 (T.H.). Author contributions: W.X. designed
the expression constructs, purified the protein complexes,
prepared the final samples for cryo-EM data collection, and
participated in figure and manuscript preparation. L.Wu performed
the EM data processing and structure determination. S.L. and
Y.G. performed structure comparison analyses and molecular
docking and prepared related figures. X.L., Q.T., and L.Wa.
performed the functional assays. X.C. performed the BRET assays.
C.Z. assisted with construct optimization. J.L., L.J., and Z.F.
expressed proteins. J.Z. assisted with the cryo-EM sample
preparation. Y.F. assisted with cryo-EM data processing and
movie preparation. Y.P. and J.Y. assisted with cryo-EM data
processing. J.C. assisted with the docking poses analysis. Y.W. and
S.Z. assisted with the strychnine binding pose analysis. X.H.
performed strychnine analog synthesis. Z.-J.L. and T.H. managed
and supervised the overall project, analyzed the structures, and
wrote the manuscript. Competing interests: The authors
declare no competing interests. Data and materials availability:
Coordinates and structures factors have been deposited in the
Protein Data Bank for strychnine-TAS2R46-miniGs/gust (PDB
ID 7XP6, EMDB-33366), ligand-free TAS2R46-miniGs/gust (PDB
ID 7XP5, EMDB-33365), and apo-TAS2R46-miniGs/gust (PDB ID
7XP4, EMDB-33364). License information: Copyright © 2022 the
authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.science.org/about/science-
licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abo1633
Materials and Methods
Figs. S1 to S11
Tables S1 to S3
References (51–79)
MDAR Reproducibility Checklist
Movies S1 and S2
41. S. Born, A. Levit, M. Y. Niv, W. Meyerhof, M. Behrens,
View/request a protocol for this paper from Bio-protocol.
J. Neurosci. 33, 201–213 (2013).
42. X. Huang, H. Chen, K. Michelsen, S. Schneider, P. L. Shaffer,
Nature 526, 277–280 (2015).
Submitted 18 January 2022; accepted 16 August 2022
10.1126/science.abo1633
Xu et al., Science 377, 1298–1303 (2022)
16 September 2022
6 of 6
An Editorial Expression of Concern was posted on 22 November 2022.
|
10.1126_science.abq3773
|
RES EARCH
R E S E A R C H A R T I C L E
◥
CORONAVIRUS
Broadly neutralizing antibodies target the
coronavirus fusion peptide
Cherrelle Dacon1†, Courtney Tucker1,2†, Linghang Peng3†, Chang-Chun D. Lee4†, Ting-Hui Lin4,
Meng Yuan4, Yu Cong5, Lingshu Wang6, Lauren Purser1, Jazmean K. Williams7, Chul-Woo Pyo8,
Ivan Kosik9, Zhe Hu9, Ming Zhao10, Divya Mohan1, Andrew J. R. Cooper1, Mary Peterson11,
Jeff Skinner11, Saurabh Dixit5, Erin Kollins5, Louis Huzella5, Donna Perry5, Russell Byrum5,
Sanae Lembirik5, David Drawbaugh5, Brett Eaton5, Yi Zhang6, Eun Sung Yang6, Man Chen6,
Kwanyee Leung6, Rona S. Weinberg12, Amarendra Pegu6, Daniel E. Geraghty8, Edgar Davidson7,
Iyadh Douagi13, Susan Moir14, Jonathan W. Yewdell9, Connie Schmaljohn5, Peter D. Crompton11,
Michael R. Holbrook5, David Nemazee3, John R. Mascola6, Ian A. Wilson4,15, Joshua Tan1*
The potential for future coronavirus outbreaks highlights the need to broadly target this group of pathogens.
We used an epitope-agnostic approach to identify six monoclonal antibodies that bind to spike proteins
from all seven human-infecting coronaviruses. All six antibodies target the conserved fusion peptide region
adjacent to the S2′ cleavage site. COV44-62 and COV44-79 broadly neutralize alpha- and betacoronaviruses,
including severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) Omicron subvariants BA.2 and
BA.4/5, albeit with lower potency than receptor binding domain–specific antibodies. In crystal structures
of COV44-62 and COV44-79 antigen-binding fragments with the SARS-CoV-2 fusion peptide, the fusion
peptide epitope adopts a helical structure and includes the arginine residue at the S2′ cleavage site.
COV44-79 limited disease caused by SARS-CoV-2 in a Syrian hamster model. These findings highlight the
fusion peptide as a candidate epitope for next-generation coronavirus vaccine development.
C oronaviruses consist of four genera that
infect birds and mammals (1). Seven
coronaviruses are known to cause human
disease: the alphacoronaviruses HCoV-
229E (human coronavirus 229E) and
HCoV-NL63 (human coronavirus NL63), as
well as the betacoronaviruses HCoV-OC43
(human coronavirus OC43), HCoV-HKU1 (hu-
man coronavirus HKU1), SARS-CoV (severe
acute respiratory syndrome coronavirus),
MERS-CoV (Middle East respiratory syn-
drome coronavirus), and SARS-CoV-2 (severe
acute respiratory syndrome coronavirus 2).
Whereas the first four coronaviruses gener-
ally cause mild disease, the latter three have
brought about serious outbreaks in recent
years. In particular, the ongoing COVID-19
pandemic caused by SARS-CoV-2 has resulted
in more than 6 million deaths since the first
cases were identified in 2019 (2). The cur-
rently dominant SARS-CoV-2 Omicron BA.2,
BA.2.12.1, BA.4, and BA.5 subvariants are at
least partially resistant to most available vac-
cines and antibody therapeutics (3–6). Further-
more, two coronaviruses previously linked only
to animal infection were recently detected in
individuals with flu-like symptoms (7, 8). These
developments highlight the importance of tar-
geting conserved and functionally essential
sites on coronaviruses.
Coronavirus infection is a multistep process
that involves enzymatic cleavage and rearrange-
ment of the surface spike protein (9). The SARS-
CoV-2 spike contains two cleavage sites: a furin
cleavage site at the boundary of the S1 and S2
subunits, and an S2′ site that is conserved in
coronaviruses. The spike protein is thought to
be cleaved at the S1-S2 site during virus as-
sembly, leaving the S1 and S2 subunits non-
covalently linked. During entry, the SARS-CoV-2
spike protein uses the receptor binding domain
(RBD) on the S1 subunit to engage angiotensin-
converting enzyme 2 (ACE2) on target cells.
After receptor binding, the S1 subunit is shed
and the S2′ site is cleaved by the membrane
enzyme transmembrane serine protease 2
(TMPRSS2) or endosomal cathepsins (1), lead-
ing to insertion of the fusion peptide into the
cell membrane and viral fusion.
Much of the protection provided by COVID-19
vaccines arises from neutralizing antibodies that
target the RBD (10). Likewise, all currently
available therapeutic monoclonal antibodies
(mAbs) target this domain (3). However, spike
elements that participate in the subsequent
stages of infection involve the more complex
S2 fusion machinery with many moving parts,
and these elements are more conserved than
the RBD, which so far has been capable of
retaining or even increasing binding to ACE2
despite a variety of mutations (11). Therefore,
these sites are worth exploring as targets for
novel COVID-19 vaccines and therapeutics
that retain efficacy against new variants and
protect against a wider range of coronaviruses.
Progress in this direction has started with re-
cent studies identifying several mAbs that tar-
get the conserved stem helix (12–16) and using
unbiased approaches to screen for mAbs of
interest (17–20). In this study, we carried out
a large-scale survey of the binding landscape
of broadly reactive mAbs against coronaviruses.
Identification of broadly reactive mAbs from
COVID-19–convalescent donors
To identify individuals who are likely to har-
bor B cells that produce broadly reactive mAbs,
we used a multiplex bead-based assay to ex-
amine plasma samples of 142 donors from a
previously described cohort of COVID-19–
convalescent individuals (20). We assessed
plasma immunoglobulin G (IgG) reactivity
toward spike glycoproteins of the seven human
coronaviruses: SARS-CoV-2 (Wuhan-Hu-1 strain),
SARS-CoV, MERS-CoV, HCoV-HKU1, HCoV-
OC43, HCoV-NL63, and HCoV-229E. Nineteen
donors were selected for mAb isolation and char-
acterization on the basis of plasma IgG reactivity
to the spike proteins of SARS-CoV-2 and at least
two other betacoronaviruses (fig. S1A).
We next investigated human IgG+ memory
B cells (MBCs) from the selected donors by
using a two-stage screen to prioritize isolating
mAbs with the greatest possible breadth of
reactivity. First, we screened supernatants
from 673,671 stimulated IgG+ B cells for binding
1Antibody Biology Unit, Laboratory of Immunogenetics, National Institute of Allergy and Infectious Diseases, National Institutes of Health, Rockville, MD 20852, USA. 2Department of Biology, The
Catholic University of America, Washington, DC 20064, USA. 3Department of Immunology and Microbiology, The Scripps Research Institute, La Jolla, CA 92037, USA. 4Department of Integrative
Structural and Computational Biology, The Scripps Research Institute, La Jolla, CA 92037, USA. 5Integrated Research Facility, Division of Clinical Research, National Institute of Allergy and
Infectious Diseases, National Institutes of Health, Frederick, MD 21702, USA. 6Vaccine Research Center, National Institute of Allergy and Infectious Diseases, National Institutes of Health,
Bethesda, MD 20892, USA. 7Integral Molecular, Philadelphia, PA 19104, USA. 8Clinical Research Division, Fred Hutchinson Cancer Research Center, Seattle, WA 98109, USA. 9Cellular Biology
Section, Laboratory of Viral Diseases, National Institute of Allergy and Infectious Diseases, National Institutes of Health, Bethesda, MD 20892, USA. 10Protein Chemistry Section, Research
Technologies Branch, National Institute of Allergy and Infectious Diseases, National Institutes of Health, Rockville, MD 20852, USA. 11Malaria Infection Biology and Immunity Section, Laboratory
of Immunogenetics, National Institute of Allergy and Infectious Diseases, National Institutes of Health, Rockville, MD 20852, USA. 12New York Blood Center, Lindsley F. Kimball Research Institute,
New York, NY 10065, USA. 13Flow Cytometry Section, Research Technologies Branch, National Institute of Allergy and Infectious Diseases, National Institutes of Health, Bethesda, MD 20892,
USA. 14B Cell Immunology Section, Laboratory of Immunoregulation, National Institute of Allergy and Infectious Diseases, National Institutes of Health, Bethesda, MD 20892, USA. 15The Skaggs
Institute for Chemical Biology, The Scripps Research Institute, La Jolla, CA 92037, USA.
*Corresponding author. Email: tanj4@nih.gov
†These authors contributed equally to this work.
Dacon et al., Science 377, 728–735 (2022)
12 August 2022
1 of 8
RES EARCH | R E S E A R C H A R T I C L E
A
B
C
SARS-CoV-2
-CoVs ( 2)
-CoV
n = 2136
n = 211
(9.9%)
n = 24
(1.1%)
All Hu-CoVs
n =14
0.65%
mAb VH Gene
COV91-27
3-30
COV44-62
COV44-79
COV77-04
COV77-39
COV78-36
S2P6
1-2
3-30
4-59
4-59
5-10-1
1-46
2
-
V
o
C
1
-
V
o
C
S
R
E
M
1
U
K
H
3
4
C
O
3
6
L
N
E
9
2
2
V
o
C
C
V
o
C
D
P
A
H
1
H
AUC
0.0
1.5
3.0
4.5
6.0
7.5
D
E
NT50 (µg/mL)
AssayNIH
AssayScripps
CoV-2 CoV-1 MERS NL63 229E OC43 CoV-2 CoV-1 MERS NL63
COV91-27 65.62 27.37 >100 2.09 54.65 >300 >100 35.63† >100 46.37
COV44-62 20.92 6.37 25.48 1.06 2.11 22.87 17.68 55.30 39.90† 44.00
COV44-79 27.54 20.91 >100 3.24 56.19 28.40 21.02 46.79 >100 59.53
†
COV77-04 >100 >100 >100 >100 22.69 >300 >100 >100 >100 49.10
COV77-39 >100 >100 >100 >100 24.39 >300 >100 >100 >100 >100
COV78-36 >100 >100 >100 1.67 23.74 19.18 >100 >100 >100 59.00
Neg ctrl mAb >100 >100 >100 >100 >100 >100 >100 >100 >100 >100
100
80
60
40
20
n
o
i
t
a
z
i
l
a
r
t
u
e
N
%
0
10-1
100
80
60
40
20
n
o
i
t
a
z
i
l
a
r
t
u
e
N
%
0
10-1
COV44-62
100
102
101
Concentration (µg/mL)
103
COV44-79
100
102
101
Concentration (µg/mL)
103
WT
Alpha
Beta
Gamma
Delta
Mu
Omicron BA.1
Omicron BA.2
Omicron BA.4/5
WT
Alpha
Beta
Gamma
Delta
Mu
Omicron BA.1
Omicron BA.2
Omicron BA.4/5
NT50
(µg/mL)
9.80
11.24
7.06
12.80
13.28
6.52
10.38
20.26
51.89
NT50
(µg/mL)
21.53
24.21
19.66
30.64
45.33
16.51
33.02
55.44
49.81
Fig. 1. Broadly neutralizing antibodies target coronaviruses associated with
human disease. (A) Analysis of the frequency of MBCs expressing broadly
reactive antibodies from 19 donors. Values in parentheses represent the
percentage of SARS-CoV-2–reactive supernatants that also bind the specified
subsets of nonsarbecovirus coronavirus spikes. A total of 10,356 MBC culture
supernatants (50 to 100 B cells per well) were screened. n, number of MBC
culture wells. (B) Phylogenetic relationships across the coronavirus spike
proteins targeted by the broadly reactive mAbs were inferred by the neighbor-
joining method in MEGA11 using full-length amino acid sequences of CoV spike
proteins. Branch lengths are drawn to scale, and bootstrap values from 500
samplings are shown on the branches. The scale bar represents the number of
amino acid substitutions per site. (C) Heatmap representing the binding of
broadly reactive mAbs to spike proteins from coronaviruses across the alpha-,
beta-, and deltacoronavirus genera. H1 hemagglutinin was included as a negative
control for mAb binding experiments, and area under the curve (AUC) values for
each antigen are shown after subtraction with values for the negative control
antigen CD4. (D) Values represent antibody titer at 50% neutralization (NT50)
against SARS-CoV-2 (Wuhan-Hu-1 strain), SARS-CoV (indicated as “CoV-1”
in the figure), MERS-CoV, HCoV-NL63, and HCoV-229E envelope-pseudotyped
lentivirus, as well as authentic HCoV-OC43. For AssayScripps, values are average
of two experiments. For values with the † symbol, one NT50 was determinable
and one was not (i.e., >100 mg/ml), and the determinable NT50 is shown.
Negative control mAbs were anti–CoV-2 RBD CV503 (for OC43 assay) (20),
anti-influenza HA CR9114 (for AssayNIH except OC43) (42), and anti-dengue DEN3
(for AssayScripps) (43). NT50 values were calculated using the dose-response-
inhibition model with five-parameter Hill slope equation in GraphPad Prism 9.
(E) Neutralization of SARS-CoV-2 variants of concern (pseudovirus, AssayScripps)
by COV44-62 and COV44-79.
to the coronavirus spike panel used in the
plasma screen. Supernatants from only 2%
(n = 211) of the MBC culture wells met our
criteria for broad reactivity by binding at
least three betacoronavirus spike proteins
(Fig. 1A). Next, we developed an optofluidics
assay to isolate individual MBCs of interest
using the Berkeley Lights Beacon system
(fig. S1B). Candidate MBCs identified in the
supernatant screen were sorted individually
into nanoliter-volume pens and assessed in
real time for secretion of mAbs that bound to
beads coated with a cocktail of MERS-CoV and
HCoV-OC43 spikes, followed by beads coated
with SARS-CoV-2 spike. Double-positive MBCs
were exported for single-cell reverse transcrip-
tion polymerase chain reaction and antibody
expression as recombinant IgG1. In total, we
obtained 60 IgG mAbs with reactivity to at
least three coronaviruses.
To fully examine their breadth, we tested
the 60 mAbs for binding to spikes from the
seven human coronaviruses. Only six mAbs—
COV91-27, COV44-62, COV44-79, COV77-04,
COV77-39, and COV78-36—bound to spike
proteins from all seven coronaviruses (Fig. 1,
B and C). Notably, four of the six also bound
to spike from two new coronaviruses that have
recently been associated with human disease:
canine CoV HuPn-2018 (CCoV-HuPn-2018) and
Dacon et al., Science 377, 728–735 (2022)
12 August 2022
2 of 8
RES EARCH | R E S E A R C H A R T I C L E
porcine deltacoronavirus 0081-4 (PDCoV-0081-
4) (7, 8) (Fig. 1C). The six broadly reactive mAbs
were isolated from four different donors and
were encoded by four different heavy-chain
variable (VH) genes (VH1-2, VH3-30, VH4-59,
and VH5-10-1) (table S1). Five mAbs were highly
mutated, with VH nucleotide mutation frequen-
cies ranging from 10 to 13% (fig. S1C). Given that
these mAbs were isolated from COVID-19–
convalescent individuals in New York ~1 month
after the first outbreak in March 2020, these
mutation levels suggest that the B cells were
primed during an earlier seasonal coronavirus
infection and possibly reactivated during SARS-
CoV-2 infection.
COV44-62 and COV44-79 broadly
neutralize coronaviruses
We assessed the neutralizing potency of the six
mAbs against SARS-CoV-2, SARS-CoV, MERS-
CoV, HCoV-NL63, and HCoV-229E envelope
pseudotyped viruses, as well as authentic
HCoV-OC43. COV44-62 and COV44-79 showed
the broadest functional reactivity, neutralizing
the betacoronaviruses SARS-CoV-2, SARS-CoV,
and HCoV-OC43, as well as the alphacorona-
viruses HCoV-NL63 and HCoV-229E (Fig. 1D
and fig. S1D). Moreover, both mAbs neu-
tralized SARS-CoV-2 variants of concern,
including the Omicron BA.2 and BA.4/5 sub-
variants, as well as authentic SARS-CoV-2
(Fig. 1E and fig. S1E). COV44-62 also neutral-
ized MERS-CoV, whereas no other mAbs neu-
tralized this virus within the concentrations
tested.
Broadly reactive mAbs target the coronavirus
fusion peptide
To determine the domain of SARS-CoV-2 spike
that was targeted by the six broadly reactive
mAbs, we assessed mAb binding to the SARS-
CoV-2 S2 subunit as well as the RBD and N-
terminal domain (NTD) of the S1 subunit. All
six mAbs bound only to the S2 subunit (Fig. 2A).
VJ germline-reverted versions of the broadly
neutralizing mAbs COV44-62 and COV44-79
SARS-CoV-2 RBD
SARS-CoV-2 NTD
SARS-CoV-2 S2
I
) 106
F
M
(
I
) 106
F
M
(
I
) 106
F
M
(
i
g
n
d
n
b
i
b
A
m
105
104
103
10-5
i
g
n
d
n
b
i
b
A
m
10-410-310-2 10-1
100 101 102
103
Concentration (µg/mL)
105
104
103
10-5
10-410-310-210-1
100 101 102
103
Concentration (µg/mL)
Spike-2P
kd
s-1
ka
M-1s-1
COV44-62 7.22E+03 8.99E-05
COV44-79 1.27E+03 1.08E-04
COV91-27 6.63E+03 8.32E-05
COV77-04 9.53E+03 3.05E-05
COV77-39 9.28E+03 4.86E-05
COV78-36 6.22E+03 6.01E-05
KD
nM
14.40
84.40
13.50
3.61
6.03
9.21
ka
M-1s-1
2.3E+04
9.2E+03
2.2E+04
1.0E+04
2.0E+04
2.3E+04
S2 subunit
kd
s-1
2.2E-05
1.0E-05
1.0E-05
1.0E-05
1.0E-05
2.6E-05
KD
nM
0.95
1.10
0.49
0.98
0.53
1.13
i
g
n
d
n
b
i
b
A
m
)
C
U
A
(
C
i
g
n
d
n
b
i
b
A
m
105
104
103
10-5
12
10
8
6
4
2
0
10-410-310-2 10-1
100 101 102
103
Concentration (µg/mL)
SARS-CoV-2 S2 binding
S2 WT
S2 2P
S2
HexaPro
COV44-62
COV44-79
COV77-04
COV77-39
COV78-36
COV91-27
S2P6
L9 (neg control)
A
B
D
9
3
-
7
7
V
O
C
9
7
-
4
4
V
O
C
6
3
-
8
7
V
O
C
4
0
-
7
7
V
O
C
7
2
-
1
9
V
O
C
2
6
-
4
4
V
O
C
6
P
2
S
S2P6 0.22 2.18 1.50 1.97 1.82 1.07 1.44
COV77-39 1.65 0.27 0.24 0.18 0.24 0.29 0.17
COV44-79 1.73 0.19 0.21 0.14 0.16 0.10 0.15
COV78-36 1.92 0.22 0.06 0.17 0.09 0.10 0.07
COV77-04 10.14 7.53 0.29 1.41 1.55 0.33 0.25
COV91-27 1.83 0.25 0.10 0.17 0.20 0.09 0.07
COV44-62 1.87 0.42 0.14 0.33 0.28 0.26 0.11
Fig. 2. Broadly reactive mAbs target the same region within the SARS-CoV-2
S2 subunit. (A) Titration curves for mAb binding to selected regions within the
SARS-CoV-2 spike protein: the RBD, NTD, and S2 subunit. Interconnected data
points are shown without curve fitting. L9 is a malaria-specific mAb used as a
negative control (44). MFI, median fluorescence intensity. (B) On rates, off rates, and
dissociation constants of the six fusion peptide Fabs for binding to SARS-CoV-2
prefusion-stabilized spike (2P) with an unmodified furin cleavage site and the
nonstabilized S2 subunit. ka, association rate constant; kd, dissociation rate constant;
KD, equilibrium dissociation constant. (C) Fusion peptide mAb binding (AUC) to
wild-type (WT) SARS-CoV-2 S2 subunit and S2 subunit constructs modified with two
(2P) or six (HexaPro) stabilizing proline mutations. (D) Epitope binning of broadly
reactive antibodies versus the S2 stem-helix targeting mAb S2P6. All included
antibodies were tested as both ligands and analytes. Solid lines indicate two-way
competition, whereas the dashed line indicates one-way competition. Red boxes
indicate competing antibody pairs, green boxes indicate noncompeting antibody
pairs, and hatched boxes indicate self-competition.
Dacon et al., Science 377, 728–735 (2022)
12 August 2022
3 of 8
RES EARCH | R E S E A R C H A R T I C L E
showed weaker binding to S2 than the mutated
versions (fig. S2A), suggesting that somatic mu-
tations were important for improving binding
to the target site. A surface plasmon resonance
(SPR) kinetics assay determined the binding
affinity of antigen-binding fragments (Fabs)
derived from these mAbs for prefusion-stabilized
whole SARS-CoV-2 spike (2P, with an intact
S1-S2 cleavage site) and the unmodified S2
subunit. The Fabs bound with low-to-moderate
nanomolar affinity to both proteins, but their
affinity for the S2 subunit was 3- to 76-fold
higher than their affinity for whole spike (Fig. 2B
and fig. S2B). There were no substantial dif-
ferences between the six Fabs in their affinity
for the S2 subunit. The mAbs showed reduced
binding to a form of S2 that had been sta-
bilized with two proline mutations (S2-2P) and
bound more poorly still to a further stabilized
version with six proline mutations (S2-HexaPro)
(Fig. 2C). SPR-based competition showed that
the six mAbs competed for the same binding
site on the S2 subunit (Fig. 2D), but none com-
peted with S2P6, a control mAb targeting the
stem helix region (12), indicating their specific-
ity for a distinct site on S2.
To further investigate the specificity of these
mAbs, we performed SPR-based peptide
mapping using an array of 15–amino acid
(15-mer) overlapping peptides that spanned
the entire SARS-CoV-2 S2 subunit (S686 to
K1211, accession no. YP_009724390.1). All
six mAbs bound to peptides 42 to 44, which
share the 815RSFIEDLLF823 motif (Fig. 3A).
This motif is located within the SARS-CoV-2
fusion peptide region, directly C-terminal
to the S2′ cleavage site. To determine the
diversity of this region across corona-
viruses, we selected 34 viral isolates repre-
senting each of the four coronavirus genera
(Fig. 3B). Nearly all amino acid positions in
the 815RSFIEDLLF823 motif were conserved
in >85% of viruses selected; the exception
was F817, which was conserved in <50% of
isolates examined (Fig. 3, B and C). The
fusion peptide appears partially surface-
exposed in a range of coronavirus spike
proteins, including those of SARS-CoV-2,
SARS-CoV, MERS-CoV, and MHV (mouse
hepatitis virus) (21, 22) (Fig. 3C and fig. S3A).
However, antibody access to this site may be
partially occluded by the S1 subunit on an
adjacent protomer, consistent with stronger
binding to the S2 subunit relative to the
SARS-CoV-2 spike (Fig. 2B and fig. S3B).
The mAb specificity toward the fusion pep-
tide is consistent with their reduced binding
to the HexaPro S2 construct (Fig. 2C), which
includes a nonconservative F817P mutation at
this site (23). To identify key amino acids for
mAb binding, we performed an alanine scan
on a peptide encompassing residues 810 to
830 and focused on residues targeted by the
broadly neutralizing mAbs COV44-62 and
COV44-79 (Fig. 3D). Four amino acids—E819,
D820, L822, and F823—were important for
binding of COV44-62, where mutation of the
F823 residue abolished binding. Similarly,
residues critical for the binding of COV44-79
were E819, D820, and F823 but also included
R815 at the S2′ cleavage site (Fig. 3D). All five
residues identified as important for COV44-
62 or COV44-79 binding are among the most
conserved residues in the coronavirus spike
protein. D820 and L822 are completely con-
served, whereas R815, E819, and F823 are
conserved in 33 of 34 coronaviruses (Fig. 3,
B and E). Amino acid mutations at the pep-
tide level may have different effects from
mutations in the intact spike protein, where
modified interactions with surrounding resi-
dues may also affect antibody binding. There-
fore, we screened the six mAbs using a shotgun
alanine mutagenesis approach, whereby every
amino acid in the S2 subunit of intact spike
was individually mutated to generate a panel
of spike mutants (Fig. 3F and fig. S3, C and D).
In general, this assay identified a greater num-
ber of residues as important for mAb binding,
including some with a more intermediate
phenotype. For COV44-62, D820, L822, and
F823 were again crucial for binding, and K825,
D830, and R815 were also identified as im-
portant (Fig. 3F). For COV44-79, the results
closely matched the peptide alanine scan, with
the same four amino acids (R815, E819, D820,
and F823) identified as the most critical. When
the six mAbs were analyzed as a group, we
found that only the four broadest neutraliz-
ing mAbs were negatively affected by the
R815A mutation (fig. S4, A and B), suggesting
that binding to the S2′ cleavage site may be a
distinguishing property of broadly neutralizing
mAbs against this site.
Crystal structures of anti–fusion
peptide antibodies
To elucidate the molecular characteristics of
anti–fusion peptide antibodies that neutralize
SARS-CoV-2, the Fabs of the three broadest
neutralizing mAbs (COV44-62, COV44-79, and
COV91-27) were complexed with 15-mer pep-
tides containing the fusion peptide sequence
(Fig. 4). Crystal structures were determined to
1.46-, 2.8-, and 2.3-Å resolution, respectively
(Fig. 4, fig. S5, and table S2). Fourteen of the 15
peptide residues were visible in the electron
density map for COV44-62 (fig. S5A), 13 of
which have a buried surface area (BSA) >0 Å2
in complex with antibody. For COV44-79, 12 of
the 15 peptide residues were visible (fig. S5A),
10 of which have a BSA >0 Å2. Similarly, in
COV91-27, 12 peptide residues had interpretable
density (Fig. 4C and fig. S5A), with nine ex-
hibiting a BSA >0 Å2. The fusion peptide forms
a helix as in the prefusion state of the SARS-
CoV-2 spike (fig. S5B). All three complementarity-
determining regions (CDRs) of the heavy chain
(HC) of all three Fabs are involved in peptide
recognition, whereas CDR1 and CDR3 of the
light chain (LC) of COV44-62, and only LCDR3
of COV44-79 and COV91-27, contact the peptide
(Fig. 4, A to C). The BSA on each Fab is
dominated by the HC and is 791 Å2 for COV44-
62 (627 Å2 by HC and 164 Å2 by LC), 634 Å2 for
COV44-79 (505 Å2 by HC and 129 Å2 by LC), and
573 Å2 for COV91-27 (447 Å2 by HC and 126 Å2
by LC). The fusion peptide makes side-chain
and backbone H-bonds and salt bridges with
COV44-62 mainly through K814, R815, E819,
D820, L822, F823, and N824, and hydropho-
bic interactions through I818, L822, and F823
(Fig. 4D). These residues include the key resi-
dues (E819, D820, L822, and F823) identified
by site-directed mutagenesis (Fig. 3D). The fu-
sion peptide did not form as many interactions
with COV44-79 and COV91-27 (Fig. 4, E and F).
However, R815, E819, and D820 contributed
H-bonds and salt bridges, and I818, L822, and
F823 made hydrophobic interactions. In all
three antibodies, R815, S816, I818, E819, D820,
L822, and F823 contributed the most BSA to
the interaction (Fig. 4, G to I). There was par-
tial overlap between the residues of COV44-62
and COV44-79 that interacted with the fusion
peptide and those that were mutated from the
germline (fig. S6), consistent with the reduc-
tion in binding of germline-reverted versions
of these mAbs (fig. S2A). The structural results
are consistent with the peptide and spike pro-
tein mutagenesis data that identify the key
binding residues (Fig. 3, D to F, and fig. S3D).
Notably, the arginine at the S2′ cleavage site is
involved in recognition by these anti–fusion
peptide antibodies. Although the antibodies
all interact with one face of the fusion peptide
helical structure (fig. S5D), their approach
angles to the peptide differ. Superimposition
of the fusion peptide structures onto an intact
SARS-CoV-2 spike trimer structure in the
prefusion state showed a potential clash with
the S protein, suggesting that a conformational
change or conformational dynamics around
the fusion peptide is required to accommo-
date antibody targeting (fig. S5C). Nevertheless,
these antibodies have neutralization activity
against SARS-CoV-2 and therefore are able to
interact with the fusion peptide on the virus
(Fig. 1D).
Response to the fusion peptide after
vaccination and infection
We compared the binding of polyclonal IgG
from mRNA-1273–vaccinated donors (fig. S7A),
COVID-19–convalescent individuals, and COVID-
19–naïve individuals to the SARS-CoV-2 fusion
peptide (peptide 43) (Fig. 3A and fig. S7B). All
COVID-19–naïve donors showed minimal bind-
ing to the peptide, indicating a minor contribu-
tion by previous seasonal coronavirus infections
to circulating fusion peptide–specific IgG. Al-
though there was an increased response in
Dacon et al., Science 377, 728–735 (2022)
12 August 2022
4 of 8
RES EARCH | R E S E A R C H A R T I C L E
A
B
D
e
d
i
t
p
e
P
F
i
e
k
p
s
l
e
o
h
W
686
816
855
910
984
1034
1067
0411
1611
1211
FP
HR1
C Heli
CD
SH
HR2
S2
COV91-27
COV44-62
COV44-79
COV77-04
COV77-39
COV78-36
42 43 44
FP
C
CCoV-HuPn-2018
K R K Y R S A I E D L L F D K V V T S G L G T V D E D Y K R C T G G
HCoV-229E
HCoV-NL63
HCoV-HKU1
HCoV-OC43
R V A G R S A I E D I L F S K L V T S G L G T V D A D Y K K C T K G
R I A G R S A L E D L L F S K V V T S G L G T V D V D Y K S C T K G
- S S S R S L L E D L L F N K V K L S D V G F V E A - Y N N C T G G
K A S S R S A I E D L L F D K V K L S D V G F V E A - Y N N C T G G
Mouse Hepatitis Virus
A I R G R S A I E D L L F D K V K L S D V G F V E A - Y N N C T G G
Percent
identity
100
90
70
50
BtCoV-HKU3
BtCoV-RaTG13
BtCoV-WIV1
K P T K R S F I E D L L F N K V T L A D A G F M K Q - Y G D C L G D
50%
K P S K R S F I E D L L F N K V T L A D A G F I K Q - Y G D C L G D
K P T K R S F I E D L L F N K V T L A D A G F M K Q - Y G E C L G D
Civet SARS-CoV-007/2004 K P T K R S F I E D L L F N K V T L A D A G F M K Q - Y G E C L G D
PCoV-GX/P2V
SARS-CoV-1 Urbani
K P S K R S F I E D L L F N K V T L A D A G F I K Q - Y G D C L G D
K P T K R S F I E D L L F N K V T L A D A G F M K Q - Y G E C L G D
SARS-CoV-2 Wuhan Hu-1 K P S K R S F I E D L L F N K V T L A D A G F I K Q - Y G D C L G D
B.1.1.7
B.1.351
P.1
B.1.617.2
BA.1
BA.2
BA.2.12.1
BA.4/5
K P S K R S F I E D L L F N K V T L A D A G F I K Q - Y G D C L G D
K P S K R S F I E D L L F N K V T L A D A G F I K Q - Y G D C L G D
K P S K R S F I E D L L F N K V T L A D A G F I K Q - Y G D C L G D
K P S K R S F I E D L L F N K V T L A D A G F I K Q - Y G D C L G D
K P S K R S F I E D L L F N K V T L A D A G F I K Q - Y G D C L G D
K P S K R S F I E D L L F N K V T L A D A G F I K Q - Y G D C L G D
K P S K R S F I E D L L F N K V T L A D A G F I K Q - Y G D C L G D
K P S K R S F I E D L L F N K V T L A D A G F I K Q - Y G D C L G D
BtCoV-HKU4
S S S Y R S A I E D L L F D K V T I A D P G Y M Q G - Y D D C M K Q
MERS-CoV-EMC/2012
S R S A R S A I E D L L F D K V T I A D P G Y M Q G - Y D D C M Q Q
BtCoV-HKU9
AIBV
AIBV-Beaudette
TCoV
WhaleCoV-SW1
BuCoV HKU11-796
-
MuCoV HKU13-3514
-
PDCoV-0081-4/2014
PDCoV-0329-4/2015
ThCoV HKU12-600
-
-
WiCoV HKU20
A T T Y R S A F S D L L Y D K V R I T D P G F M Q S - Y Q K C I D S
S P R R R S F I E D L L F T S V E S V G L P T D D A - Y K N C T A G
S R R K R S L I E D L L F T S V E S V G L P T N D A - Y K N C T A G
Q N Q G R S T I E D L L F D K V V T L G V S E V D Q N Y D K C I A S
P R D A R S T I E D I L F D K V T T V G L G T V D A D Y D K C T K G
K A G G R S A I E D L L F N K V V T N G L G T V D Q D Y K A C S K D
K I G E K S V I E D L L F N K V V T N G L G T V D Q D Y K A C S K D
R L G G R S A I E D L L F N K V V T S G L G T V D Q D Y K A C S R D
R L G G R S A I E D L L F N K V V T S G L G T V D Q D Y K A C S R D
K Q G G R S A I E D L L F D K V V T N G L G T V D Q D Y K E C T K G
V K Q G R S A L E D L L F T K V V T A G L G T V D A D Y E K C A K G
Response units
0
8000
Peptides
42 PSKPSKRSFIEDLLF
43
44 RSFIEDLLFNKVTLA
PSKRSFIEDLLFNKV
Sequence conservation
Low
High
COV44-62
COV44-79
200
100
)
%
(
T
W
o
t
0
WTS K P S K R S F I E D L L F N K V T L
A WTS K P S K R S F I E D L L F N K V T L A
E
1.0
y
t
i
l
i
0.5
b
a
b
o
r
0.0P
COV44-62
COV44-79
815
820
825
830
835
840
SARS-CoV-2 spike amino acid numbering
800
600
) 200
%
(
T
W
o
t
100
0
WTS K P S
K R S F I E D L L F N K V T L A
D
WTS K P S
K R S F I E D L L F N K V T L A D
e
v
i
t
a
e
r
l
i
g
n
d
n
B
i
e
v
i
t
a
e
r
l
i
g
n
d
n
B
i
Fig. 3. Broadly neutralizing antibodies target the conserved fusion pep-
tide. (A) Heatmap of SARS-CoV-2 S2 peptide array. Binding responses were
assessed by SPR using a 15-mer peptide array with 12–amino acid overlay
covering the entire S2 subunit. Each column within the map represents a
single peptide. The white triangle denotes the S1-S2 cleavage site, and the
black triangle indicates the S2′ cleavage site. FP, fusion peptide; HR1, heptad
repeat 1; C Helix, central helix; CD, connector domain; SH, stem helix; HR2,
heptad repeat 2. (B) Sequence alignment of the fusion peptide from 34 viral
isolates representing a diverse group of coronaviruses across four genera.
This analysis was performed using MAFFT v7 with a BLOSUM62 scoring
matrix and the L-INS-I algorithm. Single-letter abbreviations for amino
acids: A, Ala; C, Cys; D, Asp; E, Glu; F, Phe; G, Gly; H, His; I, Ile; K, Lys; L, Leu;
M, Met; N, Asn; P, Pro; Q, Gln; R, Arg; S, Ser; T, Thr; V, Val; W, Trp; and
Y, Tyr. (C) Sequence conservation of prefusion SARS-CoV-2 spike protein (PDB
ID 6VSB); the fusion peptide (amino acids 816 to 843) is outlined in black. The
inset shows a magnified view of this region. (D) Alanine scan evaluating the
binding of COV44-62 and COV44-79 to the SARS-CoV-2 fusion peptide.
Responses were normalized to the wild-type sequence. A cutoff of 20% (brown
dashed line) was used to identify residues that were critical for binding.
(E) Sequence logo plot of diversity within the fusion peptide region of
coronaviruses from 34 isolates, built using WebLogo 3. Letter height is
proportional to the probability of an amino acid at a given position, and amino
acid residues are colored by charge. Narrow stacks (amino acids) indicate
deletions or gaps in the sequences. Numbering is based on the SARS-CoV-2
Wuhan-Hu-1 sequence. The key residues in the epitope footprints of mAbs
COV44-62 (red) and COV44-79 (blue), based on peptide alanine scanning, are
highlighted above the logo plot. (F) Amino acids critical for the binding of
COV44-62 and COV44-79, identified by shotgun alanine mutagenesis of S2
residues on whole spike protein. Only fusion peptide residues are shown. Key
residues were identified on the basis of a <20% signal relative to wild-type spike
(brown dashed line), with no corresponding loss of signal for a control mAb,
which targets the spike protein but does not bind to this site (fig. S3C).
Dacon et al., Science 377, 728–735 (2022)
12 August 2022
5 of 8
RES EARCH | R E S E A R C H A R T I C L E
A
D
B
C
812
PSKRSFIEDLLFNKV
826
809
PSKPSKRSFIEDLLF
823
809
PSKPSKRSFIEDLLF
823
Peptide 43
Peptide 42
Peptide 42
E
F
G
)
2
Å
(
a
e
r
a
e
c
a
f
r
u
s
d
e
i
r
u
B
250
200
150
100
50
0
S H
H
S
H
S
HH H
2
1
8
P
3
1
8
S
4
1
8
K
5
1
8
R
6
1
8
S
7
1
8
F
8
1
8
I
9
1
8
E
0
2
8
D
1
2
8
L
2
2
8
L
3
2
8
F
4
2
8
N
5
2
8
K
H
250
200
150
100
50
0
H
H
S
H
H
2
1
8
P
3
1
8
S
4
1
8
K
5
1
8
R
6
1
8
S
7
1
8
F
8
1
8
I
9
1
8
E
0
2
8
D
1
2
8
L
2
2
8
L
3
2
8
F
I
250
200
150
100
50
0
H
H H
H
S
H
2
1
8
P
3
1
8
S
4
1
8
K
5
1
8
R
6
1
8
S
7
1
8
F
8
1
8
I
9
1
8
E
0
2
8
D
1
2
8
L
2
2
8
L
3
2
8
F
Fig. 4. Crystal structures of COV44-62, COV44-79, and COV91-27 in complex with SARS-CoV-2
fusion peptide. (A to C) Overall interactions of (A) COV44-62, (B) COV44-79, and (C) COV91-27 with the
fusion peptide. Fabs are shown in molecular surface representation, and the CDRs and peptides are
represented as tubes. Cyan and yellow denote the heavy and light chains of the Fabs, respectively. Peptides
are shown in green. H1, H2, H3, L1, and L3 denote CDRs in the heavy (H) and light (L) chains. The
resolutions of the crystal structures are 1.46, 2.8, and 2.3 Å for the COV44-62, COV44-79, and COV91-27
complexes, respectively. Peptide residues observed in the crystal structures are in bold; residues that interact
with antibodies (BSA >0 Å2) are in red. (D to F) Details of the interactions between (D) COV44-62,
(E) COV44-79, and (F) COV91-27 with the fusion peptide. VH and VL indicate the variable domains of the
heavy and light chains, respectively. Kabat numbering was used for the Fabs; numbering in the native
spike protein was used for the fusion peptide. Colors for the heavy chain, light chain, and fusion peptide are
as in (A). (G to I) BSA (gray) and accessible surface area (white) of each residue of the fusion peptide in
complex with antibody are shown in the stacked bar chart. Residues that form polar interactions with
COV44-62, COV44-79, and COV91-27 are denoted atop the corresponding bar with “H” for a hydrogen bond
or “S” for a salt bridge. Buried and accessible surface areas were calculated with PISA (45).
several vaccine recipients after the second dose
(P = 0.025), this was not enhanced by ad-
ministration of a booster dose. The COVID-19–
convalescent donors did not have significantly
higher responses than vaccinated donors after
the second dose (P = 0.864). However, several
convalescent donors had the highest re-
sponses in all three cohorts, suggesting that
SARS-CoV-2 infection triggers a strong fusion
peptide–specific antibody response in some
individuals.
COV44-62 and COV44-79 inhibit
membrane fusion
Spike-mediated cell fusion relies on insertion
of the fusion peptide into the target cell mem-
brane and thus might be inhibited by antibody
binding to the fusion peptide. Consistent with
this, COV44-62 and COV44-79 inhibited the
fusion of cells expressing SARS-CoV-2 spike
and cells expressing the ACE2 receptor in an
imaging-based assay (Fig. 5A). We further
tested the six fusion peptide–specific mAbs
in a more quantitative assay wherein fusion
would trigger the release of an enzyme that
cleaves a chromogenic substrate. Only the
mAbs that neutralized SARS-CoV-2 were able
to strongly inhibit fusion (Fig. 5B).
COV44-79 limits disease in the Syrian
hamster model
We evaluated the in vivo efficacy of COV44-62
and COV44-79 against SARS-CoV-2 infection in
the Syrian hamster model, a well-established
model that recapitulates features of moderate-
to-severe COVID-19 in humans (24–26). We
converted the Fc regions of the two mAbs to
hamster IgG2 to allow optimal Fc function.
The mAbs were administered intraperitoni-
ally at 16 mg per kilogram of body weight
(mg/kg), followed by intranasal administra-
tion of 5 log10 plaque-forming units of SARS-
CoV-2 WA1 24 hours later (Fig. 5, C and D,
and fig. S8). Hamsters treated with COV44-79
and, to a lesser extent, COV44-62 had a smaller
decrease in body weight and recovered more
quickly than untreated hamsters (P < 0.01 from
days 3 to 7 for COV44-79; P < 0.05 from days 5
to 7 for COV44-62) (Fig. 5C). Similar results
were observed in a second experiment in which
COV44-79 was tested in comparison with a
hamster IgG2 isotype control (fig. S8B). Fur-
thermore, semiquantitative scoring revealed
that hamsters treated with COV44-79 had less
interstitial pneumonia than untreated ham-
sters on day 7 (P < 0.05) (Fig. 5D). COV44-79
was also able to slightly reduce lung viral
titers relative to control hamsters, on the
basis of subgenomic RNA quantification and
plaque assay analysis (fig. S8C).
Discussion
The broad conservation and functional im-
portance of the fusion peptide highlight the
potential of this site as a candidate for corona-
virus vaccine development. These findings
have parallels to studies of HIV-1 gp120, where
the surface-exposed fusion peptide was iden-
tified as a target of neutralizing mAbs (27).
This discovery led to the investigation of the
HIV-1 fusion peptide as a candidate immuno-
gen to elicit broadly neutralizing antibodies in
animals (28); subsequent animal studies have
increased the potency and breadth of the
fusion peptide–targeting antibody response
(29). Despite these potential advantages, the
coronavirus fusion peptide has not been a
major focus for development of therapeutic
mAbs and COVID-19 vaccines. The main draw-
back of the mAbs described here is their com-
paratively low in vitro neutralization potency.
These mAbs fit into a wider trend of a trade-off
between potency and breadth: Highly potent
mAbs that target the RBD are restricted to
sarbecoviruses and most do not neutralize all
variants of SARS-CoV-2 (3, 5, 6, 20, 30), whereas
mAbs that target the stem helix (12–16) and
Dacon et al., Science 377, 728–735 (2022)
12 August 2022
6 of 8
RES EARCH | R E S E A R C H A R T I C L E
A
ACE2
CoV-2 Spike
Merged
Control
(no mAb)
COV44-62
COV44-79
B
100
80
60
40
20
0
)
%
(
n
o
i
t
i
i
b
h
n
i
i
n
o
s
u
F
10 -2
Fusion inhibition
10-1
100
Concentration (µg/mL)
101
COV44-62
COV44-79
COV77-04
COV77-39
COV78-36
COV91-27
L9 (neg control)
C
)
%
(
t
i
h
g
e
w
y
d
o
b
e
g
a
r
e
v
A
110
105
100
95
90
85
102
ns
ns
ns ns
ns
**
ns
***
*
***
**
****
*
***
D
e
r
o
c
s
y
g
o
o
h
l
t
a
P
4
3
2
1
0
Interstitial pneumonia
ns
ns
ns
*
0
1
2
3
4
Day
5
6
7
Day 3
Day 7
Untreated, SARS-CoV-2 naive
PBS-treated, SARS-CoV-2 exposed
COV44-62 (16 mg/kg)
COV44-79 (16 mg/kg)
Fig. 5. COV44-62 and COV44-79 inhibit SARS-CoV-2 spike–mediated
fusion and COV44-79 limits disease in a Syrian hamster model. (A) Images
of fusion between HeLa cells stably expressing SARS-CoV-2 spike (red
fluorescent protein) and HeLa cells stably expressing the ACE2 receptor (green
fluorescent protein) after counterstaining with Hoechst (blue). Cells were
cocultured in the presence of COV44-62 or COV44-79, or without a mAb
(control). Scale bars, 500 mm. (B) Fusion inhibition of six fusion peptide–specific
mAbs in a quantitative assay. (C) Weight change for SARS-CoV-2–naïve
animals versus virus-exposed animals that were mock-treated or treated with
16 mg/kg of mAb. Statistical significance for average body weight was analyzed
across the 7-day time course using a mixed-effects repeated measures model
with Dunnett’s post-test multiple comparison (n = 12 animals from days 0 to 3;
n = 6 animals from days 4 to 7). Error bars show mean ± SD. (D) Pathology
scores for SARS-CoV-2–naïve animals versus virus-exposed animals that were
mock-treated or treated with 16 mg/kg of mAb. Scores for interstitial pneumonia
pathology (days 3 and 7) based on gross pathology observations were
statistically analyzed by a Kruskal-Wallis test with Dunn’s post-test multiple
comparison (n = 6 to 12 animals per condition) between the mAb-treated and
mock-treated groups on each day. Bars show median and interquartile range.
*P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001; ns, not significant.
those identified here have greater breadth
but are less potent. However, as previously
reported for at least one antistem helix mAb
(15), COV44-79 performed better than expected
in the hamster model, which suggests that it
may function in a way that is not captured
effectively in the neutralization assay. For
instance, Fc effector functions may enhance
activity, as observed with anti–SARS-CoV-2
antibodies with other specificities (31–33).
Moreover, there is substantial scope for im-
provement for mAbs with this specificity, and
subsequent studies may discover mAbs with
higher potency, as with the HIV-1 fusion peptide
(34). Techniques to improve antibody affinity
and potency could also be useful (35, 36).
Additionally, vaccination with fusion peptide
constructs may trigger a polyclonal response of
greater magnitude and potency. We found that
three doses of the mRNA-1273 vaccine did not
produce strong antibody responses to the
fusion peptide, although several COVID-19–
convalescent individuals exhibited strong anti-
body responses to this site. This observation
is consistent with greater exposure of the S2
subunit to B cells during natural infection as
a result of S1 uncoupling, which likely occurs
less frequently with prefusion-stabilized spike
protein. Furthermore, depletion of fusion
peptide–specific antibodies from the serum
Dacon et al., Science 377, 728–735 (2022)
12 August 2022
7 of 8
RES EARCH | R E S E A R C H A R T I C L E
of COVID-19–convalescent patients resulted
in a 20% reduction in SARS-CoV-2 neutral-
ization (37), and reactivity to the fusion peptide
was correlated with neutralization titer (38),
indicating that polyclonal antibodies to the
fusion peptide can play an appreciable role. Thus,
our findings are consistent with previous studies
that have highlighted the potential utility of
the fusion peptide as a target epitope (37–41),
and the fusion peptide–targeted mAbs provide
additional tools to combat COVID-19 and enhance
pandemic preparedness.
I. Ullah et al., Immunity 54, 2143–2158.e15 (2021).
31.
32. R. Yamin et al., Nature 599, 465–470 (2021).
33. Y. C. Bartsch et al., Nat. Med. 27, 454–462 (2021).
34. C. H. Shen et al., Cell Host Microbe 27, 531–543.e6 (2020).
35. S. Kratochvil et al., Immunity 54, 2859–2876.e7 (2021).
36. A. Wellner et al., Nat. Chem. Biol. 17, 1057–1064 (2021).
37. C. M. Poh et al., Nat. Commun. 11, 2806 (2020).
38. W. R. Morgenlander et al., J. Clin. Invest. 131, e146927 (2021).
39. X. Sun et al., Nat. Microbiol. 7, 1063–1074 (2022).
40. N. Vanderheijden et al., Front. Immunol. 12, 790415 (2022).
41. D. L. N. F. Maeda et al., Proc. Natl. Acad. Sci. U.S.A. 118,
e2025622118 (2021).
42. C. Dreyfus et al., Science 337, 1343–1348 (2012).
43. T. F. Rogers et al., Science 369, 956–963 (2020).
44. L. T. Wang et al., Immunity 53, 733–744.e8 (2020).
45. E. Krissinel, K. Henrick, J. Mol. Biol. 372, 774–797 (2007).
RE FE RENCES AND N OT ES
ACKN OWLED GMEN TS
1. P. V’kovski, A. Kratzel, S. Steiner, H. Stalder, V. Thiel,
Nat. Rev. Microbiol. 19, 155–170 (2021).
2. E. Dong, H. Du, L. Gardner, Lancet Infect. Dis. 20, 533–534
(2020).
3. S. Iketani et al., Nature 604, 553–556 (2022).
4. N. Andrews et al., N. Engl. J. Med. 386, 1532–1546 (2022).
5. E. Takashita et al., N. Engl. J. Med. 10.1056/NEJMc2207519
(2022).
6. L. A. VanBlargan et al., Nat. Med. 28, 490–495 (2022).
J. A. Lednicky et al., Nature 600, 133–137 (2021).
7.
8. A. N. Vlasova et al., Clin. Infect. Dis. 74, 446–454 (2022).
9. C. B. Jackson, M. Farzan, B. Chen, H. Choe, Nat. Rev. Mol.
Cell Biol. 23, 3–20 (2022).
10. J. S. Tregoning, K. E. Flight, S. L. Higham, Z. Wang, B. F. Pierce,
Nat. Rev. Immunol. 21, 626–636 (2021).
11. T. N. Starr et al., Cell 182, 1295–1310.e20 (2020).
12. D. Pinto et al., Science 373, 1109–1116 (2021).
13. M. M. Sauer et al., Nat. Struct. Mol. Biol. 28, 478–486 (2021).
14. C. Wang et al., Nat. Commun. 12, 1715 (2021).
15. P. Zhou et al., Sci. Transl. Med. 14, eabi9215 (2022).
16. W. Li et al., Cell Rep. 38, 110210 (2022).
17. F. Amanat et al., Cell 184, 3936–3948.e10 (2021).
18. K. A. Huang et al., PLOS Pathog. 17, e1009352 (2021).
19. W. N. Voss et al., Science 372, 1108–1112 (2021).
20. H. Cho et al., Sci. Transl. Med. 13, eabj5413 (2021).
21. A. C. Walls et al., Nature 531, 114–117 (2016).
22. Y. Yuan et al., Nat. Commun. 8, 15092 (2017).
23. C. L. Hsieh et al., Science 369, 1501–1505 (2020).
24. J. F. Chan et al., Clin. Infect. Dis. 71, 2428–2446 (2020).
25. M. Imai et al., Proc. Natl. Acad. Sci. U.S.A. 117, 16587–16595
(2020).
26. S. F. Sia et al., Nature 583, 834–838 (2020).
27. R. Kong et al., Science 352, 828–833 (2016).
28. K. Xu et al., Nat. Med. 24, 857–867 (2018).
29. R. Kong et al., Cell 178, 567–584.e19 (2019).
30. M. A. Tortorici et al., Nature 597, 103–108 (2021).
We thank the blood sample donors at the New York Blood
Center; S. Bangaru, G. Ozorowski, A. Torrents de la Peña, and
A. Ward for providing HCoV-OC43 and MERS-CoV; G. Wright and
N. Muller-Sienerth (Wright lab, University of York) for providing
recombinant CD4; L. Wang and R. Seder for providing the L9
antibody; the Burton lab (The Scripps Research Institute) for
providing BA.2 and BA.4/5 gene fragments; and M. Cohen and
J. Laux for assistance with cell sorting. We also thank N. Vaughan,
K. Cooper, R. Reeder, M. St Claire, K. Hadley, D. Drawbaugh,
A. Hischak, R. Hart, N. Isic, M. Murphy, E. Postnikova, M. Anantpadma,
and E. Eudy for assistance with hamster experiments. We are grateful
to the staff of Advanced Photon Source and Stanford Synchrotron
Radiation Lightsource (SSRL) Beamline 12-1 for assistance with
synchrotron data collection. Funding: This work was supported by the
Division of Intramural Research and the Vaccine Research Center,
National Institute of Allergy and Infectious Diseases (NIAID), National
Institutes of Health (NIH) (I.D., S.M., J.W.Y., P.D.C., M.R.H., J.R.M., and
J.T.). This project has been funded in whole or in part with federal
funds from the NIAID, NIH, US Department of Health and Human
Services (DHHS), under contract HHSN272201800013C. This work
was also supported by NIH grant R01AI132317 (D.N. and L.P.), HHSN
contract 75N93019C00073 (J.K.W. and E.D.), and Bill and Melinda
Gates Foundation grant INV-004923 (I.A.W.). This research used
resources of the Advanced Photon Source, a US Department of
Energy (DOE) Office of Science user facility operated for the DOE
Office of Science by Argonne National Laboratory under contract DE-
AC02-06CH11357. GM/CA@APS has been funded by the National
Cancer Institute (ACB-12002) and the National Institute of General
Medical Sciences (AGM-12006 and P30GM138396). Extraordinary
facility operations were supported in part by the DOE Office of
Science through the National Virtual Biotechnology Laboratory, a
consortium of DOE national laboratories focused on the response to
COVID-19, with funding provided by the Coronavirus CARES Act. Use
of the SSRL, SLAC National Accelerator Laboratory, is supported by
the DOE, Office of Science, Office of Basic Energy Sciences, under
contract DE-AC02-76SF00515. The SSRL Structural Molecular Biology
Program is supported by the DOE Office of Biological and
Environmental Research and by the National Institute of General
Medical Sciences of the NIH (P30GM133894). Author contributions:
Conceptualization: J.T. and C.D. Methodology: C.D., C.T., L.Pe.,
C.-C.D.L., T.-H.L., M.Y., Y.C., L.W., D.D., B.E., I.K., Z.H., R.S.W., A.P.,
E.D., D.E.G., I.D., and S.M. Formal analysis: C.D., C.T., L.Pe., M.Y.,
Y.C., L.W., C.-C.D.L., T.-H.L., C.-W.P., J.S., and E.D. Data curation:
M.P. Investigation: C.D., C.T., L.Pe., M.Y., L.W., C.-C.D.L., L.Pu., C.-W.P.,
J.K.W., T.-H.L., M.Z., M.P., D.M., A.J.R.C., S.D., E.K., L.H., D.P., R.B.,
S.L., D.D., B.E., Y.Z., E.S.Y., M.C., and K.L. Resources: R.S.W. and
S.M. Writing – original draft: C.D., C.T., M.Y., and J.T. Writing –
review & editing: all authors. Visualization: C.D., C.T., L.Pe., M.Y.,
Y.C., L.W., C.-C.D.L., J.K.W., E.D., and J.T. Supervision: E.D., D.E.G.,
S.M., J.W.Y., C.S., P.D.C., D.N., M.R.H., J.R.M., I.A.W., and J.T.
Funding acquisition: C.S., P.D.C., D.N., M.R.H., J.R.M., I.A.W., and
J.T. Competing interests: J.T. and C.D. are coinventors on a
provisional patent (US Patent Application no. 63/308,898) filed on
the mAbs described in this study. J.K.W. and E.D. are employees
of Integral Molecular. Y.C., S.D., E.K., L.H., D.P., R.B., S.L., D.D.,
B.E., and M.R.H. performed this work as employees of Laulima
Government Solutions, LLC. The content of this publication does
not necessarily reflect the views or policies of the DHHS or of the
institutions and companies with which the authors are affiliated.
All other authors declare no competing interests. Data and
materials availability: All data associated with this manuscript are
available in the main text or the supplementary materials. Crystal
structures have been deposited into the Protein Data Bank (PDB
IDs 8D36 for COV44-62, 8DAO for COV44-79, and 8D6Z for
COV91-27). Antibody sequences have been deposited in GenBank
(accession numbers ON695901 to ON695912). Materials described
in this manuscript will be available through a material transfer
agreement with the NIAID. License information: This work is
licensed under a Creative Commons Attribution 4.0 International
(CC BY 4.0) license, which permits unrestricted use, distribution,
and reproduction in any medium, provided the original work is
properly cited. To view a copy of this license, visit https://
creativecommons.org/licenses/by/4.0/. This license does not
apply to figures/photos/artwork or other content included in the
article that is credited to a third party; obtain authorization from
the rights holder before using such material.
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abq3773
Materials and Methods
Figs. S1 to S8
Tables S1 and S2
References (46–71)
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
Submitted 4 April 2022; accepted 6 July 2022
Published online 12 July 2022
10.1126/science.abq3773
Dacon et al., Science 377, 728–735 (2022)
12 August 2022
8 of 8
|
10.1126_science.abq6100
|
mal state (24, 25). A lack of granular quasipar-
ticles would naїvely be expected to suppress
shot noise, because the flow of a continuous
fluid should have no fluctuations.
Despite their ubiquity, strange metals have
yet to be examined through shot-noise mea-
surements for several technical reasons, and
only a few relevant theoretical predictions ex-
ist for any quantum critical systems (26, 27). In
many materials, strange metallicity is cut off at
low temperatures by the onset of superconduc-
tivity, which complicates matters because shot-
noise measurements also require an electrical
bias eV, where e is the charge of the electron and
V is the applied voltage, that is large compared
with the thermal scale kBT to distinguish from
thermal noise, where kB is the Boltzmann constant.
Tunneling transport into a strange metal faces
the challenge that only discrete, individual elec-
trons can be added or removed, which likely leads
to noise dominated by single-electron effects.
Fortunately shot noise can be measured within a
material using a diffusive mesoscopic wire, which
requires the nanofabrication of such structures
without affecting electronic properties—a ma-
jor challenge for many materials.
RES EARCH
HEAVY FERMIONS
Shot noise in a strange metal
Liyang Chen1, Dale T. Lowder2, Emine Bakali3, Aaron Maxwell Andrews4, Werner Schrenk5,
Monika Waas3, Robert Svagera3, Gaku Eguchi3, Lukas Prochaska3, Yiming Wang2, Chandan Setty2,
Shouvik Sur2, Qimiao Si2, Silke Paschen3, Douglas Natelson2,6,7*
S trange-metal behavior has been observed in materials ranging from high-temperature superconductors
to heavy fermion metals. In conventional metals, current is carried by quasiparticles; although it has
been suggested that quasiparticles are absent in strange metals, direct experimental evidence is lacking.
We measured shot noise to probe the granularity of the current-carrying excitations in nanowires
of the heavy fermion strange metal YbRh2Si2. When compared with conventional metals, shot noise in
these nanowires is strongly suppressed. This suppression cannot be attributed to either electron-phonon
or electron-electron interactions in a Fermi liquid, which suggests that the current is not carried by
well-defined quasiparticles in the strange-metal regime that we probed. Our work sets the stage
for similar studies of other strange metals.
port of ordinary “granular” charge carriers
of magnitude e with an average current Ih i.
Shot noise has revealed fractionalization of
charge in the fractional quantum Hall liquid
(20, 21), fractional effective charges in quan-
tum dot Kondo systems (22, 23), and pairing
in superconducting nanostructures in the nor-
S trange metals are non-Fermi liquids
that exhibit a low temperature (T)
electrical resistivity contribution that is
directly proportional to T (1). This re-
sponse has been reported across many
materials families, including cuprate (2–4)
and pnictide (5) superconductors, ruthenates
(6), heavy fermion metals (7–9), and twisted
bilayer graphene (10). Strange-metal properties
typically arise at finite temperature above a
quantum critical point (QCP), often in prox-
imity to antiferromagnetic order (11). There
are two broad classes of theories on metallic
QCPs. Within the standard Landau approach
of order parameter fluctuations, quasiparticles
retain their integrity (12, 13). By contrast, in
approaches beyond the Landau framework
(14–18), no long-lived quasiparticles are ex-
pected to remain. Thus, determining the nature
of the low-energy current-carrying excitations
is an important means to elucidate the nature of
strange metals near QCPs.
(cid:1)
How can we determine whether the current
carriers in strange metals are quasiparticles?
Shot noise in electrical conduction (19) is a
distinctive probe of mesoscopic systems in
(cid:3)
which the current noise, SI ¼ I (cid:2) Ih i
, in
a system driven out of equilibrium accesses
the nature of the charge-carrying excitations.
Here, I is the instantaneous current and Ih i is
the average current. The Fano factor, F, gives
the ratio between the measured noise SI and
2e Ih i, the expectation for Poissonian trans-
Þ2
ð
1Applied Physics Graduate Program, Rice University,
TX 77005, USA. 2Department of Physics and Astronomy,
Rice Center for Quantum Materials, Rice University, Houston,
TX 77005, USA. 3Institute of Solid State Physics, TU Wien,
Wiedner Hauptstraße 8-10, 1040 Vienna, Austria. 4Institute
of Solid State Electronics, TU Wien, Gußhausstraße 25-25a,
Gebäude CH, 1040 Vienna, Austria. 5Center for Micro and
Nanostructures, TU Wien, Gußhausstraße 25-25a, Gebäude
CH, 1040 Vienna, Austria. 6Department of Electrical and
Computer Engineering, Rice University, Houston, TX 77005,
USA. 7Department of Materials Science and
NanoEngineering, Rice University, Houston, TX 77005, USA.
*Corresponding author. Email: natelson@rice.edu
Fig. 1. YbRh2Si2 nanowire device preparation and characterization. (A) YbRh2Si2 nanowire between two large-
area, thick-sputtered Au contacts on top of the unpatterned YbRh2Si2 film, which were deposited to ensure that
the measured voltage is dominated by the nanowire. (B) Higher-magnification view of the indicated region in (A).
Sample fabrication is discussed in detail in (33). (C) Normalized resistance as a function of temperature for both the
unpatterned film and the etched nanowire. The inset shows that resistivity in the low-temperature limit in both the film
and the wire is linear in temperature (with black dashed line as a linear-in-temperature guide to the eye), as seen
previously (31). Unpatterned film resistance at 100 K is 17.8 ohms. Nanowire resistance at 100 K is 164.7 ohms.
(D) Normalized resistance as a function of the in-plane magnetic field for both the unpatterned molecular-beam epitaxy
film and the etched nanowire (magnetic field B is oriented transverse to the nanowire), with curves shifted vertically for
clarity. Zero-field resistances for the film at 10, 7, 5, and 3 K (top to bottom) are 6.5, 5.5, 4.8, and 4.1 ohms,
respectively. Zero-field resistances for the wire at 10, 7, 5, and 3 K are 57.8, 49.0, 42.2, and 35.5 ohms, respectively. The
nearly identical response between the nanowire and the unpatterned film confirms that patterning did not substantially
alter the electronic properties of the epitaxial YbRh2Si2 material and that the resistance is dominated by the wire.
Chen et al., Science 382, 907–911 (2023)
24 November 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Noise characteriza-
tion of a YbRh2Si2 nanowire.
(A) Differential resistance
dV/dI as a function of bias
current at 10, 7, 5, and 3 K
(top to bottom). Comparison
with theoretical shot-noise
expectations requires this
information (see Eqs. 1 and 2).
(B) Averaged voltage noise
spectra (with zero-bias spectra
subtracted) of a YbRh2Si2
nanowire device at different
bias levels at T = 3 K, over a
bandwidth between 300 and
600 kHz. This spectral range is
free of extrinsic features, and
these voltage noise spectra are analyzed (Eq. 2) to determine the shot noise at each bias. Each spectrum shown is an average of 4500 spectra with 10-kHz bandwidth.
We have successfully made mesoscopic wires
for noise measurements from epitaxial films of
the heavy fermion material YbRh2Si2, a par-
ticularly well-defined strange metal (9, 28).
YbRh2Si2 has a zero-temperature magnetic field–
induced continuous quantum phase transition
from a low-field antiferromagnetic heavy Fermi
liquid metal to a paramagnetic one. The Hall
effect displays a rapid isothermal crossover
that extrapolates to a jump at the QCP in the
zero-temperature limit, which provides evi-
dence for a sudden reconstruction of the Fermi
surface across the QCP and an associated change
in the nature of the quasiparticles between
the two phases (29), as expected in the Kondo
destruction description (14–16) for a beyond-
Landau QCP. At finite temperatures, a quan-
tum critical fan of strange metallicity extends
over a broad range of temperature and mag-
netic field (28, 30). Recent time-domain THz
transmission measurements (31) of the optical
conductivity of epitaxial films of YbRh2Si2 re-
veal the presence of quantum critical charge
fluctuations below 15 K, supporting the Kondo
destruction picture in this system.
Measuring shot noise in YbRh2Si2 wires di-
rectly examines how current flows in a system
thought to lack discrete charge excitations;
these results can then be compared with predic-
tions in Fermi liquids. We report measurements
of shot noise in mesoscopic wires patterned
from epitaxial films of YbRh2Si2, examined
below 10 K, in the strange-metal regime where
phonon scattering is not expected to be relevant
to the conductivity. The measured shot noise
is found to be far smaller than both weak- and
strong electron-electron scattering expectations
for Fermi liquids and also smaller than the val-
ues measured on a gold nanowire for compari-
son. Furthermore, the electron-phonon coupling
that was determined experimentally using long
YbRh2Si2 nanowires rules out strong electron-
phonon scattering as a noise-suppression mech-
anism. Therefore, the suppressed shot noise is
evidence that current-carrying excitations in this
strange metal defy a quasiparticle description
in the examined temperature range. The no-
quasiparticle model described in previous work
(26, 27), despite being derived using conformal
field theory for different kinds of QCP and as-
sociated phases than those of YbRh2Si2, predicts
nontrivial bias- and temperature-dependent
noise that is qualitatively consistent with the
observed trends.
Measuring shot noise in YbRh2Si2 devices
High-quality epitaxial films of YbRh2Si2 were
grown by molecular beam epitaxy on germa-
nium substrates (31, 32) [see section 3 in (33)
for details]. The temperature dependence of
the resistivity of these films above 3 K shows
strange-metal properties (r = r0 + ATa, where
a ≈ 1 in the low temperature limit, r is is the
electrical resistivity, and A is the temperature
coefficient) as in the bulk material (Fig. 1C).
The films are patterned into nanowires through
a combination of electron-beam lithography
and reactive ion etching (see fig. S2 and the ac-
companying discussion). The nanowire shown
in Fig. 1B is 60 nm thick, 660 nm long, and
240 nm in width. Thick source and drain con-
tact pads ensure that the dominant voltage
measured under bias is across the nanowire;
the contact pads also act as thermal sinks (34).
An important concern in fabricating nano-
structures from strongly correlated materials is
that the patterning process does not alter the
underlying physics. As shown in Fig. 1C, the
resistance R(T) of the nanowire closely matches
that of the unpatterned film, including a domi-
nant linear-in-T dependence at low tempera-
tures. Similarly, in Fig. 1D, the magnetoresistance
(field in-plane, perpendicular to the current)
in the nanowire is nearly identical to that of
the unpatterned film, showing that the fabrica-
tion process did not alter the material’s proper-
ties. This consistency also shows that the total
R is dominated by the wire, because the large
contacts are coated in thick gold and would
not exhibit such a magnetoresistance. Three
nanowires patterned from this same film all
show essentially identical transport and noise
properties (data from devices 2 and 3 are
shown in fig. S5).
The noise-measurement technique is well de-
veloped (21, 24, 34). A current bias is applied to
the device by means of a heavily filtered voltage
source and ballast resistors. Using a custom
probe, the voltage across the device is measured
through two parallel sets of amplifiers and a
high-speed data acquisition system (fig. S1C).
The time-series data are cross-correlated and
Fourier transformed to yield the voltage noise SV
across the device, with the correlation mitigating
the amplifier input noise [see sections 1 and 2 of
(33) for a detailed discussion of calibration and
averaging]. Figure 2A shows the variation of the
differential resistance, dV/dI, as a function of
bias current, whereas Fig. 2B gives examples of
voltage noise spectra at several bias currents
at a base temperature of 3 K. At the maxi-
mum bias currents that were applied, the
voltage drop across the wire is several mV, a
bias energy scale that considerably exceeds
kBT (0.25 meV at 3 K), as is needed for shot
noise measurements.
Theoretical expectations for the shot noise
and Fano factor
To understand the measured noise in YbRh2Si2,
we first considered the expected current shot
noise result for a diffusive metallic constriction.
This is a long-established calculation within the
Landauer-Büttiker formalism (35–39) for a Fermi
gas (i.e., without any electron-electron interac-
tions). A metal with well-defined quasiparticles is
assumed in the source and drain, which obey the
Fermi-Dirac (FD) distribution with a tempera-
ture set by the contacts, T0. In the noninteracting,
nanoscale limit, conduction takes place through
spin-degenerate quantum channels with various
transmittances, ti. Each channel contributes to SI
by an amount proportional to ti(1 − ti). By av-
eraging over the distribution of transmittances
(35–38), one finds a predicted Fano factor
Chen et al., Science 382, 907–911 (2023)
24 November 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Noise versus bias current characteristics. (A) Noise versus bias
current for a YbRh2Si2 wire at 10, 7, 5, and 3 K (top to bottom), with fits to
Eq. 1 to extract effective Fano factors. Error bars are the standard deviation from
15 repeated bias-sweep measurements. Also shown for illustrative purposes
=4; 1=3 and 0 (indicated by the gray dot-dashed
are expectations for F ¼
curves from top to bottom, respectively), which were calculated by using the
measured differential resistance at each temperature and Eq. 2. At all
p
ffiffiffi
3
temperatures, the measured voltage noise is far below the theoretical expect-
ations for shot noise in a diffusive nanowire of a Fermi liquid, even in the weak
electron-electron scattering limit. (B) Analogous data for a gold wire over the
same temperature range, as discussed in section 9 of (33). Data here are much
closer to the F ¼ 1=3 Fermi liquid expectation, with the small deviation at the
highest temperatures being attributed to electron-phonon scattering effects.
Error bars are the standard deviation from 15 repeated bias-sweep measurements.
F ≡ SI =2e Ih i ¼ 1=3. When inelastic electron-
electron scattering is added to the otherwise
noninteracting Fermi system, such that the
system size along the direction of the current
exceeds the electron-electron scattering length,
L > Lee (see fig. S11) but is smaller than the
electron-phonon scattering length, Lph, there is
a redistribution of energy and effective ther-
malization among the carriers (40, 41). There
is a local quasi-thermal FD distribution within
the wire described by a local electronic temper-
ature Te(x) that is higher than the lattice tem-
perature, T0 = T, which is assumed to be uniform
and equal to the temperature of the contacts.
This approach leads to a prediction of F ¼
p
=4 ≈ 0:433 (40, 41). Fano factor predic-
tions in both L < Lee < Lph and Lph > L > Lee
limits have been confirmed in experiments in
mesoscopic metal wires (34, 42–44).
ffiffiffi
3
In the present context, an important ques-
tion is what happens in a Fermi liquid state
when the electron-electron interactions are so
strong that the quasiparticle weight is orders
of magnitude smaller than the noninteract-
ing case (equal to 1) for a free electron (though
still nonzero) and the Landau parameters are
correspondingly large. It was recently shown
that charge conservation constrains the Fano
factor to be independent of the quasiparticle
weight, and the combination of instantaneous
electronic interactions and Poissonian charge
transport dictates that the shot noise and av-
erage current get renormalized identically by
the Landau parameters (45). As a result, for
this regime (see solid line in fig. S11) that per-
tains to a strongly correlated Fermi liquid of
interest here, the Fano factor would be F ¼
p
ffiffiffi
3
=4 ≈ 0:433. [For further details, see section
13 of (33) and (45).]
Within the Fermi liquid quasiparticle pic-
ture, the only way to suppress shot noise below
these levels is through strong electron-phonon
scattering, which perturbs the electronic dis-
tribution function. In the limit of very strong
electron-phonon coupling, the electronic dis-
tribution is constrained to be in equilibrium
with the lattice temperature, T0, and only
Johnson-Nyquist noise at T0 remains.
Comparison to theoretical expectations
Figure 3 shows the measured voltage noise as
a function of bias current for a YbRh2Si2 nano-
wire, and its counterpart for a gold nanowire
for comparison. Shown as gray dot-dashed lines
are the F ¼ 1=3 expectations based on the mea-
sured differential resistance, dV =dI. Indepen-
dent of any detailed analysis, the measured noise
in the YbRh2Si2 wire is clearly suppressed
well below the Fermi liquid expectation at all
temperatures. Additional data on two more
wires (devices 2 and 3) are essentially iden-
tical [see section 7 of (33) and fig. S5]. By
contrast, the gold nanowire data [discussed
further in section 9 of (33) and fig. S7] are
consistent with Fermi liquid predictions, with
a slight suppression of the noise above 10 K as
electron-phonon scattering becomes relevant
(fig. S7D).
The electron-phonon coupling may be ex-
tracted experimentally by analyzing the noise
as a function of bias in a wire sufficiently long
that electron-phonon scattering is dominant
(42). As detailed in section 5 of (33), we per-
formed this analysis using a 30-mm-long YbRh2Si2
Chen et al., Science 382, 907–911 (2023)
24 November 2023
wire to determine the effective electron-phonon
coupling in this material and found a value
sufficiently small that strong electron-phonon
scattering is ruled out as a mechanism for sup-
pressing the noise in the much shorter YbRh2Si2
nanowire constrictions.
Extracting effective Fano factors from the
measured noise requires analysis in terms of
finite temperature expressions for the shot noise.
Subtleties about thermal noise can arise when
the device is non-ohmic, as discussed in sec-
tion 10 of (33), but corrections from the ohmic
case are small for the measured nonlinearities
shown in Fig. 2A. The expected form for the
current shot noise in an ohmic system with Fano
factor F and differential resistance dV =dI is (19)
SI ¼ F (cid:3) 2e Ih icoth
(cid:6)
(cid:5)
eV
2kBT
ð
þ 1 (cid:2) F
Þ4kBT
(cid:5) (cid:6)(cid:2)1
dV
dI
ð1Þ
(cid:7)
This expression reduces to the Johnson-Nyquist
(cid:8)(cid:2)1
current noise SI;JN ¼ 4kBT dV
I¼0 in the zero-
dI
bias limit and becomes SI ¼ F (cid:3) 2e Ih i as ex-
pected in the high-bias limit eV ≫kBT . In the
experiment, we measured voltage noise, and, for
ease of comparison, we subtracted off the zero-
bias Johnson-Nyquist noise so that effective
Fano factors may be estimated by fitting to the
voltage-based expression for the shot noise:
(cid:9)
(cid:5) (cid:6)
2
dV
dI
SV ¼
F (cid:3) 2e Ih icoth
ð
1 (cid:2) F
Þ4kBT
(cid:2) 4kBT
(cid:10)
I
(cid:5) (cid:6)(cid:2)1
dV
dI
1
þ
(cid:5)
(cid:6)
eV
2kBT
(cid:5) (cid:6)
dV
dI
I¼0
ð2Þ
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Fano factors and context for their interpretation. Fano factors found
from fitting the data in Fig. 3 are shown. Error bars are the standard error
from fitting 15 repeated bias sweep measurements. In a Fermi liquid, current is
carried by individual quasiparticle excitations, and the current as a function of
time fluctuates with the arrival of each discrete transmitted carrier. Carriers
scatter diffusively through static disorder (brown dots). When electron-electron
scattering is weak (sample length L < Lee), the expected Fano factor is F = 1/3
(green mark in the graph), whereas electron-phonon coupling can suppress
this at higher temperatures. When electron-electron scattering is strong (L > Lee),
p
ffiffiffi
the expected Fano factor is F ¼
=4 (blue mark in the graph). In a system
3
without well-defined quasiparticles, charge transport is more continuous,
which leads to suppressed current fluctuations; in the extreme limit, where
electronic excitations are entirely nondispersive, the Fano factor is expected to
vanish (red mark). Dashed lines are guides to the eye.
The fitted Fano factors of a YbRh2Si2 device
and a gold nanowire device are shown in
Fig. 4, which provides a direct comparison
between YbRh2Si2 and a Fermi liquid diffu-
sive wire. Corrections stemming from the non-
ohmic response lead to lower inferred Fano
factors (fig. S8).
p
ffiffiffi
3
Our detailed thermal modeling of the present
system under the standard Fermi liquid assump-
tions [see section 6 of (33)] confirms that, includ-
ing electronic thermal transport through the
Wiedemann-Franz relation and the measured
electron-phonon coupling, F ¼
=4 would
be expected in the present high-bias limit. This
is in sharp contrast to the experimental data
shown in Fig. 3A. Experiments on bulk YbRh2Si2
crystals do not show large deviations from the
Wiedemann-Franz relation in this temperature
range (46, 47). Electronic transport measure-
ments and thermodynamic measurements of
YbRh2Si2 in this temperature regime, as well
as THz optical conductivity measurements (31)
in these films, show that phonons are not con-
tributing strongly to the electronic properties
in YbRh2Si2 below 15 K. As shown in section 6 of
(33) and fig. S4, the measured electron-phonon
coupling in YbRh2Si2 is too small by more than
a factor of 35 to be responsible for the observed
noise suppression.
To interpret these results, it is important to
consider the nature of quasiparticles in terms
of the single-particle spectral function and
distribution functions. For a Fermi gas, the
single-particle spectral function A(k, D) at a
given wavevector k is a delta function in en-
ergy D at D = Ek, where Ek is the quasiparticle
energy as a function of k, meaning that a par-
ticle excitation at (k, Ek) in the zero tempera-
ture limit is perfectly well defined in energy and
has an infinite lifetime with a spectral weight
Z = 1. Correspondingly, the particle excitations
follow the FD distribution, and the Fermi sur-
face is a perfectly sharp boundary at T = 0. In
a Fermi liquid, the spectral function retains a
peak for k near the Fermi surface, which de-
scribes a quasiparticle with a nonzero spectral
weight Z < 1. The distribution function near
the Fermi surface is smeared but still has a
nonzero discontinuity at T = 0 K (48).
In the case of the particular type of non-
Fermi liquid with a complete destruction of
quasiparticles, one has Z = 0 everywhere on the
Fermi surface. With such a complete smearing
of the Fermi surface, there is no discontinuity
in the distribution function even at T = 0 K.
In this limit, when driven by a bias that does
not greatly perturb the non-FD distribution
function, there are no granular quasiparticles
that carry the electrical current. We can then
expect a much-reduced shot noise, as we observe
in the form of a Fano factor that is consider-
ably smaller than not only the strong electron-
electron scattering expectation F ¼
=4 but
even the weak electron-electron scattering
counterpart F = 1/3. We highlight this contrast
in Fig. 4 and discuss it further in section 13 of
(33). For reference, in the extreme case when
the electron spectral function at any given k as
a function of energy is entirely featureless, the
continuous electron fluid would have no shot
ffiffiffi
3
p
noise at all (F = 0). Interestingly, one approach to
a quantum critical system with no quasiparticles
(26, 27) predicts nonzero noise with a trend in
bias and temperature that is quantitatively sim-
ilar to that shown in Fig. 3A (as seen in fig. S10),
though that model is based on a different form of
quantum criticality (a superconductor-insulator
transition) than that in YbRh2Si2. This is dis-
cussed further in section 12 of (33).
Discussion and outlook
Shot noise is a probe that gives special access to
the nature of charge carriers. The suppressed
noise shown in Fig. 3A and summarized in Fig. 4
is evidence that current in this strange-metal
regime is not governed by the transport of
individual, granular quasiparticles. A Fano fac-
tor of zero is expected only for the most ex-
treme case of a non-Fermi liquid that has a
completely flat spectral function. A non-Fermi
liquid that still has residual dispersive spec-
tral features, despite a vanishing quasiparticle
weight Z, would lead to a nonzero Fano factor.
Any residual dispersive spectral features are
naturally expected to somewhat sharpen as
T → 0 K, leading to a rise in F as temperature is
lowered, but, in that case, would never reach the
F ¼
=4 expectation for a strongly correlated
Fermi liquid (where Z is finite). The present ex-
periment takes place firmly in the non-Fermi
liquid regime (down to nearly a factor of 10 be-
low the effective single-ion Kondo temperature)
that is already seen to exhibit critical scaling of
the optical conductivity (31). As discussed further
in section 13 of (33), an effective Fermi liquid
ffiffiffi
3
p
Chen et al., Science 382, 907–911 (2023)
24 November 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
finite temperature correction to the expected
Fano factor is not a naturally self-consistent
explanation for the trend in Fig. 4. Even so, the
present data are limited to 3 K and above; it is
desirable to extend our measurements to below
3 K, which would allow for a direct compari-
son with the T = 0 K theoretical expectations.
Although scattering techniques show the in-
coherent, nonquasiparticle electronic response
as a diffuse continuum across (k, D), shot noise
specifically targets the current-carrying excita-
tions. The shot noise probes both the equilib-
rium non-Fermi liquid distribution function
and its nonequilibrium evolution when per-
turbed by the difference in source and drain
chemical potentials. In Fermi liquids, this
approach has provided insights into inelastic
electron-electron scattering and the evolution
of the nonequilibrium distribution function.
This technique will provide crucial experimen-
tal constraints on such processes in materials
in the strange-metal regime. Moreover, strange
metallicity as inferred from the resistivity is ob-
served across many systems with quite disparate
underlying microscopic physics (2–8, 10, 28).
Shot noise provides an opportunity to test the
extent to which these phenomenologically
similar strange metals can fit within a single
paradigm. We expect our work to trigger ex-
tensive further theoretical studies.
RE FE RENCES AND N OT ES
1. C. M. Varma, Rev. Mod. Phys. 92, 031001 (2020).
2. A. Legros et al., Nat. Phys. 15, 142–147 (2019).
3. R. L. Greene, P. R. Mandal, N. R. Poniatowski, T. Sarkar,
Annu. Rev. Condens. Matter Phys. 11, 213–229
(2020).
J. Ayres et al., Nature 595, 661–666 (2021).
I. M. Hayes et al., Nat. Phys. 12, 916–919 (2016).
J. A. N. Bruin, H. Sakai, R. S. Perry, A. P. Mackenzie, Science
339, 804–807 (2013).
4.
5.
6.
7. Hv. Löhneysen et al., Phys. Rev. Lett. 72, 3262–3265 (1994).
8. T. Park et al., Nature 440, 65–68 (2006).
9. D. H. Nguyen et al., Nat. Commun. 12, 4341 (2021).
10. Y. Cao et al., Phys. Rev. Lett. 124, 076801 (2020).
11. P. Coleman, A. J. Schofield, Nature 433, 226–229 (2005).
12. J. A. Hertz, Phys. Rev. B 14, 1165–1184 (1976).
13. A. J. Millis, Phys. Rev. B 48, 7183–7196 (1993).
14. Q. Si, S. Rabello, K. Ingersent, J. L. Smith, Nature 413,
804–808 (2001).
48. D. Pines, P. Nozieres, The Theory of Quantum Liquids Volume I:
Normal Fermi Liquids (Basic Books, 1989).
49. L. Chen et al., Data and code for: Shot noise in a strange metal.
15. P. Coleman, C. Pépin, Q. Si, R. Ramazashvili, J. Phys. Condens.
Zenodo (2023); https://doi.org/10.5281/7800018.
Matter 13, R723–R738 (2001).
16. T. Senthil, M. Vojta, S. Sachdev, Phys. Rev. B 69, 035111
AC KNOWLED GME NTS
(2004).
17. R. A. Davison, K. Schalm, J. Zaanen, Phys. Rev. B 89, 245116
(2014).
18. S. A. Hartnoll, A. Lucas, S. Sachdev, Holographic Quantum
Matter (MIT Press, 2018).
19. K. Kobayashi, M. Hashisaka, J. Phys. Soc. Jpn. 90, 102001
(2021).
20. R. de-Picciotto et al., Nature 389, 162–164 (1997).
21. L. Saminadayar, D. C. Glattli, Y. Jin, B. Etienne, Phys. Rev. Lett.
79, 2526–2529 (1997).
22. T. Delattre et al., Nat. Phys. 5, 208–212 (2009).
23. O. Zarchin, M. Zaffalon, M. Heiblum, D. Mahalu, V. Umansky,
Phys. Rev. B 77, 241303 (2008).
24. P. Zhou et al., Nature 572, 493–496 (2019).
25. K. M. Bastiaans et al., Science 374, 608–611 (2021).
26. A. G. Green, J. E. Moore, S. L. Sondhi, A. Vishwanath,
Phys. Rev. Lett. 97, 227003 (2006).
27. J. Sonner, A. G. Green, Phys. Rev. Lett. 109, 091601
(2012).
28. J. Custers et al., Nature 424, 524–527 (2003).
29. S. Paschen et al., Nature 432, 881–885 (2004).
30. P. Gegenwart et al., Phys. Rev. Lett. 89, 056402
(2002).
31. L. Prochaska et al., Science 367, 285–288 (2020).
32. E. Bakali et al., J. Cryst. Growth 595, 126804 (2022).
33. Materials and methods are available as supplementary
materials.
34. M. Henny, S. Oberholzer, C. Strunk, C. Schönenberger,
Phys. Rev. B 59, 2871–2880 (1999).
35. C. W. J. Beenakker, M. Büttiker, Phys. Rev. B 46, 1889–1892
(1992).
36. K. E. Nagaev, Phys. Lett. A 169, 103–107 (1992).
37. B. L. Altshuler, L. S. Levitov, A. Y. Yakovets, JETP Lett. 59,
857–863 (1994).
38. Y. V. Nazarov, Phys. Rev. Lett. 73, 134–137 (1994).
39. D. B. Gutman, Y. Gefen, Phys. Rev. B 64, 205317
(2001).
40. V. I. Kozub, A. M. Rudin, Phys. Rev. B 52, 7853–7856
(1995).
41. K. E. Nagaev, Phys. Rev. B 52, 4740–4743 (1995).
42. M. Henny et al., Appl. Phys. Lett. 71, 773–775 (1997).
43. F. Liefrink, J. I. Dijkhuis, M. J. M. de Jong, L. W. Molenkamp,
H. van Houten, Phys. Rev. B 49, 14066–14069 (1994).
44. A. H. Steinbach, J. M. Martinis, M. H. Devoret, Phys. Rev. Lett.
76, 3806–3809 (1996).
45. Y. Wang et al., Shot noise as a characterization of strongly
correlated metals. arxiv:2211.11735 [cond-mat.str-el]
(2022).
46. H. Pfau et al., Nature 484, 493–497 (2012).
47. M. Taupin et al., Phys. Rev. Lett. 115, 046402 (2015).
We acknowledge helpful conversations with M. Foster and A. Lucas.
Funding: This work was funded by US Department of Energy, Basic
Energy Sciences, Experimental Condensed Matter Physics award
DE-FG02-06ER46337 (D.N., L.C., D.T.L.); NSF award DMR-1704264
(D.N.); European Research Council ERC Advanced Grant
101055088 (G.E., S.P.); Austrian Science Fund FWF I4047 (E.B.,
S.P.); Austrian Research Promotion Agency FFG 2156529 (S.P.);
Austrian Science Fund FWF SFB F 86 (G.E., S.P.); European
Union Horizon 2020 grant agreement 824109-EMP (E.B., G.E., L.P.,
M.W., R.S., S.P.); Austrian Research Promotion Agency FFG 883941
(W.S., A.M.A.); US Air Force Office of Scientific Research
FA8665-22-1-7170 (W.S., A.M.A.); NSF award DMR-2220603
(Y.W., C.S., S.S., Q.S.); Robert A. Welch Foundation Grant C-1411
(Y.W., C.S., S.S., Q.S.); and Vannevar Bush Faculty Fellowship
ONR-VB N00014-23-1-2870 (Q.S.). Some of the noise
measurement hardware was acquired through NSF award
DMR-1704264. Author contributions: Growth and characterization
of the YbRh2Si2 films: E.B., W.S., M.W., R.S., G.E., L.P., A.M.A.,
S.P.; Patterning, device fabrication, noise measurements: L.C., D.T.L.;
Noise data analysis and interpretation: L.C., D.N., Q.S., S.P.;
Theoretical analysis of Fermi liquid effects: Y.W., C.S., S.S., Q.S.;
Critical insights on heavy fermion strange metals: Q.S., S.P.;
Writing – initial draft: L.C., D.N., Q.S., S.P.; Writing – editing:
L.C., D.T.L., E.B., W.S., M.W., R.S., G.E., L.P., A.M.A., Y.W.,
C.S., S.S., Q.S., S.P. Competing interests: The authors declare
no competing financial interests. Data and materials
availability: All data presented in this paper and the
supplementary materials and the relevant analysis code are
available through Zenodo (49). License information: Copyright
© 2023 the authors, some rights reserved; exclusive licensee
American Association for the Advancement of Science. No claim
to original US government works. https://www.science.org/
about/science-licenses-journal-article-reuse. This research was
funded in whole or in part by the Austrian Science Fund (F 86), a
cOAlition S organization. The author will make the Author
Accepted Manuscript (AAM) version available under a CC BY
public copyright license.
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abq6100
Materials and Methods
Supplementary Text
Figs. S1 to S11
References (50–52)
Submitted 19 April 2022; resubmitted 5 December 2022
Accepted 12 October 2023
10.1126/science.abq6100
Chen et al., Science 382, 907–911 (2023)
24 November 2023
5 of 5
|
10.1126_science.abp9979
|
RES EARCH
SUPERCONDUCTIVITY
Superconducting vortices carrying a temperature-
dependent fraction of the flux quantum
Yusuke Iguchi1,2*, Ruby A. Shi1,2,3, Kunihiro Kihou4, Chul-Ho Lee4, Mats Barkman5,
Andrea L. Benfenati5, Vadim Grinenko6, Egor Babaev5, Kathryn A. Moler1,2,3,7
Magnetic field penetrates type-II bulk superconductors by forming quantum vortices that enclose a
magnetic flux equal to the magnetic flux quantum. The flux quantum is a universal quantity that depends
only on fundamental constants. In this study, we investigated isolated vortices in the hole-overdoped
Ba1−xKxFe2As2 (x = 0.77) by using scanning superconducting quantum interference device (SQUID)
magnetometry. In many locations, we observed objects that carried only part of a flux quantum, with a
magnitude that varied continuously with temperature. We demonstrated mobility and manipulability of
these objects and interpreted them as quantum vortices with nonuniversally quantized (fractional)
magnetic flux whose magnitude is determined by the temperature-dependent parameters of a
multicomponent superconductor.
A classical electrically conducting fluid al-
lows the creation of arbitrary vortices.
By contrast, the fundamental property
of an ordinary superconductor is the
quantization of vorticity (1). This implies
the quantization of the magnetic flux (2) that
these vortices induce. The magnetic flux F
(2–4) is defined as the integral of magnetic
field B over the surface of the sample F ¼
∫dxdyB (cid:2) n^ , where n^ is a unit vector normal
to the surface. This quantization means that
magnetic fields inside a superconductor can
change only in discrete steps: An externally
applied magnetic field penetrates a type-II
superconductor in increments of F0 associated
with the entrance of quantum vortices from
the boundary of the sample (5). The magnetic
flux quantum F0 = hc/2e (in CGS units) de-
pends only on fundamental constants: the
electron charge e, the speed of light c, and the
Planck constant h. Such quantum vortices
are mobile objects that can move through the
superconductor subject to thermal excitation,
a pinning landscape, or forces applied by cur-
rents or electromagnetic fields. Vortex stabil-
ity is dictated by a nontrivial winding of the
phase q of the complex superconducting order
parameter. The phase changes by 2p around
the vortex core (1).
Despite the diversity of superconducting ma-
terials, vortices carry an integer number (usu-
ally one) of flux quanta, irrespective of the other
properties of superconducting materials such
as composition, temperature, degree of purity,
1Geballe Laboratory for Advanced Materials, Stanford
University, Stanford, CA 94305, USA. 2Stanford Institute for
Materials and Energy Sciences, SLAC National Accelerator
Laboratory, Menlo Park, CA 94025, USA. 3Department
of Physics, Stanford University, Stanford, CA 94305, USA.
4National Institute of Advanced Industrial Science and
Technology (AIST), Tsukuba, Ibaraki 305-8568, Japan.
5Department of Physics, Royal Institute of Technology, SE-106
91 Stockholm, Sweden. 6Tsung-Dao Lee Institute, Shanghai
Jiao Tong University, Shanghai 201210, China. 7Department of
Applied Physics, Stanford University, Stanford, CA 94305, USA.
*Corresponding author. Email: yiguchi@stanford.edu
etc. The experimental evidence for flux quanti-
zation comes from a number of measurements,
which in addition to scanning superconduct-
ing quantum interference device (SQUID) mag-
netometry include scanning Hall probes (6)
and magnetic force microscopy. Other probes
that do not measure magnetic flux directly but
allow assessment of the scaling of vortex den-
sity with applied field, such as Bitter decora-
tion (7) and scanning tunneling microscopy
(8), are also consistent with integer flux quan-
tization. Under certain conditions, immobile
objects carrying a half of the flux quantum
were observed in an intrinsic Josephson junc-
tion formed on grain boundaries of d-wave
superconductors (9) and on small rings of
Sr2RuO4 (10) and b-Bi2Pd (11). In supercon-
ductors consisting of almost decoupled layers
(12) as well as artificial layered structures (13),
the vortex core can move a great distance be-
tween consecutive layers, leading to partial
flux despite the standard integer phase quan-
tization of the phase windings in each indi-
vidual layer. Here we report scanning SQUID
magnetometry on Ba1−xKxFe2As2 (x = 0.77).
We show that, along with the standard vortices
carrying single flux quantum, the material
has vortex excitations that carry a fraction of
the flux quantum. Notably, the fraction is not
a rational number but instead smoothly varies
with temperature, and hence the flux quanti-
zation is nonuniversal, meaning that it is not a
function of only fundamental constants.
Temperature-dependent fractional vortex flux
To explore the magnetic properties of
Ba1−xKxFe2As2, we used the scanning SQUID
susceptometer to microscopically image the
magnetic flux on the cleaved ab plane of single
crystals (Fig. 1, A and B) (14, 15). We conducted
measurements below the superconducting crit-
ical temperature Tc (fig. S1). By cooling the
sample in a small uniform magnetic field of
3.5 mG, we observed that below Tc, the sys-
tem formed dilute configurations of conven-
tional vortices that coexisted with isolated
vortex-like objects carrying a small fraction of
magnetic flux (fig. S2). To determine whether
these objects are vortices and to estimate the
total flux trapped within them, further mea-
surements were performed. By cooling through
Tc with a small local magnetic field from the
scanning SQUID field coil, we also observed
the conventional vortices and the fractional
vortex–like objects (Fig. 1, C to E, and fig. S3).
We observed both the conventional vortex and
the fractional vortex–like objects appearing at
the same location in different cooling cycles,
where the sample was cooled down to 10.5 K
from 25 K with magnetic fields (Fig. 1, C and
E). Hence the specific area in the sample did
not determine the character of a flux-carrying
object. To characterize further the nature of
the object, a fractional antivortex object was
created at the same position during different
cooling cycles (Fig. 1D).
To obtain the simplest estimate for the mag-
netic field penetration depth l, we simulated
the magnetic field of the conventional single-
quantum vortex with a point source magnetic
monopole field with the total flux F0 (16). We
roughly estimated the magnetic field pene-
tration depth l from this point source mod-
el. We fitted it with the spacing between the
sample surface and the pickup loop center, z0 =
1250 nm (Fig. 1H), yielding a magnetic field
penetration depth l ≈ 2.3 mm. A fitting of the
vortex object in Fig. 1C with l ≈ 2.3 mm and
the fractional point source yielded the frac-
tion of the flux quantum FF ≈ 0.3F0 at T =
11.0 K (Fig. 1F).
By measuring the magnetic flux at different
temperatures, we observed that the flux of this
object was temperature dependent. The mag-
netic field amplitude was also very different
from that of a conventional vortex and half-
quantum objects that form on grain boun-
daries of a d-wave superconductor (17). We
observed that the amplitude of the peak mag-
netic flux from the fractional vortex–like ob-
ject decreased with decreasing temperature
(Fig. 2A). This temperature-dependent flux is
in strong contrast to the conventional vortex. The
usual vortex carries temperature-independent
single flux quantum. Hence the peak amplitude
of magnetic field increases with decreasing tem-
perature, reflecting the temperature dependence
of the magnetic field penetration depth (Fig. 2A).
To estimate the temperature dependence of
the fractional flux FF, we fitted the cross sections
of the fractional vortex–like object at several
temperatures (Fig. 2, B and C), the same way as
in Fig. 1, F and G. For the fitting of the fractional
vortex–like object, we used the same penetra-
tion depth obtained from the fitting of the
conventional vortex at the same region. We
observed fractional vortex–like objects in dif-
ferent areas of the sample (figs. S4 and S5).
Iguchi et al., Science 380, 1244–1247 (2023)
23 June 2023
1 of 4
RES EARCH | R E S E A R C H A R T I C L E
A
]
m
µ
[
y
Ba0.23K0.77Fe2As2
0
4
3
2
1
C
40
Fractional vortex
0
] 20
Φ
m
[
Φ
0
-100
B
-100
x [µm]
0
-20
-90
-70
x [µm]
-40
-30
y [µm]
Pickup
loop
Field coil
shield
1 µm
F
40
0.3Φ0 point source model
0
] 20
Φ
m
[
Φ
0
-20
-10
0
x [µm]
10
-10 0 10
y [µm]
D
]
0
Φ
m
[
Φ
40
20
0
-20
G
]
0
Φ
m
[
Φ
40
20
0
-20
Fractional vortex
-90
-70
x [µm]
-40
-30
y [µm]
-0.3Φ0 point source model
-10
0
]mµ[ x
10
-10 0 10
y [µm]
E
40
20
0
]
0
Φ
m
[
Φ
-20
-90
H
]
0
Φ
m
[
Φ
40
20
0
-20
Conventional
vortex
region 2
T = 11.0 K
Φ [mΦ0]
45
-70
x [µm]
-40
-30
y [µm]
-45
Φ0 point source
model
Simulation
λ = 2.3 µm
Φ [mΦ0]
45
10
-10
0
]mµ[ x
10
-10 0
y [µm]
-45
Fig. 1. Scanning SQUID imaging fractional vortices and conventional
vortices in the same area of Ba1−xKxFe2As2. (A) Optical image of the sample
with scan regions 1 to 4. (B) Pickup loop and field coil of the SQUID
susceptometer are covered with superconducting shields, except for the loop
area to detect local magnetic flux. (C to E) SQUID measurements. Isolated
(C) fractional vortex carrying ~0.3 of the flux quantum, (D) fractional antivortex,
and (E) conventional vortex appear in different cooling cycles in the same area
at 11 K after local field (C) 6 mA, (D) −6 mA, and (E) 1 mA cooling to 10.5 K
from 25 K and heating to 11 K. (D) The conventional vortex was moved
to this location by applying the local field at 10.8 K in the same way as in Fig. 4D.
(F to H) Simulations of (F) 0.3 F0, (G) −0.3 F0, and (H) F0 point source
models show similarity to the data in (C) to (E), respectively.
Fractional vortex
region 2
Conventional vortex
region 2
A
100
]
0
Φ
m
[
Φ
k
a
e
P
80
60
40
20
region 2
Conventional
vortex
Fractional
vortex
Tc
B
]
0
Φ
m
[
Φ
x
u
l
f
c
i
t
e
n
g
a
M
35
30
25
20
15
10
5
0
0
9
10
T [K]
11
-10
T [K]
11.0
10.0
9.0
0
10
-5
Cross section [µm]
5
C
300
200
]
0
Φ
m
[
Φ
100
T [K]
11.0
10.8
10.6
10.4
10.0
9.0
0
-10
0
-5
10
Cross section [µm]
5
Fig. 2. Temperature dependence of the magnetic field of a fractional versus conventional vortex.
(A) Comparison of the temperature dependence of the maximum flux through the pickup loop with the
loop above the centers of the fractional vortex and the conventional vortex. The measurement indicates that
the carried fraction of the flux quantum drops continuously as temperature decreases. (B and C) Measured
cross sections of (B) the fractional vortex and (C) the conventional vortex along the x axis, where the
y position is indicated by horizontal arrows in Fig. 1, C and E. Solid lines are fitting results of point source
models. Note that London model–based ansatz for fractional vortices can overestimate the field at the
vortex center and underestimate the field at the vortex tail owing to nonlinear effects important for fractions
other than one-half (18).
Alternatively, we obtained the fractionally
quantized flux in the fractional vortices by
fitting a fractional vortex with fitting parame-
ters of the penetration depth and the fraction
(fig. S9) or by integrating flux over a frac-
tional vortex without use of the penetration
depth (fig. S10). We also verified that the fit-
ting with the F0 point source model does not
work for the observed fractional vortices (fig.
S8). The obtained fraction FF/F0 showed a
similar temperature dependence in three
tested regions (Fig. 3), suggesting that these
objects are unconventionally quantized frac-
tional vortices.
Mobility of fractional vortex
The fraction of the flux quantum gradually
decreased with decreasing temperature, and
below T/Tc = 0.8, it could not be resolved
from the background noise. This temperature
dependence could be caused by the fraction
of carried flux further dropping to a very small
value at low temperatures and/or by the effect
of magnetic flux delocalization of fractional
vortices (18). When we cooled to 3 K, well be-
low the temperature where we can resolve
fractional vortex objects, then heated back
above 9 K, we observed the reemergence of
the fractional vortex objects with similar mag-
netic flux pinned at the same location (the
fractional vortex object in Figs. 1D and 2B is
the reemergent one; the initially created vortex
is not shown), but sometimes cooling and sub-
sequent heating resulted in disappearance of
the object from the scanned area.
To determine whether the observed vortex-
like object was a true quantum vortex, the mo-
bility of the object needed to be demonstrated.
We checked this by monitoring the positions
of these objects when we cooled or heated the
sample. We found that fractional vortex–like
objects sometimes randomly moved to other
positions (Fig. 4, A and B). The observed pro-
cess was akin to the mobility of conventional
vortices moving between pinning positions.
Next, we were able to manually change the
position of the fractional vortex–like objects
by applying a local repulsive force created by
the scanning SQUID field coil (Fig. 4, C and D).
We demonstrated that the scanning SQUID
manipulated both the fractional vortex–like
object and the conventional vortex at the same
time (Fig. 4D).
In superconductors with very weak inter-
layer coupling, magnetic probe manipulation
can break a vortex into two small magnetic
excitations that are described as a broken
stack of pancake vortices (12). To investigate
this scenario, we manipulated integer flux
vortices in a similar way but did not observe
such behavior, which is consistent with the
Iguchi et al., Science 380, 1244–1247 (2023)
23 June 2023
2 of 4
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Temperature and spatial
dependence of magnetic flux in a
vortex. (A) The total magnetic flux
FF measured in fractional vortices located
in three different regions is obtained
from fittings in Fig. 2B and fig. S4 and
shows similar temperature dependence
in all three regions. (B) Temperature
dependence of the penetration depth
obtained from the fittings in Fig. 2C and
fig. S5 is used to fit the profiles of the
fractional vortices in the same region.
Similar temperature-dependent penetra-
tion depth is obtained by the scanning
SQUID susceptibility measurements
on region 1 (fig. S1). The error bars in
the susceptibility measurements
represent the uncertainty of the SQUID
height, ~±0.1 mm. The Tc ~11.2 K is
obtained from the scanning SQUID
susceptometry (fig. S1).
A
0
/
F
Φ
n
o
i
t
c
a
r
F
B
]
m
µ
[
much weaker anisotropy of the upper critical
fields in our sample compared with strongly
layered systems (19). Moreover, some of the
fractional vortex–like objects appeared with-
out any other fractional vortex–like objects
within a 50-mm area (figs. S2 and S3). Finally,
the temperature dependence of the fractional
vortices flux is inconsistent with the pancake
vortex in layered systems (Fig. 2B). The ob-
served robustness and mobility established
that the observed objects represented quan-
tum vortices.
To ensure that the local variation of mag-
netic field penetration depth does not play a
large role, we created integer and fractional
vortices at the same positions during different
cooling cycles. Figure S7 demonstrates that the
two types of vortices have similar localization
but very different magnetic field amplitude.
We also observed homogeneous susceptibil-
ity and its monotonic increase at low temper-
ature over the whole scan area (fig. S6), which
ruled out the local variation of magnetic field
penetration depth scenario. We numerically
estimated the penetration depth from the
susceptibility at region 1 (20), which had a sim-
ilar temperature dependence to that obtained
from fitting vortex field (Fig. 3B). We did not
observe any magnetic defects above Tc.
Discussion
The material Ba1−xKxFe2As2 has a multiband
electronic structure (21, 22). Furthermore,
1.0
0.8
0.6
region 1, antivortex
region 2, vortex
region 2, antivortex
region 3, vortex
0.4
Fractional vortex
0.2
0
5
4
3
2
1
0
0.8
0.85
0.9
0.95
1.0
T/Tc
Susceptibility at region 1
region 1
region 2
region 3
Conventional vortex
0.8
0.85
0.9
0.95
1.0
T/Tc
previous experiments demonstrated that
Ba1−xKxFe2As2 spontaneously breaks the time-
reversal symmetry (23–25). Systems with mul-
tiple broken symmetries are described by
multicomponent order parameters. This sug-
gests that, in principle, there are necessary
degrees of freedom to support several types
of vortices associated with phase windings in
different components of the order parameter.
A multiband superconductor with bands la-
beled by a band index j is described by multi-
(cid:1)
(cid:1)
(cid:1)eiqj. In such a case, in
(cid:1)
ple complex fields yj ¼ yj
the simplest model, the current will be given by a
sum of contributions from different bands J ¼
(cid:1)
X
(cid:1)2A ,
ð
j ∇yj (cid:4) yj∇y(cid:3)
ð
jÞ (cid:4) 2e2=mc
eℏ=i2m
(cid:1)
(cid:1)
Þ yj
Þðy(cid:3)
j
where A is the vector potential, and m is the
mass parameter. It was discussed at the level
of effective fields theory that vortex excitations
are possible with phase winding in only one of
the bands (26); the above expression can then
be used to calculate the magnetic flux enclosed
in the resulting vortex. This can be done by
integrating the vector potential over a path s
located far away from vortex where J = 0.
That gives the magnetic flux of a vortex: F ¼
j2=
j2=
∮sA·dl ¼ ∮s∇q1·dl ℏc=e
Þ y
j
(cid:1)
X
1
1
(cid:1)2. The flux is thus no longer a function
of just fundamental physical constants; in-
stead, it depends on microscopic details and
temperature. This is because the ratios of the
(cid:1)
(cid:1) associated with superconducting
fields yj
components in different bands have, in general,
(cid:1)
(cid:1)
(cid:1)2 ¼ F0 y
(cid:1)
yj
j
j
(cid:1)
(cid:1)
j yj
X
(cid:1)
(cid:1)
ð
different temperature dependencies. In cer-
tain cases, multicomponent theories arise for
fields that are associated with linear combi-
nations of superconducting gaps in different
bands (27), but the argument remains rather
similar.
There are factors that can suppress, or in
some cases prohibit, the formation of frac-
tional vortices in multiband superconduc-
tors. In multicomponent superconductors, the
components are electrically charged. The un-
avoidable electromagnetic intercomponent in-
teraction tends to confine fractional vortices
into integer flux vortices in the bulk of a multi-
band superconductor (26). This is the principal
difference compared with multicomponent
superfluids where vortices carry no magnetic
flux and no such electromagnetic confine-
ment of vortices exists. Similarly, interband
interactions, such as interband Josephson
coupling, tend to lock phases in different bands,
making one-flux-quanta composite objects
the energetically cheapest type of vortex
(26). Hence, although fractional vortices are
theoretically possible in multiband mate-
rials, they are more energetically expensive
(26). Nonetheless, a minority fraction of frac-
tional vortices are theoretically allowed to
form in the bulk of a superconductor owing
to fluctuations, quenches, or pinning, akin
to the effect of remanent vorticity in super-
fluids. However, fractional vortices have not
been previously observed in multiband super-
conductors such as MgB2 (28–30) and pnictides
(31–33) nor in multiband materials where
time-reversal symmetry breaking was reported
(34–37).
This raises the question of what is special
about Ba1−xKxFe2As2 at x = 0.77. First, we note
that experiments indicate that Ba1−xKxFe2As2
is a so-called s + is (or similar) superconductor
(23–25, 38). This is a multiband state where
the time-reversal symmetry is broken. The pre-
cise microscopic model for superconducting
state in this material is not known. However,
the calculations in the framework of the mini-
mal phenomenological Ginzburg-Landau model
suggested that, while it is not a necessary
ingredient, under certain conditions, the broken
time-reversal symmetry in a multiband system
helps the formation of mobile fractional vorti-
ces in individual bands (39). Second, the dop-
ing x = 0.77, where we could create fractional
vortices on demand, corresponds to the spe-
cial point where the critical temperature of
the time-reversal symmetry breaking is max-
imal (24, 25). This maximum suggests that
the components are well developed, and hence
vortices carry a substantial fraction of mag-
netic flux quantum near the temperature of
the superconducting phase transition at x =
0.77. By contrast, using the same simple ap-
proach, we could not observe fractional vorti-
ces in the Ba1−xKxFe2As2 samples with different
Iguchi et al., Science 380, 1244–1247 (2023)
23 June 2023
3 of 4
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Mobility of the fractional vortex.
(A and B) The fractional vortex (FV) moves
while (A) cooling and (B) heating. This motion
happens randomly. (C and D) Scanning
SQUID susceptometer manipulates (C) the
FV and (D) both the FV and the conventional
vortex (CV) by applying a repulsive local
field. The schematic image of the field coil
indicates the location where the local field is
applied. Note that the peak fluxes of FV
in (B) and (C) at 11 K are ~13 mF0, from
which a fraction of ~0.3 is obtained. Vortices
were created by local field [(A) to (C)]
6 mA and (D) uniform field 3.5 mG cooling
to 10.5 K from 25 K. The “After 3rd”
image has two scans combined, separated
by a vertical black line.
doping levels (table S1). However, we cannot
exclude that fractional vortices can be cre-
ated at other doping levels by using different
approaches. Also, we cannot exclude that at
different doping levels there are fractional
vortices that carry too small a fraction of flux
quantum to be detected or too large a fraction
of the flux quantum to be distinguished from
single-quantum vortices.
We demonstrated the existence of stable
vortices with phase winding only in one of the
bands in a numerical solution of the three-band
Bogoliubov–de Gennes model. The solutions are
given in (40), using techniques presented in (41).
Although our solutions provide a fully micro-
scopic demonstration of the principle of stable
fractional vortices, further theoretical studies are
needed to connect the observation of fractional
vortices with concrete microscopic models of
Ba1−xKxFe2As2. Of special interest is the tem-
perature dependence of the enclosed fraction
of flux quantum, which contains information
that allows one to constrain microscopic mod-
els of pairing.
Our observations demonstrate that vortices
carrying a temperature-dependent fraction of
the flux quantum are possible in supercon-
ductors. Dynamics, control and manipulation,
the precise configuration of their magnetic
field, and the mechanisms of pinning of these
objects are promising directions that will give
insights into both the basic questions of super-
conductivity and the nature of superconduc-
tivity in Ba1−xKxFe2As2. They may also enable
the use of these objects in fluxonics-based cryo-
genic computing (42, 43).
RE FERENCES AND NOTES
L. Onsager, Nuovo Cim. 6 (suppl. 2), 279–287 (1949).
1.
2. F. London, Phys. Rev. 74, 562–573 (1948).
3. F. London, Superfluids, Volume I: Macroscopic Theory of
Superconductivity (Dover Publications, ed. 2, 1960)
4. L. Onsager, Phys. Rev. Lett. 7, 50 (1961).
5. A. A. Abrikosov, Sov. Phys. JETP 5, 1174 (1957).
6. A. M. Chang et al., Appl. Phys. Lett. 61, 1974–1976 (1992).
7. P. L. Gammel et al., Phys. Rev. Lett. 59, 2592–2595 (1987).
8. H. F. Hess, R. B. Robinson, R. C. Dynes, J. M. Valles Jr.,
J. V. Waszczak, Phys. Rev. Lett. 62, 214–216 (1989).
J. R. Kirtley et al., Phys. Rev. Lett. 76, 1336–1339 (1996).
9.
10. J. Jang et al., Science 331, 186–188 (2011).
11. Y. Li, X. Xu, M.-H. Lee, M.-W. Chu, C. L. Chien, Science 366,
238–241 (2019).
12. L. Luan et al., Phys. Rev. B 79, 214530 (2009).
13. H. Bluhm, N. C. Koshnick, M. E. Huber, K. A. Moler, Phys. Rev.
Lett. 97, 237002 (2006).
14. K. Kihou et al., J. Phys. Soc. Jpn. 85, 034718 (2016).
15. J. R. Kirtley et al., Rev. Sci. Instrum. 87, 093702 (2016).
16. J. R. Kirtley et al., Supercond. Sci. Technol. 29, 124001 (2016).
17. J. R. Kirtley, C. C. Tsuei, K. A. Moler, Science 285, 1373–1375
(1999).
18. E. Babaev, J. Jäykkä, M. Speight, Phys. Rev. Lett. 103, 237002 (2009).
19. Y. Liu et al., Phys. Rev. B 89, 134504 (2014).
20. J. R. Kirtley et al., Phys. Rev. B 85, 224518 (2012).
21. S. Maiti, A. V. Chubukov, Phys. Rev. B 87, 144511 (2013).
22. J. Böker, P. A. Volkov, K. B. Efetov, I. Eremin, Phys. Rev. B 96,
014517 (2017).
23. V. Grinenko et al., Phys. Rev. B 95, 214511 (2017).
24. V. Grinenko et al., Nat. Phys. 16, 789–794 (2020).
25. V. Grinenko et al., Nat. Phys. 17, 1254–1259 (2021).
26. E. Babaev, Phys. Rev. Lett. 89, 067001 (2002).
27. J. Garaud, M. Silaev, E. Babaev, Physica C 533, 63–73 (2017).
28. M. R. Eskildsen et al., Phys. Rev. Lett. 89, 187003 (2002).
29. T. Nishio et al., Phys. Rev. B 81, 020506(R) (2010).
30. J. Gutierrez et al., Phys. Rev. B 85, 094511 (2012).
31. L. J. Li, T. Nishio, Z. A. Xu, V. V. Moshchalkov, Phys. Rev. B 83,
224522 (2011).
32. B. Kalisky et al., Phys. Rev. B 83, 064511 (2011).
33. I. P. Zhang et al., Phys. Rev. B 100, 024514 (2019).
34. P. G. Björnsson, Y. Maeno, M. E. Huber, K. A. Moler, Phys. Rev. B
72, 012504 (2005).
35. J. R. Kirtley et al., Phys. Rev. B 76, 014526 (2007).
36. C. W. Hicks et al., Phys. Rev. B 81, 214501 (2010).
37. Y. Iguchi et al., Phys. Rev. B 103, L220503 (2021).
38. V. L. Vadimov, M. A. Silaev, Phys. Rev. B 98, 104504 (2018).
39. J. Garaud, E. Babaev, Phys. Rev. Lett. 112, 017003 (2014).
40. See supplementary materials.
41. A. Benfenati, M. Barkman, E. Babaev, Phys. Rev. B 107, 094503
(2023).
42. K. Miyahara, M. Mukaida, K. Hohkawa, Appl. Phys. Lett. 47,
754–756 (1985).
43. T. Golod, A. Iovan, V. M. Krasnov, Nat. Commun. 6, 8628 (2015).
44. Y. Iguchi et al., Data for “Observation of superconducting
vortices carrying a temperature-dependent fraction of the flux
quantum,” version 1, Zenodo (2023); https://doi.org/10.5281/
zenodo.7644830.
AC KNOWLED GME NTS
The authors thank N. Nandi and J. Garaud for fruitful discussions.
Funding: This work was primarily supported by the Department of
Energy, Office of Science, Basic Energy Sciences, Materials Sciences
and Engineering Division, under contract no. DE- AC02-76SF00515.
E.B. was supported by Swedish Research Council grants 2016-06122
and 2022-04763 and by the Wallenberg Initiative Materials Science
for Sustainability (WISE) funded by the Knut and Alice Wallenberg
Foundation. Author contributions: Y.I. carried out the scanning
SQUID microscopy, analyzed data, and wrote the manuscript. R.A.S.
carried out the scanning SQUID microscopy. K.K. and C.-H.L.
synthesized the crystals. M.B. and A.L.B. wrote BdG code and carried out
and analyzed the BdG calculations. V.G. supervised the project and
selected and characterized the samples. E.B. conceptualized, contributed
to writing, and supervised the project. K.A.M. supervised the project. All
the authors discussed the results and implications and commented
on the manuscript. Competing interests: The authors declare no
competing financial interests. Data and materials availability: All data
needed to evaluate the conclusions in the paper are present in the
paper and/or the supplementary materials and available in Zenodo
(44). License information: Copyright © 2023 the authors, some
rights reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government works.
https://www.science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abp9979
Materials and Methods
Figs. S1 to S11
Table S1
References
Submitted 10 March 2022; resubmitted 26 August 2022
Accepted 15 May 2023
Published online 1 June 2023
10.1126/science.abp9979
Iguchi et al., Science 380, 1244–1247 (2023)
23 June 2023
4 of 4
|
10.1126_science.abq6753
|
RES EARCH
QUANTUM GASES
Universal scaling of the dynamic BKT transition in
quenched 2D Bose gases
Shinichi Sunami1*, Vijay Pal Singh2,3, David Garrick1, Abel Beregi1, Adam J. Barker1, Kathrin Luksch1,
Elliot Bentine1, Ludwig Mathey4,5, Christopher J. Foot1
The understanding of nonequilibrium dynamics in many-body quantum systems is a fundamental issue in
statistical physics. Experiments that probe universal properties of these systems can address such
foundational questions. In this study, we report the measurement of universal dynamics triggered by a
quench from the superfluid to normal phase across the Berezinskii-Kosterlitz-Thouless transition in a
two-dimensional (2D) Bose gas. We reduced the density by splitting the 2D gas in two, realizing a quench
across the critical point. The subsequent relaxation dynamics were probed with matter-wave
interferometry to measure the local phase fluctuations. We show that the time evolution of both the
phase correlation function and vortex density obeys universal scaling laws. This conclusion is supported
by classical-field simulations and interpreted by means of real-time renormalization group theory.
the treatments of nonequilibrium dynamics
[e.g., (11)], and this motivates in-depth experi-
mental studies to guide and test theories.
For this purpose, ultracold gases have emerged
as a platform of unprecedented control and
tunability, serving as quantum simulators for
the investigation of many-body dynamics.
This approach has led to the observation of
Kibble-Zurek (KZ) scaling (12–15) and univer-
sal scaling laws (16–18) after a quench. Despite
theoretical interest (19–21), universal critical
dynamics in nonequilibrium continuous two-
dimensional (2D) quantum gases remain elusive
because of the lack of precise experimental
probes. We show how fluctuations of the many-
body 2D system can be probed directly through
the extension of local matter-wave interferometry
(22) similar to that previously used to probe the
local phase fluctuations of near-integrable 1D
T he relaxation dynamics of a many-body
system that is quenched out of equilib-
rium display a wide range of scenarios,
from simple exponential decay to relax-
ation through metastable or prethermal-
ized states (1, 2), including phenomena such as
pattern formation (3), and the absence of ther-
malization (4). Systems that are quenched across
a phase transition are particularly intriguing be-
cause of their universal self-similar behavior, ex-
pected in systems as diverse as superfluid helium
(5), liquid crystals (6), biological cell membranes
(7), the early universe (8), and cold atoms (9, 10).
There are numerous theoretical challenges in
1Clarendon Laboratory, University of Oxford, Oxford OX1
3PU, UK. 2Institut für Theoretische Physik, Leibniz
Universität Hannover, 30167 Hannover, Germany 3Quantum
Research Centre, Technology Innovation Institute, Abu Dhabi,
United Arab Emirates. 4Zentrum für Optische Quantentechnologien
and Institut für Laserphysik, Universität Hamburg, 22761
Hamburg, Germany. 5The Hamburg Centre for Ultrafast
Imaging, Hamburg 22761, Germany.
*Corresponding author. Email: shinichi.sunami@physics.ox.ac.uk
Fig. 1. Observation
of nonequilibrium
dynamics in 2D
Bose gases with
matter-wave inter-
ferometry. (A) A 2D
superfluid is split
into two daughter
clouds, thereby
quenching through
the BKT transition.
The two clouds
evolve for time t and
are released to pro-
duce matter-wave
interference after a
TOF. Local phase
fluctuations are
observed by opti-
cally pumping the
slice (red sheet) and
then performing
absorption imaging.
(B) Equilibrium
phase diagram of
trapped 2D Bose
gases (27). The
quench forces the
system out of equi-
librium toward the
normal phase. (C) (Top) Examples of interference images. Phase dislocation
caused by a vortex is visible in the image at 510 ms. (Bottom) The histograms
show the phase differences, Df ¼ f xð Þ (cid:2) f x′ð Þ, at x (cid:2) x′
45 experimental runs. The decreasing height and increasing width indicate increased
Þ (continuous lines) in
phase fluctuations. (D) (Left) Free energies Fi nvð
equilibrium for the initial and final conditions of the quench, with their minima
j ¼ 5 mm, from
Þ and Ff nvð
j
indicated by red points (38). After the quench, the system undergoes nonequilibrium
dynamics (dashed lines) and relaxes toward the state with nonzero free-vortex
density, nv. (Right) Illustration of the dynamics, showing the transition between
a scale-invariant phase supported by bound vortex-antivortex pairs and the broken
scale-invariant phase characterized by free vortices with the mean vortex-vortex
distance Lv, where tc is the crossover time.
Sunami et al., Science 382, 443–447 (2023)
27 October 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Nonequilibrium scaling dynamics of the correlation function.
(A) Relaxation dynamics of the phase correlation function C (cid:1)xð Þ measured after
the quench for the initial reduced temperature,
average over 45 realizations or more, and the error bars denote standard error.
The data immediately after the critical time, tc, are indicated by the red curve.
For comparison, the measured data for equilibrium 2D gases (22) are shown by
the gray curves:
in the normal regime. (B) Linear scaling of length x′ ¼ (cid:1)xt=tc, according to time t=tc,
~
T ¼ 0:41 (top) in the superfluid regime and
~
T ¼ 0:34, where C (cid:1)xð Þ is an
~
T ¼ 0:61 (bottom)
Fig. 3. Universal behavior
across the dynamic BKT
transition. (A) Measured
algebraic exponent h tð Þ
after the quench at differ-
results in a collapse toward a common curve for time evolution up to t (cid:3) 1 s. This
universal function is compatible with the expected behavior at the crossover in
equilibrium (red dashed line), including the effect of inhomogeneity (27). At long
times, deviations are observed (green points), indicating the breaking of scale
invariance. (Inset) The c2 values for the algebraic and exponential fit functions
(27). (C) Rescaling of the distance ~x ¼ (cid:1)x
nv tð Þ, results in a collapse of the curves for times t > 1 s. Scaling behavior is
compatible with an exponential decay (black solid line). (Inset) nv tð Þ.
, according to the vortex density
ffiffiffiffiffiffiffiffiffiffi
nv tð Þ
p
~
T (filled circles),
ent initial
where the equilibrium critical
~
Tc ¼ 0:53, is
temperature,
crossed by the quench
(22). Open circles indicate
the corresponding simu-
lation results (27). Solid
lines indicate linear fits to
the experimental data, and
error bars denote standard
fit errors. (B) Time evolu-
tion of the measured (filled
circles) and simulated
(open circles) vortex den-
sity, nv tð Þ. The solid lines
are power-law fits to the
experimental data. Error
bars are statistical, given
by the square root of
observed vortex number.
(C) Time evolution h t=tc
Þ,
~
scaled according to the
T-dependent crossover time, tc. The horizontal error bars arise from the uncertainty in tc. This universal behavior is used to determine the critical
exponent, hc
(D) Scaled time evolution nv t=tc
(Inset) Dependence of tc on
Þ, plotted on a log-log scale, displays a universal growth after tc. Black solid line is the fit with power law nvºt2n, which yields n ¼ 1:1 1ð Þ.
¼ 0:13 1ð Þ (horizontal dotted line), at t=tc ¼ 1. The gray shaded curve indicates the simulation result at
ð
~
T, with a solid line as a guide for the eye.
~
T (cid:3) 0:4, and the solid line is a guide for the eye.
ð
systems (1, 2). A feature of 2D systems is that,
unlike 1D systems, they exhibit phase transi-
tions with associated critical phenomena. An
especially interesting case is that of the crit-
ical dynamics across the Berezinskii-Kosterlitz-
Thouless (BKT) transition (23–25), when the
system is quenched from the superfluid to
the normal phase, the opposite of commonly
used quenches that move from the disordered
to ordered phase. Real-time renormalization-
group (RG) theory and truncated Wigner sim-
ulations (19) predict that the relaxation occurs
Sunami et al., Science 382, 443–447 (2023)
27 October 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
Þ
Fig. 4. Real-time renormalization
group flow and measurements. Gray
lines with arrows are the RG flow of
the parameters x ¼ 1= 2hð
Þ (cid:2) 1= 2hc
ð
ffiffiffi
p
and y ¼
pg [Eqs. 1 and 2 and (27)].
2
These results are compared with the
experimental data for six different initial
temperatures, with the time scaled by the
corresponding
time, tc. Error bars denote standard errors
propagated from the values for h and nv.
The results from a numerical simulation
~
T ¼ 0:47 are shown as black crosses,
at
where the deviation from experimental data
at small x is attributed to the slow trap-
induced heating (27).
~
T-dependent crossover
through a reverse-Kibble-Zurek–type mecha-
nism in which delayed vortex proliferation
results in a metastable supercritical phase.
For equilibrium systems, the BKT transition
is driven by the unbinding of vortex-antivortex
pairs (22, 26), underscoring the topological
nature of the transition. This transition is
characterized by a change of the functional
form of the correlation function from a power
Þ ¼
law deep in the superfluid regime, g1 r; r′
hY rð Þ†
Y r′ð Þiº r (cid:2) r′j(cid:2)h
j
, to exponential deep
j=r0,
in the normal regime, g1 r; r′
where Y rð Þ is the bosonic field operator at
position vector r, h is the algebraic exponent,
and r0 is a correlation length. A similar change
of the functional form of the single-particle
correlation function is also expected for the
dynamical BKT transition. However, this change
is expected to occur as a smooth crossover be-
tween the two phases (19, 27).
Þ º e(cid:2) r(cid:2)r′
ð
ð
j
Here, we studied the critical dynamics across
the BKT transition by quenching a 2D Bose
gas from the superfluid to the normal phase
by splitting it in two. Using spatially resolved
matter-wave interferometry, we measured the
first-order correlation function and vortex
density to analyze their relaxation dynamics.
We found that relaxation occurs through a
two-step process involving phonon relaxation
followed by dynamical vortex proliferation.
We demonstrated universal scaling laws for
the algebraic exponent and vortex density by
performing measurements at different initial
conditions. Both real-time RG theory (19, 28)
and classical-field simulations are in good
agreement with our measurements.
Matter-wave interferometry after a quench
Our experiments began with a single, pancake-
shaped quasi-2D Bose gas in the superfluid re-
gime, consisting of N ≈ 9 (cid:4) 104 atoms of 87Rb
at reduced temperatures in the range ~T ¼
0:34 (cid:2) 0:5; here, ~T ≡ T =T0 is the ratio of the
initial temperature, T , and the critical temper-
Þ ≈ 120 nK of a non-
ℏwr=pkB
ature T0 ¼
ffiffiffiffiffiffiffi
6N
p
ð
ð
ð
C x (cid:2) (cid:1)x=2; x þ (cid:1)x=2
Þ ¼ 0 denotes uncorrelated
phases. For quantitative analysis, we average
C x (cid:2) (cid:1)x=2; x þ (cid:1)x=2
Þ over x within the region
of clear interference fringes to obtain C (cid:1)xð Þ (33).
In (22), C (cid:1)xð Þ of 2D Bose gases at equilibrium
was found to decay exponentially in the normal
regime, whereas in the superfluid regime it was
found to decay algebraically with spatially vary-
ing exponent for inhomogeneous 2D gases in
the superfluid regime, which is in agreement
with the prediction based on local correlation
approximation in (34).
In Fig. 2A, we show the time evolution of
C (cid:1)xð Þ after the quench at t ¼ 0. Initially, there
is almost no spatial correlation decay because
the two clouds have nearly identical phases;
their phases decouple within ∼100 ms (27). After
this, we observe a temporal decay of C (cid:1)xð Þ for
all (cid:1)x. At longer times, we observe rapid spatial
decay of correlation, indicating vanishing co-
herence at large distances. To determine the
nature of this dynamic transition, we fit C (cid:1)xð Þ
with the algebraic and exponential functions
that are used to characterize the equilibrium
BKT transition (22). As the system relaxes from
the initial superfluid state to the normal state, it
is expected that C (cid:1)xð Þ will evolve smoothly from
algebraic to exponential scaling in the short- and
long-time limits, respectively. At intermediate
times, both power-law and exponential fitting
can be used to extract physical properties of the
nonequilibrium state; within the real-time RG
picture, nonequilibrium systems in the crossover
region are away from the fixed point and there
is no direct correspondence to equilibrium sys-
tems (19, 35). Indeed, at short and interme-
diate times, the spatial decay of the correlation
function is compatible with algebraic scaling,
including the effect of inhomogeneity of the
system, and with exponential scaling for long
times (27). We identified the crossover time, tc,
as the time at which the correlation function be-
comes better described by exponential scaling
rather than algebraic (Fig. 2B, inset). Even after
tc, the c2 values for the power-law model remain
small, supporting the description of the system
with the power-law exponent h beyond tc (27).
The dynamic BKT transition is expected to have
self-similar dynamics with a length scale that
depends linearly on time (19, 36, 37). Motivated by
this, we have plotted the correlation functions
with rescaled length x′ ¼ (cid:1)xt=tc, using tc (cid:3) 0:5 s
(Fig. 2B). This shows convincingly that, except
for long times, the fluctuations in the system
depend only on the rescaled parameter x′
through a universal function, which we found
to be close to the expected function at the equi-
librium BKT crossover (Fig. 2B). We found the
same behavior independent of the initial condi-
tion (temperature) of the system, demonstrating
the robustness of the scale-invariant behavior
near the critical point (fig. S4).
At long times, scale invariance is broken by
vortex excitations, which results in an emergent
p
p
ffiffiffiffiffi
8p
interacting trapped gas (29), where wr=2p ¼
11 Hz is the radial trapping frequency, ℏ is
the reduced Planck constant, and kB is the
Boltzmann constant. The quench is imple-
mented by a rapid splitting of the system in a
multiple-RF–dressed potential (27, 30–32)
(Fig. 1A), which results in a pair of decoupled
clouds, each with atom number N ′ ¼ N =2,
trapped in the two minima of a double-well
potential. Each well has a vertical-trap frequency
of wz=2p ¼ 1 kHz, hence the dimensionless
2D interaction strength is ~g ¼
as=‘0 ¼
0:076 (27), where as is the 3D s-wave scattering
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
length and ‘0 ¼
Þ
ð
ℏ= mwz
is the harmonic
oscillator length along z for an atom of mass m.
The initial reduced temperature, ~T , is chosen
to be close to the equilibrium critical point at
~T c;eq ¼ 0:53 so that the value after splitting,
~T ′=~T ¼ 1:67, corresponds to the normal phase
(27). To investigate the dynamics, we let each
cloud evolve independently for time t before
performing a time-of-flight (TOF) expansion
of tTOF ¼ 16 ms, after which we detected the
matter-wave interference that encodes the
in situ relative phase fluctuation, f xð Þ, of two
clouds along a line that goes through the center
of the cloud. Interference images and histo-
grams of spatial phase fluctuations Df show
stronger fluctuations at long evolution times
(Fig. 1C). The dynamics across the BKT tran-
sition are expected to be scale invariant until
the bound vortex-antivortex pairs dissociate to
disrupt the phase coherence (Fig. 1D).
Relaxation dynamics
To analyze the relaxation dynamics, we used the
interference pattern to determine both the cor-
relation function and the vortex density (22). The
phase correlation of the system between two
points at locations x (cid:2) (cid:1)x=2 and x þ (cid:1)x=2, with
spatial separation(cid:1)x, is determined byC x (cid:2) (cid:1)x=2;
i
h
Nr
x þ (cid:1)x=2Þ ¼ Re
,
j¼1
where the index, j, runs over Nr ¼ 45 exper-
imental repeats; C x (cid:2) (cid:1)x=2; x þ (cid:1)x=2
Þ ¼ 1 in-
ð
dicates perfect correlation of phases, whereas
½
ei fj x(cid:2)(cid:1)x=2
ð
Þ(cid:2)fj xþ(cid:1)x=2
Þ
ð
X
1
Nr
ð
(cid:5)
Sunami et al., Science 382, 443–447 (2023)
27 October 2023
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
p
ffiffiffiffiffi
nv
p
ffiffiffiffiffi
nv
length scale, r0 ≈ 1=
, where nv is the
vortex density (27). To demonstrate this, we
plotted the correlation function at long times
against the rescaled distance ~x ¼ (cid:1)x
(Fig.
2C) (37). We obtained nv from the occurrence
of sudden jumps of phases, which indicate the
presence of a vortex core (22, 26, 27). These
transformed correlation functions are nearly
time independent, showing that the system
can be characterized by the vortex density deep
in the normal regime.
Varying the initial conditions
Having verified the behavior of the dynamic
BKT transition, we next analyzed its universal
characteristics by varying ~T . The time evolu-
tion of the algebraic exponent, h , determined
by using an algebraic fit to C (cid:1)xð Þ, exhibits a
linear increase in which the increase is faster
for higher ~T (Fig. 3A). This shows that the dy-
namics are accelerated at higher ~T and that
the system quickly crosses over to the normal
phase. This is also reflected in the measurements
of the vortex density,nv, showing a faster growth
at higher ~T (Fig. 3B). We found that the vortex
growth follows a power-law scaling as ex-
pected from the RG predictions (fig. S8). We
compared the measurements of h and nv with
the corresponding results of classical-field
simulations, which give consistent dynamics
(Fig. 3, A and B).
To confirm universal scaling, h and nv have
been plotted as a function of scaled time, t=tc
(Fig. 3, C and D). The time evolutions for various
initial values of ~T collapse onto a single curve,
showing the robustness of dynamical scaling.
We found a linear increase of h across t ¼ tc.
In equilibrium theory, h scales approximately
linearly with temperature in the superfluid
regime (i.e., h º T =4TBKT) (25), thus connect-
ing the temperature scale with phase fluctua-
tions. From this linear estimate, we obtained the
critical exponent hc ¼ 0:13 1ð Þ at t=tc ¼ 1, close
to the value hc ¼ 0:17 3ð Þ found for a finite-size
equilibrium system with similar experimental
parameters (22) and different from the value
of hBKT ¼ 0:25 for the equilibrium BKT tran-
sition in the thermodynamic limit. The linear
increase of h after t=tc ¼ 1 is a precursor of
nonequilibrium superheated superfluid (28)
as a consequence of a delayed vortex prolifera-
tion. The vortex growth after t=tc ¼ 1 exhib-
ited universal power-law scaling nv º t2n, with
n (cid:3) 1; this agrees with the RG prediction, as
shown in the next section.
Comparison to renormalization-group theory
We next compared the experimental results
with predictions based on the real-time RG
equations (19, 28). These equations describe
the time evolution of parameters character-
izing the system from arbitrary nonequilibrium
states flowing toward fixed points, which rep-
resent possible equilibrium states. For the
dynamic BKT transition, the real-time RG equa-
tions are (19, 27, 28)
(cid:3)
¼ 2 (cid:2)
(cid:4)
1
2h
g
t
dg
dt
dh
dt
¼ pg2
16ht
þ g
ð1Þ
ð2Þ
Þ
ð
½
ð
p
ffiffiffiffiffiffi
n~g
where the vortex fugacity, g, is related to the
vortex density by nv tð Þ ¼ n ~g exp 2ln g=2
Þ=
2 (cid:2) 1= 2nð
Þ(cid:5). n is the mean density within the
ð
Þ
region of interest used for fringe analysis, and
~g is the interaction strength as defined pre-
viously; these parameters also determine the
healing length, x ¼ 1=
, characterizing
the length scale of the vortex cores (27, 38).
This RG flow derives from the dynamical sine-
Gordon model, serving as a dual model for
describing the BKT transition; we have added
a phenomenological heating term, g, to ac-
count for the slow, trap-induced heating of
the system (27). For h < 1=4, the fugacity is
strongly suppressed, resulting in a linear dis-
persion, wk ¼ ck. As h increases in time, the
vortex fugacity becomes relevant and increases,
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
resulting in a dispersion, wk ¼ c
0 tð Þ
k2 þ 1=r2
.
As we have argued and demonstrated in the
previous sections, this is indeed supported by
the two-step scaling behavior. Furthermore,
at long times, we have 1= 2hð
Þ ≪ 2, therefore
nv ºg ºt2 (Fig. 3D).
p
p
ffiffiffi
2
Þ and y ¼
ð
Þ (cid:2) 1= 2hc
In Fig. 4, the experimental observations are
plotted together with the RG flow diagram of
Eqs. 1 and 2. For this representation, we defined
x ¼ 1= 2hð
pg. This en-
sures that xc ¼ 0 at h ¼ hc independent of
system sizes, where hc ¼ 0:13 1ð Þ for our finite-
sized system, and hc ¼ hBKT ¼ 1=4 is the the-
oretical prediction in the thermodynamic limit.
Our results follow a universal trajectory in the
flow diagram. The quenched system begins at
large x, where vortex excitations are suppressed
and fugacity is small. Later on, nonequilibrium
phonon creation drives the system toward
smaller x, however still with suppressed y.
As the system approaches the critical point
xc ¼ 0, the onset of vortex excitation drives
the transition.
Discussion and outlook
Our work provides a comprehensive under-
standing of nonequilibrium dynamics across
the BKT transition. The experimental mea-
surements support the real-time RG picture of
universality out of equilibrium, indicating that
it is an excellent starting point for the theoret-
ical study of a wide range of many-body dynam-
ics within the framework of RG. The results
also show that our matter-wave interference
technique is ideally suited for further in-depth
investigation of universal dynamics in 2D sys-
tems, such as the Kibble-Zurek scaling (8) and
nonthermal fixed points (39).
REFERENCES AND NOTES
1. T. Langen et al., Science 348, 207–211 (2015).
2. T. Schweigler et al., Nature 545, 323–326 (2017).
3. H. P. Zahn et al., Phys. Rev. X 12, 021014 (2022).
4. T. Kinoshita, T. Wenger, D. S. Weiss, Nature 440, 900–903
(2006).
5. W. H. Zurek, Nature 317, 505–508 (1985).
6.
I. Chuang, R. Durrer, N. Turok, B. Yurke, Science 251,
1336–1342 (1991).
7. S. L. Veatch, O. Soubias, S. L. Keller, K. Gawrisch, Proc. Natl.
Acad. Sci. U.S.A. 104, 17650–17655 (2007).
8. T. W. B. Kibble, J. Phys. A Math. Theor. 9, 1387 (1976).
9. Q. Zhou, T.-L. Ho, Phys. Rev. Lett. 105, 245702 (2010).
10. A. Polkovnikov, K. Sengupta, A. Silva, M. Vengalattore, Rev.
Mod. Phys. 83, 863–883 (2011).
11. J. Eisert, M. Friesdorf, C. Gogolin, Nat. Phys. 11, 124–130 (2015).
12. N. Navon, A. L. Gaunt, R. P. Smith, Z. Hadzibabic, Science 347,
167–170 (2015).
13. L. W. Clark, L. Feng, C. Chin, Science 354, 606–610 (2016).
14. J. Beugnon, N. Navon, J. Phys. At. Mol. Opt. Phys. 50, 022002
(2017).
15. A. Keesling et al., Nature 568, 207–211 (2019).
16. M. Prüfer et al., Nature 563, 217–220 (2018).
17. S. Erne, R. Bücker, T. Gasenzer, J. Berges, J. Schmiedmayer,
Nature 563, 225–229 (2018).
18. J. A. P. Glidden et al., Nat. Phys. 17, 457–461 (2021).
19. L. Mathey, A. Polkovnikov, Phys. Rev. A 81, 033605 (2010).
20. H.-C. Chu, G. A. Williams, Phys. Rev. Lett. 86, 2585–2588 (2001).
21. A. J. Groszek, P. Comaron, N. P. Proukakis, T. P. Billam,
Phys. Rev. Res. 3, 013212 (2021).
22. S. Sunami et al., Phys. Rev. Lett. 128, 250402 (2022).
23. V. Berezinskiˇı, Sov. J. Exp. Theor. Phys. 34, 610 (1972).
24. J. M. Kosterlitz, D. J. Thouless, J. Phys. C Solid State Phys. 6,
1181–1203 (1973).
25. D. R. Nelson, J. M. Kosterlitz, Phys. Rev. Lett. 39, 1201–1205
(1977).
26. Z. Hadzibabic, P. Krüger, M. Cheneau, B. Battelier, J. Dalibard,
Nature 441, 1118–1121 (2006).
27. Materials and methods are available as supplementary
materials.
28. L. Mathey, K. J. Günter, J. Dalibard, A. Polkovnikov, Phys. Rev. A
95, 053630 (2017).
29. For the range of parameters used in this work, the ratio T=T0
has a one-to-one mapping to the peak phase-space density
of the trapped gas and sufficiently represents the effect of the
quench; see (27).
30. A. J. Barker et al., New J. Phys. 22, 103040 (2020).
31. T. L. Harte et al., Phys. Rev. A 97, 013616 (2018).
32. A. J. Barker et al., J. Phys. At. Mol. Opt. Phys. 53, 155001 (2020).
33. We analyzed the fringes within a region of interest
corresponding to 80% of the Thomas-Fermi radius of the
quasicondensate. As shown in fig. S3, this is where the
interference fringes are clear and allow reliable measurements.
34. I. Boettcher, M. Holzmann, Phys. Rev. A 94, 011602 (2016).
35. In fact, a combined fitting function can be used that has
power-law behavior at short distances and exponential scaling
at long distances, where the exponential length scale can be
infinite.
36. P. Comaron, F. Larcher, F. Dalfovo, N. P. Proukakis, Phys. Rev.
A 100, 033618 (2019).
37. To motivate the scaling regimes, we considered a linear
ð
q
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
0 tð Þ
k2 þ 1=r2
spectrum, wk ¼ ck. With this dispersion, dynamical phase
Þ are kept invariant if a z ¼ 1
factors of the form exp iwkt
scaling is applied that leaves ckt invariant. Furthermore, for a
spectrum that includes an additional length, r0, such as
wk ¼ c
, for long times, only the modes with
w ≈ c=r0 tð Þ contribute to the dynamics. Here the scaling is
replaced by keeping ct=r0 tð Þ fixed, as it was demonstrated for
the long-time dynamics in Fig. 2C. As discussed in the main
text and shown in Fig. 4, the length scale r0 tð Þ is related to
the vortex fugacity increasing at long times.
38. X. G. Wen, Quantum Field Theory of Many-Body Systems:
From the Origin of Sound to an Origin of Light and Electrons
(Oxford Univ. Press, 2010).
39. J. Schole, B. Nowak, T. Gasenzer, Phys. Rev. A 86, 013624
(2012).
40. S. Sunami et al., Dataset for “Universal scaling of the dynamic
BKT transition in quenched 2D Bose gases,” version 3, Zenodo
(2023); https://zenodo.org/records/8385524.
AC KNOWLED GME NTS
We acknowledge discussions with J. Okamoto on theoretical
analysis and thank J. Chalker for comments on our manuscript.
Sunami et al., Science 382, 443–447 (2023)
27 October 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
Funding: The experimental work was supported by the EPSRC
grant reference EP/S013105/1. S.S. acknowledges the Murata
Overseas Scholarship Foundation, the Ezoe Memorial Recruit
Foundation, the Daishin Foundation, and St. Hilda’s
College, Oxford, for financial support. D.G., A.B., A.J.B., and
K.L. thank the EPSRC for doctoral studentships. L.M. acknowledges
funding by the Deutsche Forschungsgemeinschaft (DFG) in the
framework of SFB 925 – project ID 170620586 and the excellence
cluster “Advanced Imaging of Matter” – EXC 2056 – project ID
390715994. V.P.S. acknowledges funding by the Cluster of Excellence
“QuantumFrontiers” – EXC 2123 – project ID 390837967. Author
contributions: S.S. performed the experiments and data
analysis. V.P.S. and L.M. developed numerical and analytical
models and contributed to the interpretation of our experimental
data. S.S. and V.P.S. wrote the manuscript. L.M. and C.J.F.
supervised the project. All authors contributed to the discussion
and interpretation of our results. Competing interests: The
authors declare that they have no competing interests. Data and
materials availability: All data presented in this paper and
numerical simulation codes are available at Zenodo (40). License
information: Copyright © 2023 the authors, some rights
reserved; exclusive licensee American Association for the Advancement
of Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abq6753
Materials and Methods
Figs. S1 to S9
References (41–54)
Submitted 11 May 2022; accepted 16 September 2023
10.1126/science.abq6753
Sunami et al., Science 382, 443–447 (2023)
27 October 2023
5 of 5
|
10.1126_science.abq7487
|
RES EARCH
HUMAN EVOLUTION
Genomic inference of a severe human bottleneck
during the Early to Middle Pleistocene transition
Wangjie Hu1,2†‡, Ziqian Hao3†, Pengyuan Du1,3, Fabio Di Vincenzo4, Giorgio Manzi5, Jialong Cui2,
Yun-Xin Fu6,7, Yi-Hsuan Pan2*, Haipeng Li1,8*
Population size history is essential for studying human evolution. However, ancient population size history
during the Pleistocene is notoriously difficult to unravel. In this study, we developed a fast infinitesimal time
coalescent process (FitCoal) to circumvent this difficulty and calculated the composite likelihood for present-
day human genomic sequences of 3154 individuals. Results showed that human ancestors went through a
severe population bottleneck with about 1280 breeding individuals between around 930,000 and 813,000 years
ago. The bottleneck lasted for about 117,000 years and brought human ancestors close to extinction. This
bottleneck is congruent with a substantial chronological gap in the available African and Eurasian fossil record.
Our results provide new insights into our ancestry and suggest a coincident speciation event.
proach is needed to improve the inference ac-
curacy of population size history.
Population size changes that occurred hun-
dreds of thousands of years ago affected the
rates of coalescence and thus have left their
signatures in the site frequency spectrum (SFS)
of genomic sequences. The SFS is the distri-
bution of allele frequencies in the sequences,
randomly collected from the present-day hu-
man population. Each SFS category contains
a certain number of mutations of the same
size. Because SFS is crucial for demographic
inference (6, 8–12) and construction of key
summary statistics (13), many efforts have been
devoted to deriving its analytical formulas (14–17).
However, these formulae may not achieve the
required accuracy because of propagation and
accumulation of numerical errors resulting from
their dependence on the joint probability den-
sity function of coalescent times (16–18).
In this study, to circumvent this numerical
problem, we developed the fast infinitesimal
time coalescent process (FitCoal) (Fig. 1) that
analytically derives expected branch length
for each SFS category under arbitrary demo-
graphic models. FitCoal does not need phased
haplotype data and prior information on de-
mography. The effects of sequencing errors or
hitchhiking due to positive selection can be
circumvented, largely by focusing on a subset
of SFS that are less influenced by those factors.
We used FitCoal to analyze a large number of
present-day human genomic sequences from 10
African and 40 non-African populations. Results
showed that our ancestors experienced a sev-
ere population bottleneck between about 930
and 813 kyr BP, most likely because of climatic
changes. The average number of breeding indi-
viduals was only about 1280 during the bottle-
neck period. Our findings indicate that the
severe bottleneck brought the ancestral human
population close to extinction and completely
reshaped present-day human genetic diversity.
Results
FitCoal
We developed FitCoal to determine the expected
branch lengths for an SFS (Fig. 1). During
FitCoal analysis of a sample, the time period
in which the most recent common ancestor
A lthough the lineage of humans is esti-
mated to have separated from that of
chimpanzees and bonobos more than
6 million years ago, anatomically mod-
ern humans (Homo sapiens) are estimated
to have originated around 300 thousand to
200 thousand years before the present (kyr BP)
in Africa (1–3). On the basis of present-day hu-
man genomic sequences, the recent population
size histories (i.e., the dynamics of population
size since the emergence of modern humans)
have been intensively studied, revealing the
worldwide spread of our ancestors (4–8). How-
ever, ancient population size history of the genus
Homo during the Pleistocene is still poorly
known, although it is essential for understand-
ing the origin of the human lineage. It is likely
to be very difficult or impossible to obtain an-
cient DNA from African Homo samples dated
before the emergence of H. sapiens. It would
be particularly notable if present-day human
genomic sequences could be used to robustly
infer both the recent and ancient population
size histories of humankind. Thus, a new ap-
1CAS Key Laboratory of Computational Biology, Shanghai
Institute of Nutrition and Health, University of Chinese Academy
of Sciences, Chinese Academy of Sciences, Shanghai, China.
2Key Laboratory of Brain Functional Genomics of Ministry of
Education, School of Life Science, East China Normal University,
Shanghai, China. 3College of Artificial Intelligence and Big Data
for Medical Sciences, Shandong First Medical University &
Shandong Academy of Medical Sciences, Jinan, China. 4Natural
History Museum, University of Florence, Florence, Italy.
5Department of Environmental Biology, Sapienza University of
Rome, Rome, Italy. 6Department of Biostatistics and Data
Science, School of Public Health, University of Texas Health
Science Center at Houston, Houston, TX, USA. 7Key Laboratory
for Conservation and Utilization of Bioresources, Yunnan
University, Kunming, China. 8Center for Excellence in Animal
Evolution and Genetics, Chinese Academy of Sciences,
Kunming, China.
*Corresponding author. Email: yxpan@sat.ecnu.edu.cn (Y.-H.P.);
lihaipeng@sinh.ac.cn (H.L.)
†These authors contributed equally to this work.
‡Present address: Department of Genetics and Genomic Sciences,
Icahn School of Medicine at Mount Sinai, New York, NY, USA.
Fig. 1. Illustration of FitCoal. (Left) The backward process in which four lineages (represented by the
four solid black circles at the bottom) coalesce into one (represented by the single solid black circle at
the top) after passing through millions of infinitesimal time intervals (Dt). The area highlighted in blue shows the
backward transformation process of different coalescent states with tiny probability changes in an infinitesimal
time interval. Thick arrows indicate high transformation probabilities, and thin arrows indicate low transformation
probabilities. The blue and purple arrows correlate to the two events in the middle pane represented by
blue- and purple-colored lines. Each state is indicated with a box, in which one circle indicates one lineage.
The boxes with solid black circles represent the states with the probability of 1. The boxes with empty
circles represent the states with the probability of 0. The probabilities between 0 and 1 are represented
by gray circles. (Middle) Hypothetical coalescent trees with branches of different states, indicating the
number of lineages. Blue branches represent a transformation from four to three lineages. Purple branches
indicate that no coalescent event occurred. (Right) The size of a theoretical population over time. The
width of shadowed area denoted as N(t) indicates the effective population size (i.e., the number of breeding
individuals) at time t. It is assumed that the effective population size remains unchanged within a Dt.
Hu et al., Science 381, 979–984 (2023)
1 September 2023
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
A
100
e
z
s
i
Constant size
n = 30, L = 10 Mb
)
s
d
n
a
s
u
o
h
t
n
i
(
n
o
i
t
l
a
u
p
o
p
e
v
i
t
c
e
f
f
E
10
1
D
e
z
s
i
10,000
)
s
d
n
a
s
u
o
h
t
n
i
(
l
n
o
i
t
a
u
p
o
p
e
v
i
t
c
e
f
f
E
100
1
Bottleneck
n = 170, L = 30 Mb
B
100
10
Instantaneous increase
n = 30, L = 10 Mb
C
100
10
1
10
100
1000
10
100
1000
10
100
1000
Exponential growth I
n = 30, L = 10 Mb
Exponential growth II
n = 30, L = 10 Mb
F
500
100
10
E
10,000
100
1
10
10
100
Time (thousand years ago)
Time (thousand years ago)
Exponential growth III
n = 170, L = 10 Mb
10
Time (thousand years ago)
100
1000
True Model
FitCoal
PSMC
Stairway Plot
SMC++
Fig. 2. Population size histories inferred by FitCoal, PSMC, Stairway Plot, and
SMC++ with simulated samples. (A) Constant size model. (B) Instantaneous
increase model. (C) Bottleneck model. (D) Exponential growth I model. (E) Exponential
growth II model. (F) Exponential growth III model. In all panels, thin black lines indicate
the true models. Thick red lines indicate the medians of FitCoal-inferred population size
histories; thin red lines represent 2.5 and 97.5 percentiles of FitCoal-inferred population
size histories. Yellow, green, and blue lines indicate results obtained with PSMC, Stairway
Plot, and SMC++, respectively. The mutation rate is assumed to be 1.2 × 10–8 per
base per generation, and a generation time is assumed to be 24 years. n is the number of
simulated sequences, and L is the length in Mb of each simulated sequence.
originated was partitioned into as many time
intervals as needed, such that each time inter-
val (Dt) was very small (e.g., 1 month or 1 year).
During each time interval, the population size
was assumed to be constant. The probabilities
of all coalescent states (i.e., all possible ances-
tral lineages) were calculated backward in
time. For each state, the branch length during
a time interval was calculated by multiplying
its probability with population size and then
transformed to determine the expected branch
lengths. Because the expected branch length
of an SFS category during a time interval was
precalculated, FitCoal could be very fast.
FitCoal demographic inference
After the expected branch lengths were de-
termined, the composite likelihood of an SFS
(6, 9, 19) was calculated. FitCoal is effective
for a wide range of sample sizes in the calcu-
lation of the composite likelihood of a given
SFS and is much more accurate than simu-
lation approaches (fig. S1). When inferring
population size history, the likelihood was
maximized in a wide range of demographic
scenarios. Moreover, both instantaneous and
long-term exponential population changes were
considered. Similar to previous studies (6, 19),
the likelihood of the SFS was first maximized
with the constant size model, followed by re-
peated maximization of the likelihood with
increased number of inference time intervals
until the best model was found.
Demographic inference from simulated data
The accuracy of FitCoal demographic infer-
ence was evaluated by simulation and com-
parison of results with those of PSMC (pairwise
sequentially Markovian coalescent), Stairway
Plot, and SMC++ methods (6–8) (Fig. 2). To
ensure fair comparisons, we tested six demo-
graphic models by simulating 200 independent
datasets for each model, as described previ-
ously (6), with the assumption that a generation
time is 24 years (6, 20) and that the mutation
rate is 1.2 × 10–8 per site per generation (6, 21).
Results showed that the medians of FitCoal-
inferred population size histories were almost
identical to the true models, and the 95% con-
fidence intervals of FitCoal inference were nar-
rower than those of PSMC, Stairway Plot, and
SMC++ (Fig. 2). FitCoal inference accuracy
could be further improved by increasing sam-
ple size and lengths of sequences (fig. S2). The
proportion of the most recent changes in popu-
lation size inferred from the six models showed
that FitCoal could distinguish between instan-
taneous and exponential changes (table S1).
Overall, our results confirmed that SFSs could
be used to estimate population size histories (22).
It has been suggested that a population size
history could be inferred by using a subset of
SFS or a collapsed SFS (6, 19); the latter is an
SFS with high frequency mutations combined
into one category. Results of simulations showed
that the FitCoal could still accurately determine
a population size history even when a portion
(10 to 90%) of an SFS was truncated (i.e., ex-
cluded for analysis) (figs. S3 to S5), thus reduc-
ing the impact of confounding factors, such as
hitchhiking effect due to positive selection (fig.
S6) or sequencing errors on FitCoal analyses.
Demographic inference of African populations
To infer population size histories of African
populations, seven African populations in the
1000 Genomes Project (1000GP) (23) and three
African populations in the Human Genome
Diversity Project–Centre d’Etude du Polymor-
phisme Humain (HGDP-CEPH) panel (24) were
analyzed by the FitCoal (tables S2 to S4). Only
autosomal noncoding regions were used to
partially avoid the effect of purifying selection.
To avoid hitchhiking effect due to positive
Hu et al., Science 381, 979–984 (2023)
1 September 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Histories of human populations in 1000GP and HGDP-CEPH genomic
datasets inferred by FitCoal, SMC++, Stairway Plot, PSMC, and Relate. The
mutation rate is assumed to be 1.2 × 10–8 per base per generation, and a generation
time is assumed to be 24 years. (A) Inferred population size histories of 26
populations in 1000GP. (B) Linear-scaled estimation of sizes over time of populations
in 1000GP during the severe bottleneck period. (C) Inferred population size histories
of 24 populations in the HGDP-CEPH panel. (D) Linear-scaled estimation of sizes
over time of populations in the HGDP-CEPH panel during the severe bottleneck period.
(E and F) Comparison of population size histories of African YRI and Yoruba
populations inferred by FitCoal, SMC++, Stairway Plot, PSMC, and Relate. Only
the population size histories up to 200 kyr BP were analyzed by Stairway
Plot. In (A) to (D), colored lines indicate the following: red, African populations;
brown, Middle East populations; yellow, European populations; blue, East Asian
populations; green, Central or South Asian populations; and gray, American
populations. Black dashed circles with arrow heads represent the discrepancy in
the population size between African agriculturalist and non-African populations.
In (E) and (F), red, blue, yellow, green, and gray lines indicate results obtained
with FitCoal, SMC++, PSMC, Stairway Plot, and Relate, respectively.
selection (25), high-frequency mutations were
excluded from the analysis. Results showed that
all 10 African populations went through a sev-
ere bottleneck (Fig. 3 and figs. S7 and S8). The
bottleneck was estimated to persist for 117 kyr,
from 930 ± 23.52 (SEM) (range, 854 to 1042)
to 813 ± 11.02 (SEM) (range, 772 to 864) kyr BP.
The average effective population size (i.e., the
number of breeding individuals) (26) during
the bottleneck period was determined to be
1280 ± 131 (SEM) (range, 770 to 2030), which
was only 1.3% of its ancestral size (98,130 ±
8720; range, 58,600 to 135,000). To evaluate the
impact of the bottleneck on current human
genetic diversity, we analyzed the expected
pairwise nucleotide diversity. Results showed
that 65.85% of current human genetic diver-
sity was lost because of the bottleneck.
Demographic inference of non-African populations
the 21 non-African populations in the HGDP-
CEPH panel (Fig. 3, A to D; figs. S7 and S8; and
tables S2, S5, and S6). The average ancestral
population sizes of the populations in the two
datasets were 20,260 (range, 18,850 to 22,220)
and 20,030 (range, 19,060 to 21,850), respec-
tively, similar to those determined in previous
studies (7, 8, 24). The estimated population
size started to decline around 368 (range, 175 to
756) and 367 (range, 167 to 628) kyr BP, respec-
tively, which is consistent with previous find-
ings that African and non-African divergence
occurred much earlier than the out-of-Africa
dispersal (7, 8, 23, 24). The inferred out-of-
Africa dispersal and the recent population size
expansion and reduction are consistent with
those of previous studies (5–8, 23, 24).
Severe bottleneck during the Early to Middle
Pleistocene transition
The severe bottleneck was not directly detected
in all 19 non-African populations in 1000GP and
The ancient severe bottleneck was directly de-
tected in each of the 10 African populations
but in none of the 40 non-African populations.
To investigate this discrepancy, we performed
simulations with three demographic models,
designated bottlenecks I, II, and III (Fig. 4, A
to C, and figs. S9 and S10). Bottleneck I sim-
ulated the population size history of African
agriculturalist populations with the ancient
severe bottleneck, and bottlenecks II and III
simulated that of non-African populations with-
out and with the ancient severe bottleneck,
respectively. Both bottlenecks I and II were
inferred precisely in all simulated data sets
(tables S7 to S9). However, no ancient severe
bottleneck was detected in bottleneck III sim-
ulations, which indicates that the out-of-Africa
dispersal hinders the chance of discovering the
ancient severe bottleneck. Furthermore, the
ancient severe bottleneck was found to cause a
discrepancy in the estimation of the population
size between the bottleneck III model and the
inferred population size history after the bottle-
neck was relieved, which suggests a hidden effect
Hu et al., Science 381, 979–984 (2023)
1 September 2023
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Verification of the severe bottleneck. (A) Bottleneck I model,
mimicking the true population size history of African agriculturalist populations.
(B) Bottleneck II model, mimicking the inferred population size history of non-
African populations. (C) Bottleneck III model, mimicking the true population
size history of non-African populations. The gap between actual and FitCoal
estimated population size is indicated by the black dashed circle and arrowhead.
(D) Bottleneck IV model, mimicking a population with an exponential reduction
in size 1.5 million years ago. (E) Bottleneck V model, mimicking a population
with a moderate bottleneck. (F) Bottleneck VI model, mimicking a population
with a weak bottleneck. Black lines represent the true models. Thick red lines
represent the medians of FitCoal estimated population sizes over time; thin
red lines represent 2.5 and 97.5 percentiles of FitCoal estimated population
sizes over time. Blue, green, yellow, and gray lines represent the medians of
10 runs each of SMC++, Stairway Plot, PSMC, and Relate. The mutation rate is
assumed to be 1.2 × 10–8 per base per generation, and a generation time is
assumed to be 24 years. Numbers of simulated sequences are 188 in bottlenecks
I, IV, V, and VI and 194 in bottlenecks II and III. The lengths of simulated
sequences are 800 Mb each.
of the ancient severe bottleneck on non-African
populations (Fig. 4C and figs. S9C and S10C).
After the bottleneck was relieved, the average
population size of non-African populations in
1000GP was 20,260, and that of those in HGDP-
CEPH was 20,030. For African agriculturalist
populations, the average population size in
1000GP was 27,080, and that of those in HGDP-
CEPH was 27,440. This population size difference
of 7020 (Fig. 3, A and C) is likely due to the
hidden effect of the ancient severe bottleneck
on non-African populations. Because the out-of-
Africa dispersal existed in non-African popu-
lations but not in African populations, African
populations had more lineages remaining to
be traced back to the ancient severe bottle-
neck (fig. S11). In the analysis of the African YRI
(Yoruba in Ibadan, Nigeria) population, the
minimum sample size was three individuals
for detection of the severe bottleneck (fig. S12).
Because the signal for the existence of the
severe bottleneck was too weak to be detected
in non-African populations by FitCoal, we per-
formed an extended FitCoal analysis. To eli-
minate noise effects resulting from problems
such as sequencing or overfitting errors on the
inference of population size history, we used
the FitCoal-inferred recent population size his-
tory as a starting point for size estimation of
an ancient population. With this modifica-
tion, all 19 non-African populations in 1000GP
were found to have gone through the severe
bottleneck with approximately 1450 individ-
uals between 921 and 785 kyr BP (fig. S13 and
table S10). This result is consistent with that
obtained with African populations.
To further examine the severe bottleneck, we
simulated a slow population reduction start-
ing 1.5 million years ago (Fig. 4D). The FitCoal-
inferred population size histories were different
from those observed in 1000GP and HGDP-
CEPH populations, which supports the hypoth-
esis that a sudden size reduction occurred at
the beginning of the bottleneck. Results of sim-
ulations were similar to that of the observed
cases (Fig. 3, E and F) in that they also showed
that PSMC, SMC++, and Relate methods (7, 8, 27)
underestimated the severity of the ancient bot-
tleneck (Fig. 4 and figs. S14 to S17). More-
over, the inferred population declines shown
in Fig. 4A were more severe than those inferred
from the real data because the simulated data
were generated under the assumption of a neu-
trally evolved single population and a homoge-
neous recombination rate, whereas humans
evolved with subpopulations and a heterogeneous
recombination rate (28, 29). FitCoal was not
found to overestimate such severity (Fig. 4) or
to falsely detect a bottleneck when examining
the effects of continuous and pulsed introgres-
sions among existing and ghost populations
(figs. S18 to S21). Therefore, the discovery of
the ancient severe bottleneck was not due to
overfitting the data in FitCoal analyses.
Discussion
In this study, we developed FitCoal, a model-
flexible method for demographic inference.
One key feature of FitCoal is that the expected
Hu et al., Science 381, 979–984 (2023)
1 September 2023
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. Schematic diagram of
human population size history.
Both African (light green) and non-
African (light blue) populations
are presented. The width of the boxes
represents the effective population
size (i.e., the number of breeding
individuals) with naturally occurred
fluctuations. The occurrence time of
the out-of-Africa dispersal and the
divergence between African and
non-African populations are indicated.
The gray-shaded time duration
indicates the Early to Middle Pleis-
tocene transition between 1250 and
700 kyr BP. The red arrow indicates
the peak of glaciation during the
transition (i.e., the 0.9 Ma event).
The ancient severe bottleneck
inferred in this study is highlighted. The gap in the available African hominin fossil record and an indicative
chronology for H. erectus, the LCA, and H. sapiens are shown. The estimated time period in which two
ancestral chromosomes (chromosome, Chr.) fused to become one is also shown on the right.
branch lengths can be accurately determined
for an SFS under an arbitrary demographic
model, which allows precise calculation of the
likelihood. Analyses by FitCoal are, in most
cases, less time consuming than those by other
methods such as PSMC, SMC++, Stairway Plot,
and Relate (28/32 = 87.5%) (tables S11 and S12).
By discarding rare and high-frequency muta-
tions, FitCoal can avoid the effects of sequencing
errors or hitchhiking due to positive selection
without losing its inference accuracy. Because
both instantaneous and exponential changes
are allowed within each inference time inter-
val, FitCoal can reveal the dynamic of population
size precisely. Because coalescent events be-
come rare when tracing backward in time, the
length of inference time interval is usually set
to increase progressively (6–8). Although this
strategy can capture recent population size
changes, it may miss ancient ones. Therefore,
FitCoal inference time intervals are allowed to
vary during demographic inference, and FitCoal
can make a fast and accurate inference of recent
and ancient population size histories.
The most important discovery with FitCoal
is that human ancestors went through a se-
vere bottleneck in the late Lower Pleistocene
(Fig. 5). This ancient severe bottleneck was
directly found in all 10 African populations,
but only a weak signal of the existence of such
was detected in all 40 non-African popula-
tions. This observation is consistent with the
coalescent theory and the occurrence of the
out-of-Africa dispersal. Results of our large-
scale simulations demonstrated that FitCoal
did not falsely infer the bottleneck because of
positive selections (figs. S6 and S22) or pop-
ulation structure (fig. S18 to S21). Because we
observed no overfitting cases and results ob-
tained by examining different sets of genomic
regions (Fig. 3, A and C, and fig. S23) were sim-
ilar, the existence of the ancient severe bottle-
neck was ascertained.
Our results indicate that the ancient severe
bottleneck lasted for approximately 117 kyr
(Fig. 5), and that about 98.7% of human an-
cestors were lost at the beginning of the bot-
tleneck, thus threatening our ancestors with
extinction. The estimated effective population
size during the bottleneck period was only
1280 breeding individuals, which was compa-
rable to the effective population sizes of other
endangered mammals (30, 31). This size (1280)
might have been overestimated because of hid-
den population structure (32). Naturally occur-
ring population size fluctuations might have
further increased the extinction risk for our
ancestors during the bottleneck. The bottleneck
could also have increased the inbreeding level
of our ancestors, thus contributing to the 65.85%
loss in present-day human genetic diversity.
The ancient population size reduction that
occurred around 930 kyr BP was likely driven
by climatic changes during the Early to Mid-
dle Pleistocene transition (33, 34). During this
transition period known as the “0.9 Ma event”
(Ma, million years ago), glaciations were changed
from predominantly short-term to long-term
events with more extreme thermic intensity,
especially at the peak of glaciation. This event
resulted in a decrease in marine surface tem-
perature to the lowest that occurred during the
entire transition period (33), with an inferred
long period of drought and extensive wildlife
turnover in Africa and Eurasia (35).
The existence of the ancient severe bottle-
neck could explain the extreme scarcity of the
available hominin fossil record in Africa and
Eurasia between 950 and 650 kyr BP (Fig. 5 and
fig. S24). In Africa, only a few fossil specimens
dated in this time period have been found, in-
cluding the cranial fragments from Gombore
in Ethiopia and the fossil samples from Tighenif
in Algeria (36, 37). Although the taxonomic sta-
tuses of these fossils are still not clear, they have
features resembling those of later fossils at-
tributed to Homo heidelbergensis. They are dif-
ferent from the coeval Homo antecessor from a
paleoanthropological site in Spain (Atapuerca,
Gran Dolina), and some scholars considered
H. antecessor as a possible alternative for the last
common ancestor (LCA) (38). During the same
chronological interval, the East Asian fossil
record contains specimens identified as Homo
erectus (39). It does not appear that East Asian
H. erectus is connected to the ancient severe
bottleneck because it is unlikely to have contrib-
uted to the lineage leading to modern humans
(38). In addition, coincident with this bottleneck,
two ancestral chromosomes are believed to have
fused to form chromosome 2 in humans around
900 to 740 kyr BP (40, 41). Therefore, the ancient
severe bottleneck possibly marks a speciation
event leading to the emergence of the LCA
shared by Denisovans, Neanderthals, and mod-
ern humans, whose divergence has been dated
to about 765 to 550 kyr BP (38, 42, 43).
A rapid population recovery was detected in
all 10 African populations with a 20-fold in-
crease in size around 813 kyr BP. Control of
fire could be part of the explanation for this
population expansion, which is shown by the
early archaeological evidence found in Israel
dated about 790 kyr BP (44). Other factors, such
as climatic changes (33, 34), might also be a
driving force for this rapid population recovery.
The ancient severe bottleneck was not de-
tected in previous SFS-based analyses (6, 10, 12, 14).
This failure might be due to the use of prede-
fined demographic models. In this study, we
found that the likelihood must be accurately
calculated to detect the severe bottleneck (fig.
S1). The use of other methods such as Stairway
Plot—which may not have sufficient resolu-
tion power for estimation of ancient popula-
tion size history—is another possible reason for
the failure (6).
Our study revealed that an extremely small
human population lasted for about 117 kyr
around 930 to 813 kyr BP. Many questions
remain unanswered, such as where these indi-
viduals lived, how they overcame the cata-
strophic climate changes, and how the ancient
population remained so small for so long. Fur-
ther studies are warranted to investigate these
matters to obtain a more detailed picture of
human evolution during the Early to Middle
Pleistocene transition.
REFERENCES AND NOTES
J. Galway-Witham, C. Stringer, Science 360, 1296–1298 (2018).
1.
2. C. M. Schlebusch et al., Science 358, 652–655 (2017).
3. A. Meneganzin, T. Pievani, G. Manzi, Evol. Anthropol. 31, 199–212 (2022).
4. M. Stoneking, J. Krause, Nat. Rev. Genet. 12, 603–614 (2011).
5. S. Ramachandran et al., Proc. Natl. Acad. Sci. U.S.A. 102,
15942–15947 (2005).
6. X. Liu, Y. X. Fu, Nat. Genet. 47, 555–559 (2015).
7. H. Li, R. Durbin, Nature 475, 493–496 (2011).
Hu et al., Science 381, 979–984 (2023)
1 September 2023
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
J. Terhorst, J. A. Kamm, Y. S. Song, Nat. Genet. 49, 303–309 (2017).
8.
9. H. Li, W. Stephan, PLOS Genet. 2, e166 (2006).
10. L. Excoffier, I. Dupanloup, E. Huerta-Sánchez, V. C. Sousa,
M. Foll, PLOS Genet. 9, e1003905 (2013).
11. R. C. Griffiths, S. Tavaré, Math. Comput. Model. 23, 141–158 (1996).
12. R. N. Gutenkunst, R. D. Hernandez, S. H. Williamson,
C. D. Bustamante, PLOS Genet. 5, e1000695 (2009).
13. Y. X. Fu, Theor. Popul. Biol. 145, 95–108 (2022).
14. J. Jouganous, W. Long, A. P. Ragsdale, S. Gravel, Genetics 206,
1549–1567 (2017).
15. Y. X. Fu, Theor. Popul. Biol. 48, 172–197 (1995).
16. A. Polanski, M. Kimmel, Genetics 165, 427–436 (2003).
17. D. Zivković, T. Wiehe, Genetics 180, 341–357 (2008).
18. A. Polanski, A. Bobrowski, M. Kimmel, Theor. Popul. Biol. 63,
33–40 (2003).
19. A. Bhaskar, Y. X. R. Wang, Y. S. Song, Genome Res. 25,
268–279 (2015).
20. A. Scally, R. Durbin, Nat. Rev. Genet. 13, 745–753 (2012).
21. A. Kong et al., Nature 488, 471–475 (2012).
22. A. Bhaskar, Y. S. Song, Ann. Stat. 42, 2469–2493 (2014).
23. The 1000 Genomes Project Consortium, Nature 526, 68–74 (2015).
24. A. Bergström et al., Science 367, eaay5012 (2020).
25. J. C. Fay, C. I. Wu, Genetics 155, 1405–1413 (2000).
26. H. C. Harpending et al., Proc. Natl. Acad. Sci. U.S.A. 95,
1961–1967 (1998).
27. L. Speidel, M. Forest, S. Shi, S. R. Myers, Nat. Genet. 51,
1321–1329 (2019).
28. E. M. L. Scerri et al., Trends Ecol. Evol. 33, 582–594 (2018).
29. Z. Hao, P. Du, Y. H. Pan, H. Li, Hum. Genet. 141, 273–281 (2022).
30. H. Li et al., Proc. Natl. Acad. Sci. U.S.A. 113, 14079–14084 (2016).
31. Y. X. Chen et al., Zool. Res. 43, 523–527 (2022).
32. J. Wakeley, Genetics 153, 1863–1871 (1999).
33. M. J. Head, P. L. Gibbard, in Early-Middle Pleistocene
Transitions: The Land–Ocean Evidence, M. J. Head, P. L. Gibbard,
Eds. (Geological Society of London, 2005), pp. 1–18.
34. P. U. Clark et al., Quat. Sci. Rev. 25, 3150–3184 (2006).
35. M. J. Head, B. Pillans, S. A. Farquhar, Episodes 31, 255–259 (2008).
36. A. Profico, F. Di Vincenzo, L. Gagliardi, M. Piperno, G. Manzi,
J. Anthropol. Sci. 94, 41–63 (2016).
37. A. Mounier, F. Marchal, S. Condemi, J. Hum. Evol. 56, 219–246
(2009).
38. A. Bergström, C. Stringer, M. Hajdinjak, E. M. L. Scerri,
P. Skoglund, Nature 590, 229–237 (2021).
39. J.-J. Bahain et al., Anthropologie 121, 215–233 (2017).
40. B. Poszewiecka, K. Gogolewski, P. Stankiewicz, A. Gambin,
BMC Genomics 23, 616 (2022).
41. T. R. Dreszer, G. D. Wall, D. Haussler, K. S. Pollard, Genome Res. 17,
1420–1430 (2007).
42. K. Prüfer et al., Nature 505, 43–49 (2014).
43. D. Reich et al., Nature 468, 1053–1060 (2010).
44. N. Goren-Inbar et al., Science 304, 725–727 (2004).
45. H. Li, Genomic inference of a severe human bottleneck during
the Early to Middle Pleistocene transition, version 3, Mendeley
Data, 2023; https://doi.org/10.17632/xmf5r8nzrn.3.
46. W. Hu et al., FitCoal v1.1., Zenodo, 2023; https://doi.org/
10.5281/zenodo.7857456.
ACKN OWLED GMEN TS
We thank D. Živković for sharing his codes to calculate the expected
branch lengths, X. Liu for providing simulation results, and Bio-Med
Big Data Center at Shanghai Institute of Nutrition and Health for
providing a computing facility. Funding: National Natural Science
Foundation of China grants 32270674, 91131010, and 91731304
(H.L.); National Natural Science Foundation of China grant 31100273
(Y.H.P.); National Natural Science Foundation of China grants
91631304 and 32130011 (Y.X.F.); National Natural Science
Foundation of China grant 82171801 (Z.H.); Strategic Priority Research
Program of the Chinese Academy of Sciences grant XDPB17 (H.L.);
National Key Research and Development Project grant 2022YFF1203202
(H.L.); National Institutes of Health grant R01HG009524 (Y.X.F.);
Education Bureau of Jinan and Shandong First Medical University grant
JNSX2021046 (Z.H.); Key Laboratory of Brain Functional Genomics at
East China Normal University grant 20SKBFGK2 (Y.H.P. and H.L.);
Shanghai Institute of Nutrition and Health grant JBGSRWBD-SINH-2021-
10 (H.L.); China Postdoctoral Science Foundation grant
2022M711978 (Z.H.); Shandong Provincial Natural Science
Foundation grant ZR2022QC062 (Z.H.); and Shandong Provincial
Postdoctoral Innovation Talent Support Program grant
SDBX2022012 (Z.H.). Author contributions: Conceptualization:
W.H., Z.H., F.D.V., G.M., Y.X.F., Y.H.P., and H.L.; Methodology: W.H.,
Z.H., Y.H.P., and H.L.; Software: W.H., Z.H., and H.L.; Investigation:
W.H., Z.H., P.D., F.D.V., G.M., J.C., Y.X.F., Y.H.P., and H.L.; Visualization:
W.H., Z.H., P.D., J.C., F.D.V., and G.M.; Funding acquisition: Z.H., Y.X.F.,
Y.H.P., and H.L.; Supervision: G.M., Y.H.P., and H.L.; Writing: W.H.,
Z.H., P.D., F.D.V., G.M., Y.X.F., Y.H.P., and H.L. Competing interests:
The authors declare that they have no competing interests. Data
and materials availability: Raw data are deposited at Mendeley (45),
and FitCoal is archived at Zenodo (46). The download links for
1000GP and HGDP-CEPH data are available in table S2. License
information: Copyright © 2023 the authors, some rights reserved;
exclusive licensee American Association for the Advancement of
Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abq7487
Materials and Methods
Supplementary Text
Figs. S1 to S60
Tables S1 to S23
References (47–96)
MDAR Reproducibility Checklist
Submitted 28 April 2022; resubmitted 7 February 2023
Accepted 11 July 2023
10.1126/science.abq7487
Hu et al., Science 381, 979–984 (2023)
1 September 2023
6 of 6
|
10.1126_science.abq7768
|
RES EARCH
CONSERVATION
The value of private properties for the conservation
of biodiversity in the Brazilian Cerrado
Paulo De Marco Jr.1*, Rodrigo A. de Souza2, André F. A. Andrade1, Sara Villén-Pérez3,
Caroline Corrêa Nóbrega4, Luiza Motta Campello5, Marcellus Caldas6
Areas set aside for conservation within private lands may be key to enhancing biodiversity-friendly
landscapes. This conservation strategy should be especially effective in highly threatened regions that
are poorly protected by public lands, such as the Brazilian Cerrado. Brazil’s Native Vegetation Protection
Law has included set-aside areas within private properties, but their relevance to conservation has
not been evaluated. We assess whether private lands are contributing to biodiversity in the Cerrado,
a global biodiversity conservation priority and major region for food production, where land use
conflicts are often at odds with conservation objectives. We determined that private protected areas
accommodate up to 14.5% of threatened vertebrate species ranges, which increases to 25% when
considering the distribution of remaining native habitat. Moreover, the spatial spread of private
protected areas benefits a large number of species. Ecological restoration of private protected lands
would improve the benefits of this protection system, especially in the Southeastern Cerrado, where a
large economic hub meets a threat hotspot.
P rotected areas are the cornerstones
for the long-term conservation of bio-
diversity. They cover about 15% of the
terrestrial surface and 7.3% of the ocean
surface (1), and global analyses show
that they are still insufficient to protect bio-
diversity (2). The need for complementary
strategies to join or reinforce protection
networks is especially urgent to deal with
the lack of connectivity due to habitat frag-
mentation. A promising approach is to make
landscapes that are now occupied by eco-
nomic activity more “biodiversity-friendly”
(3). Biodiversity-friendly landscapes seek to
preserve habitat patches in human-dominated
areas to favor the persistence of native species
(3), including beneficial animals such as pol-
linators, predators, and fruit dispersers (4).
Most human-dominated areas are under pri-
vate ownership, and this represents a large
proportion of global land, varying from 44.2% in
Brazil (5) to 52% in Germany, 75% in the United
States (excluding Alaska) (6), and nearly 80%
in the United Kingdom and Spain (7). Thus,
improving the biodiversity-friendliness of pri-
vate landholdings could amplify the benefits
of the existing protection system by increas-
1Departamento de Ecologia, Universidade Federal de Goiás,
Goiânia, GO 74 690-720, Brazil. 2Centro Nacional de
Informações Ambientais (CENIMA), Instituto Nacional de
Meio Ambiente e Recursos Naturais Renováveis (IBAMA),
SCEN Ibama, Ed. Sede, Bloco F, Brasília, DF 70818-900,
Brazil. 3Universidad de Alcalá, GloCEE – Global Change
Ecology and Evolution Research Group, Departamento de
Ciencias de la Vida, 28805, Alcalá de Henares, Madrid,
Spain. 4Aliança da Terra, Av. das Indústrias, 601, Quadra 151
Lote 47 Sala 301, Santa Genoveva, Goiânia, GO 74670-600,
Brazil. 5Universidade Brasília, Instituto de Geociências,
Campus Universitário Darcy Ribeiro ICC, Ala Central, Brasília-
DF 70910-900, Brazil. 6Department of Geography and
Geospatial Sciences, Kansas State University, Manhattan,
KS 66502, USA.
*Corresponding author. Email: pdemarco@ufg.br
ing both the total habitat available and the
connectivity among remaining habitats (8),
ensuring population persistence and richer
biodiversity (9). However, conservation pri-
oritization efforts have often overlooked
the role of private lands while focusing on
public protection networks (10). Here, we
provide an evaluation of the relevance of
set-aside areas of private land in one of the
most important and vulnerable worldwide
arenas for the conflict between food pro-
duction and biodiversity conservation: the
Brazilian Cerrado (11). In addition, we present
potential scenarios for restoration priorities
to optimize the protection of 103 threatened
terrestrial vertebrates in the biome.
Biodiversity-friendly landscapes must be de-
signed to increase connectivity among habitat
patches and to maintain sufficient habitat to
assure the long-term persistence of biodiversity
(8, 12). The strategy to implement this ap-
proach varies widely across the land-sharing
and -sparing continuum and from voluntary
to mandatory actions implemented by differ-
ent countries (13). In parts of Australia, for
example, there is a well-established model in
which some rights are voluntarily relinquished
in favor of conservation under a binding legal
agreement and in exchange for economic in-
centives (14). Similar approaches are also found
in the United States and Canada (15). In Latin
America, the scheme is similar but shows a
larger participation of nongovernmental or-
ganizations (NGOs) in land purchase for con-
servation, especially in Costa Rica, Ecuador,
Argentina, and Chile (16). Otherwise, man-
datory regulations to protect a portion of
every rural property may represent a mean-
ingful strategy for conservation in largely
human-dominated landscapes. One of the
best-established examples of this policy was
implemented in the Brazilian Forest Code
almost a century ago (federal decree no. 23.793/
1934 and federal law no. 4.771/1965). It was
originally conceived under a utilitarian view
that focused on the importance of vegetation
to water resources, soil fertility, and wood
storage within rural properties (17). Even so,
the Forest Code has important implications
for biodiversity conservation today.
To enforce the sustainable use of natural
resources, the Brazilian Forest Code required
rural owners to select patches to become
legal reserves within their rural properties.
The legal reserve area varies between 20%
(criteria for most of Brazil) to 80% (for the
Amazon) of the property area. By contrast,
the location of permanent protection areas
is not eligible because they are designed to
protect geological stability (e.g., topographic
slope higher than 45°) and water resources
(e.g., areas around streams, rivers, and springs).
The existence of legal reserves and permanent
protection areas has suffered long-standing
pressure from political and economic sectors,
with reiterated attempts to change this legis-
lation during its history. As a consequence,
some changes were implemented in the 2012
Native Vegetation Protection Law (federal
law no. 12.651/2012) to maintain the general
definition for existent categories but allow
new deforestation, mainly in the Cerrado
biome. The 2012 Forest Code also demands
that landowners provide georeferenced infor-
mation about land uses and protected areas
in their rural properties through the Rural
Environmental Registry [Cadastro Ambiental
Rural (CAR)]. Here, we take the opportunity
created by CAR to analyze the spatial dis-
tribution of all private protected areas from
684,942 rural properties registered in the
Cerrado biome to assess its potential value
for the conservation of threatened vertebrate
species and to predict the potential benefits
of fully restoring set-aside areas. The Cerrado
is a wooded grassland, or savanna, covering
about 20% of Brazil. It is home to distinctive
and threatened species, such as the maned wolf
(Chrysocyon brachyurus), the giant anteater
(Myrmecophaga tridactyla), and the highly en-
dangered blue-eyed ground dove (Columbina
cyanopis). By 2018, natural vegetation loss
reached 90 million ha—45% of Cerrado area—
mostly in private property (17, 18).
All analyses are based on the overlap be-
tween model predictions of species’ ranges
and the proportion of private protected areas
with a 10-km–by–10-km cell from CAR’s polygons.
We used conservative estimates of species’
ranges based on advanced ecological niche
modeling techniques that account for dis-
persal constraints (19). Landscape-cell rele-
vance to conservation was estimated by giving
higher weight to smaller-ranged species, which
are proportionally more affected by habitat
De Marco et al., Science 380, 298–301 (2023)
21 April 2023
1 of 4
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. The proportion of distributional range of threatened vertebrate spe-
cies that falls within legal reserves and permanent protection areas in
relation to its range size in the Cerrado biome. (A to F) Historical species’
ranges that overlap legal reserves (LR) (A) and permanent protection areas
(PPA) (B), as estimated by the regression through the origin (black lines), are
13.01 and 3.21%, respectively. The null expectation (red dashed line) is that each
species overlaps both categories according to the proportion of those classes
across the whole Cerrado area (12.80% for legal reserves and 4.20% for
permanent protection areas). The predicted mean proportion of species
distribution within total private protected areas (C) is 14.48%. The portion of
species’ ranges with available remaining habitat that overlaps with legal reserves
(D) and permanent protection areas (E) is larger (black lines; 23.06 and
5.49%, respectively) and presents a larger interspecific variation but is close to
the null expectation (red dashed line). After discounting habitat loss, the
predicted mean proportion of species distributions within total private protected
areas (F) is high (25.04%) and close to the null expectation. The estimate
of the overlap (slope of black line) is indicated in each plot, together with the R2
and its statistical significance.
loss. Moreover, we assumed that species can
persist in the private protected patches inde-
pendently of their size, isolation, or type of
surrounding matrix because species-specific
sensitivity to these variables is unknown for
most of the species we evaluated (20). We
start by assuming that private protected areas
are fully restored, though a considerable part
of those set-aside areas is, at present, not well
preserved and suffers from human interfer-
ence (11). This assumption is relevant because
their restoration is mandatory even under the
current Forest Code (21). Thus, our analysis
assesses the conservation value if restoration
is properly implemented. Finally, we explore
this further by indicating where restoration
will bring higher benefits to biodiversity.
We show that an average of 13.01% of the
range of threatened species falls within legal
reserves [slope of the regression of range
within legal reserves and total species’ range
in Cerrado, forced through the origin; coeffi-
cient of determination (R2) = 0.987; Fig. 1A].
This value is only slightly higher than the null
expectation, which is the percentage cover of
legal reserves in Cerrado (12.86%). The sim-
ilarity to the null expectation suggests that
these areas are representative of the envi-
ronmental variation in the Cerrado favoring
better representation of species’ distribution
ranges of threatened vertebrates. This hypothesis
was supported both by the frequency distribution
of public and private areas in relation to first
climatic principal components analysis (PCA)
and by the overlap of the entire environmental
variation of the Cerrado (figs. S2-1 and S2-2).
The mean proportion of the predicted species
ranges that fall within permanent protection
areas is lower (3.21%; the slope of the re-
gression; R2 = 0.944; Fig. 1B). This prediction
is also slightly lower than the null expectation
that species overlap is only determined by
the proportion of this category in the whole
Cerrado (4.26%). Applying the same analysis
for the current public (federal and state lev-
els) protected area system in the Cerrado
shows that only 13.78% of species’ ranges is
protected, slightly lower than the null expec-
tation based on the coverage of public pro-
tected network in this region (15.30%; fig. S2-3).
Private protected land is more evenly distrib-
uted across the Cerrado and thus is better
suited to benefit a larger number of species.
This is a desirable quality that contrasts with
the public protection system, which is biased
toward less-favorable lands for agriculture and
does not represent the distribution of most
threatened species (10). Otherwise, private
land may represent an even higher proportion
of threatened species’ ranges if we restrict the
analysis to the available remnants of native
vegetation in the Cerrado. In that case, the
predicted mean proportion of species’ ranges
within legal reserves is 23.06% and, for per-
manent protection areas, is 5.49% (R2 = 0.851
for legal reserves and R2 = 0.756 for permanent
protection areas; Fig. 1, D and E). The general
agreement to the null expectation still holds,
but there is an increased scattering that
suggests higher interspecies variation in their
level of protection. This variation supports the
nonrandom distribution of habitat loss in the
De Marco et al., Science 380, 298–301 (2023)
21 April 2023
2 of 4
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Priority areas for the restoration of private lands based on their
contribution for threatened species conservation. (A and B) Identification
of restoration priorities within private protected areas (A) and their
distribution in the Cerrado (B) based on their relevance to threatened species
conservation, considering both the biodiversity relevance index and private
protected area size. (C) Prediction of the conservation milestones that would
be reached as an increased number of private protected areas are
restored after the prioritization established in (A). Conservation milestones
are determined in terms of the number of species that would benefit and
the percentage of their range that would be preserved. The dotted, dashed,
and solid blue lines represent the conservation targets of 10, 15, and 20% of
species range size, respectively.
Cerrado, which causes different levels of ex-
posure among species (20).
The combined effect of legal reserves and
permanent protection areas may protect up to
14.48% of species’ distribution ranges in the
Cerrado (R2 = 0.964; Fig. 1C). This prediction
increases to 25.04% after discounting for cur-
rent habitat loss outside private protected
assigned areas (R2 = 0.791; Fig. 1F). This cov-
erage is consistently higher than that expected
from the distribution of total private protected
area in the Cerrado (9.7%) and the remain-
ing habitat in these protected areas (19.7%).
Our results show that private lands may pro-
tect nearly 25% of the remaining climatically
suitable habitats for threatened vertebrates
in the Cerrado, so that its relevance for con-
servation is much higher than is now as-
sumed. This also evidences the importance
of habitat restoration within set-aside pri-
vate lands, which is expected to occur under
the current legal system. First, an increase in
habitat amount due to restoration of private
protected areas is expected to increase spe-
cies’ population sizes and reduce their risk
of extinction (22). In addition, an increase in
connectivity among remaining habitats is
expected to favor species’ dispersion and per-
sistence in the landscape (3).
Different private lands are not equally im-
portant to species conservation across the
Cerrado region. The spatial variance of the
relevance to the conservation index (Fig. 2, A
and B) shows that a small set of 192 cells,
mostly distributed in the most highly affected
São Paulo state, has a disproportional impor-
tance to conservation. Those areas are part of
the distribution of at least 70 small-ranged
species and still bear a relatively large amount
of protected land to restore within its pre-
dicted distribution. The entire set retains
nearly 145,000 ha of protected private land,
with an estimated cost of restoration not
higher than $60 million based on assisted
regeneration methods, which is only 0.02%
of the exports value of the Brazilian agri-
business sector in 2021 (https://indicadores.
agricultura.gov.br/agrostat/index.htm). An
analysis of potential scenarios for the prior-
itization of areas within set-aside private
lands shows that after ordering all Cerrado
cells according to their relevance for conser-
vation, the cumulative protected private land
area points to a positive scenario. Restoration
of the 10% top-priority cells will achieve the
goal of 10% range protection for 10 threat-
ened species, a 25% restoration will achieve
the same goal for 26 species, and a 50% res-
toration will achieve the same goal for 49
species. More ambitious conservation targets
(15 or 20% of species’ range protection; Fig. 2C)
are attained only for a small number of spe-
cies or under optimistic restoration scenarios.
For instance, restoring 75% of the protected
private land would protect 20% of the range
of 17 species, which includes many small-
ranged species that are well represented in
the Cerrado biome. We argue that an explicit
policy to assure restoration will return clear
conservation benefits based on those scenarios.
Our results support the importance of pri-
vate lands to the protection of threatened
Cerrado species. They indicate that restoring
private protected areas is an important con-
servation goal that deserves special funds and
attention. We show that, at least for the con-
servation of threatened terrestrial vertebrate
species, it is possible to devise a prioritization
scheme to guide the restoration efforts of those
areas. In addition, private protected area res-
toration would also have direct effects on
De Marco et al., Science 380, 298–301 (2023)
21 April 2023
3 of 4
RES EARCH | R E S E A R C H A R T I C L E
essential ecosystem services. Based on recent
calculations of carbon storage in Cerrado areas
(23), we made a conservative estimate that
shows that the restoration of private pro-
tected areas could capture 12 × 106 tonnes of
carbon, which is a substantial contribution
toward the 2°C climate target (24). Restor-
ing private set-aside areas may also improve
pollination services for major crops, such
as soybean, and other relevant croplands of
fruits and vegetables. Although those services
are not always recognized by private owners
(25), private land protection still carries the
possibility of increasing the visibility of its
benefits, thereby boosting restoration efforts.
The choice to dedicate land and resources for
biodiversity conservation is political and in-
fluenced by the value that people place on
biodiversity (26). Conservation in private lands
may increase the perception of ecosystem ser-
vices and promote willing-to-conserve atti-
tudes (27), thus reinforcing society’s positive
view of biodiversity conservation.
RE FE RENCES AND N OT ES
1. UN Environment World Conservation Monitoring Centre
(UNEP-WCMC), International Union for Conservation of
Nature (IUCN), National Geographic Society (NGS), “Protected
Planet Report 2018” (UNEP-WCMC, IUCN, and NGS, 2018).
2. D. Leclère et al., Nature 585, 551–556 (2020).
3. F. P. L. L. Melo, V. Arroyo-Rodríguez, L. Fahrig,
M. Martínez-Ramos, M. Tabarelli, Trends Ecol. Evol. 28,
462–468 (2013).
4. M. Duru et al., Agron. Sustain. Dev. 35, 1259–1281 (2015).
5. G. Sparovek et al., Land Use Policy 87, 104062 (2019).
6. Headwaters Economics, “A profile of land use” (Headwaters
Economics, 2018).
J. I. Watling et al., Ecol. Lett. 23, 674–681 (2020).
7. L. A. Powell, Anim. Biodivers. Conserv. 35, 295–306 (2012).
8. V. Arroyo-Rodríguez et al., Ecol. Lett. 23, 1404–1420 (2020).
9.
10. U. Oliveira et al., Sci. Rep. 7, 9141 (2017).
11. R. Rajão et al., Science 369, 246–248 (2020).
12. L. Fahrig, J. Biogeogr. 40, 1649–1663 (2013).
13. S. Kamal, M. Grodzińska-Jurczak, G. Brown, J. Environ.
Plann. Manage. 58, 576–597 (2015).
14. C. L. Archibald et al., Environ. Sci. Policy 115, 99–107 (2021).
15. J. Owley, A. R. Rissman, Land Use Policy 51, 76–84 (2016).
16. Environmental Law Institute, Legal Tools and Incentives for
Private Lands Conservation in Latin America: Building Models
for Success (Environmental Law Institute, 2003).
17. B. Soares-Filho et al., Science 344, 363–364 (2014).
18. Projeto MapBiomas, Coleção 7 da Série Anual de Mapas de
Cobertura do Uso do Solo do Brasil (2022); https://
mapbiomas-br-site.s3.amazonaws.com/Estatísticas/1_-_
TABELA_GERAL_COL7_MAPBIOMAS_BIOMAS_UF_FINAL.xlsx.
19. P. Mendes, S. J. E. Velazco, A. F. A. de Andrade, P. De Marco,
Ecol. Modell. 431, 109180 (2020).
20. P. De Marco Jr. et al., Biodivers. Conserv. 29, 1637–1658 (2020).
21. K. de Mello et al., Environ. Sci. Policy 120, 1–10 (2021).
22. J. J. O’Grady, D. H. Reed, B. W. Brook, R. Frankham,
Biol. Conserv. 118, 513–520 (2004).
23. B. Zimbres et al., For. Ecol. Manage. 499, 119615 (2021).
24. Y. M. Wei et al., Nat. Clim. Chang. 11, 112–118 (2021).
25. F. P. Lima, R. P. Bastos, Ecosyst. Serv. 40, 101029 (2019).
26. A. R. Whittaker, People Nat. 2, 450–467 (2020).
27. F. Pereira Lima, R. Pereira Bastos, Ecosyst. Serv. 44, 101121 (2020).
28. P. De Marco Jr. et al., Species distribution data and distribution
of private and public protected areas in Cerrado, Brazil. Dryad
(2023); https://doi.org/10.5061/dryad.7pvmcvdz5.
AC KNOWLED GME NTS
This work is dedicated to G. Fonseca, who spent a lifetime
dedicated to biodiversity conservation in Brazil and worldwide.
We thank E. Bruna and G. Fonseca for their invaluable
contributions to an early version of this manuscript. Funding:
P.D.M. is continuously supported by a CNPq productivity grant
(310547/2020-2). S.V.-P. was supported by Comunidad de
Madrid and Universidad de Alcalá (2017-T2/AMB-6035; 2022-
T1/AMB-237), Sustainability Victoria (CM/JIN/2019), and
Ecological Transition and Demographic Challenge (2020/00085/
001). M.C. is supported by National Science Foundation
NSF-BCS 2117533: Agri-environmental Conservation Incentives in
the Extreme Wildfire Context of the U.S. Southern Plains.
A.F.A.A. is supported by Conselho Nacional de Desenvolvimento
Científico e Tecnológico (CNPq - 165174/2020-0). Author
contributions: Conceptualization: P.D.M., R.A.S., C.C.N., M.C.;
Methods (geographical data management): R.A.S., L.M.C.;
Methods (ecological models): A.F.A.A., R.A.S., S.V.-P., C.C.N.;
Data analysis: P.D.M., A.F.A.A.; Writing – first draft: P.D.M.,
C.C.N., R.A.S.; Writing – review: P.D.M., M.C., S.V.-P. Competing
interests: The authors declare that they have no competing
interests. Data and availability: All data (species distribution
models and spatialized private and public conservation
distribution) are available from the Dryad dataset repository
(28). License information: Copyright © 2023 the authors, some
rights reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government
works. https://www.science.org/about/science-licenses-journal-
article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abq7768
Material and Methods
Figs. S1 and S2
References (29–49)
MDAR Reproducibility Checklist
Submitted 8 June 2022; accepted 1 March 2023
10.1126/science.abq7768
De Marco et al., Science 380, 298–301 (2023)
21 April 2023
4 of 4
|
10.1126_science.abq6948
|
RES EARCH
NANORIBBONS
Topologically localized excitons in single
graphene nanoribbons
Song Jiang1*, Tomáš Neuman1,2, Alex Boeglin1, Fabrice Scheurer1, Guillaume Schull1*
Intrinsic optoelectronic properties of atomically precise graphene nanoribbons (GNRs) remain largely
unexplored because of luminescence quenching effects that are due to the metallic substrate on which
the ribbons are grown. We probed excitonic emission from GNRs synthesized on a metal surface
with atomic-scale spatial resolution. A scanning tunneling microscope (STM)–based method to transfer
the GNRs to a partially insulating surface was used to prevent luminescence quenching of the
ribbons. STM-induced fluorescence spectra reveal emission from localized dark excitons that are associated
with the topological end states of the GNRs. A low-frequency vibronic emission comb is observed and
attributed to longitudinal acoustic modes that are confined to a finite box. Our study provides a path to
investigate the interplay between excitons, vibrons, and topology in graphene nanostructures.
S ince their first on-surface synthesis (1),
atomically precise graphene nanoribbons
(GNRs) have attracted tremendous in-
terest in the nanoscience and technology
communities for their topology-related
physical properties (2–6). Indeed, their specific
edge conformations host peculiar electronic
states that in turn lead to unconventional trans-
port or magnetic properties (7–12). In addition,
their optical properties hold great promise for
realizing robust and controllable atomically
thin optoelectronic devices (13). Indeed, GNRs
combine many of the outstanding character-
istics of graphene with an electronic gap, which
is a necessary property for many applications,
including light-emitting devices. Whereas the-
oretical studies discuss in great detail how
the optical properties of GNRs may be ad-
vantageously controlled through atomic-scale
variations of their width, length, and edge
shapes (14–21), experiments reporting on the
excitonic properties of GNRs are scarce (22–27),
especially those that focus on fluorescence of
on-surface grown GNRs. These experiments
either are limited to ensemble averaging mea-
surements where the light emission is domi-
nated by the response of defects (28–30) or focus
on individual GNRs in direct contact with me-
tallic electrodes that alter the GNR excitonic
properties (31). Indeed, because the synthesis
of these GNRs is performed directly on metal-
lic surfaces, which in turn causes luminescence
quenching, the intrinsic emission properties
of atomically precise GNRs remain an almost
unexplored territory.
Here, we build on a strategy that involves
using a scanning tunneling microscope (STM)
tip to transfer individual seven-atom-wide
armchair edge GNRs (7-AGNRs) from the bare
1Université de Strasbourg, CNRS, IPCMS, UMR 7504, F-67000
Strasbourg, France. 2Institut des Sciences Moléculaires
d’Orsay (ISMO), UMR 8214, CNRS, Université Paris-Saclay,
91405 Orsay Cedex, France.
*Corresponding author. Email: song.jiang@ipcms.unistra.fr (S.J.);
guillaume.schull@ipcms.unistra.fr (G.S.)
part of a Au(111) surface to a neighboring thin
insulating NaCl layer (32). Using STM-induced
luminescence (STML), we then address the
fluorescence properties of single GNRs that
are isolated from any contact with metallic
electrodes. Our STML data reveal a sharp emis-
sion line at an energy that is lower than the
excitonic emission expected for an infinitely
long ribbon and that is traced back to dark
excitons that involve topological states local-
ized at the GNR termini. This emission line is
accompanied by a rich and complex vibronic
emission spectrum.
Luminescence from decoupled GNRs
Recent STML works have demonstrated that
the fluorescence properties of individual or-
ganic molecules can be excited when they are
sufficiently decoupled from a metallic sub-
strate (33–38). A common strategy consists of
evaporating the molecule on a thin insulating
layer of oxide or salt that is adsorbed on a
metal surface. This efficient approach is,
however, not suited to long organic structures
such as GNRs whose synthesis requires a cat-
alytic reaction step at the surface of metal-
lic substrates (39). In Fig. 1, we detail our
method to measure the luminescence pro-
perties of single GNRs with subnanometer--
scale precision. We first follow the usual
on-surface synthesis approach (Fig. 1A) to
form seven-atom-wide and m-atom-long
AGNRs [(7, m)AGNRs] on a Au(111) surface
from a 10,10′-dibromo-9,9′-bianthryl (DBBA)
precursor (1) and subsequently evaporate
NaCl so as to form three-monolayer-thick
(3ML) NaCl islands on Au(111) [see section
I of (40)]. An STM image of the substrate
after such a preparation (Fig. 1B) shows a
clean NaCl island (bottom left) and several
(7, m)AGNRs of different lengths and orien-
tations located on the bare gold area. In Fig.
1C, we schematically explain how the tip of
the STM is used to transfer a (7, m)AGNR
adsorbed on the gold surface onto a NaCl
cluster (32) [see section II of (40) for details]:
(i) The tip approaches a (7, m)AGNR extremity
until contact is reached. A weak bond between
the last tip atom and the reactive ribbon ter-
minus allows the ribbon in the junction to
be lifted by retracting the tip by a few nano-
meters. (ii) The tip is then laterally displaced
on top of a NaCl cluster where (iii) the rib-
bon is released by applying a voltage pulse
of 3.5-V amplitude and 30-ms width. Figure
1D shows an STM image of a (7, 28)AGNR
that was deposited on 3ML NaCl using this
method. No modification of the ribbon can
be observed in this image (41), indicating that
the procedure does not affect the structure of
the GNR.
In Fig. 1E, we display an STML spectrum that
was acquired for the STM tip located at the
position marked by a black dot in Fig. 1D. This
spectrum reveals an intense and complex sig-
nal composed of sharp lines of an excitonic
nature that are absent for ribbons directly
adsorbed on Au(111) [see section III of (40)
for details]. This attests to the success of
our decoupling procedure. In this STML
spectrum, one first identifies an intense
0-0 line at ≈1.45 eV (855 nm), with a sub-meV
spectral width (Fig. 1F), followed by several
features of weaker intensities assigned to
vibronic emission. Whereas the vibronic
peaks at high energy (>1000 cm−1) are rem-
iniscent of Raman-like patterns that are fre-
quently observed in STML spectra (33, 36, 42, 43),
the series of equally spaced peaks at low
energy (<500 cm−1) is unusual; we discuss
this later. The energy of the 0-0 line (hn =
1.45 eV, where h is Planck’s constant and n
is frequency) is intriguingly low, because the
lowest excitonic transition is expected at
≈2 eV for a (7, ∞)AGNR (22). However, no
fluorescence contribution is observed at ener-
gies higher than the one of the 0-0 line, ir-
respective of the voltage bias used (up to 2.8 V)
or of the GNR length.
Influence of topological end states
on optical properties
To identify the origin of the 0-0 line, we display
in Fig. 2A a series of STML spectra recorded
along the main axis of a decoupled (7, 24)AGNR.
With the exception of a peak at 677 cm−1 (dis-
cussed later), all spectral contributions fade
rapidly when the tip is moved away from the
ribbon terminus. A differential conductance
(dI/dV, where I is current and V is voltage)
spectrum recorded at a GNR terminus (red
spectrum in Fig. 2B) reveals electronic states
at V = (cid:1)0.6 and 1.9 V, which correspond to
localized states of a topological nature (44)
that are absent from spectra recorded at the
center of the (7, 24)AGNR (blue spectrum in
Fig. 2B). These topological end states result
from the unsaturated electronic structure of
the sp2-hybridized carbon atoms located at
Jiang et al., Science 379, 1049–1053 (2023)
10 March 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
the center of the zigzag (7, 24)AGNR termini
(32). The similar spatial dependencies of the
optical and electronic (dI/dV) signals (i.e.,
intense at the GNR extremities and weak in
the middle) suggest that the end states are
involved in the fluorescence process. To con-
firm this hypothesis, we investigated the STML
properties of a decoupled ribbon (Fig. 2C) that
has the central carbon atom of one of its termini
bonded with two hydrogen atoms (labeled as
“CH2 terminus”). This configuration, which nat-
urally occurs for a fraction of the (7, m)AGNRs
synthesized on Au(111) surfaces, is known to
saturate the ribbon electronic structure, lead-
ing to a sp3 hybridization of the central carbon
atom and to the absence of topological state
on this side (45, 46). The STML spectrum (in
red) acquired at this CH2 terminus reveals broad
emission resonances similar to the plasmonic
emission that is measured with the same tip on
top of the Au(111) substrate (in black). By con-
trast, the spectrum acquired at the opposite
CH terminus (in blue) reveals an excitonic
emission signature. All these observations
indicate a prominent role of the topological
end states in the fluorescence process.
To elucidate the role of these end states, we
performed time-dependent density-functional
theory (TDDFT) calculations of a (7, 16)AGNR
whose left edge is “saturated,” as described
above. Owing to the open-shell nature of the
electronic structure of this half-saturated rib-
bon, the ground-state density is obtained from
a spin-unrestricted doublet DFT calculation.
We plot the corresponding frontier Kohn-Sham
orbitals and their ground-state occupations
for both spin channels in Fig. 2D. The three
orbitals on the left are shown for the majority
spin (up), and those on the right for the minority
spin (down). The Kohn-Sham energies that
correspond to these orbitals are schematically
shown in the center. The pair of occupied orbi-
tals (bottom) is reminiscent of the valence band
of infinite GNRs; conversely, the two unoccupied
orbitals (top) correspond to the GNR conduction
band. The singly occupied and unoccupied or-
bitals (middle) are clearly localized on the
unsaturated (right) edge (Fig. 2D) and can be
associated with the topological states.
We next calculated the excited states of the
ribbon using linear-response TDDFT as imple-
mented in Gaussian 16 (47) [see section IV of
(40) for more details]. We identified low-lying
excited states that can be seen as linear com-
binations of configurations where a spin-up
electron, promoted from the edge state, ap-
pears in the conduction band, and a spin-down
electron, promoted from the valence band,
appears in the empty edge state, in agreement
with earlier calculations (31). The transition
of lowest energy (1.59 eV) carries a negligible
transition dipole moment that results from the
destructive interference between the spin-up
and spin-down transition channels, and we
Fig. 1. STML from decoupled (7, m)AGNRs. (A) On-surface synthesis of (7, m)AGNRs from molecular
precursor DBBA. (B) Typical STM image (V = 0.05 V, I = 5 pA) of (7, m)AGNRs after subsequent deposition of
3ML-NaCl islands on Au(111). (C) Sketch of the STM manipulation procedure to transfer (7, m)AGNRs from
the Au(111) surface to a 3ML-NaCl island: (1) picking up the ribbon at one terminus with the STM tip,
(2) laterally moving the ribbon to the 3ML-NaCl island, (3) releasing the ribbon from the tip with a bias pulse
(V = 3.5 V, duration of 30 ms), and (4) positioning the tip on the ribbon to measure the STML spectra.
(D) STM image (V = −2.3 V, I = 5 pA) of a (7, 28)AGNR transferred onto a 3ML-NaCl island through
STM manipulation. (E) STML spectrum (V = 2.3 V, I = 100 pA, t = 120 s) acquired for the tip located on
one end [black dot in (D)] of the decoupled (7, 28)AGNR. The weak signal >500 cm−1 is shown magnified
by a factor of five in the inset (red spectrum). Raw and smoothed data appear in gray and red, respectively,
in the magnified spectrum. The vertical blue bars indicate the vibronic peak energies. (F) Highly resolved
STML spectrum (V = 2.3 V, I = 200 pA, t = 120 s) acquired from the same decoupled (7, 28)AGNR and
for the same position of the tip with a 1200–grooves mm−1 grating and a 5-mm slit (corresponding to a
spectral resolution of 0.35 meV at 1.45 eV).
therefore denote this transition as dark. The
absence of a transition dipole moment be-
comes apparent upon inspection of the cor-
responding transition charge density (D0 → D1)
shown on the left side of Fig. 2E, which also
demonstrates that the excitation is localized
on the ribbon’s nonsaturated terminus.
Interestingly, the oscillator strength of this
dark state can be activated through efficient
coupling with the picocavity plasmon con-
fined at the tip, as was suggested in previous
works (34, 48) and as we further detail in sec-
tion V of (40). We therefore assign this transi-
tion to the experimentally observed 0-0 peak at
1.45 eV. This assignment is also consistent with
the particularly narrow width of the 0-0 line
that reflects the long-lived nature of the D1 dark
state. Conversely, the same spin-dependent tran-
sition channels can also interfere constructively
and give rise to a bright transition between the
ground state and a higher-lying excited state D4
(1.82 eV). Because of the involvement of an or-
bital localized close to the nonsaturated end of
the ribbon in this excitation, its transition den-
sity (D0 → D4) is also localized on this terminus
(right side of Fig. 2E). Other dark excitations
appear in the TDDFT calculations between D1
and D4 and are discussed in section IV of (40).
Building on our TDDFT calculations and ex-
perimental observations, in Fig. 2F, we propose
a model based on a many-body representa-
tion of the GNR states to explain the mech-
anisms of the reported GNR fluorescence. If
one considers only what happens on one side
of the ribbon, the (7, m)AGNR on NaCl/Au(111)
is in a neutral ground state of doublet char-
acter, D0. At a negative voltage of ≈−0.6 V, an
electron can tunnel from the (7, m)AGNR to the
Jiang et al., Science 379, 1049–1053 (2023)
10 March 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Local excitonic emission from (7, m)AGNRs. (A) STML spectra (V = 2.7 V, I = 200 pA, t = 300 s) acquired
over the line running along the long axis of a decoupled (7, 24)AGNR, an image of which is shown in the inset (V = 2.3 V,
I = 3 pA). (B) dI/dV spectra acquired in the center (blue) and at one extremity (red) of the (7, 24)AGNR
imaged in (A). The blue and red asterisks in the inset images mark the tip positions used to record the dI/dV
spectra. Constant height dI/dV maps that were acquired at voltages corresponding to dI/dV resonances are
displayed in the insets. The voltage dependency of the 0-0 emission efficiency appears as black squares
overlaid on the red spectrum. (C) STM image (V = −2.5 V, I = 5 pA) (top) and sketch (middle) of a decoupled
(7, 20)AGNR with one CH terminus (i.e., leading to a sp2 carbon atom) and one CH2 terminus (i.e., leading
to a sp3 carbon atom). STML spectra (V = 2.0 V) acquired on each terminus (as marked with asterisks in the top
STM image) and on top of the bare Au(111) are shown at the bottom. The gray shading represents the raw data,
and the solid lines represent the smoothed data. (D) Frontier Kohn-Sham orbitals and their corresponding ground-
state occupations for both spin channels. The color represents the phase (sign) of the wave function visualized
as an isosurface (red is negative, and blue is positive). (E) Transition electron density associated with the D0 → D1
(left) and D0 → D4 (right) transitions calculated using TDDFT. The transition density, which is the oscillating
component of the electron density that is associated with the electronic transition, is shown as an isosurface, where
the color represents the sign (phase) of the density. (F) Fluorescence excitation model: At sufficiently high positive
voltage (1.85 V), the GNR can be transiently driven to its negative charge state (S
the tip to one of its topological end states. Subsequent tunneling of this charge to the substrate may leave the GNR in
one of the excited neutral states (D1 to D4) that nonradiatively relax to the lowest-lying state D1. The molecule
eventually relaxes to its ground state D0 by emitting a photon. This simple scheme assumes a negligible voltage drop
between the molecule and the substrate and remains qualitatively valid even when this drop is moderate.
(cid:1)
0 ) by charge tunneling from
0 (D0 → Sþ
tip, driving the ribbon into a positively charged
state of singlet character Sþ
0 blue
arrow). This state is only transiently populated
because the (7, m)AGNR is rapidly neutralized
back to its original state (D0) by tunneling of
an electron from the substrate (Sþ
→ D0 red
0
arrow). The topological end-state resonance at
V = −0.6 V in Fig. 2B (red spectrum) therefore
corresponds to the D0 → Sþ
0 transition. The
same reasoning applies for the other end-state
resonance at V = +1.9 V, which therefore cor-
responds to a D0 → S(cid:1)
0 transition and a tran-
siently negatively charged (7, m)AGNR. Here
as well, the ribbon may rapidly return to the
initial D0 state by tunneling of an electron to
the substrate. But because S(cid:1)
0 has an energy
(1.9 eV) that is higher than those of the excited
states (D1 to D4), S(cid:1)
→ D1 to D4 transitions
0
may occur. The fact that eventually only the
D1 → D0 emission is observed in the experi-
ment indicates fast nonradiative transitions
from Di (i ≥ 2) to the lowest excited dark state
D1 and explains why the fluorescence of the
7-AGNR appears intrinsically low in usual
photoluminescence measurements (28). The
bias onset of the D1 → D0 emission [≈1.85 V;
black squares in Fig. 2B and section VI in
(40)] matches the voltage that is required to
tunnel into the S(cid:1)
0 state, consistent with the
proposed mechanism.
Luminescence of confined acoustic vibronic
modes of the GNRs
A notable advantage of GNRs over usual chromo-
phores is that one can envisage tuning their
optoelectronic properties by changing their
length. In Fig. 3, we investigate how this parame-
ter affects the STML properties of (7, m)AGNRs
by studying ribbons made of 4 [(7, 16)AGNR]
to 15 [(7; 60)AGNR] DBBA units (Fig. 3A). For
all these (7, m)AGNRs, the energy of the 0-0
line remains essentially constant [see section
VII of (40) for more details], in agreement with
TDDFT simulations [section IV of (40)] and
with the observation of an optical transition
determined by excitons localized at the ribbon
termini. The series of vibronic peaks at low
energy (<500 cm−1) presents a different be-
havior. As shown in Fig. 3B, the energy separa-
tion between successive peaks decreases with
ribbon length; for the (7, 60)AGNR, separat-
ing the peaks from each other becomes dif-
ficult. Additionally, the number of peaks in
the comb increases up to ≈10 for the longest
ribbons. These behaviors reflect the confine-
ment of acoustic modes, hereafter referred to
as longitudinal acoustic modes (LAMs), in
the GNRs that act as one-dimensional boxes
of controllable length. The first-order LAM
(a) was identified in Raman measurements
that were performed on ensembles of length-
selected AGNRs (49); here, we report the higher-
order modes and monitor their dispersion as
a function of the GNR length. In Fig. 3C, we
report the evolution of the energy separation
between successive vibronic lines as a func-
tion of the number of carbon atoms, m, in
the long axis of the ribbons. As expected, Fig.
3C reveals a linear dispersion of the modes
with a slope corresponding to the speed of
sound (v = 18.7 km s−1) in GNRs and graphene
(49–51). Notably, DFT simulations [blue tri-
angles in Fig. 3C; see also section IV of (40)]
of the different (7, m)AGNRs reproduce al-
most perfectly the dispersion observed in
our STML spectra. This confirms the vibro-
nic peak assignment and allows us to repre-
sent the first three LAMs (a, b, and g in Fig.
3D) for a (7, 16)AGNR.
Interestingly, the envelope of the vibronic
comb is very similar for all ribbons, showing
attenuated emission for the peaks closest to
the 0-0 line, a maximum intensity reached for
peaks from 100 to 200 cm−1, and a slow decay
at higher energy. At the same time, the max-
imum intensity of the vibronic peaks reaches
only up to 50% of the 0-0 line intensity. In a
Franck-Condon picture, multiple excitations
of the same vibrational mode would also re-
sult, as in Fig. 3, in regularly spaced vibronic
peaks (33). However, the overall shape of the
vibronic comb (i.e., a dominant 0-0 line and
vibronic peaks whose intensity increases and
later decreases with detuning from the 0-0 line)
cannot be reconciled with a Franck-Condon
distribution. The data of Fig. 3 therefore in-
dicate that each vibronic peak of the comb
corresponds to single excitations of LAMs
of different orders. Besides, only modes that
have even symmetry (such as a and g) have
Jiang et al., Science 379, 1049–1053 (2023)
10 March 2023
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
been predicted to be optically active in far-
field Raman spectroscopy (49) because only
those result in a change of polarizability. This
selection rule is clearly not respected in our
experiments because all the modes, regard-
less of their parity, appear intense in the
spectra. This somewhat surprising behavior
can be attributed to the localization of the
exciton to only one end of the GNR (i.e., low-
ering the symmetry of the system), an effect
that we associate with the interaction with
the tip in our experiment. This results in a
net oscillator strength for modes of both
even and odd symmetries. Moreover, fitting
the relative intensities of the LAMs with a
Franck-Condon model [section VIII of (40)]
allows us to retro-engineer the effective de-
formation that is associated to the exciton
creation. These data suggest that the exciton
confinement is even stronger than expected
from the gas-phase TDDFT calculations, which
likely also reflects the presence of the tip. The
overall emerging picture is that of emitters
strongly localized at the topological ends of
the GNRs and Franck-Condon–coupled to de-
localized acoustic modes of the ribbon struc-
ture. Here, the spectroscopic vibronic pattern
reflects an end-localized deformation in the
excited state with respect to the ground-state
geometry.
High-energy vibronic peaks
Lastly, we discuss the length-dependency of
the high-energy (>500 cm−1) vibronic peaks
(Fig. 4). This spectral section is characteristic
of the probed material and is often referred
to as the fingerprint region. In contrast to the
low-energy peaks, the most-intense peaks here
do not shift with ribbon length. With the ex-
ception of the 677-cm−1 peak, all these peaks
can be identified based on a comparison with
calculations of the vibronic activities (Fig. 4A)
and literature (52). The 1238-cm−1 peak can be
assigned to in-plane bond-bending vibrational
modes of the C–H bonds at the GNR edges.
The 1348- and 1610-cm−1 peaks are the so-called
D and G modes, respectively, which are asso-
ciated to E2g and A1g deformations of the car-
bon rings that are found in all vibronic spectra
of sp2 carbon materials. Their intensity ratio
D/G is generally used to determine the pres-
ence of defects in meso- and macroscopic scale
systems; a low ratio indicates an overall good
structural quality of the material (53). At the
level of an individual ribbon, this ratio may be
used to tag the bonding motive of the carbon
atoms that surround the exciton. More pre-
cisely, the D mode that is a second-order process
in graphene becomes first-order in the case of
GNRs (52), which explains its high intensity in
our STML spectra.
The 677-cm−1 peak, whose intensity varies
with tip position differently from the rest of
the spectra (Fig. 2A), is not reproduced by
Fig. 3. Coupling of excitons with longitudinal acoustic modes for (7, m)AGNRs of increasing length.
(A) STM images (−2 V < V < −3 V) of decoupled (7, m)AGNRs. (B) STML spectra (2.3 V < V < 2.7 V) acquired from
the ribbons in (A). The vertical bars indicate equally spaced peaks. The first three peaks are labeled a, b, and g.
The gray shading represents the raw data, and the solid lines represent the smoothed data. (C) The experimental
(red dots) and calculated (blue triangles) average energy separation, DE, between successive vibronic peaks as a
function of m. For clarity, the error bars, which are smaller than the size of the dots, are not shown. L refers to the ribbon
length. (D) The first three LAMs (a, b, and g) calculated by DFT [see section IV of (40)] for a (7, 16)AGNR. The red
arrows indicate the normalized atomic displacement profile (the arrow lengths were scaled by a factor of 10).
Fig. 4. Fingerprint region of the (7, m)AGNR vibronic spectra. (A) High-energy region of the STML spectra
(2.3 V < V < 2.7 V) acquired from the same ribbons imaged in Fig. 3A. The gray shading represents the raw
data, and the solid lines represent the smoothed data. (B) DFT calculation of the main modes identified
in the spectra in (A). A visual representation of the high-frequency vibrational modes calculated for the (7, 16)
ribbon (only half a ribbon is shown) that can be tentatively assigned to the vibronic peaks obtained from the
spectra in (A) is shown. The vectors that indicate the vibrational modes (normalized atom displacements) have
been scaled by a factor of 10 for the C–HIR, D, and G modes and by a factor of five for the C–H mode.
simple Franck-Condon simulations. Our closed-
shell DFT calculations for ribbons of various
lengths consistently reveal a normal mode at
692 cm−1 that we tentatively associate with
the experimental peak at 677 cm−1. This mode
presents a single antinode and is of odd sym-
metry (see the C–HIR mode in Fig. 4B, where IR
refers to infrared active). We therefore suggest
that this vibrational mode promotes non-
adiabatic coupling (i.e., Herzberg-Teller active
modes) between the localized exciton and
higher-lying delocalized excitons, which explains
the nonvanishing intensity of the 677-cm−1 peak
for the tip located on top of the middle of the
GNR (Fig. 2A). Furthermore, in the experimental
spectra, weak vibronic combs on the low-energy
side of the 677- and 1348-cm−1 peaks that re-
produce the shape and energy separation re-
ported in Fig. 3 can also be distinguished. These
peaks are therefore interpreted as combination
Jiang et al., Science 379, 1049–1053 (2023)
10 March 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
bands that replicate the low-energy vibronic
progression.
Conclusions and outlook
Our atomically resolved fluorescence measure-
ments reveal sharp (≈0.6 meV) emission from
a long-lived dark exciton localized at the topo-
logical ends of (7, m)AGNRs. These localized
emitting centers are coupled to one-dimensional
acoustic phonon modes that are delocalized over
the whole ribbon. Emitting centers localized in
insulators and/or semiconductors, such as
color centers or defects in solids (54), are often
used as single- or entangled-photon sources of
particular interest for quantum sensing and
quantum technology applications. An advan-
tage of the topologically localized centers in
GNRs over more conventional solid-state quan-
tum emitters is that the number and the po-
sition of the photon sources can be tailored
through chemical engineering of the short and
long edges of the GNR, which thus provides an
efficient path to tune intersource coupling and
control the classical and quantum emission
properties. An obvious next step will be to
identify the single-photon source character
of the emission at the topologically localized
centers and to characterize their perform-
ance. In addition, each topological end state
of (7, m)AGNRs hosts an unpaired electron and
is therefore spin-polarized, thereby provid-
ing organic nanoscale solutions for quantum
schemes that combine electronic, magnetic, and
photonic degrees of freedom. These GNRs can
also be viewed as ideal atomically controlled
platforms to identify, with atomic-scale spatial
accuracy, the role of exciton-phonon cou-
pling on the (de)coherence of the quantum
units. Eventually, the two ends of the GNR could
be functionalized with specifically chosen
chromophores (55) to determine whether de-
localized acoustic phonon modes can affect the
coherent coupling between the chromophore
dipoles (34, 56), as does electronic-vibrational
mixing in light-harvesting complexes.
RE FE RENCES AND N OT ES
J. Cai et al., Nature 466, 470–473 (2010).
1.
2. W. Bo, Y. Zou, J. Wang, RSC Advances 11, 33675–33691
(2021).
3. R. S. K. Houtsma, J. de la Rie, M. Stöhr, Chem. Soc. Rev. 50,
6541–6568 (2021).
46. L. Talirz et al., J. Am. Chem. Soc. 135, 2060–2063 (2013).
47. M. J. Frisch et al., Gaussian 16 Revision C.01 (Gaussian Inc.,
4. W. Miao, L. Wang, X. Mu, J. Wang, J. Mater. Chem. C Mater.
2016).
Opt. Electron. Devices 9, 13600–13616 (2021).
5. H. Wang et al., Nat. Rev. Phys. 3, 791–802 (2021).
6. S. Song et al., arXiv:2204.12880 [cond-mat.mes-hall]
(2022).
48. T. Neuman, R. Esteban, D. Casanova, F. J. García-Vidal,
J. Aizpurua, Nano Lett. 18, 2358–2364 (2018).
49. J. Overbeck et al., ACS Nano 13, 13083–13091 (2019).
50. R. Gillen, M. Mohr, J. Maultzsch, Phys. Status Solidi B Basic
7. M. Koch, F. Ample, C. Joachim, L. Grill, Nat. Nanotechnol. 7,
Res. 247, 2941–2944 (2010).
713–717 (2012).
8. P. Ruffieux et al., Nature 531, 489–492 (2016).
9. O. Gröning et al., Nature 560, 209–213 (2018).
10. D. J. Rizzo et al., Nature 560, 204–208 (2018).
11. D. J. Rizzo et al., Science 369, 1597–1603 (2020).
12. R. E. Blackwell et al., Nature 600, 647–652 (2021).
13. Z. Liu et al., Adv. Sci. 9, e2106055 (2022).
14. L. Yang, M. L. Cohen, S. G. Louie, Nano Lett. 7, 3112–3115
(2007).
15. D. Prezzi, D. Varsano, A. Ruini, A. Marini, E. Molinari, Phys. Rev. B
51. X. Cong et al., Carbon 149, 19–24 (2019).
52. P. V. Fedotov, D. V. Rybkovskiy, A. I. Chernov, E. A. Obraztsova,
E. D. Obraztsova, J. Phys. Chem. C 124, 25984–25991
(2020).
53. A. C. Ferrari, D. M. Basko, Nat. Nanotechnol. 8, 235–246 (2013).
54. G. Zhang, Y. Cheng, J.-P. Chou, A. Gali, Appl. Phys. Rev. 7,
031308 (2020).
55. J. Li et al., Nano Lett. 19, 3288–3294 (2019).
56. J. Doležal et al., ACS Nano 16, 1082–1088 (2022).
57. T. Cao, F. Zhao, S. G. Louie, Phys. Rev. Lett. 119, 076401
77, 041404 (2008).
(2017).
16. D. Prezzi, D. Varsano, A. Ruini, E. Molinari, Phys. Rev. B 84,
58. D. G. de Oteyza, T. Frederiksen, J. Phys. Condens. Matter 34,
041401 (2011).
17. J. Alfonsi, M. Meneghetti, New J. Phys. 14, 053047
(2012).
18. C. Attaccalite, E. Cannuccia, M. Grüning, Phys. Rev. B 95,
125403 (2017).
19. T. Deilmann, M. Rohlfing, Nano Lett. 17, 6833–6837
(2017).
20. C. Cardoso, A. Ferretti, D. Prezzi, Eur. Phys. J. B 91, 286
(2018).
21. X. Wu et al., Phys. Chem. Chem. Phys. 22, 8277–8283
(2020).
22. R. Denk et al., Nat. Commun. 5, 4253 (2014).
23. G. Soavi et al., Nat. Commun. 7, 11010 (2016).
24. C. Bronner, D. Gerbert, A. Broska, P. Tegeder, J. Phys. Chem. C
120, 26168–26172 (2016).
25. G. Borin Barin et al., ACS Appl. Nano Mater. 2, 2184–2192
(2019).
26. S. Zhao et al., Nano Lett. 20, 1124–1130 (2020).
27. A. Tries et al., Nano Lett. 20, 2993–3002 (2020).
28. B. V. Senkovskiy et al., Nano Lett. 17, 4029–4037
(2017).
29. M. Pfeiffer et al., Nano Lett. 18, 7038–7044 (2018).
30. C. Ma et al., ACS Nano 14, 5090–5098 (2020).
31. M. C. Chong et al., Nano Lett. 18, 175–181 (2018).
32. S. Wang et al., Nat. Commun. 7, 11507 (2016).
33. X. H. Qiu, G. V. Nazin, W. Ho, Science 299, 542–546
(2003).
34. Y. Zhang et al., Nature 531, 623–627 (2016).
35. H. Imada et al., Nature 538, 364–367 (2016).
36. B. Doppagne et al., Science 361, 251–255 (2018).
37. J. Doležal et al., Nano Lett. 19, 8605–8611 (2019).
38. K. Kimura et al., Nature 570, 210–213 (2019).
39. S. Clair, D. G. de Oteyza, Chem. Rev. 119, 4717–4776
(2019).
443001 (2022).
59. S. Jiang, T. Neuman, A. Boeglin, F. Scheurer, G. Schull, Data
for “Topologically localized excitons in single graphene
nanoribbons.” Zenodo (2022); https://doi.org/10.5281/
zenodo.7611826.
AC KNOWLED GME NTS
We thank A. Rosławska for fruitful discussions and V. Speisser and
M. Romeo for technical support. Funding: This project has
received funding from the European Research Council (ERC) under
the European Union’s Horizon 2020 research and innovation
program (grant agreement no. 771850). The International Center
for Frontier Research in Chemistry (FRC) is also acknowledged
for financial support. This work has been supported by the University
of Strasbourg’s IdEx program and by the “Investissements
d’Avenir” LabEx PALM (ANR-10-LABX-0039-PALM). Author
contributions: S.J., F.S., and G.S. conceived, designed, performed,
and analyzed the experiments. T.N. and A.B. developed, performed,
and analyzed the theoretical simulations. All the authors discussed
the results and contributed to the revision of the paper.
Competing interests: The authors declare no competing interests.
Data and materials availability: All data needed to evaluate the
conclusions in the paper are present in the paper or the
supplementary materials. Data for all figures presented in this
study are available on Zenodo (59). The following software was
used: COMSOL 5.6, ACDC module to calculate the plasmonic
potential of the tip, Gaussian 16 version C.01 to perform ab initio
calculations on the GNRs, Jmol 14.32.15 for visualization, and
Matlab R2021a for data postprocessing. License information:
Copyright © 2023 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science. No
claim to original US government works. https://www.science.org/
about/science-licenses-journal-article-reuse
40. See supplementary materials.
41. Y.-Y. Sung et al., ACS Nano 16, 10281–10291 (2022).
42. B. Doppagne et al., Phys. Rev. Lett. 118, 127401
(2017).
43. F.-F. Kong et al., Nat. Commun. 12, 1280 (2021).
44. These states are termed topological because the one-
dimensional nanoribbon has a nontrivial Z2 invariant that
originates from its Zak phase (57, 58) and the end states thus
emerge at the interface between the ribbon and a topologically
trivial vacuum.
45. J. van der Lit et al., Nat. Commun. 4, 2023 (2013).
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abq6948
Materials and Methods
Supplementary Text
Figs. S1 to S8
Table S1
References (60–65)
Submitted 16 May 2022; accepted 6 February 2023
10.1126/science.abq6948
Jiang et al., Science 379, 1049–1053 (2023)
10 March 2023
5 of 5
|
10.1126_science.adc9487
|
RES EARCH
MAGNETISM
Proximate deconfined quantum critical point
in SrCu2(BO3)2
Yi Cui1†, Lu Liu2,3†, Huihang Lin1†, Kai-Hsin Wu4, Wenshan Hong2, Xuefei Liu1, Cong Li1, Ze Hu1,
Ning Xi1, Shiliang Li2,5,6, Rong Yu1,7*, Anders W. Sandvik4,2*, Weiqiang Yu1,7*
The deconfined quantum critical point (DQCP) represents a paradigm shift in quantum matter
studies, presenting a “beyond Landau” scenario for order-order transitions. Its experimental
realization, however, has remained elusive. Using high-pressure 11B nuclear magnetic resonance
measurements on the quantum magnet SrCu2(BO3)2, we here demonstrate a magnetic field–
induced plaquette singlet to antiferromagnetic transition above 1.8 gigapascals at a notably low
temperature, Tc ≃ 0.07 kelvin. First-order signatures of the transition weaken with increasing
pressure, and we observe quantum critical scaling at the highest pressure, 2.4 gigapascals.
Supported by model calculations, we suggest that these observations can be explained by
a proximate DQCP inducing critical quantum fluctuations and emergent O(3) symmetry
of the order parameters. Our findings offer a concrete experimental platform for investigation
of the DQCP.
T he theoretically proposed deconfined
quantum critical point (DQCP) (1) con-
nects two different ordered ground states
of quantum matter by a continuous quan-
tum phase transition (QPT). This type of
criticality, which has been explored primarily
in the context of two-dimensional (2D) quan-
tum magnets (2), lies beyond the conventional
paradigm of discontinuous (first-order) tran-
sitions between ordered phases with unrelated
symmetries. The DQCP is associated with un-
conventional phenomena, including fractional
spinon excitations and deconfined gauge fluc-
tuations (3–5). Further investigations have
extended the concept, introducing emergent
symmetries (6–11) and exotic first-order tran-
sitions (12, 13). In a very recent scenario, the
DQCP is proposed to be a multicritical point
(14, 15) connected to a gapless quantum spin
liquid (QSL) (16–20).
Although DQCP phenomena are broadly
relevant in quantum materials (21), there
has been no supportive experimental identi-
fication in any system. Quantum magnets in
which the interactions can be varied over a wide
enough range to realize two phases bordering
a DQCP are rare. An exception is the layered
1Department of Physics and Beijing Key Laboratory of Opto-
electronic Functional Materials and Micro-nano Devices,
Renmin University of China, Beijing 100872, China. 2Beijing
National Laboratory for Condensed Matter Physics and
Institute of Physics, Chinese Academy of Sciences, Beijing
100190, China. 3School of Physics, Beijing Institute of
Technology, Beijing 100081, China. 4Department of Physics,
Boston University, Boston, MA 02215, USA. 5School of
Physical Sciences, Graduate University of the Chinese
Academy of Sciences, Beijing 100190, China. 6Songshan
Lake Materials Laboratory, Dongguan, Guangdong 523808,
China. 7Key Laboratory of Quantum State Construction and
Manipulation (Ministry of Education), Renmin University
of China, Beijing 100872, China.
*Corresponding author. Email: rong.yu@ruc.edu.cn (R.Y.);
sandvik@bu.edu (A.W.S.); wqyu_phy@ruc.edu.cn (W.Y.)
†These authors contributed equally to this work.
material SrCu2(BO3)2 (22–24), in which anti-
ferromagnetic (AFM) Heisenberg interactions
between the S = 1/2 Cu2+ spins (Fig. 1A) pro-
vide a faithful realization of the 2D Shastry-
Sutherland model (SSM) (25). In the SSM, three
different T = 0 phases are well established to
form as a function of the ratio g = J/J′ of the
inter- to intradimer couplings (26, 27): an exact
dimer-singlet phase (DS, with singlets on the J′
bonds), a twofold degenerate plaquette-singlet
(PS) phase (Fig. 1B), and a Néel AFM phase (Fig.
1C). At ambient pressure, SrCu2(BO3)2 is well de-
scribed by the g ≃ 0.63 SSM with a DS ground
state (24). An applied pressure increases g, driv-
ing the system into a PS phase at P ≃ 1.8 GPa
(28, 29), which persists with transition tem-
perature TP ≃ 2 K up to P ≃ 2.6 GPa (30, 31).
An AFM phase with Néel temperature TN from
2.5 to 4 K has been detected between 3.2 and
4 GPa (30).
Here, we report a 11B nuclear magnetic reso-
nance (NMR) study of SrCu2(BO3)2 in a mag-
netic field H up to 15 T and at pressures up to
2.4 GPa, aiming to characterize the field-
driven PS-AFM transition. At 2.1 GPa, PS
and AFM transitions are resolved using their
NMR signatures and merge at a common
point (HC,TC), with HC ≃ 6 T and TC ≃ 0.07 K
(Fig. 1D). Such a low TC in relation to TP and
TN farther away from HC indicates proximity
to a TC = 0 QPT. First-order discontinuities at
(HC,TC) weaken with increasing pressure, and
we observed quantum-critical scaling of the
spin-lattice relaxation at 2.4 GPa for T > TC.
Our results support the existence of a multi-
critical DQCP controlling the quantum fluctu-
ations at 2.4 GPa, with TC on the associated
first-order line suppressed by an emergent O(3)
symmetry of the combined scalar PS and O(2)
AFM order parameters (7, 8). By synthesizing
past and present experiments on SrCu2(BO3)2
and model calculations, we arrived at the global
phase diagram depicted in Fig. 2. Before fur-
ther discussing the DQCP scenario, we present
our NMR detection of the various phases and
transitions.
NMR identification of phases
We performed 11B NMR measurements on
SrCu2(BO3)2 single crystals at pressures of up
to 2.4 GPa in fields between 0.2 and 15 T and
temperatures down to 0.07 K. Experimental
details are provided in the supplementary ma-
terials (33). We first discuss NMR line shifts
to detect the relevant quantum phases and tran-
sitions, followed by results of the spin-lattice
relaxation rate 1/T1.
A typical 11B NMR spectrum, shown in Fig.
3A, has a central peak with four satellite peaks
on either side, from inequivalent sites B1 to B4
(Fig. 1A) caused by a small tilt angle between
the field and the crystalline c axis (33). The
satellites are sensitive to changes of the lattice
structure because of the local coupling be-
tween the nuclear quadrupole moment and
the electric field gradient (33). As shown at a
low field and P = 2.1 GPa in Fig. 3B, the full-
width at half maximum (FWHM) height of the
satellites increases on cooling <10 K until a
maximum at T ≃ 3 K, reflecting increasing lat-
tice fluctuations when the spins form fluc-
tuating plaquette singlets above the ordered
PS phase (31). This PS liquid crosses over to
the trivial paramagnetic (PM) state at higher
temperature.
Below 1.8 K, the FWHM in Fig. 3B rises sharp-
ly and saturates around 1 K. As explained in
section S2 of the supplementary materials (33),
the rapid broadening follows from an orthog-
onal lattice distortion when a full-plaquette
(FP) PS state (Fig. 1B) forms. The FWHM as a
proxy for the PS order parameter is further
corroborated by the consistency of TP ≃ 1.8 K
at the low field applied in Fig. 3B with the lo-
cation of a sharp specific-heat peak (30, 31),
marked in Fig. 1D.
Figure 3C shows the evolution of the central
peak with T at P = 0.9 GPa and H = 4 T. The
negative Knight shift at the higher tempera-
tures reflects the hyperfine coupling Ahf ≃
–0.259 T/mB [see section S3 of the supple-
mentary materials (33)] for H∥c (34, 35). The
shift increases rapidly below T* ≃ 7 K when
dimer singlets form in the DS state. At 2.1 GPa
(Fig. 3E), PS order forms < 2 K, but the Knight
shift changes rapidly at T ≃ 4 K when the PS
liquid forms.
The first-order transition between the DS
phase and the PS or PS liquid phase termi-
nates at an Ising-type critical point, which at
H = 0 is located at P ≃ 1.9 GPa, T ≃ 3.3 K (31).
At low T, the DS-PS transition takes place
between 1.7 and 1.8 GPa (30). The first-order
DS line must therefore bend slightly, as indi-
cated in Fig. 2A, and can be crossed versus T at
fixed P and H. Indeed, at 1.85 GPa (Fig. 3D),
Cui et al., Science 380, 1179–1184 (2023)
16 June 2023
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
A
D
)
K
(
T
B
C
J'
J
T * (11T -1
1 )
T * (11Kn)
TP (11T -1
1 )
TP (FWHM)
TN (11T -1
1 )
PM
Cv data
2.10 GPa
2.00 GPa
2.20 GPa
2.28 GPa
2.1 GPa
PS liquid
PS
AFM
5
10
H (T)
15
5
4
3
2
1
0
0
Fig. 1. Experimental overview. (A) Atomic struc-
ture of a SrCu2(BO3)2 plane. Pairs of Cu2+ ions form
spin dimers (ellipses) with Heisenberg intradimer
(J′) and interdimer (J) interactions (black dashed
line). Each unit cell contains four B ions, for which
we investigated the NMR response. (B) The PS
phase in the equivalent square lattice of J (blue)
and J′ bonds (red). In SrCu2(BO3)2, the singlets
(shading) form on the full (J′) plaquettes in one of
two symmetry-equivalent patterns, whereas in
the SSM, the singlets form on half of the empty
plaquettes. (C) The AFM phase, which breaks O(3)
symmetry when H = 0 and O(2) symmetry when
H ≠ 0. For SrCu2(BO3)2 in a c-axis field, we found
that the moments ordered along the a or b axis.
(D) Field-temperature phase diagram at 2.1 GPa,
showing the PM, PS liquid, ordered PS, and
AFM phases resolved by our NMR measurements
(Figs. 3 to 5). The transition temperatures TP
and TN and the crossover temperature T* are
compared with specific-heat measurements
(30, 31). The data for 2.1 GPa come from (30)
and the rest from (31). The red box marks the
regime analyzed in Fig. 5F.
the central peak is split at temperatures be-
tween 3 and 4 K, indicating phase coexistence.
Previously, a different splitting was reported at
2.4 GPa (35, 36), which was perhaps caused by
pressure inhomogeneity, but we observed the
double peak only at 1.85 and 1.95 GPa [see
section S3 of the supplementary materials (33)].
Outside of this pressure range, T* likely only
marks a rapid crossover between the PM and
PS liquid, with associated sharp specific heat
peaks (30, 31) which were observed also away
from the critical point and reproduced (31) by
SSM calculations. We have found no NMR
signatures of a structural transition here or
at higher temperatures [see section S3 of the
supplementary materials (33)].
Above 1.95 GPa, AFM order emerges at high
fields and leads to splitting of the NMR central
A
H
emergent O(3)
DQCP
DS
PS
AFM
g (P)
T
B
PS liquid
Ising critical
point
PM
T
PM
PS liquid
PS
AFM
H
quantum
critical fan
emergent O(3)
AFM
g (P)
Fig. 2. Schematic phase diagram and DQCP
scenario. (A) Phases in the space of coupling
[g = J/J′ in the SSM, P in SrCu2(BO3)2], temper-
ature, and magnetic field. A multicritical DQCP
separates a line of first-order QPTs and either
a QSL phase (17) or a line of generic DQCPs
(38); the region marked with dashed lines repre-
sents this undetermined feature. The first-order DS
transition at fixed H (solid green line) terminates
at an Ising critical point [green circle, based on
previous experiments (31)]. The green dashed lines
indicate crossovers at T*(g) into the PM phase.
The dashed orange line shows how the slightly
curved first-order DS transition line can be
crossed versus T at fixed P (based on present
experimental data). The ordered AFM phase
at T > 0 requires interlayer couplings, as in SrCu2
(BO3)2. The magnetization plateau states at
larger H (28, 32) are not shown. (B) Phase diagram
drawn to highlight (H,T) planes exemplified
by Fig. 1D. Red crosses indicate TC > 0 caused
by weak 3D effects and violations of O(3)
symmetry. The shading represents the “fan” in
which quantum critical scaling is expected
(supported by present experimental data). The
blue dashed lines indicate the plane of highest-
pressure (2.4 GPa) measurements.
peak by alternating positive and negative hyper-
fine fields, as shown at 2.1 and 2.4 GPa in Fig. 4,
A and B, respectively, both at T = 0.07 K. The
sudden rise with field of the peak-splitting fR – fL
(a proxy AFM order parameter), shown in Fig.
4C, signals a discontinuous onset of AFM order
at HC(P), with the discontinuity much weaker
at the higher pressure.
In the AFM state, the uniform magnetization
does not exhibit any obvious discontinuity at
HC and remains < 2% of the saturated moment
at our highest field of 15 T (fig. S5). A crossover
temperature T * persists also at high fields,
where the PS liquid develops increasing spin
fluctuations (discussed further below).
Spin-lattice relaxation rate
1/T1 is a direct probe of low-energy spin fluc-
tuations and can detect the PS and AFM
transitions more precisely than the line shifts;
the two probes give consistent results. Figure
5, A and B, show 1/T1 at P = 2.1 GPa for a
wide range of applied fields that we grouped
into those below and >6.2 T, corresponding,
respectively, to the low-T PS and AFM phases;
Fig. 5, C and D, show the same at 2.4 GPa
with the separation at 5.8 T.
At 2.1 GPa (Fig. 5B), we found a sharp drop
of 1/T1 at T* ≃ 3 to 4 K and a broad peak or
sharper kink <2 K. At low fields in Fig. 5A, the
latter feature appeared up to 6.1 T and clearly
marked the opening of a spin gap below TP. At
P = 2.4 GPa, we did not find a peak at TP (Fig.
5C), but rather a sharp crossover from a low-T
gapped regime to a window with power-law
behavior that is analyzed in Fig. 5E and will be
further discussed below. At the higher fields in
Fig. 5, B and D, the low-T features (<0.8 K) are
much sharper and coincide with the NMR peak
splitting in Fig. 4, A and B. Thus, we can safely
identify these peaks for H ≥ 6.33 T as TN (37).
The minimum in 1/T1 around 1.5 K in Fig. 5B
increases with the field, indicating increasing
spin fluctuations in the PS liquid state.
Figure 5F shows very clear field-induced PS-
AFM transitions revealed by these signals at
both P = 2.1 and 2.4 GPa. The PS and AFM
boundaries, TP(H) and TN(H), respectively,
meet at a very low TC. Given phase coexistence
(Fig. 4, A and B), the QPT at HC is clearly first
order. The proxy AFM order parameter fR – fL
in Fig. 4C is consistent with HC determined
from 1/T1 at both pressures. The much smaller
first-order discontinuity of fR – fL at the higher
pressure indicates the approach toward a con-
tinuous QPT.
We extracted the PS spin gap D by fitting 1/T1
below TP to a semi-empirical form T (cid:1)ae(cid:1)D=kBT
with a ≈ 1 [see section S5 of the supplementary
materials (33)]. As expected, a linear decrease
with H of D at both pressures is revealed in
Fig. 6A and is caused by the field lowering of
the S = 1 (Sz = 1) state above the singlet PS
ground state. These results are compatible with
previously determined H = 0 gap estimates
(29, 30) and the known g factor.
At a first-order transition into the AFM
phase, the PS gap should jump discontinuously
Cui et al., Science 380, 1179–1184 (2023)
16 June 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
Center
1.8 GPa
4 T
1.54 K
11
C
10
f1
6
4
2
0
A
)
s
t
i
n
u
.
b
r
a
(
y
t
i
s
n
e
t
n
I
1
1
B
)
z
H
k
(
M
H
W
F
SL4
SL3&2
SL1
SR2&3
SR1
SR4
-1.4 -1.2
-0.2 0.0
f - H (MHz)
0.2
1.2
1.4
2.1 GPa
0.2 T Satellites
TP
SR1
SR2
SR3
SR4
SR1 (DR)
25
20
15
10
)
s
t
i
n
u
.
b
r
a
(
T
*
y
t
i
s
n
e
t
n
I
1
1
9
8
7
6
5
4
3
2
1
D
0.9 GPa
4 T
Center
10.0 K
9.0
8.0
7.0
6.0
5.0
4.0
3.0
2.5
2.0
f1
f2
1.85 GPa
4 T
Center
6.0 K
5.0
4.0
3.9
3.8
3.7
3.6
3.5
3.4
3.3
3.2
3.1
3.0
2.9
E
f2
2.1 GPa
5 T
Center
6.0 K
5.0
4.5
4.2
4.0
3.5
3.0
2.5
2.3
2.0
1.5
1.0
5
0
1
2
3
4
5
T (K)
6
7
8
9 10
0
-0.2
-0.1
0.0
f - H (MHz)
0.1
-0.2
-0.1
0.0
0.1
f - H (MHz)
0.2
-0.2
0.0
0.2
f - H (MHz)
0.4
Fig. 3. NMR spectra and line shifts. (A) NMR 11B spectrum at H = 4 T and
P = 1.8 GPa, with the field applied at 8.6° from the crystalline c axis, showing the
center line and two sets of satellites associated with the four B sites (Fig. 1A).
(B) FWHM of satellites SR1 to SR4 shown as a function of temperature at
P = 2.1 GPa and H = 0.2 T. The line at 1.8 K marks the onset of an upturn with
further cooling. SR1 was measured in the dilution refrigerator in addition to the
regular helium cryostat used for all cases [see section S2 of the supplementary
materials (33)]. (C to E) NMR center line for a range of temperatures (curves
shifted vertically for clarity) at P = 0.9 GPa and H = 4 T (C), 1.85 GPa and
4 T (D), and 2.1 GPa and 5 T (E). The peaks in the DS phase (C) and in and above
the PS phase (E) are marked f1 and f2, respectively. The split peak in (D) reflects
phase coexistence.
A
)
s
t
i
n
u
.
b
r
a
(
y
t
i
s
n
e
t
n
I
1
1
4
3
2
1
0
f
L
f
R
2.1 GPa
0.07 K
15.0 T
13.5
7.00
6.50
6.40
6.30
6.28
6.22
6.20
6.10
6.00
1.0
-0.5
0.0
0.5
f - H (MHz)
B
)
s
t
i
n
u
.
b
r
a
(
y
t
i
s
n
e
t
n
I
1
1
8
6
4
2
0
2.4 GPa
0.07 K
f
L f
R
C
6.5 T
0.15
6.4
6.3
6.1
6.0
5.9
5.8
5.7
5.6
-0.5
0.0
f - H (MHz)
0.5
)
z
H
M
(
L
f
-
R
f
0.10
0.05
2.1 GPa
2.4 GPa
0.00
0.0 0.2 0.4 0.6 0.8 1.0
H - H
c (T)
Fig. 4. AFM transition. Splitting of the NMR center line with increasing H at T = 0.07 K is shown in
(A) and (B) for P = 2.1 and 2.4 GPa, respectively. The two peaks marked fL and fR (red bars) indicate AFM
order developing above H ≃ 6 T. A center peak (blue bars) remaining at fields up to 6.1 T indicates phase
coexistence. (C) Proxy AFM order parameter fR – fL versus H – HC, where HC is determined using the
spin-lattice relaxation rate (Fig. 5).
to zero at HC (given that the AFM state is
gapless), but despite the clear first-order sig-
nals described above (Fig. 4C), we found that
D(HC) values were indistinguishable from
zero within statistical errors. We will discuss
the anomalously small gap discontinuity in
the context of the proximate DQCP scenario
further below.
Deconfined quantum criticality
The SSM at H = 0 has been a candidate for a
DQCP separating its coupling-induced PS and
AFM ground states (7, 38). The singlets in the
PS phase of the model occupy the empty pla-
quettes, in contrast to the FP state in SrCu2(BO3)2
(Fig. 1B). This aspect of the PS state depends
sensitively on other possible weak interactions
beyond the SSM (11, 39), and the SSM descrip-
tion of the global phase diagram of SrCu2(BO3)2
should remain valid.
There is mounting theoretical evidence that
a gapless QSL phase can exist between a PS
state (or closely related spontaneously dimer-
ized state) and the AFM state in frustrated 2D
quantum spin systems (16, 20, 40–42) and that
these QSL phases generically end at multicrit-
ical DQCPs (15, 17, 18). Beyond such a point, the
transition without intervening QSL is expected
to be first order, with the coexistence state at
H = 0 inheriting (and breaking) the emergent
O(4) or SO(5) symmetry, depending on the
type of singlet-ordered phase (7, 8, 10, 12, 13) of
the DQCP.
In the H = 0 SSM, early calculations indicated
a first-order PS-AFM transition (27), and a
recent calculation suggested an O(4) [from the
O(3) AFM and scalar PS order parameters]
multicritical DQCP in an extended parameter
space (11). A generic O(4) DQCP had previously
been proposed (38). The intervening gapless
QSL between the PS and AFM phases was iden-
tified very recently (17, 19) and may be explained
by an instability of the conventional DQCP (15).
These theoretical insights, along with our NMR
results for SrCu2(BO3)2, suggest the scenario in
Fig. 2. Because no experiment so far (including
ours) has explicitly confirmed a QSL phase,
the possibility remains that there is instead
Cui et al., Science 380, 1179–1184 (2023)
16 June 2023
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
E
A
B
1
1
1
P
N
1
-
1
-
3
8
4
6
5
2
4
T
T
0.8
1.2
0.0
0.4
H
b
0.1
T *
T
1
1
T
1
1
T
1
1
100
101
102
)
1
-
s
(
)
1
-
s
(
)
1
-
s
(
~ T 0.2
H
b
+
1
-
2.1 GPa
2.1 GPa
2.4 GPa
1
T (K)
6.22 T
6.2
6.1
6.0
5.8
5.5
5.0
4.0
3.0
4.0 4.5 5.0 5.5
H (T)
8.50
7.00
6.50
6.35
6.33
6.22
6.20
Fig. 5. Spin-lattice relaxa-
tion. 1/11T1(T) was measured
at 2.1 GPa [(A) and (B)]
and 2.4 GPa [(C) and (D)],
which are separated to show
the PS [(A) and (C)] and
AFM [(B) and (D)] states.
The drop in 1/T1 at a T* ≃ 4 K
[(A) and (B)] indicates the
sharp crossover into the PS
liquid. The peaks at lower
T mark TP and TN, with
uncertainties indicated by the
horizontal bars. At 2.4 GPa,
no low-T PS peak is observed
(C), but TP can be extracted
from the sudden change
from thermally activated to
quantum critical behavior,
1/T1 = aTh – bH. (E) Inset:
Power-law scaling of the
offset, bH º (HC – H)d with
d ≈ 0.8, close to HC. Main
panel: The common scaling
form with constant a and
h ≈ 0.2 is demonstrated,
and bH has been added. (F) Low-temperature phase diagrams at 2.1 and 2.4 GPa. The solid and dotted lines indicate the phase boundaries modeled by,
respectively, a logarithmic form of TP and near-critical forms of both TP and TN [see section S6 of the supplementary materials (33)]. The latter fits give the TC
values indicated with circles.
15.0 T
12.0
11.0
9.50
1
T (K)
5.7 T
5.6
5.5
5.4
5.0
4.5
4.0
3.0
5.7 T
5.6
5.5
5.4
5.0
4.5
4.0
7.5
8.0
8.5
9.0
10.0
5.8 T
5.9
6.0
6.1
0.4 0.5 0.6 0.7
4
H (T)
6.2
6.3
6.5
6.8
7.0
F
2.0
2.4 GPa
2.1 GPa
2.4 GPa
2.4 GPa
T (K)
T
P
T
N
T
P
T
N
1
0.1
T (K)
T (K)
~T 0.2
)
1
-
s
(
)
1
-
s
(
AFM
(
T
T
1
1
T
1
1
PS
0.2
0.0
0.5
0.1
1.0
1.5
0.1
0.3
10
6
5
10
D
C
K
T
T
1
1
2
2
8
0
6
1
2
4
3
1
-
1
-
)
N
P
1
1
another line of PS–AFM transitions. Although
the dashed regions in the phase diagrams in
Fig. 2 can represent either possibility, specific
heat measurements at H = 0 (30, 31) found no
phase transition between 2.6 and 3.2 GPa, con-
sistent with a QSL ground state evolving into
the T > 0 PS liquid.
A putative multicritical DQCP at H > 0 should
evolve from a corresponding H = 0 DQCP with
emergent O(4) symmetry (7, 38). Although this
O(4) point exists only in an extended parameter
space outside of the (g, H, T) cube in Fig. 2, the
fact that the field-induced magnetization is very
small at HC [see section S4 of the supplementary
materials (33)] suggests that the putative H > 0
DQCP still hosts an approximate O(4) symmetry,
with stronger O(3) character developing on
the first-order line. Strictly speaking, at H > 0,
the DQCP may evolve into a near-critical triple
point with first-order signatures at the lowest
energy scales.
Closer proximity of SrCu2(BO3)2 to some con-
tinuous QPT with increasing pressure is cer-
tainly supported by our observation of a weaker
discontinuity of the AFM order parameter at
2.4 GPa than at 2.1 GPa (Fig. 4C). Moreover, at
a clearly first-order transition, corresponding-
ly high TC values would normally be expected.
The low TC at both pressures then point to a
mechanism suppressing long-range order rather
far away from the QPT. The DQCP scenario
offers this possibility through its emergent
continuous symmetry inherited (at least up
to some large length scale) by the first-order
line. An ideal 2D coexistence state with con-
tinuous order parameter symmetry must have
TC = 0, but weak violations of the symmetry
[in combination with 3D effects (43)] would
imply a low TC > 0, as observed in SrCu2(BO3)2.
In the scenario of a first-order transition with
emergent O(3) symmetry, the Ising-type PS
order can be understood as an uniaxial defor-
mation of the O(3) order parameter. A logarith-
mic form of the PS transition temperature is then
expected: TP º ln–1 [a(HC – H)] for some value
of a (7, 44). Fits of the experimental data to this
form [see section S6 of the supplementary
materials (33)] are shown with solid curves in
Fig. 5F and indeed describe the behavior close
to HC.
To describe TN(H), we note again that inter-
layer interactions are required for TN > 0 in a
spin-isotropic system. These 3D couplings also
change a continuous QPT (TC = 0) into a first-
order line extending to a bicritical or triple
point at TC > 0 (38, 43) (red crosses in Fig. 2B).
Given the extremely low TC values in SrCu2(BO3)2,
a modified critical form with the same exponent
f governing both transitions above TC may be
expected from DQCP dualities (13, 45): TP,N =
TC + aP,N|H – HC|f. Fits with independent ex-
ponents f for the PS and AFM transitions [see
section S6 of the supplementary materials (33)]
indeed support a common value and motivate
joint fitting with a single f. Such fits are shown
by the dashed curves in Fig. 5, where TC is in
the range of 0.05 to 0.07 K at both pressures.
At 2.1 GPa, HC = 6.183 ± 0.007 and f = 0.57 ±
0.02, and at 2.4 GPa, HC = 5.719 ± 0.007 and f =
0.50 ± 0.04. These fits in which f is close to
estimates for both SO(5) (12, 14) and O(4) (45)
DQCPs [see section S6A of the supplementary
materials (33)] do not rule out the alternative
logarithmic form of TP but do further validate
the very low TC values and common transition
field HC for both order parameters.
Quantum-critical relaxation
As shown in Fig. 5C, 1/T1 at 2.4 GPa exhibits T h
scaling with h ≈ 0.2 within a window of tem-
peratures for several fields close to HC on the PS
side. The ensemble of fits is further analyzed in
Fig. 5E using the expected quantum-critical form
1/T1 = aT h – bH (46), where a is a constant and
bH > 0 for H < HC. The fact that scaling behav-
ior is not observed at 2.1 GPa (Fig. 5A) suggests
that only the system at 2.4 GPa is sufficiently
close to a continuous QPT that it realizes the
quantum critical fan (46). This is depicted in
Fig. 2B, where T is the largest energy scale (but
low enough so that the correlation length is
well above the lattice constant). The value of h is
compatible with an estimate for an O(4) DQCP
(45) and slightly below the SO(5) value (2, 12).
On the AFM side (Fig. 5D), 1/T1 is dominated
by the 3D effects causing T > 0 AFM order, with
Cui et al., Science 380, 1179–1184 (2023)
16 June 2023
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
A
)
K
(
B
K
D
)
p
m
(
P
8
6
4
2
0
3
2
1
2.1 GPa
2.4 GPa
2.15 GPa
2.4 GPa
B
h
0.8
0.4
0.0
Q = 5
AFM
PS
C
0.6
0.5
0.4
0.3
0.2
0.1
0 1 2 3 4 5 6
H (T)
0.16
0.18
g
0.20
0.22
0.0
0.0
0.2
0.6
0.4
h
h = 0.55
E
)
p
m
(
P
3
2
1
h = 0.62
h = 0.69
F
)
p
m
(
P
3
2
1
0
-0.4 -0.2 0.0 0.2 0.4
mp
0
-0.4 -0.2 0.0 0.2 0.4
mp
0
-0.4 -0.2 0.0 0.2 0.4
mp
Fig. 6. Spin gap and emergent symmetry. (A) Field-dependent gap of SrCu2(BO3)2 from experimental
data. The lines show the expected form D(H) = D(0) – ~gmBH, where D(0) represents the reported zero-
field gaps with pressure at 2.15 GPa (29) and 2.4 GPa (30), and ~g = 2.28 is the known g factor [see
section S5 of the supplementary materials (33)]. The vertical dashed lines represent HC values from
Fig. 5F. (B) Calculated ground-state phase diagram of the CBJQM versus g = J/Q and field h. The PS and
AFM phases are separated by a first-order transition. The vertical line and closely spaced points mark
the parameters in (D) to (F). (C) Calculated spin gap of the CBJQM at g = 0.2. The dashed vertical
line indicates hC, and the solid line is a fit to D(h) = D(0) – h. (D to F) Calculated distribution of the
plaquette order parameter. Double-peak (D), plateau (E), and single-peak (F) distributions are found,
respectively, in the PS phase (h = 0.55), at the transition (h = 0.62), and in the AFM phase (h = 0.69).
the associated peak in 1/T1 masking any 2D
quantum criticality, unlike the PS side, where
the spin correlations and 3D effects are much
weaker. We lack 2.4 GPa data at temperatures
higher than those shown in Fig. 5C. At 2.1 GPa,
no scaling is observed between TN and T* in
Fig. 5B, where a sharp drop below T* is imme-
diately followed by strong precursors to AFM
ordering.
Quantum spin model
We now turn to the checkerboard J – Q model
(CBJQM), in which four-spin interactions Q
replace J′ in the SSM. The CBJQM is amenable
to quantum Monte Carlo simulations and hosts
PS and AFM phases separated by a first-order
transition with emergent O(4) symmetry at zero
field (7). We here simulate [see section S1 of the
supplementary materials (33)] the same model
in a field, defining g = J/Q and h = H/J with J = 1.
In the phase diagram in Fig. 6B, the field-
driven PS-AFM transition is first order. The PS
gap D(h) obtained from the low-temperature
susceptibility (fig. S22) is shown in Fig. 6C at
g = 0.2, below the h = 0 transition at gC(0) ≈
0.217. The expected linear form D(h) = D(0) – h
for an Sz = 1 excitation is observed for h < hC,
with hC slightly less than D(0), implying a small
gap discontinuity at hC. We also observed a
very small magnetization jump, ~0.002 per
spin. These behaviors are reminiscent of the
well-known “spin-flop” transitions from Ising
to canted XY AFM phases, but with anoma-
lously small magnetization discontinuity. In
section S8 of the supplementary materials
(33), we posit that the small magnetization
and gap discontinuities, which decrease fur-
ther upon moving closer to gC(0), reflect an
approximate emergent O(3) symmetry in the
CBJQM at h > 0.
The emergent symmetry can also be studied
directly. At h = 0, the O(3) AFM order pa-
rameter (mx,my,mz) combines with the scalar
PS order parameter mp into an O(4) vector
(mx,my,mz,mp) at the T = 0 transition (7, 43).
To detect the putative O(3) symmetry of (mx,
my,mp) at h > 0, we studied the distribution
P(mp) along the vertical line in Fig. 6B. In the
PS phase (Fig. 6D), P(mp) exhibits the expected
double peak, reflecting the Z2 symmetry that is
broken in the thermodynamic limit. In the
AFM phase (Fig. 6F), there is a single central
peak, reflecting the lack of PS order.
At a conventional first-order transition, a
three-peak distribution would follow from
coexisting PS and AFM orders. By contrast,
the distribution in the coexistence state in
Fig. 6E is nearly uniform over a range of mp
values (with finite-size rounded edges). The
distribution P(mp) obtained by integrating an
O(3) symmetric P(mp,mx,my) over mx and my
should indeed be uniform for mp ∈ [–R,R],
where R ≡ max(|mp|); therefore, the approx-
imately flat distribution demonstrates emergent
O(3) symmetry in the presence of finite-size fluc-
tuations of R. Although this symmetry cannot
be exact, i.e., it exists up to some finite length
scale, it is responsible for suppressing TC and
the gap at HC; see fig. S19, where we also show
supporting results for cross-correlations be-
tween the PS and AFM order parameters.
We expect the same O(3) emergent symmetry
at the PS-AFM transition in SrCu2(BO3)2, in
which the ordered coexistence state breaks
the symmetry. The symmetry should be vio-
lated on long length scales because of the dis-
tance to the DQCP and also by 3D couplings.
One of the Goldstone modes associated with
the coexistence state then develops a small
gap. Studies of the CBJQM with interlayer cou-
plings suggest that the symmetry is unexpected-
ly robust (43).
Emergent O(3) symmetry on large length
scales in SrCu2(BO3)2 is supported, in partic-
ular, by our results at 2.1 GPa, where Fig. 4C
shows a large discontinuity in the AFM order
parameter, but TC is low and the gap (Fig. 6A)
is very small at HC. Moreover, the uniform mag-
netization is extremely small and does not ex-
hibit a discernible discontinuity (fig. S5). These
behaviors are analogous to those in the CBJQM
for g close to gC(0).
Discussion
Our high-pressure NMR experiments on
SrCu2(BO3)2 in a magnetic field establish an
example of a quantum magnet realizing DQCP
phenomenology, which thus far had existed
only in the realm of field theories and model
studies. We have demonstrated PS and AFM
transitions, with TP(H) and TN(H) merging at
TC ≃ 0.07 K and HC ≃ 6 T. The PS-AFM tran-
sition at HC is first order, with discontinuity
weakening with increasing pressure.
We have argued that the suppression of TC
and absence of statistically significant PS gap
discontinuity are consequences of emergent
O(3) symmetry generated by a nearby DQCP.
At the highest pressure, 2.4 GPa, 1/T1 exhibits
critical scaling for T between 0.2 and 2 K, in-
dicating sufficient proximity to the DQCP
[which is likely of the multicritical type
(14, 15, 17, 18)] for realizing the characteristic
quantum-critical fan (46) on the gapped PS side
Cui et al., Science 380, 1179–1184 (2023)
16 June 2023
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
of the transition. Strong 3D AFM ordering
effects on the gapless side of the transition
mask putative quantum criticality in 1/T1 there,
but the AFM ordering temperature TN vanishes
in a way very similar to the PS ordering tem-
perature TP, again in support of emergent
symmetry of the order parameters.
The H = 0 AFM phase was previously de-
tected in the specific heat between 3.2 and 4 GPa
(30), with TN from 2 to 3.5 K. Subsequently, re-
sults at H > 0 were also reported (31). However,
whereas TP(H) from the specific heat agrees well
with our PS transitions shown in Fig. 1D, the
heat capacity peak assumed to signal the
AFM transition did not drop below 1 K (31),
extending above the PS phase at fields as
low as 3 T. It may be difficult to detect the
small specific-heat peak signaling the AFM
transition (30) in high-field measurements
at low temperatures.
Beyond the highest pressure reached here,
a plausible scenario (15, 17) is a QSL between
the PS and AFM phases (Fig. 2). Our experi-
ments do not directly address the putative QSL,
and further investigations should elucidate the
low-T, H = 0 state between 2.6 and 3 GPa [where
no order has been detected (30, 31)] and its
evolution as H approaches 5.7 T, where our
current experiments point to a DQCP slightly
above 2.4 GPa.
RE FE RENCES AND N OT ES
1. T. Senthil, A. Vishwanath, L. Balents, S. Sachdev, M. P. A. Fisher,
Science 303, 1490–1494 (2004).
2. A. W. Sandvik, Phys. Rev. Lett. 98, 227202 (2007).
3. T. Senthil, L. Balents, S. Sachdev, A. Vishwanath,
M. P. A. Fisher, Phys. Rev. B 70, 144407 (2004).
4. H. Shao, W. Guo, A. W. Sandvik, Science 352, 213–216 (2016).
5. N. Ma et al., Phys. Rev. B 98, 174421 (2018).
6. A. Nahum, P. Serna, J. T. Chalker, M. Ortuño, A. M. Somoza,
Phys. Rev. Lett. 115, 267203 (2015).
7. B. Zhao, P. Weinberg, A. W. Sandvik, Nat. Phys. 15, 678–682
(2019).
8. P. Serna, A. Nahum, Phys. Rev. B 99, 195110 (2019).
9. G. J. Sreejith, S. Powell, A. Nahum, Phys. Rev. Lett. 122,
080601 (2019).
11. N. Xi, H. Chen, Z. Y. Xie, R. Yu, arXiv:2111.07368
42. Y. Nomura, M. Imada, Phys. Rev. X 11, 031034
[cond-mat.str-el] (2021).
(2021).
12. A. Nahum, J. T. Chalker, P. Serna, M. Ortuño, A. M. Somoza,
43. G. Sun, N. Ma, B. Zhao, A. W. Sandvik, Z. Y. Meng, Chin. Phys. B
Phys. Rev. X 5, 041048 (2015).
13. C. Wang, A. Nahum, M. A. Metlitski, C. Xu, T. Senthil,
30, 067505 (2021).
44. V. Y. Irkhin, A. A. Katanin, Phys. Rev. B 57, 379–391
Phys. Rev. X 7, 031051 (2017).
(1998).
14. B. Zhao, J. Takahashi, A. W. Sandvik, Phys. Rev. Lett. 125,
45. Y. Q. Qin et al., Phys. Rev. X 7, 031052
257204 (2020).
15. D.-C. Lu, C. Xu, Y.-Z. You, Phys. Rev. B 104, 205142
(2021).
16. W.-Y. Liu et al., Sci. Bull. 67, 1034–1041
(2022).
17. J. Yang, A. W. Sandvik, L. Wang, Phys. Rev. B 105, L060409
(2017).
46. A. V. Chubukov, S. Sachdev, J. Ye, Phys. Rev. B 49, 11919–11961
(1994).
47. Y. Cui et al., Data and simulation codes for: Proximate
deconfined quantum critical point in SrCu2(BO3)2, Zenodo
(2023); https://doi.org/10.5281/zenodo.7600028.
(2022).
18. W.-Y. Liu et al., Phys. Rev. X 12, 031039
(2022).
19. A. Keleş, E. Zhao, Phys. Rev. B 105, L041115
(2022).
20. H. Shackleton, A. Thomson, S. Sachdev, Phys. Rev. B 104,
045110 (2021).
21. Y.-H. Zhang, S. Sachdev, Phys. Rev. B 102, 155124
(2020).
22. H. Kageyama et al., Phys. Rev. Lett. 82, 3168–3171
(1999).
23. S. Miyahara, K. Ueda, Phys. Rev. Lett. 82, 3701–3704
(1999).
24. S. Miyahara, K. Ueda, J. Phys. 15, R327–R366
(2003).
25. B. S. Shastry, B. Sutherland, Physica B+C 108, 1069–1070
(1981).
26. A. Koga, N. Kawakami, Phys. Rev. Lett. 84, 4461–4464
(2000).
27. P. Corboz, F. Mila, Phys. Rev. B 87, 115144 (2013).
28. S. Haravifard et al., Nat. Commun. 7, 11956 (2016).
29. M. E. Zayed et al., Nat. Phys. 13, 962–966 (2017).
30. J. Guo et al., Phys. Rev. Lett. 124, 206602 (2020).
31. J. L. Jiménez et al., Nature 592, 370–375 (2021).
32. Y. H. Matsuda et al., Phys. Rev. Lett. 111, 137204
(2013).
33. Materials and methods, additional NMR data, data analysis, and
supporting theoretical results are available as supplementary
materials.
34. K. Kodama et al., J. Phys. 14, L319
(2002).
35. T. Waki et al., J. Phys. Soc. Jpn. 76, 073710
(2007).
36. M. Takigawa, T. Waki, M. Horvatić, C. Berthier,
J. Phys. Soc. Jpn. 79, 011005 (2010).
37. T. Moriya, J. Phys. Soc. Jpn. 18, 516–520
(1963).
AC KNOWLED GME NTS
We thank B. Normand for extensive suggestions and constructive
criticism and W. Guo, F. Mila, M. Takigawa, Y. Wang, Z.-Y.
Xie, and Y.-Z. You for helpful discussions. Funding: This work
was supported by the Ministry of Science and Technology
of China (grants 2022YFA1402700, 2022YFA1403402, and
2021YFA1400401), the National Natural Science Foundation of
China (grants 12134020, 12104503, 12174441, and 51872328),
the Simons Foundation (investigator grant 511064), the
Strategic Priority Research Program(B) of the Chinese Academy
of Sciences (grant XDB33010100), the K. C. Wong Education
Foundation (grant GJTD2020-01), the Beijing Institute of
Technology Research Fund Program for Young Scholars, the
Fundamental Research Funds for the Central Universities, and
the Research Funds of Renmin University of China (grants
22XNH096 and 21XNLG18). Some of the numerical calculations
were performed on the Shared Computing Cluster managed
by Boston University’s Research Computing Services. Part of
the measurements were conducted at the Cubic Anvil Cell
Station, the Synergetic Extreme Condition User Facility
(SECUF). Author contributions: Y.C. performed NMR
measurements and data analysis with assistance from C.L., Z.H.,
and W.Y.. W.H. and S.L. provided single crystals. X.L. and H.L.
performed the Bayesian fitting analysis. L.L., H.L., N.X., and
K.H.W. performed numerical simulations with guidance from R.Y.
and A.W.S.. W.Y., A.W.S., and R.Y. guided the project and wrote the
manuscript with input from all authors. Competing interests:
The authors declare no competing interests. Data and materials
availability: The data and simulation codes needed to evaluate the
conclusions in this study have been deposited at Zenodo (47).
License information: Copyright © 2023 the authors, some rights
reserved; exclusive licensee American Association for the Advancement
of Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse
38. J. Y. Lee, Y.-Z. You, S. Sachdev, A. Vishwanath, Phys. Rev. X 9,
SUPPLEMENTARY MATERIALS
041037 (2019).
39. C. Boos et al., Phys. Rev. B 100, 140413
(2019).
40. S.-S. Gong, W. Zhu, D. N. Sheng, O. I. Motrunich,
M. P. A. Fisher, Phys. Rev. Lett. 113, 027201
(2014).
science.org/doi/10.1126/science.adc9487
Materials and Methods
Figs. S1 to S22
Table S1
References (48–65)
Submitted 10 May 2022; accepted 12 May 2023
10.1126/science.adc9487
10. J. Takahashi, A. W. Sandvik, Phys. Rev. Res. 2, 033459
41. L. Wang, A. W. Sandvik, Phys. Rev. Lett. 121, 107202
(2020).
(2018).
Cui et al., Science 380, 1179–1184 (2023)
16 June 2023
6 of 6
|
10.1126_science.abq8271
|
RES EARCH
COMPUTER NETWORKS
Delocalized photonic deep learning on the internet’s edge
Alexander Sludds1*, Saumil Bandyopadhyay1, Zaijun Chen1†, Zhizhen Zhong2, Jared Cochrane1,3,
Liane Bernstein1, Darius Bunandar1‡, P. Ben Dixon3, Scott A. Hamilton3, Matthew Streshinsky4§,
Ari Novack4§, Tom Baehr-Jones4§, Michael Hochberg4§, Manya Ghobadi2,
Ryan Hamerly1,5*, Dirk Englund1*
Advanced machine learning models are currently impossible to run on edge devices such as smart
sensors and unmanned aerial vehicles owing to constraints on power, processing, and memory. We
introduce an approach to machine learning inference based on delocalized analog processing across
networks. In this approach, named Netcast, cloud-based “smart transceivers” stream weight data to
edge devices, enabling ultraefficient photonic inference. We demonstrate image recognition at ultralow
optical energy of 40 attojoules per multiply (<1 photon per multiply) at 98.8% (93%) classification
accuracy. We reproduce this performance in a Boston-area field trial over 86 kilometers of deployed
optical fiber, wavelength multiplexed over 3 terahertz of optical bandwidth. Netcast allows milliwatt-class
edge devices with minimal memory and processing to compute at teraFLOPS rates reserved for high-
power (>100 watts) cloud computers.
A dvances in deep neural networks (DNNs)
are transforming science and technology
(1–4). However, the increasing compu-
tational demands of the most powerful
DNNs limit deployment on low-power
devices, such as smartphones and sensors—
and this trend is accelerated by the simulta-
neous move toward Internet of Things (IoT)
devices. Numerous efforts are underway to
lower power consumption, but a fundamental
bottleneck remains because of energy consump-
tion in matrix algebra (5), even for analog ap-
proaches including neuromorphic (6), analog
memory (7), and photonic meshes (8). In all
these approaches, memory access and multiply-
accumulate (MAC) functions remain a stubborn
bottleneck near 1 pJ per MAC (5, 9–12). Edge
devices typically use chip-scale sensors, occupy
millimeter-scale footprints, and consume milli-
watts of power. Their small footprint and low
power budget mean that performance is limited
by the size, weight, and power (SWaP) of com-
puting systems integrated on the device.
To make advanced DNNs at all feasible on
low-power devices, industry has resorted to
offloading computationally heavy DNN infer-
ence to cloud servers. For instance, a smart
home device may send a voice query as a
vector U to a cloud server, which returns the
inference result V to the client (Fig. 1). This
offloading architecture adds a ∼200-ms latency
to voice commands (13), which makes services
such as self-driving impossible. Moreover, off-
loading poses security risks in both the edge
and the cloud: Hacking of the communication
1Research Laboratory of Electronics, Massachusetts Institute
of Technology, Cambridge, MA 02139, USA. 2Computer
Science and Artificial Intelligence Laboratory, Massachusetts
Institute of Technology, Cambridge, MA 02139, USA. 3Lincoln
Laboratory, Massachusetts Institute of Technology,
Lexington, MA 02421, USA. 4Nokia Corporation, New York,
NY 10016, USA. 5Physics and Informatics Laboratories, NTT
Research Inc., Sunnyvale, CA 94085, USA.
*Corresponding author. Email: asludds@mit.edu (A.S.);
rhamerly@mit.edu (R.H.); englund@mit.edu (D.E.)
†Present address: Ming Hsieh Department of Electrical and
Computer Engineering, University of Southern California, Los
Angeles, CA 90089, USA. ‡Present address: Lightmatter Inc.,
Boston, MA 02110, USA. §Present address: Luminous Computing
Inc., Mountain View, CA 94041, USA.
A
B
C
D
Fig. 1. Netcast concept. (A) Smart transceivers integrated alongside cloud
computing infrastructure including servers, data storage, network switches, and
edge nodes. The smart transceiver sequentially encodes layers of a neural
network model onto the intensity of distinct optical wavelengths using
digital-to-analog converters (DACs), optical modulators (mod), and lasers.
Wavelength-division multiplexers (WDMs) combine the separate wavelengths
from each modulator to the smart transceiver output. (B) U and V highlight
current solutions to large model deployments on the edge, with edge device
data communicated back to cloud computers. In our solution, smart transceivers
have connections to many devices at the edge of the communications network,
including cellular networks, smart sensors, content delivery networks, and
aircraft. (C) The edge client encodes input activation data onto a single
broadband optical modulator, modulating all weight wavelengths simultaneously.
Wavelengths are separated with a WDM, and the result of matrix-vector
multiplication is computed on time-integrating receivers. (D) Matrix-vector
products between an M-element input vector and (M,N) weight matrix
are time-frequency (t-w) encoded, with each wavelength accumulating its results
on a time-integrating receiver.
Sludds et al., Science 378, 270–276 (2022)
21 October 2022
1 of 7
RES EARCH | R E S E A R C H A R T I C L E
A
D
B
C
F
E
G
Fig. 2. Experimental demonstration of Netcast system. (A) Smart trans-
ceiver composed of a 48-modulator silicon photonic transmitter with 2.4 Tbps of
total bandwidth. (B) Optical spectrum of smart transceiver output, showing
16 laser sources across 3 THz of bandwidth with >25 dB optical SNR. (C) An
example of high-speed operation of the smart transceiver modulators, with a
50 GHz open eye. (D) Weights are sent over 86 km of deployed optical fiber
connecting the smart transceiver to the client. (E) Client receiver composed of a
broadband, high-speed optical modulator, a WDM demultiplexer, and custom
time-integrating receivers. (F) The client input modulator also achieves an open
eye of 10 GHz (test equipment limited). (G) Example time-integrating receiver
waveform showing constant optical power being accumulated over 10 ms and
resetting. Satellite imagery in (D) taken using a deployed satellite (Planet.com).
of client data (in vector U) has led to security
breaches of private data.
To address these problems, we introduce here
a photonic edge computing architecture, named
Netcast, to minimize the energy and latency of
large linear algebra operations such as general
matrix-vector multiplication (GEMV) (5). In
the Netcast architecture, cloud servers stream
DNN weight data (W) to edge devices in an
analog format for ultraefficient optical GEMV
that eliminates all local weight memory ac-
cess (14).
Servers containing a “smart transceiver”
(15)—which may be in the standard pluggable
transceiver format represented in Fig. 1A—
periodically broadcast the weights (W) of
commonly used DNNs to edge devices, using
wavelength division multiplexing (WDM)
to leverage the large spectrum available
at the local access layer. Specifically, the
(M,N)-sized weight matrix of one DNN layer
may be encoded in a time-frequency basis
by the amplitude-modulated field Wn tð Þ ¼
XM
j¼1
Þ, where the optical
amplitude wnj at frequency wn and time step j
wnje(cid:2)iwntd t (cid:2) jDT
ð
represents the nth row of the weight matrix
(Fig. 1D), and d is the impulse response function.
Suppose now that a camera in Fig. 1 requires
inference on an image X. To do so, it waits for
the server to stream the “image recognition”
DNN weights, which it modulates with X tð Þ ¼
XM
Þ using a broadband optical
j¼1
xjd t (cid:2) jDT
ð
modulator and subsequently separates the
wavelengths to N time-integrating detectors
to produce the vector-vector dot product
Yn tð Þ ¼
Þ. This architec-
ð
wnjxjd t (cid:2) jDt
XM
j¼1
ture minimizes the active components at
the client, requiring only a single optical
modulator, digital-to-analog converter (DAC),
and analog-to-digital converter (ADC).
Experimental implementation of Netcast
We demonstrate the Netcast protocol with a
smart transceiver (Fig. 2A), made in a com-
mercial silicon-photonic CMOS foundry (OpSIS/
IME, described in supplementary text section 2).
The smart transceiver is composed of 48 Mach-
Zehnder modulators (MZMs), each capable of
modulation up to 50 Gbps for a total band-
width of 2.4 Tbps (16). The smart transceiver
supports WDM, with Fig. 2B showing 16 WDM
lasers simultaneously transmitting through the
chip with ≈−10 dBm (100 mW) power per wave-
length. Figure 2C shows an open eye diagram
at 50 GHz (supplementary text section 8).
Weights are transmitted over 86 km of de-
ployed optical fiber from the Massachusetts
Institute of Technology (MIT) main campus to
MIT Lincoln Laboratory and back to the main
campus (Fig. 2D). The client (Fig. 2E) applies
input activation values to the incoming weight
data using a high-speed (20-GHz) broadband
lithium niobate MZM, with Fig. 2F showing
an open eye diagram at 10 GHz (limited by
testing equipment). A passive wavelength
demultiplexer separates each wavelength chan-
nel for detection onto an array of custom time-
integrating receivers, with an example of time
integration shown in Fig. 2G (supplementary
text section 6). After integration, the gener-
ated voltages from the receivers are measured
by a digitizer and stored in memory. Addi-
tional postprocessing steps, such as the non-
linear activation function, are performed using
a computer. Multiple neural network layers
Sludds et al., Science 378, 270–276 (2022)
21 October 2022
2 of 7
RES EARCH | R E S E A R C H A R T I C L E
A
B
C
D
Fig. 3. Computational accuracy of Netcast system. (A) Weight data from
multiple wavelength channels is simultaneously modulated by input data. After
wavelength multiplexing, the generated photocurrent is time-integrated.
(B) Floating-point computing accuracy comparing the results of 10,000 scalar-
scalar floating point multiplications. Electrical floating point results are
designated as y and optical results are designated as ^y. The difference y (cid:2) ^y
has a standard deviation of srms = 0.005 or ≈8-bit accuracy. (C) Example
output activation data from the optical setup correctly classifying the digit “3.”
(D) Computing results of image classification over both local links and the
86-km deployed fiber link.
Table 1. Device contributions to receiver performance assuming conventional technology.
Device energy consumption is amortized by either a spatial fan-out factor (N) or time-domain
fan-out factor (M). We assume a carrier depletion modulator in silicon is used and that a single
high-speed (gigahertz) ADC reads out from an array of N slow integrators. See supplementary text
section 19 for derivation of nonlinearity energy consumption.
Netcast client energy consumption
rable with the model’s baseline accuracy of
98.7%. Using the same test images, we utilize
3 THz of bandwidth over the deployed fiber
and classify MNIST digits with 98.8% accu-
racy. This result shows the potential for this
architecture to support ultrahigh bandwidths
in real-world deployed systems using conven-
tional components.
Device
Number of devices
Fan-out
Energy per device
Energy per MAC
Energy efficiency
N
Modulator (16)
.....................................................................................................................................................................................................................
N
DAC (37)
.....................................................................................................................................................................................................................
M
ADC (38)
.....................................................................................................................................................................................................................
M
Integrator (39)
.....................................................................................................................................................................................................................
M
Nonlinearity
.....................................................................................................................................................................................................................
–
Total
.....................................................................................................................................................................................................................
~(1/N) pJ
~(1/N) pJ
~(1/M) pJ
~(1/M) fJ
~(1/M) fJ
~(1/N) pJ
~1 pJ
~1 pJ
~1 pJ
~1 fJ
<100 fJ
–
1
1
1
N
N
–
are run by taking the resulting output acti-
vations of the previous layer and encoding
them onto the input modulator while the next
layer’s weights are transmitted.
We show the flow of data through the exper-
imental setup and the accuracy it can achieve
in Fig. 3A. Weight data are encoded to multi-
ple modulators simultaneously. For clarity, we
show a single row of the digit “3” being encoded
and the resulting time trace from a single
wavelength. We demonstrate computing with
high accuracy, with Fig. 3B showing 8 bits of
precision, more than the ≈5 bits of precision
required for neural network computation (17, 18).
After calibrating the system, we perform image
classification by running a benchmark hand-
written digit classification task [Modified Na-
tional Institute of Standards and Technology
(MNIST)], which was trained on a digital com-
puter (supplementary text sections 14 and 16).
Figure 3C illustrates an example of the system’s
computing result for classifying the digit “3.”
We then test the system’s performance both
locally and over deployed fiber using a bench-
mark three-layer MNIST model with 100 neurons
per hidden layer (supplementary text section 14).
Using 1000 test images locally, we demon-
strate 98.7% accurate computation, compa-
Netcast is designed to minimize the power
used at the client. To enable this, we make sure
every component at the client is performing a
large number of MACs (M or N) for modulation
and electrical readout, respectively. Only a single
MZM and DAC are used to encode input data
across N wavelengths, enabling N MACs of
work for every voltage applied to the modula-
tor. While the energy costs of these individual
components can be high, they have high paral-
lelism, performing many MACs of work per
time step. For encoding input activations, the
client only uses a single broadband optical
modulator, allowing for ≈(1/N) pJ per MAC of
energy consumption using standard compo-
nents. Furthermore, the integrator and ADC
can be much slower than the speed of modu-
lated weights, because readout occurs after M
timesteps. As a result, the integrator and ADC
can be M times slower, decreasing the cost of
electrical readout components to ≈(1/M) pJ
per MAC. Assuming near-term values of N =
M = 100, client energy consumption can reach
Sludds et al., Science 378, 270–276 (2022)
21 October 2022
3 of 7
RES EARCH | R E S E A R C H A R T I C L E
A
B
Fig. 4. Thermal noise limited optical sensitivity of Netcast system. (A) Experimentally measured
sensitivity of optical receivers. Standard amplified photoreceivers are shown on the right side of the plot,
with performance limited by electrical amplifier thermal noise, giving a typically optical energy of 10 to
100 fJ per MAC. The center of the plot shows linear avalanche photodiodes, which use intrinsic gain to
lower the energy per MAC, but at the cost of increased energy consumption and lower-bandwidth
time-integrating receivers, which lower the effective thermal noise floor by performing many MAC operations
for each readout. Time-integrating receivers using off-the-shelf technology can achieve high accuracy with
<100 aJ per MAC of optical sensitivity on the benchmark neural network task. (B) Confusion matrices
for labeled points in (A), showing how each digit in the MNIST dataset is classified by the optical hardware
(on-diagonal elements correspond to correct classification; columns add to 1, but rows do not have to).
≈10 fJ per MAC, which is three orders of mag-
nitude lower than is possible in existing digital
CMOS. The scaling of the client energy con-
sumption is summarized in Table 1.
In our experimental demonstration, we have
fabricated a 48-channel silicon smart transceiver
to deploy weights to the client. The modulators
used in this smart transceiver can operate at a
data rate of 50 Gbps. The client uses a fiber
lithium niobate modulator with a bandwidth
of 20 GHz and energy efficiency of 18 pJ per bit
(supplementary text section 1). Sharing this
input modulator over 48 wavelengths, we find
that our input modulator uses 370 fJ per MAC
of energy. Simple changes to the client, such as
making use of the same modulator at the client as
we do at the smart transceiver (≈450 fJ per bit),
would enable <10 fJ per MAC energy efficiency.
Our integrating receivers have a 20 mW power
consumption per channel, leading to an energy
efficiency of 1 pJ per MAC and the potential to
improve orders of magnitude with commercial
technology (see Discussion section).
Receiver sensitivity
Applications of Netcast, including free-space
deployment to drones or spacecraft, can oper-
ate in deeply photon-starved environments.
For example, recent satellite optical commu-
nication demonstrations, such as NASA’s Lunar
Laser Communication Demonstration, have
shown ≈100 Mbps communications to satel-
lites orbiting the Moon with link losses in
excess of 70 dB (19). To enable high-speed and
energy-efficient machine learning on these de-
ployments, optical receivers must have the
lowest possible noise floor, ideally operating at
the shot noise limit with ≈1 photon per MAC.
Modern photoreceivers are limited by either
thermal noise of readout electronics [also called
Johnson-Nyquist noise (20)], shot noise, flicker
(1/f ) noise, or relative intensity noise of the
laser; of these, for integrated optoelectronics,
thermal and shot noise are dominant in Net-
cast (see supplementary text sections 13 and 23).
We overcome this problem with time-integrating
receivers, which accumulate partial results from
vector-matrix multiplication. We compare the
sensitivity of different photoreceivers. Amplified
photoreceivers (Fig. 4A, right) have typical
sensitivities of ≈10 to 100 fJ per MAC. Am-
plified linear mode avalanche photodetectors
(Fig. 4A, middle) overcome some of the ther-
mal noise of the amplifier and achieve ≈1 fJ per
MAC. Our custom time-integrating receivers
(Fig. 4A, left) perform M MACs per measurement
window before readout, lowering the required
optical power per readout by M. Amplified
photodetectors, in contrast, read out after each
MAC, acquiring thermal noise for each mea-
surement and adding the results of each MAC
together to create the resulting output activa-
tion value. For time-integrating receivers, the
resulting output activation signal is measured
while measuring thermal noise once, giving a 1
M
optical energy per MAC scaling. For amplified
photodetectors, the partial-product signal terms
add together linearly, while thermal noise adds
p
p scaling. In our
in quadrature, giving a
experiment, we demonstrate that with M =
100, only 10 aJ per MAC (100 photons) of optical
energy is required (two orders of magnitude less
than for similar amplified photodetectors). This
result brings Netcast close to the fundamental
quantum limit of optical computation (21, 22),
which we can reach by engineering the receiver
to lower thermal noise.
ffiffiffiffi
M ¼ 1ffiffiffiffi
M
M
p
ffiffiffiffiffiffiffiffiffiffiffi
kBTC
Thermal noise is a hardware-dependent
noise source, originating from the thermal
motion of charge carriers in an electrical con-
ductor. In a resistor-capacitor (RC) circuit, ther-
mal noise manifests in a fluctuation in the
number of readout electrons in a circuit given
by sth ¼
=q, where kB is the Boltzmann
constant, T is temperature, q is the electron
charge, and C is the capacitance of the receiver
(23). Conventional amplified photodetectors
read out on every MAC operation and add
partial-product results to generate an output
activation value. Adding together each MAC
adds together the measured signal linearly,
and noise terms add in quadrature. This re-
sults in a signal-to-noise ratio (SNR) of
SNR ¼
Signal Electrons
Noise Electrons
r
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
XM
i¼1
ffiffiffiffiffiffiffiffiffiffiffi
Ms2
th
s2
th
¼ MPopthTclk=
q
¼ MPopthTclk=
¼
ffiffiffiffiffi
M
p
PopthTclk=sth
where Popt is photon flux incident on the
detector (units of photons per second), h is
detector quantum efficiency, and Tclk is the
time period for each MAC. In contrast, our
Sludds et al., Science 378, 270–276 (2022)
21 October 2022
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
A
C
B
D
Fig. 5. Forward looking performance of Netcast. (A) Fundamental noise bounds
of time-integrating receivers from thermal noise of an integrator and shot noise to
achieve 50% accuracy on MNIST task. Decreasing the capacitance of the time
integrator lowers thermal readout noise, enabling access to the single photon–per-
MAC regime. (B) Proof-of-concept experimental setup consisting of input and weight
modulators and superconducting nanowire single-photon detectors (SNSPDs),
allowing us to probe this fundamental single-photon bound. (C) We experimentally
validate the single-photon detectors by measuring shot noise on the detector
over many integration windows. (D) Using a three-layer MNIST model, we
experimentally measure computation with <1 photon per MAC with high accuracy.
time-integrating receivers only see thermal
noise once per measurement window
SNR ¼
Signal Electrons
Noise Electrons
¼ MPopthTclk=sth
As a result, we see that the required number
of photons per MAC is
times lower than
for standard amplified photoreceivers.
ffiffiffiffiffi
M
p
Improvements to time-integrating receivers
are possible by minimizing the integration ca-
pacitance of the receiver. Figure 5A shows the
thermal noise limit of time-integrating receivers
as integration capacitance is decreased. This
noise floor is fundamentally connected to the
size scale of photodetectors, readout electronics,
and their proximity of integration (10). Modern
foundry processes enable ≈1 fF–scale receivers,
lowering the thermal readout noise to the
single photon–per-MAC level (24, 25). This
single photon–per-MAC regime is fundamen-
tally limited by the quantum nature of light,
where precision is determined by the Poissonian
distribution of photons that arrive within a
measurement window. Poissonian noise, also
called shot noise, can be observed in exper-
imentally measured data in Fig. 5C. We inves-
tigate this fundamental bound of the Netcast
system through a proof-of-concept experiment
using superconducting nanowire single-photon
detectors (SNSPDs) as shown in Fig. 5B. These
photodetectors are ideal, demonstrating pure
shot noise–limited performance. We show that
the fundamental shot noise bound on the same
benchmark digit classification problem from
Fig. 4 allows the receiver to operate with high
accuracy with <1 photon per MAC (0.1 aJ per
MAC). This result may at first seem surprising
given that less than a single photon per MAC
is counterintuitive. We can understand this
measurement better by noting that at readout,
we have performed a vector-vector product with
M = 100 MACs. Each MAC can have less than
a single photon in it, but the measured signal
will have many photons in it. A graphical ex-
planation is given in supplementary text sec-
tion 18. This single photon–per-MAC regime
enables many new applications. The realiza-
tion of computing with less than one photon
per MAC could enable a new class of comput-
ing systems that protect both client input and
server weight data. Another application that
benefits from less than one photon per MAC is
deployed spacecraft that operate in a strongly
photon-starved environment. Weight data
from a directional base station could be trans-
mitted to the spacecraft and classified on the
craft, before the results are transmitted to Earth.
Discussion
The system-level demonstration shown here is
one example of an implementation of Netcast.
The cloud-based smart transceiver can reside
inside of existing networking hardware such
as network switches, servers, or edge nodes.
Our ideas can be extended to the case where
the user data are streamed through program-
mable network switches with smart trans-
ceivers, enabling in-network optical inference
(15). Modern network switches, such as Intel’s
Tofino switch, are an ideal platform for de-
veloping Netcast commercially, as they are pro-
grammable, enabling multiple streams of weights
Sludds et al., Science 378, 270–276 (2022)
21 October 2022
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
to be deployed at line rate (100 Gbps), and can
support 64 GB of memory, reaching the storage
requirements of modern neural networks. Prior
work has demonstrated the feasibility of using
programmable switches to perform layer-by-
layer inference with smart transceivers (15).
The large data storage of these network switches
enables multiple models to be stored and
queried. The client device could use its broad-
band modulator to allow for reflection-mode
communication back to the server, where the
client modulates received light and sends it
back along the fiber link for communication.
This querying communication can be slow and
lossy, as only a few bits are required to request
that a new model be sent.
Emerging photonic technologies, such as
low-power static phase shifters (26–28) and
high-speed phase shifters (29–32), can reduce
receiver electrical energy consumption to ≈10 aJ
per MAC. This energy can be further decreased
by making use of the tight integration of tran-
sistors and photonics in silicon using technol-
ogies such as receiverless detectors (10), photonic
DACs (33), and photonic ADCs (34). Detectors
such as avalanche detectors could be incor-
porated with a time integrator to provide a
benefit to the optical sensitivity of the receiver,
but at the cost of added electrical power con-
sumption (supplementary text section 21). Fur-
ther improvements in optical sensitivity are
possible by using coherent detection, which
boosts the received signal using a strong local
oscillator (21). Two examples of a Netcast system
using coherent detection to substantially im-
prove optical energy per MAC are detailed in
supplementary text section 12.
A number of companies have designed cus-
tom edge computing application-specific inte-
grated circuits (ASICs) with reduced SWaP
(7, 35), but these ASICs are hampered by the
same energy and bandwidth constraints as
larger CMOS processors. Analog accelerators,
such as memristive crossbar arrays and meshes
of photonic interferometers, hold promise for
lowering the power consumption of neural
networks compared with electronic counter-
parts, but existing commercial demonstrations
still consume watts of power (8, 36).
One obstacle to scaling bandwidth in tra-
ditional optical communication systems is dis-
persion in optical fiber. For a single smart
transceiver and client, techniques such as
wavelength-dependent delays can compensate
for dispersion at the smart transceiver. How-
ever, in systems where weights are deployed to
multiple clients from one smart transceiver
with different lengths of fiber, this technique
cannot be used. We discuss the effects of dis-
persion in supplementary text section 22 and
show that it is possible to make use of the
optical O-band to enable terahertz of bandwidth
at clock rates of 10 GHz per wavelength over
more than 10 km of optical fiber.
Outlook
We have described an edge computing ar-
chitecture that makes use of the strengths of
photonics and electronics to achieve orders of
magnitude in energy efficiency and optical
sensitivity improvements over existing digital
electronics. We have demonstrated scalable
photonic edge computing using WDM, time-
integrating receivers, scalability to milliwatt-
class power consumption, <1 photon–per-MAC
receiver sensitivity, and computing over de-
ployed fiber using 3 THz of bandwidth. On
image classification tasks, we show 98.8% ac-
curate image classification. The hardware shown
in this paper is readily mass-producible from
existing CMOS foundries, allowing for near-
term impact on our daily lives. Our approach
removes a fundamental bottleneck in edge
computing, enabling high-speed computing
on deployed sensors and drones.
RE FERENCES AND NOTES
1. T. B. Brown et al., arXiv:2005.14165 [cs.CL] (2020).
2. O. Vinyals et al., Nature 575, 350–354 (2019).
3. A. Krizhevsky, I. Sutskever, G. E. Hinton, Adv. Neural Inf.
4.
Process. Syst. 25, 1097–1105 (2012).
J. Deng et al., 2009 IEEE Conference on Computer Vision and
Pattern Recognition (IEEE, 2009), pp. 248–255.
5. V. Sze, Y.-H. Chen, T.-J. Yang, J. S. Emer, Proc. IEEE 105,
2295–2329 (2017).
6. M. Davies et al., IEEE Micro 38, 82–99 (2018).
7. Mythic, M1076 Analog Matrix Processor; https://mythic.ai/
products/m1076-analog-matrix-processor/.
8. C. Demirkiran et al., arXiv:2109.01126 [cs.AR] (2021).
9. M. Horowitz, 2014 IEEE International Solid-State Circuits
Conference Digest of Technical Papers (ISSCC) (IEEE, 2014),
pp. 10–14.
10. D. A. Miller, J. Lightwave Technol. 35, 346–396 (2017).
11. L. Bernstein et al., Sci. Rep. 11, 3144 (2021).
12. Y.-H. Chen, T. Krishna, J. S. Emer, V. Sze, IEEE J. Solid-State
Circuits 52, 127–138 (2016).
13. M. Satyanarayanan, Computer 50, 30–39 (2017).
14. R. Hamerly et al., Proc. SPIE 11804, 118041R (2021).
15. Z. Zhong et al., Proceedings of the ACM SIGCOMM 2021
Workshop on Optical Systems (OptSys ’21) (Association for
Computing Machinery, 2021), pp. 18–22.
16. M. Streshinsky et al., J. Lightwave Technol. 32, 4370–4377
(2014).
17. T. Gokmen, M. J. Rasch, W. Haensch, 2019 IEEE
International Electron Devices Meeting (IEDM) (IEEE, 2019),
pp. 22.3.1–22.3.4.
18. S. Garg, J. Lou, A. Jain, M. Nahmias, arXiv:2102.06365 [cs.LG]
(2021).
19. D. M. Boroson, J. J. Scozzafava, D. V. Murphy, B. S. Robinson,
M. I. T. Lincoln, 2009 Third IEEE International Conference on
Space Mission Challenges for Information Technology (IEEE,
2009), pp. 23–28.
20. J. B. Johnson, Phys. Rev. 32, 97–109 (1928).
21. R. Hamerly, L. Bernstein, A. Sludds, M. Soljačić, D. Englund,
Phys. Rev. X 9, 021032 (2019).
22. T. Wang et al., Nat. Commun. 13, 123 (2022).
23. J. Pierce, Proceedings of the IRE 44, 601–608 (1956).
24. M. Rakowski et al., 2020 Optical Fiber Communication
Conference (OFC), OSA Technical Digest (Optica Publishing
Group, 2020), paper T3H–3.
25. C. Sun et al., IEEE J. Solid-State Circuits 51, 893–907
(2016).
26. G. Liang et al., Nat. Photonics 15, 908–913 (2021).
27. R. Baghdadi et al., Opt. Express 29, 19113–19119 (2021).
28. M. Dong et al., Nat. Photonics 16, 59–65 (2022).
29. E. Timurdogan et al., Nat. Commun. 5, 4008 (2014).
30. C. Wang et al., Nature 562, 101–104 (2018).
31. M. Xu et al., Optica 9, 61 (2022).
32. W. Heni et al., Nat. Commun. 10, 1694 (2019).
33. S. Moazeni et al., IEEE J. Solid-State Circuits 52, 3503–3516
(2017).
34. A. Zazzi et al., IEEE Open J. Solid State Circuits Soc. 1, 209–221
(2021).
35. A. Yazdanbakhsh, K. Seshadri, B. Akin, J. Laudon,
R. Narayanaswami, arXiv:2102.10423 [cs.LG] (2021).
36. D. Fick, M. Henry, “Analog computation in flash memory
for datacenter-scale AI inference in a small chip,”
Hot Chips 2018 (HC30), Cupertino, California,
19–21 August 2018.
37. B. M. Pietro Caragiulo, C. Daigle, B. Murmann, Dac
performance survey 1996-2020, GitHub (2022); https://
github.com/pietro-caragiulo/survey-DAC.
38. B. Murmann, ADC performance survey 1997-2021, Stanford
University (2022); http://web.stanford.edu/~murmann/
adcsurvey.html.
39. E. Yang, T. Lehmann, “High gain operational amplifiers in
22 nm CMOS,” 2019 IEEE International Symposium on Circuits
and Systems (ISCAS) (IEEE, 2019).
40. A. Sludds, alexsludds/Delocalized_Photonic_Deep_Learning_
on_the_Internets_Edge: Zenodo Added, Zenodo (2022);
https://doi.org/10.5281/zenodo.6982196.
AC KNOWLED GME NTS
We acknowledge D. Lewis and A. Pennes for assistance in
machining laboratory equipment and E. Allen for discussions
related to using squeezed light to further reduce photon counts.
We are grateful to E. Bersin and B. Dixon for assistance in
coordinating usage of the deployed fiber and A. Rizzo for help in
converting and plotting eye diagram data as well as proofreading
the manuscript. We thank C. Panuski and S. Krastanov for
informative discussions on single-photon operation of Netcast.
We thank A. Pyke of Micro-Precision Technologies for wire bonding
the electrical connections from the printed circuit board to the
48-channel transmitter. We appreciate help from F. Wong for the
use of his SNSPDs and acknowledge M. Prabhu and C. Errando
Herranz for facilitating the usage of the SNSPDs, C. Freeman for
help in taking drone photography, NVIDIA for supplying a Tesla
K40 GPU that was used for simulations shown in the main text and
supplementary materials, and Planet.com for allowing us to take
custom satellite imagery. Funding: A.S. and S.B. are supported
by National Science Foundation Graduate Research Fellowship
1745302. L.B. is supported by the Natural Sciences and
Engineering Research Council of Canada (PGSD3-517053-2018).
This research was funded by a collaboration with NTT-Research
and NSF Eager (CNS-1946976). This material is based on research
sponsored by the Air Force Office of Scientific Research (AFOSR)
under award FA9550-20-1-0113, the Air Force Research Laboratory
(AFRL) under agreement FA8750-20-2-1007, the Army Research
Office (ARO) under agreement W911NF-17-1-0527, NSF RAISE-
TAQS grant 1936314, NSF C-Accel grant 2040695, NSF grant
ASCENT-2023468, NSF grant CAREER-2144766, and ARPA-E grant
ENLITENED PINE DE-AR0000843. The work was further supported
by funding from the Alfred P. Sloan foundation (FG-2022-18504)
and DARPA grant FastNICs 4202290027. Distribution Statement
A. Approved for public release. Distribution is unlimited. This
material is based upon work supported by the Under Secretary of
Defense for Research and Engineering under Air Force Contract
No. FA8702-15-D-0001. Any opinions, findings, conclusions or
recommendations expressed in this material are those of the
authors and do not necessarily reflect the views of the Under
Secretary of Defense for Research and Engineering. Author
contributions: A.S. created the experimental setup and conducted
the experiment. S.B. assisted in fiber-to-chip coupling and
discussions on the project. Z.C. assisted with high-speed
measurements of the setup and discussions. M.S., A.N., T.B.-J.,
and M.H. designed and taped out the smart transceiver. D.B.
packaged the 48-channel silicon transceiver. J.C. packaged the
time-integrating receivers and assisted in calibration. D.E.
established the fiber link between MIT and MIT Lincoln Laboratory,
and S.A.H. and P.B.D. helped with its use. L.B. helped in discussion
of fundamental noise sources. M.G. and Z.Z. assisted with
discussions on modern telecommunication networks. R.H.
conceived of the project idea. S.B., Z.C., L.B., M.S., A.N., T.B.-J.,
M.H., Z.Z., J.C., S.A.H., P.B.D., and M.G. provided feedback on the
manuscript. A.S., D.E., and R.H. wrote the manuscript. Competing
interests: M.H. is president of Luminous Computing. A.N. is vice
president of system hardware design at Luminous Computing.
T.B.-J. is vice president of engineering at Luminous Computing.
M.S. is vice president of packaging, photonics, and mixed-signal
at Luminous Computing. M.H., A.N., T.B.-J., and M.S. own stock
in Luminous Computing. D.B. is chief scientist at Lightmatter
and holds stock in the company. A.S. has interned at
Lightmatter, receiving a wage. D.E. is an adviser to and holds
shares in Lightmatter but received no support for this work.
Sludds et al., Science 378, 270–276 (2022)
21 October 2022
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
S.B. has received consulting fees from Nokia Corporation. M.G.
owns stock in companies with telecommunication interests
(Google and Microsoft). R.H. and D.E. have filed a patent related
to Netcast: PCT/US21/43593. M.G., Z.Z., L.B., A.S., R.H., and
D.E. have filed a provisional patent related to Netcast: 63/191,120. The
other authors declare that they have no competing interests. Data and
materials availability: Data required to recreate results in the
main text are available in Zenodo (40). License information:
Copyright © 2022 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science. No
claim to original US government works. https://www.science.org/
about/science-licenses-journal-article-reuse
Materials and Methods
Supplementary Text
Figs. S1 to S26
Table S1
References (41–88)
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abq8271
Submitted 3 May 2022; accepted 23 September 2022
10.1126/science.abq8271
Sludds et al., Science 378, 270–276 (2022)
21 October 2022
7 of 7
|
10.1126_science.adc8743
|
RES EARCH
ORGANIC CHEMISTRY
Closed-loop optimization of general reaction
conditions for heteroaryl Suzuki-Miyaura coupling
Nicholas H. Angello1,2†, Vandana Rathore1,2†, Wiktor Beker3, Agnieszka Wołos3,4, Edward R. Jira2,5,
Rafał Roszak3,4, Tony C. Wu6,7, Charles M. Schroeder1,2,5,8, Alán Aspuru-Guzik6,7,9,10,11,12,
Bartosz A. Grzybowski3,4,13,14*, Martin D. Burke1,2,15,16,17*
General conditions for organic reactions are important but rare, and efforts to identify them usually
consider only narrow regions of chemical space. Discovering more general reaction conditions
requires considering vast regions of chemical space derived from a large matrix of substrates
crossed with a high-dimensional matrix of reaction conditions, rendering exhaustive experimentation
impractical. Here, we report a simple closed-loop workflow that leverages data-guided matrix
down-selection, uncertainty-minimizing machine learning, and robotic experimentation to discover
general reaction conditions. Application to the challenging and consequential problem of heteroaryl
Suzuki-Miyaura cross-coupling identified conditions that double the average yield relative to a
widely used benchmark that was previously developed using traditional approaches. This study
provides a practical road map for solving multidimensional chemical optimization problems with
large search spaces.
to high yields (4). Even the application of ma-
chine learning to optimization protocols (5–9)
does not ensure generality, which is critical
for automating, accelerating, and ultimately
democratizing the small molecule–making
process. Identification of such general con-
ditions is difficult because the search space—
spanning all possible combinations of sub-
strates multiplied by all possible combinations
of reaction conditions—is enormous and
thus impractical to navigate using standard
approaches.
Heteroaryl molecular fragments are ubiqui-
tous in many industrially relevant functional
molecules, including pharmaceuticals, mate-
rials, catalysts, dyes, and natural products. In
all of these spaces, synthesis remains a key
bottleneck. Finding general conditions for
(hetero)aryl Suzuki-Miyaura cross-coupling
(SMC) is therefore an important problem. It is
also a challenging and largely unsolved prob-
lem, primarily owing to variable degrees of
both desired and undesired reactivities across
the very large and diverse range of potential
heteroaryl and aryl substrates (10–12). We re-
cently attempted, but failed, to use machine
learning (ML) to discover general reaction
conditions by mining the extensive chemical
literature on (hetero)aryl SMC (13). This is
mainly because the choices of conditions re-
ported in the literature lacked causal links to
the substrates’ structures, and because of a
lack of published (or otherwise accessibly ar-
chived) negative results.
Here, we report a simple closed-loop work-
flow that can efficiently navigate vast substrate-
condition space to discover general reaction
conditions. The approach leverages: (i) data-
guided matrix down-selection to render the
vast search space tractable while retaining val-
idity to the whole; (ii) uncertainty-minimizing
ML to efficiently drive prediction optimization;
and (iii) robotic experimentation to increase
throughput, precision, and reproducibility of
datasets recursively generated on demand
(Fig. 1). We demonstrate that this workflow
succeeds in identifying general reaction con-
ditions for the (hetero)aryl SMC reaction. The
optimized solution doubled the average yield
compared with benchmark general conditions
that had previously been developed through
traditional human-guided experimentation
(hereafter referred to as JACS 2009) (14) and
that have since been used extensively in aca-
demic and industrial laboratories worldwide
(cited in >590 papers and patent applica-
tions). This approach can thus find powerful
T he development of automated synthesis
methods for peptides (1), nucleic acids
(2), and polysaccharides (3) required dis-
covery of highly general reaction con-
ditions applicable to a wide range of
building block combinations. In contrast, in the
synthesis of small organic molecules, bespoke
reaction conditions are usually developed to
maximize the yield of each target molecule,
minimize side products, and/or minimize the
cost of the corresponding process. Reaction
optimization per target is often necessary be-
cause synthetic methods are typically opti-
mized on only one or a few pairs of substrates
and then applied to a wider range of substrate
combinations with the rarely fulfilled hope
that the same conditions will generally lead
1Department of Chemistry, University of Illinois at Urbana-
Champaign, Urbana, IL, USA. 2Beckman Institute for
Advanced Science and Technology, University of Illinois at
Urbana-Champaign, Urbana, IL, USA. 3Allchemy, Inc.,
Highland, IN, USA. 4Institute of Organic Chemistry, Polish
Academy of Sciences, Warsaw, Poland. 5Department of
Chemical and Biomolecular Engineering, University of Illinois
at Urbana-Champaign, Urbana, IL, USA. 6Department of
Chemistry, University of Toronto, Toronto, ON, Canada.
7Department of Computer Science, University of Toronto,
Toronto, ON, Canada. 8Department of Materials Science and
Engineering, University of Illinois at Urbana-Champaign,
Urbana, IL, USA. 9Vector Institute for Artificial Intelligence,
Toronto, ON, Canada. 10Canadian Institute for Advanced
Research, Toronto, ON, Canada. 11Department of Chemical
Engineering and Applied Chemistry, University of Toronto,
Toronto, ON, Canada. 12Department of Materials Science and
Engineering, University of Toronto, Toronto, ON, Canada.
13Center for Soft and Living Matter, Institute for Basic
Science, Ulsan, Republic of Korea. 14Department of
Chemistry, Ulsan Institute of Science and Technology, Ulsan,
Republic of Korea. 15Carl R. Woese Institute for Genomic
Biology, University of Illinois at Urbana-Champaign, Urbana,
IL, USA. 16Cancer Center at Illinois, University of Illinois at
Urbana-Champaign, Urbana, IL, USA. 17Carle Illinois College
of Medicine, University of Illinois at Urbana-Champaign,
Urbana, IL, USA.
*Corresponding author. Email: mdburke@illinois.edu (M.D.B.);
nanogrzybowski@gmail.com (B.A.G.)
†These authors contributed equally to this work.
Angello et al., Science 378, 399–405 (2022)
28 October 2022
1 of 7
Fig. 1. Problem definition and substrate scope for generalized heterocyclic cross-coupling. Workflow
developed in this work for the discovery of general reaction conditions.
RES EARCH | R E S E A R C H A R T I C L E
solutions that lie in vast multidimensional
search spaces and stands to accelerate the
field of organic chemistry’s march toward auto-
mated and democratized small molecule syn-
thesis (15–28), which critically requires more
general reaction conditions.
Data-guided down-selection of substrates
To enable practical pursuit of general hetero(aryl)
SMC reaction conditions, we first strategical-
ly down-selected both the matrix of possible
building block combinations and the matrix
of possible reaction conditions in a way that
preserved relevance of the subsets to their
wholes (Fig. 1). Specifically, we first data mined
the inventories of common fine chemical sup-
pliers and assembled a list of ~5400 (hetero)aryl
halide building blocks that were practically
purchasable and therefore accessible for
study [supplementary materials (SM) sec-
tion 4]. To define a representative subset of
this chemical space, we applied a stratified
clusterization strategy (fig. S21) to algorith-
mically cluster the building blocks by their
common (hetero)aromatic ring substructures
and pendant functionalities, down-selecting
54 “centroid” molecules most representative
of each section of the available chemical space.
Combining these molecules with a selection
of 54 commercially available (hetero)aryl
N-methyliminodiacetic acid (MIDA) boronates
defined a down-selected substrate scope com-
posed of 2688 representative cross-coupling
products (figs. S22 and S23). Mapping this
potential product space and comparing it to
all previously reported heteroaryl products
in the literature revealed substantial overlap
between both sets, suggesting that it is repre-
sentative of heteroaryl chemical space as a
whole (Fig. 2A). However, testing even this
initially down-selected collection of cross-
coupling products against many possible re-
action conditions is technically unfeasible.
Accordingly, we pursued a second layer of
down-selection. Specifically, we used a greedy
algorithm based on the Tanimoto similarity
(29) to identify from this larger collection a set
of 11 representative substrate pairs that max-
imize mutual dissimilarity of the resulting
products (Fig. 2B). For all of these products,
we determined liquid chromatography–mass
Fig. 2. Automated synthesis of the initial training set. (A) T-distributed
stochastic neighbor embedding (t-SNE) mapping of the substrate combinations
(2688 heteroaryl products) examined in this work versus all (hetero)aryl products
previously reported in the literature. Blue circles represent literature-reported
products, yellow stars represent products exclusively belonging in this reported search
space, and green triangles represent products present in both sets. (B) t-SNE
mapping of the product space synthesized during the training and test sets versus the
overall reaction space. Blue circles represent products belonging to the reported
search space, green triangles represent products belonging to the training set, and
yellow stars represent products belonging to the test set. (C) Reaction scheme and
chemical structures of the initial training set. Me, methyl; Et, ethyl. (D) Photo of the
automated synthesis instrument used in this work. (E) Initial training set with the
benchmark condition; all other common palladium catalysts reported in the
literature; and a condition with the most common catalyst [Pd(PPh3)4], base
(Na2CO3), temperature (100°C), and solvent (dioxane:water) used in the literature.
Yields are the average of two automated repetitions (±2% deviation), measured
by LCMS-UV/Vis with an authentic product standard (response factor to
phenanthrene; SM section 9). HPLC, high-performance liquid chromatography.
(F) Yield-based similarity between Pd-based catalysts differing only in organic
ligands. Each square quantifies the Spearman rank correlation coefficient between
yields obtained for each of the 11 substrate pairs. Two pairs of ligands (XPhos
and dppf, and SPhos and PCy3) were highly correlated and redundant. PtBu3,
tri-tert-butylphosphine; PCy3, tricyclohexylphosphine; dba, dibenzylideneacetone;
dppf, 1,1′-bis(diphenylphosphino)ferrocene.
Angello et al., Science 378, 399–405 (2022)
28 October 2022
2 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Closed-loop experimentation and analysis. (A) Convergence of the
model’s uncertainty. Dashed horizontal line depicts threshold for (cid:1)stotal (cid:1) (cid:1)strain
obtained in calibration simulations. The shaded areas correspond to 95%
confidence interval computed by repeatedly training the model 10 times.
(B) Comparison of ML-guided searches versus random searches for general
conditions. Simulations with both model selection policy [probability of
improvement (PI) in conditions space and maximum uncertainty (MaxUnc)
in substrate space for given conditions; abbreviated as PI:MaxUnc and
corresponding to green lines] and random selection of the next reactions
(red lines) were repeated 100 times to evaluate the random factor in the
algorithm (random initialization of neural network weights as well as selection
of the next step in the random baseline). The shaded areas mark the
Angello et al., Science 378, 399–405 (2022)
28 October 2022
3 of 7
RES EARCH | R E S E A R C H A R T I C L E
interquartile range. (C) Comparison of yield distribution between literature-
reported reactions and those explored in this work. (D) Color scale indicates the
percent yield per general condition as perceived by the ML model at the
conclusion of a given optimization round, 1 to 5. Along the horizontal axis,
the conditions are ranked according to the ML prediction after round 5.
(E) Ranking per general condition per round as perceived by the ML model.
(F) Uncertainty per general condition per round as perceived by the ML model
(computed from 10 repetitions). (G) Number of substrates tested per general
condition per round of closed-loop optimization. (H) Coverage of reaction
space tested by round 5 of closed-loop experimentation. A value of 1 indicates
that the condition was tested, and a value of 0 indicates that the condition
was not tested. (I) Distributions of yields measured and predicted prior to the
measurement for each round of closed-loop optimization. MAE, mean absolute
error; MSE, mean squared error.
spectrometry ultraviolet/visible spectroscopy
(LCMS-UV/Vis) response factor curves, which
enabled us to automatically measure the yields
of automated reactions (SM sections 8 and 9).
Data-guided down-selection of conditions
We considered four condition variables—
solvent, base, catalyst and ligand, and temper-
ature. As our aim was to test a broad range of
conditions, we initially down-selected repre-
sentatives of condition classes on the basis of
not only their extent of prior use from our
earlier comprehensive literature analysis (13)
but also structural and functional diversity.
For instance, whereas the two most commonly
used solvents in the literature are dioxane and
dimethoxyethane, they both belong to the same
solvent class of ethers, and so we selected only
one of them (dioxane). Similar reasoning led
us to keep only one carbonate base [in (13), we
showed that the nature of the cation did not
alter the yields]. We selected 100°C as the most
frequently used temperature in the literature,
as well as 60°C, which was used in the pre-
viously developed benchmark protocol (14). In
the end, we selected three solvents (dioxane,
toluene, and dimethylformamide, all used in
5:1 mixture with water), two bases (sodium
carbonate and potassium phosphate), two
temperatures (60° and 100°C), and seven cat-
alysts [Pd SPhos G4, Pd(PPh3)4, Pd XPhos G4,
Pd P(tBu)3 G4, Pd PCy3 G4, Pd2(dba)3, and
Pd(dppf)Cl2; G4 refers to the fourth-generation
Buchwald precatalyst] to evaluate. The down-
selected 11 building block combinations de-
scribed above were tested under an initial set
of conditions to “seed” the ML optimization
(Fig. 2C) and then tested iteratively under a
much broader set of conditions during the
ML-guided optimization phase.
Seeding experiments, reaction standardization,
and conditions space
All reactions were performed automatically on
the robotic system shown in Fig. 2D. Before
solvent addition, heating, and stirring, reac-
tion mixtures were purged with 10 automated
vacuum and argon cycles, which led to highly
reproducible reaction yields (fig. S11). This
automated Schlenk process was necessary—
even when using air-stable precatalysts and
building blocks—for reproducibility. To “seed”
the optimization procedure, we performed all
couplings between the aforementioned 11 pairs
of substrates, each under seven different con-
ditions: those corresponding to the JACS 2009
benchmark (5:1 dioxane:water, 60°C, K3PO4,
Pd SPhos G4); same base and solvents but
with the other selected palladium catalysts
[Pd XPhos G4, Pd P(tBu)3 G4, Pd PCy3 G4,
Pd2(dba)3, and Pd(dppf)Cl2]; and a condition
with the most common catalyst [Pd(PPh3)4],
base (Na2CO3), temperature (100°C), and sol-
vent (dioxane:water) used in the literature
(Fig. 2E). When each reaction was repeated
twice, the yields exhibited only ±2% devia-
tion, underscoring one of the key advan-
tages of automated experimentation [indeed,
it has previously been reported (30) that
repetition of the same reaction even by the
same human experts entails variability of
∼10 to 15%].
This initial round of experiments also allowed
us to identify catalysts that, for different sub-
strate pairs, systematically gave similar yields
and could thus be redundant. Such functional
rather than structural similarity is quantified
by the Spearman rank matrix shown in Fig. 2F
and correlating yields obtained for all 11
substrate pairs using two different catalyst
ligands—in this representation, redundant
catalysts correspond to high-correlation, off-
diagonal elements (e.g., XPhos and dppf, or
PCy3 and SPhos). On the basis of this analysis,
we eliminated PCy3 and dppf from our pool of
ligands to decrease redundancy, and we elimi-
nated Pd2(dba)3 because of poor performance
(<5% yield for 8 of 11 substrates), yielding a full
space to be explored of 528 reactions (11 sub-
strates × 2 temperatures × 2 bases × 3 solvents ×
4 catalysts).
Uncertainty-minimizing ML for generality
A reaction condition can be considered max-
imally general when it provides the highest
average yield across the widest range of chem-
ical space. Optimization for generality is an
unsolved and underexplored challenge in the
evolving field of ML. We thus considered an
alternative de novo approach, where small
sets of highly reproducible data are generated
on demand during ML-guided closed-loop
optimization, including negative data vastly
underrepresented in existing datasets. We
also decided to strategically focus the ML algo-
rithm on decreasing model uncertainty and
thereby maximize the efficiency of the learn-
ing process.
Denoting the set of possible reaction con-
ditions as C = {c}, a set of substrate pairs as S =
{s}, and reaction yield as y(s,c), our aim is to
maximize the objective function given by
f cð Þ ¼
X
1
Sj
j
Þ
y s; c
ð
s∈S
ð1Þ
Then, the general conditions cgeneral are given as
cgeneral ¼ arg max
c∈C
f cð Þ
ð2Þ
At first glance, the problem of identifying
cgeneral in the least number of experiments
resembles standard Bayesian optimization
(BO). However, there is a substantial differ-
ence: In all BO algorithms, each experiment
or measurement performed immediately pro-
vides information about the objective func-
tion desired for optimization. In contrast,
experimental evaluation of f(c) in our problem
requires multiple experiments (because sum-
mation in Eq. 1 runs over the entire set S)—that
is, determination of f (c) for given conditions
requires experiments with every pair of sub-
strates in the S set. To address this problem,
we modified the standard BO approach by
constructing a surrogate model for predict-
ing reaction yields, ^y s; c
Þ. We then used its
ð
predictions to estimate ^f cð Þ according to
Eq. 1 and using the model’s prediction for
the yet-unperformed reactions. Note that in
standard BO, we would have observed f(c) for
the “seen” conditions and estimated for the
“unseen” ones; in our case, f(c) was estimated
even for the already-tested conditions, unless
the entire substrate space S had already been
tested. On the basis of these considerations,
the optimization over C (selection of the next
conditions to examine) is performed with stan-
dard BO techniques, whereas sampling of S is
achieved using an active learning approach—
each of these techniques is known on its own
(8, 31), but the combination of the two (where
the observation of the BO objective can be in-
complete) seems to be unknown. In particular,
we decided to choose substrate pairs on the
basis of the model’s prediction of uncertainty
for given substrates under given reaction con-
ditions: the highly uncertain (low-confidence)
predictions indicate missing information, and
providing the model with the corresponding
experimental data should decrease its uncer-
tainty the most.
Angello et al., Science 378, 399–405 (2022)
28 October 2022
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Test set for
ML-discovered reaction
conditions. (A) Set of
20 diverse compounds from
outside the training set
selected to test whether the
discovered general reaction
conditions translate to other
diverse heteroaryl product
classes. JACS 2009 condition:
5:1 dioxane:water, 60°C, K3PO4,
Pd SPhos G4.. ML General
condition 1: 5:1 dioxane:water,
100°C, Na2CO3, Pd XPhos G4.
ML General condition 2:
5:1 dioxane:water, 100°C,
Na2CO3, Pd SPhos G4. ML
general condition 3: 5:1
dioxane:water, 100°C, Na2CO3,
Pd(PPh3)4. ND, not detected.
(B) Jitter plot showing the
performance of the top
ML conditions versus the
benchmark. Brackets indicate
95% confidence interval.
(C) Jitter plot showing the
relative performance in
change of yield of the top
ML conditions versus the
benchmark. Brackets indicate
95% confidence interval.
(D) The number of products
per general condition with
>10% yield measured.
(E) Relative protodeboronation
per condition, as measured
by integrated UV peak area
(UVPDB) standardized to
the internal standard (UVSTD).
(F) Relative remaining halide
per condition, as measured by
integrated UV peak area
(UVHAL) standardized to the
internal standard (UVSTD).
(G) Relative product forma-
tion per condition (UVPDT)
relative to by-product forma-
tion (UVBYPDT).
For uncertainty estimation, we sought a
model offering prediction uncertainty com-
mensurate with prediction error—for instance,
highly confident predictions with high error
are undesirable. Per the analyses of numer-
ous neural-network (NN) and Gaussian process
(GP) models (SM section 3), we ultimately
selected an ensemble of GP supplemented
with a NN kernel component [GPE(NN)]; sim-
ilar approaches were recently used in BO (32)
and interactive learning (33). Such a model is
particularly appealing because of its flexibil-
ity (the similarity metric between different
conditions will be learned from the data) and
the reliability of the prediction uncertainties
Angello et al., Science 378, 399–405 (2022)
28 October 2022
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
(e.g., it is guaranteed that the prediction of
a test sample will not be more confident
than a training sample). For the selection of
conditions to be tested, we chose as an acqui-
sition function the probability of improve-
ment (PI).
Closed-loop, ML-driven optimization with
robotic experimentation
The GPE(NN)/PI model guided the auto-
mated experiments over the down-selected
search space. We worked in experimental
batches, meaning that multiple experiments
were performed before the theoretical model
was updated. Within each batch, the algo-
rithm formed a “priority queue” of unexplored
reactions by sorting and selecting condi-
tions according to the computed PI and
substrate pairs according to the prediction
uncertainty. The batch size for rounds 1 and
2 of optimization was 36 duplicated reac-
tions, followed by 72 and 84 unduplicated
reactions for rounds 3 and 4 and round 5,
respectively.
Over the closed-loop rounds, the model’s
uncertainty decreased and converged at the
fifth round to the threshold obtained during
calibration simulations (Fig. 3A and fig. S20),
suggesting that the model gained sufficient
knowledge about the whole space, at which
point the optimization was terminated. This
strategy converges to this optimum in about
half as many reactions as does random sampling
(Fig. 3B) and with a higher likelihood of suc-
cess compared with typical BO strategies (fig.
S30). As the algorithm explored the reaction-
condition space, reaction yields for our data-
set were distributed more or less uniformly
over the range of possible values (Fig. 3C). In
other words, our protocol learned by probing
both low- and high-yielding conditions. This
contrasts with the distribution of yields in
published reaction sets—such yields are heavi-
ly skewed toward positive outcomes, which,
as we discussed in many of our previous works
on computerized synthesis (34–36), limits the
usefulness of approaches aiming to learn from
published datasets.
In this dataset, the discovered top-1 condi-
tion conferred 72% average yield across all
11 substrates, whereas the benchmark condi-
tion [found to also be the top-5 (i.e., fifth best)
condition] conferred 64% average yield. To
understand how the model arrived at this op-
timum, we examined the model’s perception
of the average yield and ranking of each gen-
eral condition per round (Fig. 3, D and E).
Within the first two rounds, the model gains
the ability to accurately categorize these con-
ditions into high, medium, and low overall
average yield and, in the subsequent rounds,
establishes the correct ranking within these
categories. The increasing accuracy of the
model over the course of the experiment is
recapitulated in Fig. 3F, which shows the
model’s ranking uncertainty decreasing be-
tween rounds and is especially apparent for
the top conditions. The model chose to test a
few substrates per round across many con-
ditions for the first three rounds, followed by
primarily filling in the top conditions in the
latter rounds (Fig. 3G). By the fifth round, the
model explored nearly all of the top-7 con-
ditions, which corresponds to every condition
with >50% overall average yield, as estimated
by the model (Fig. 3H).
Finally, we analyzed the yields of reactions
the model requested in order to gain more in-
formation about the reaction-condition space
(Fig. 3I). These values are not expected to in-
crease as the optimization progresses, because
the yield of a single experiment is not our
objective. Given the uncertainty-guided selec-
tion of substrates, one could even expect the
opposite: Once a set of suitable conditions is
identified, further exploitation should in-
volve lower-yielding reactions to verify that
the found candidate conditions are indeed
better, as well as to increase confidence of the
estimate of f (c). The results shown in Fig. 3I
indicate that after exploring good reactions
in the second iteration, the model gradually
shifted its attention toward the parts of the
reaction-condition space that can be consid-
ered as “negative examples” (and, in doing so,
improved its prediction accuracy). From these
results, it appears that (i) relatively good can-
didate solutions were identified early, (ii) the
model initially tried to look for better-yielding
reactions (to find better candidates), and (iii)
more and more attention was dedicated to
decreasing the uncertainty of its estimates as
the “loop” progressed.
Quantifying generality
After the discovery of higher-yielding gen-
eral conditions in the training set, we next
sought to determine whether the learning
would transfer to substrates outside of the
optimization—specifically, over 20 substrate
pairs chosen [by the Butina algorithm (37)]
to maximize dissimilarity to the training set
while ensuring coverage of the heterocyclic
substructure and functional group space (Fig.
2B). We then set out to synthesize and purify
all of the computer’s suggestions and test them
against the benchmark condition and the top-3
highest yielding general reaction conditions
discovered during the closed-loop optimiza-
tion (Fig. 4A), as ranked by the model after
the completion of round 5. Despite including
some very challenging building block combi-
nations, this process was 95% successful, with
only one product having no measurable yield
under all four conditions.
The ML-discovered general reaction condi-
tions performed substantially better than the
previously reported and widely used bench-
mark condition (14). The top-2 conditions
provided statistically significant increases in
average yield compared with the benchmark,
with the top condition doubling the overall
average yield from 21% to 46% (Fig. 4B). Com-
paring the relative increase in yield reveals
statistically significant differences between
the top-1 and both the top-2 and top-3 con-
ditions (Fig. 4C). Notably, the experimental
yields correlate with the predicted ranking
of the conditions such that the yield for the
top-1 is higher than that for the top-2, which,
in turn, is higher than that for the top-3. In
the context of functional discovery efforts, the
binary capacity to isolate or not isolate testable
quantities of purified targeted compounds is
arguably even more important than the spe-
cific percent yield. We estimate that the prac-
tical limit for isolating purified products is
10% yield. For the benchmark condition, only
11/20 targeted products cleared this bar, whereas
this fraction rose to 19/20 for the top-1 condi-
tion (Fig. 4D).
Extending the reaction times for couplings
that were low yielding under the benchmark
conditions did not increase yields (fig. S33).
Comprehensive analysis of by-products and
product formation for all 20 reactions (SM
section 10) demonstrated that a favorable
shift from the former to the latter accom-
panies the shift from the benchmark to the ML-
discovered reaction conditions. Specifically, the
ML-discovered conditions were associated with
a trend toward decreased protodeboronation
(Fig. 4E), increased halide conversion (Fig. 4F),
and an overall statistically significant increase
in the ratio of product to total by-products
formation (Fig. 4G) (0.30 ± 0.12 for JACS
2009 versus 0.58 ± 0.12 for ML conditions;
P = 0.0005).
Outlook
The straightforward workflow developed here
has enabled the accelerated discovery of im-
proved general reaction conditions for diffi-
cult C–C bond forming reactions, representing
a key step toward increasing the efficiency,
generality, and accessibility of small molecule
synthesis. This result also highlights the power
of down-selection as an entry point into large
multidimensional search spaces, the distinct
advantages of a de novo ML approach for
navigating such spaces by generating datasets
that evenly reflect the reality of positive and
negative data during optimization, and the
particular suitability of robotized chemistry
for generating high-quality, reproducible data.
Future studies will incorporate next-generation
ligands and reagents to yield further improved
general reaction conditions, creating an ac-
tionable path for automated small molecule
synthesis to achieve reaction efficiencies ap-
proaching that of automated peptide synthesis.
This general workflow should be applicable to
Angello et al., Science 378, 399–405 (2022)
28 October 2022
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
the optimization of reactions beyond SMC, as it
did not require extensive prior literature data
to be successful.
RE FE RENCES AND N OT ES
1. R. B. Merrifield, Science 150, 178–185 (1965).
2. M. H. Caruthers, Science 230, 281–285 (1985).
3. O. J. Plante, E. R. Palmacci, P. H. Seeberger, Science 291,
1523–1527 (2001).
I. W. Davies, Nature 570, 175–181 (2019).
4.
5. D. T. Ahneman, J. G. Estrada, S. Lin, S. D. Dreher, A. G. Doyle,
Science 360, 186–190 (2018).
6. C. W. Coley et al., Science 365, eaax1566 (2019).
7. M. Christensen et al., Commun. Chem. 4, 112 (2021).
8. B. J. Shields et al., Nature 590, 89–96 (2021).
9. A. B. Clayton et al., React. Chem. Eng. 4, 1545–1554 (2019).
10. P. A. Cox, A. G. Leach, A. D. Campbell, G. C. Lloyd-Jones,
J. Am. Chem. Soc. 138, 9145–9157 (2016).
11. P. A. Cox et al., J. Am. Chem. Soc. 139, 13156–13165 (2017).
12. H. L. D. Hayes et al., J. Am. Chem. Soc. 143, 14814–14826 (2021).
13. W. Beker et al., J. Am. Chem. Soc. 144, 4819–4827 (2022).
14. D. M. Knapp, E. P. Gillis, M. D. Burke, J. Am. Chem. Soc. 131,
6961–6963 (2009).
15. J. Li et al., Science 347, 1221–1226 (2015).
16. A. Buitrago Santanilla et al., Science 347, 49–53 (2015).
17. A.-C. Bédard et al., Science 361, 1220–1225 (2018).
18. J. M. Granda, L. Donina, V. Dragone, D.-L. Long, L. Cronin,
Nature 559, 377–381 (2018).
19. S. Lin et al., Science 361, eaar6236 (2018).
20. D. Perera et al., Science 359, 429–434 (2018).
21. M. Trobe, M. D. Burke, Angew. Chem. Int. Ed. 57, 4192–4214 (2018).
22. S. Steiner et al., Science 363, eaav2211 (2019).
23. S. Chatterjee, M. Guidi, P. H. Seeberger, K. Gilmore, Nature
579, 379–384 (2020).
24. S. H. M. Mehr, M. Craven, A. I. Leonov, G. Keenan, L. Cronin,
Science 370, 101–108 (2020).
25. C. Liu et al., Nat. Chem. 13, 451–457 (2021).
26. D. J. Blair et al., Nature 604, 92–97 (2022).
27. V. Fasano et al., ChemRxiv [Preprint] (2020). https://doi.org/
10.26434/chemrxiv-2022-zwctv-v2.
28. A. Bender et al., Nat. Rev. Chem. 6, 428–442 (2022).
29. D. Bajusz, A. Rácz, K. Héberger, J. Cheminform. 7, 20
(2015).
30. F. S. Emami et al., Angew. Chem. Int. Ed. 54, 10797–10801
(2015).
31. Y. Gong, D. Xue, G. Chuai, J. Yu, Q. Liu, Chem. Sci. 12,
14459–14472 (2021).
32. M. Wistuba, J. Grabocka, arXiv:2101.07667 [cs.LG] (2021).
33. J. Bowden, J. Song, Y. Chen, Y. Yue, T. A. Desautels, “Deep
kernel Bayesian optimization,” 37th Conference on Uncertainty
in Artificial Intelligence (UAI 2021), 27 to 30 July 2021.
34. T. Klucznik et al., Chem 4, 522–532 (2018).
35. B. Mikulak-Klucznik et al., Nature 588, 83–88 (2020).
36. F. Strieth-Kalthoff et al., Angew. Chem. Int. Ed. 61, e202204647
(2022).
37. D. Butina, J. Chem. Inf. Comput. Sci. 39, 747–750 (1999).
38. N. H. Angello et al., Closed-loop optimization of general
conditions for heteroaryl Suzuki coupling, version 0.1.2, Zenodo
(2022); https://doi.org/10.5281/zenodo.7099435.
39. N. H. Angello et al., Closed-loop optimization of general
conditions for heteroaryl Suzuki coupling, version 2, Zenodo
(2022); https://doi.org/10.5281/zenodo.7106075.
40. verysure, P. Schleich, aspuru-guzik-group/b3: Release for
Science Publication, version 1.0, Zenodo (2022);
https://doi.org/10.5281/zenodo.7094835.
41. N. H. Angello et al., Closed-loop optimization of general
conditions for heteroaryl Suzuki coupling, version 1.0, Zenodo
(2022); https://doi.org/10.5281/zenodo.6517012.
ACKN OWLED GMEN TS
Funding: This work was supported by the Defense Advanced
Research Projects Agency under the Accelerated Molecular
Discovery Program (cooperative agreement no. HR00111920027
dated 1 August 2019) to M.D.B., B.A.G., and A.A.-G. The content of
the information presented in this work does not necessarily reflect
the position or the policy of the US government. This work was
also supported by funding from the Molecule Maker Lab Institute:
An AI Research Institutes program supported by NSF under award
no. 2019897 to C.M.S. Any opinions, findings, and conclusions or
recommendations expressed in this material are those of the
author(s) and do not necessarily reflect those of the National
Science Foundation. N.H.A. was supported by the Department
of Defense (DoD) through the National Defense Science and
Engineering Graduate (NDSEG) Fellowship Program. B.A.G. was
supported by the Institute for Basic Science, Korea (project code
IBS-R020-D1). Author contributions: N.H.A., V.R., W.B., A.W.,
R.R., B.A.G., and M.D.B. conceived of the study. N.H.A. and E.R.J.
developed and validated the automated synthesizer. V.R. and
N.H.A. performed the automated synthesis experiments. V.R. and
N.H.A. purified, characterized, and generated response factor
curves for new compounds. W.B., A.W., R.R., and B.A.G. developed
the closed-loop machine-learning strategy, building block
clustering, and chemical space quantifications. N.H.A. and T.C.W.
data mined the substrate scope building block set. M.D.B.,
B.A.G., A.A.-G., and C.M.S. supervised the research. The manuscript
was written by N.H.A., V.R., W.B., A.W., R.R., B.A.G., and M.D.B.,
with contributions from all other authors. Competing interests: The
University of Illinois has filed patent applications related to MIDA
boronates (US20170002026A1, US9845317B2, WO2011103435A3,
and US20160207943A1 with M.D.B. as inventor). A.A.-G. is chief
visionary officer and board member of Kebotix, Inc., a company that
carries out closed-loop molecular materials discovery. Data and
materials availability: All data and code generated as part of this
study are freely accessible either in the supplementary materials
or in open repositories. Code for the simulation example, model
selection, and substrates clusterization as well machine- and
human-readable versions of the reaction data are freely accessible
at Zenodo (38). The automated synthesis code, machine parts list
and build guide, datamined commercially available building block
set, checklist for reporting and evaluating machine learning models,
and tabulated numerical data underlying Figs. 2 to 4 are deposited
at Zenodo (39). The code used in the building block selection
process is available at Zenodo (40). All code related to the closed-
loop optimization is freely available at Zenodo (41). License
information: Copyright © 2022 the authors, some rights reserved;
exclusive licensee American Association for the Advancement of
Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adc8743
Materials and Methods
Supplementary Text
Figs. S1 to S56
Tables S1 to S3
References (42–50)
Submitted 5 May 2022; resubmitted 23 August 2022
Accepted 29 September 2022
10.1126/science.adc8743
Angello et al., Science 378, 399–405 (2022)
28 October 2022
7 of 7
|
10.1126_science.abh0474
|
RES EARCH
BIOMEDICINE
An autonomously swimming biohybrid fish designed
with human cardiac biophysics
Keel Yong Lee1†, Sung-Jin Park1,2†, David G. Matthews3, Sean L. Kim1, Carlos Antonio Marquez1,
John F. Zimmerman1, Herdeline Ann M. Ardoña1‡§, Andre G. Kleber4,
George V. Lauder3, Kevin Kit Parker1,5,6*
Biohybrid systems have been developed to better understand the design principles and coordination
mechanisms of biological systems. We consider whether two functional regulatory features of the
heart—mechanoelectrical signaling and automaticity—could be transferred to a synthetic analog of
another fluid transport system: a swimming fish. By leveraging cardiac mechanoelectrical signaling, we
recreated reciprocal contraction and relaxation in a muscular bilayer construct where each contraction
occurs automatically as a response to the stretching of an antagonistic muscle pair. Further, to
entrain this closed-loop actuation cycle, we engineered an electrically autonomous pacing node, which
enhanced spontaneous contraction. The biohybrid fish equipped with intrinsic control strategies
demonstrated self-sustained body–caudal fin swimming, highlighting the role of feedback mechanisms
in muscular pumps such as the heart and muscles.
C irculatory systems in living organisms
are intricately designed to transport
blood throughout the body. Their most
basic function is fluid transport, and a
diversity of similar fluid pumping mech-
anisms and designs are found throughout
nature (1). Fluid pumps in vertebrates, con-
sidered broadly, range from a human circula-
tory system with closed vessels within which
fluid moves, to oscillatory fluid mechanisms in
aquatic species in which fluid is transported
along the body to generate propulsive thrust.
Inspired by these distinct but similar natural
processes, we and others have developed bio-
hybrid analogs of an external fluid pump ca-
pable of mimicking the locomotion of aquatic
species (2–4). The underlying motivation for
developing biohybrid systems capable of repro-
ducing biological behaviors is to better under-
stand the design principles and coordination
mechanisms of biological systems, although
the performance of these systems has been
lacking in comparison to natural fluid transport
pumps (4).
A key feature of aquatic species is closed-
loop actuation of antagonistic musculature that
1Disease Biophysics Group, John A. Paulson School of
Engineering and Applied Sciences, Harvard University,
Boston, MA 02134, USA. 2Biohybrid Systems Group, Coulter
Department of Biomedical Engineering, Georgia Institute of
Technology and Emory University School of Medicine,
Atlanta, GA 30322, USA. 3Museum of Comparative Zoology,
Harvard University, Cambridge, MA 02138, USA. 4Beth Israel
Deaconess Medical Center, Harvard Medical School, Boston,
MA 02115, USA. 5Wyss Institute for Biologically Inspired
Engineering, Harvard University, Boston, MA 02115, USA.
6Harvard Stem Cell Institute, Harvard University, Cambridge,
MA 02138, USA.
*Corresponding author. Email: kkparker@seas.harvard.edu
†These authors contributed equally to this work.
‡Present address: Department of Chemical and Biomolecular
Engineering, Henry Samueli School of Engineering, University of
California Irvine, CA 92697, USA.
§Present address: Sue and Bill Gross Stem Cell Research Center,
University of California, Irvine CA 92697, USA.
provides control over the direction of momen-
tum transfer from the body muscles to the
fluid, enabling efficient locomotion. Similarly,
in the circulatory system, muscles of the heart
dynamically respond to physiological demands
through internal feedback systems and impart
momentum to drive fluid motion. Mechano-
electrical signaling and cardiac automaticity
play an essential role in regulating the con-
tractile pace and strength in a closed-loop con-
trol system (Fig. 1, A to C). Mechanoelectrical
signaling (5, 6) is hypothesized to regulate
intracardiac feedback, which allows cardio-
myocytes (CMs) to adaptively respond to dy-
namic mechanical pressures (7, 8) by inducing
changes in electrophysiology through stretch-
activated mechanosensitive proteins (9, 10)
(Fig. 1B). Automaticity of the heart stems from
the sinoatrial node, which is structurally and
functionally insulated from the surrounding
myocardium (11–13) and initiates spontaneous
electrical activity in the absence of an external
stimulus and without direct neural interven-
tion (Fig. 1C).
We reasoned that using principles of cardiac
control systems to design a biohybrid platform
could result in a fluid pumping system with
comparable efficiencies to natural fishlike fluid
pumping systems. Leveraging fundamental fea-
tures of cardiac function allows for autono-
mous self-pacing and independent motion
control while providing the basis for a closed-
loop design that mimics aquatic swimming
systems. We designed, built, and tested a bio-
hybrid fish equipped with an antagonistic mus-
cular bilayer and a geometrically insulated
cardiac tissue node (G-node) with human stem
cell–derived CMs or neonatal rat ventricular
CMs (Fig. 1, D to F) to test the ability of a bio-
hybrid system to control the movement of
fluids with biological levels of performance. To
integrate mechanoelectrical signaling of CMs
in a simplified biohybrid platform, we recre-
ated asynchronous muscle contractions (Fig.
1D) based on insect muscles (fig. S1) (14). In
insects, each contraction results automat-
ically from a response to the stretching of an
antagonistic muscle pair, generating self-
sustained muscle contraction cycles. In the
muscular bilayer construct of the biohybrid
fish (Fig. 1D), CMs are electrically connected
within each side and mechanically coupled
across sides, so that the shortening of con-
tracting muscles on each side directly trans-
lates to axial stretching of the opposite side
muscle, leading to antagonistic muscle ex-
citations and contractions. To replicate the
electrically insulated structure of a sinoatrial
node (Fig. 1C) (15), we functionally isolated a
small number of CMs (the source) in the G-
node (Fig. 1E) with a single exit pathway that
allows for an electrical connection between
the G-node and muscle tissues (the sink).
This facilitated the activation of large down-
stream quiescent muscle cells (sink) with a
small number of activating CMs (source) by
reducing the impedance between source and
sink (11–13, 15, 16). Together, the muscular bi-
layer and G-node in the biohybrid fish (Fig. 1F)
enabled the generation of continuous rhythms
to regulate its antagonistic muscle pair to
produce spontaneous yet coordinated body–
caudal fin (BCF) propulsion swimming.
Antagonistic contraction of muscular
bilayer construct
We developed a muscular bilayer construct
by modifying hydrogel-based muscular thin
films (16–18). The double-sided micromolded
gelatin thin film (200 mm thick) was engi-
neered by sandwiching a gelatin and cross-
linker (microbial transglutaminase) mixture
between two polydimethylsiloxane stamps
with line groove features (25 mm ridge width,
4 mm groove width, and 5 mm groove depth).
CMs were then seeded onto both sides of
the micromolded gelatin so that they could
self-assemble as laminar, anisotropic mus-
cle with engineered cellular alignment, char-
acteristic of the ventricular myocardium (Fig.
1, G and H).
To demonstrate independent activation be-
tween the muscular bilayer tissues, we used
lentiviral transduction to express blue-light–
sensitive (ChR2) (19) and red-light–sensitive
(ChrimsonR) (20) ion channels in each mus-
cle layer (Fig. 1, H and I, fig. S2, and movie
S1). Alternating blue-and-red light stimula-
tion (15-ms pulses of 450 and 620 nm light,
respectively) activated ChR2- and ChrimsonR-
expressing muscle layers independently. The
shortening of contracting muscles on each
side was transduced to produce antagonistic
bending stress and oscillate the muscle con-
struct along the longitudinal axis (fig. S3 and
movie S2). The contractions and relaxations
Lee et al., Science 375, 639–647 (2022)
11 February 2022
1 of 8
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Design and assembly of the biohybrid fish. (A) Intrinsic autonomous
muscle control of the heart. (B) Mechano-electrical signaling that adaptively
responds to dynamic mechanical pressures by inducing changes in electrophysiology
through stretch-activated mechanosensitive proteins. (C) Automaticity of the
cardiac sinoatrial node, which is structurally and functionally insulated from
the surrounding myocardium and initiates spontaneous electrical activity.
(D to F) Autonomously swimming biohybrid fish. (D) Muscular bilayer in which
the shortening of contracting muscles on each side directly translates to
axial stretching of the opposite side muscle, leading to stretch-induced
antagonistic muscle contractions. (E) G-node, where functionally isolated cardio-
myocytes (CMs) generate spontaneous muscle activation rhythms. (F) Biohybrid fish
equipped with the muscular bilayer and the G-node. (G) Image of the biohybrid fish
made of human stem cell–derived CMs. (H and I) Muscular bilayer construct
showing representative (H) mesoarchitecture and (I) microarchitecture. The gelatin
posterior body was sandwiched by two muscle tissues expressing either a blue-light–
sensitive opsin [ChR2 (green)] or a red-light–sensitive opsin [ChrimsonR (red)].
Representative immunostaining images of both tissues (sarcomeric alpha actinin,
gray; nuclei, blue) show that Z-lines of the sarcomeres (the cell force-generating
units) are perpendicular to the antero-posterior axis. (J) Five layers of body
architecture: the body was symmetrical along the left-right axis but asymmetrical
along both the antero-posterior and dorso-ventral axis; this design was chosen
to maintain directional body stability against roll and propel the body forward.
of muscular bilayer muscles were decoupled
at low pacing frequencies (e.g., 1 and 1.5 Hz),
but at higher pacing frequencies (e.g., 2.5 and
3 Hz), the relaxation of one side started to
overlap with the subsequent contraction of
the other side (fig. S3 and movie S2). The
overlapping, fast, active contraction of the
opposite-side muscle considerably increased
the oscillating speed of the muscular bilayer
construct (fig. S3), preventing diastolic stress
development that single-layered muscular
thin films exhibit at high pacing frequencies
(17, 18). These antagonistic muscle contrac-
tions in the muscular bilayer construct per-
mitted large peak-to-peak amplitudes over
a wide range of pacing frequencies (fig. S3),
in contrast to single-layered muscular thin
films (17, 18).
Integration of the muscular bilayer into
biohybrid fish
The muscular bilayer construct was integrated
into the biohybrid fish (16) by means of tissue
engineering techniques (fig. S4). Inspired by
fish musculoskeletal structure (fig. S5), we
created an asymmetrical body along both the
antero-posterior and dorso-ventral axes while
maintaining sagittal symmetry through a five-
layered architecture. From left to right (Fig.
1J), the biohybrid fish consists of (i) a layer of
aligned muscle tissue made of human stem
cell–derived CMs, (ii) a rigid paper layer in the
anterior body and caudal fin fabricated by
laser ablation, (iii) a compliant gelatin layer in
the posterior body cast by means of a three-
dimensional elastomer polydimethylsiloxane
mold, (iv) a second paper layer, and (v) a sec-
ond aligned muscle tissue layer for forming
the antagonistic muscle pair. The passive com-
ponent of the biohybrid fish is made up of paper
(thickness 190 mm; Young’s modulus 4 GPa;
density 1.2 g/ml), the gelatin body (thickness
192.22 ± 1.95 mm; Young’s modulus 56 kPa;
density 1.5 g/ml), and a plastic floater fin (thick-
ness 1 mm; Young’s modulus 1.3 GPa; density
Lee et al., Science 375, 639–647 (2022)
11 February 2022
2 of 8
RES EARCH | R E S E A R C H A R T I C L E
0.833 g/ml) was designed to maintain direc-
tional body stability and neutral buoyancy
while minimizing drag during forward swim-
ming. The large surface area of the floater fin
combined with the relatively heavy weight of
the hydrogel insert in the anterior ventral
portion of the body helped the fish maintain
an upright orientation. Neutral buoyancy was
achieved by adjusting the size of the plastic
floater fin, thereby matching the average den-
sity of the biohybrid fish to the media in which
it was suspended. The active component of
the biohybrid fish consists of a muscular bi-
layer construct on the flexible posterior gela-
tin hydrogel body and operates as a single
self-propelling system through coordinated
contraction of muscle tissues. The final over-
all design (fig. S6) consists of 73,000 live CMs
in a hydrogel-paper composite body 14 mm in
length and 25.0 mg of total mass, including
0.36 mg muscle mass (fig. S7).
Optogenetically induced BCF propulsion
To characterize system-level kinematics of
the muscular bilayer, we controlled antago-
nistic muscle contractions in the biohybrid
fish by external optogenetic stimulation (Fig. 2).
We stimulated the muscular bilayers by alter-
nating blue and red light-emitting diode light
pulses (Fig. 2A) while the bilayers were sub-
merged in a 37°C Tyrode’s salt solution con-
taining glucose. As shown in the video-tracking
analysis (Fig. 2, B to H, and movie S3), the
biohybrid fish (i) initiated contraction of the
muscle tissue on the left side upon red light
stimulation and produced a peak oscillation
amplitude in the tail (Fig. 2, B, F, and I); (ii)
induced contraction of the muscle tissue on
the right side after blue light stimulation (180°
phase shift between red and blue lights); (iii)
recovered its tail at a near-straight position
(Fig. 2, C and G) and reached peak thrust
production (Fig. 2I); (iv) oscillated its tail with
peak amplitude right before a subsequent red
light stimulation (Fig. 2, D and I); and (v)
rebounded back to a near-straight position
(Fig. 2E) generating maximal thrust (Fig. 2I).
As shown by the lateral deflection (Fig. 2H),
the body curvature (Fig. 2J), and swimming
displacement (Fig. 2K), the biohybrid fish
generated rhythmic forward thrust repro-
ducing BCF propulsion. The biohybrid fish
deformed its posterior body with a single
bend while switching between positive and
negative posterior body curvature upon light
stimulation. The biohybrid fish oscillated its
fin instead of generating a bending body wave
because optical stimulation induced a simul-
taneous global muscle contraction. The rela-
tively stiff anterior body and caudal fin resisted
deformation from fluid forces. This allowed
the biohybrid fish to exhibit asymmetric body
deformation in which the largest lateral de-
flections and curvatures occurred in the pos-
terior body between 0.5 and 0.8 of total length
(Fig. 2J), in a manner reminiscent of BCF
swimmers (fig. S8).
Antagonistic muscle contractions of the
biohybrid fish generated a hydrodynamic
signature similar to those of wild-type BCF
swimmers—specifically, the water flow in the
wake of and around the fish bodies, which
we visualized with particle image velocimetry
(PIV) (Fig. 2, B to H, fig. S8). The biohybrid
fish shed two vortex pairs per tail-beat cycle
and one pair per lateral tail excursion (movie
S3), one of the key characteristic flow patterns
of BCF swimmers (movies S4 to S6). Each
lateral tail excursion from bent to near-straight
positions induced strong wake flows that
formed a visible vortex pair with the oppo-
site rotational direction (Fig. 2, B and C, vorti-
ces 1 and 2′; Fig. 2, D and E, vortices 2 and 3′;
and Fig. 2, F and G, vortices 3 and 4). When
the vortex pair reached the tail from the pos-
terior body, it was shed (Fig. 2D, vortices 1 and
2′; Fig. 2F, vortices 2 and 3′) and continuously
moved away from the fish body under its own
momentum (Fig. 2, D to G, vortices 1 and 2′;
and Fig. 2, F and G, vortices 2 and 3′).
The inclusion of a muscular bilayer archi-
tecture improved the high-frequency swim-
ming of the biohybrid fish. Our optogenetically
controlled biohybrid fish (6.4-mm-long muscle
tissue body) responded up to 3 to 4 Hz (fig. S9
and movie S7), maintained a large tail-beat
amplitude and angle, and exhibited a positive
pacing frequency and tail beat angular-speed
relationship (fig. S9). The biohybrid fish made
of human stem cell–derived CMs (movie S7)
and primary neonatal rat ventricular CMs
(fig. S10 and movie S8) exhibited increased
swimming speeds with increasing pacing fre-
quencies (Fig. 2L) reminiscent of the force-
frequency relationship of the heart. By contrast,
a previous biohybrid stingray (3) exhibited
reduced swimming speeds at high pacing fre-
quencies as it had single-layered muscle and
lacked antagonistic muscle contractions. The
upper limit of optogenetic pacing frequency
that induces a 1:1 stimulus response is also
affected by body dimensions: a longer-bodied
fish (8.2 mm; fig. S6C) exhibited oscillatory
motion up to 2 Hz, but not at 2.5- and 3-Hz
stimulation (movie S9).
Autonomous BCF propulsion
We tested whether reconstructing antagonistic
muscle contractions with CMs could sustain
spontaneous rhythmic contractions by means
of mechanoelectrical signaling (Fig. 3A). Spon-
taneous activation and contraction on one side
of the 49-day-old biohybrid fish led to a sub-
sequent antagonistic contraction on the oppo-
site side through mechanical coupling between
muscle tissues (Fig. 3B and movies S10 and
S11). These spontaneous antagonistic con-
tractions led to alternating bending motions
of the posterior body (Fig. 3, C to E), resulting
in rhythmically sustained forward displace-
ment (Fig. 3D) as shown in optogenetically
triggered body–caudal fin propulsions (Fig. 2).
Notably, biohybrid fish with a larger tail-beat
angle had a higher probability to induce a sub-
sequent muscle contraction (Fig. 3F), suggest-
ing that the lengthening of one muscle layer
caused by a shortening of the other muscle
layer directly induced subsequent contractions
through cardiac mechanoelectrical signaling.
We treated the biohybrid fish with stretch-
activated ion channel inhibitors [streptomycin
(21) and gadolinium (Gd3+) (22) (movies S12
and S13). We observed that these inhibitors
disrupted antagonistic contractions in the
biohybrid fish by breaking the positive rela-
tionship between peak tail-beat angle and
probability of antagonistic contractions (Fig.
3F and fig. S11). Further, frequency-domain
analysis showed that the spontaneous fre-
quencies of streptomycin- and Gd3+-treated
muscular bilayer tissues were not harmonic
(fig. S12). Stretched-activated ion channel
inhibition decreased swimming speeds (Fig.
3G), which demonstrated that mechano-
electrical signaling mediates self-sustainable
spontaneous rhythmic contractions in mus-
cular bilayers.
We tested whether reconstructing a geo-
metrically distinct and electrically insulated
node could initiate spontaneous electrical ac-
tivity as a result of the automaticity of CMs in
the absence of an external stimulus. Inspired
by the partial electrical insulation of a sino-
atrial node (15), we created the G-node (Fig. 1E),
where a small number of CMs are structurally
and functionally isolated with a single exit
pathway. The G-node is electrically coupled
by gap junctions (12, 23) to muscle tissues and
facilitates progressive activation of large qui-
escent neighboring muscle cells (sink) by a
small number of activating CMs (source). The
geometrical design of both the G-node and the
sink is crucial in determining the leading mus-
cle activation site, because the electrical current
being exchanged between individual CMs of
different membrane potentials can be reflected
at the tissue edges (12, 16). Thus, we hypothe-
sized that the reflection of intracellular currents
at the perimeter of the G-node would synchro-
nize the spontaneous activity and initiate coor-
dinated pacemaking from the G-node.
To decouple the effect of antagonistic mus-
cle contractions from muscle activation at
the G-node, we mechanically restricted muscle
movement with laboratory tape on a glass slide
and determined muscle activation through
calcium imaging (24). CMs in the G-node and
four corners [anterior ventral corner (AV),
anterior dorsal corner (AD), posterior ventral
corner (PV), and posterior dorsal corner (PD)
(figs. S13 and S14)] of the muscle tissue over-
came source-sink mismatch and initiated
Lee et al., Science 375, 639–647 (2022)
11 February 2022
3 of 8
RES EARCH | R E S E A R C H A R T I C L E
A
B
104 ms
C
200 ms
D
304 ms
E
400 ms
F
504 ms
G
600 ms
15
H
red light
left
I
l
e
g
n
a
t
a
e
b
-
l
i
a
t
)
e
e
r
g
e
d
(
d
e
e
p
s
)
s
/
m
m
(
e
c
n
a
t
s
d
i
)
m
m
(
90
0
0
-90
40
20
0
-20
2
0
-2
2222
2’2’2’2’2’2
11111111
3333333
3’3’3’333
222
3333333
22
2222’2’222
111111
2’22
1111
111111
22222222
11111
blue light
right
B
C
200
l
e
s
u
p
t
h
g
i
l
d
e
r
l
e
s
u
p
t
h
g
i
l
e
u
b
l
200
400
y
d
o
b
g
n
o
a
l
n
o
i
t
i
s
o
p
0.5
J
)
a
P
k
(
-1
0 (head)
curvature (1/mm)
K
1
E
D
400
20
0
s
s
e
r
t
s
-20
15
10
5
0
)
m
m
(
e
c
n
a
t
s
d
i
200
400
time (ms)
1 (tail)
0
1
2
time (s)
3
-5
0
1
2
time (s)
3
head
tail
)
s
/
1
(
y
t
i
c
i
t
r
o
v
44444
333333
3’3’33333
22222
2’222
5 mm
11111
-15
5 mm
L
)
s
/
m
m
(
d
e
e
p
s
12
8
4
0
biohybrid fish
(muscular bilayer,
human)
biohybrid fish
(muscular bilayer,
rat)
stingray (single layer, rat)
1
3
pacing freq (Hz)
2
Fig. 2. Optogenetically induced BCF propulsion. (A) Upon alternating blue and
red light stimulation, the biohybrid fish induces contraction of the ChR2- and
ChrimsonR-expressing muscles, respectively. (B to G) Body kinematics and
hydrodynamics of the biohybrid fish during one and a half tail-beat cycles. (B and
F) Peak contraction of left muscles. (C and G) recovery to straight position.
(D) peak contraction of right muscle. (E) recovery to straight position. Left and
right muscles work antagonistically against each other, leading to rhythmically
sustained body and caudal fin (BCF) propulsion. PIV flow measurements highlight
the shedding of the positive and negative vortex pair at every lateral tail excursion.
(H) Corresponding midline kinematics (time step: 50 ms). (I to K) Kinematic
analysis of seven strokes; correlation between optogenetic muscle activation and
BCF locomotion (n = 7 strokes; data represent mean ± SEM). (J) aCurvature
of the midline; (K) moving distance. (L) Positive relationship between pacing
frequency and moving speed of optogenetically stimulated biohybrid fish
[n = 31 videos from seven stingrays (3); 27 videos from six rat fish; and 54 videos
from 12 human fish.] Data represent mean ± SEM).
muscle activation (Fig. 3, H and I, figs. S13
and S14, and movies S14 and S15). As we hy-
pothesized, the G-node predominantly acti-
vated the muscle construct over the other four
corners of the muscle tissue (Fig. 3, I and J,
and figs. S13 and S14). Comparing two G-node
sizes showed that a larger G-node containing
~1700 cells increased the probability of ini-
tial muscular activation at the G-node com-
pared with the smaller G-node with a pointed
node (~600 cells) (fig. S13 and movie S15),
suggesting that a group of geometrically dis-
tinct CMs are needed to initiate muscular
activation. Additionally, rounding the sink’s
corners decreased the probability of activa-
tion at the corners (fig. S14A) by increas-
ing the number of downstream cells at each
respective corner (fig. S15), but the G-node’s
corner design did not affect the probability
of activation at the G-node (fig. S14B), which
indicates that acute angles in small source
tissue such as the G-node are not critical in
determining the activation site. Rather, this
suggests that a larger perimeter-to-area ratio
of the G-node synchronized electrical inter-
action between the geometrically distinct CMs
through reflections of electrotonic currents
and produced a relatively fast and synchron-
ized activation over the sink tissue. Acute
angled anterior corners of the fish body in-
creased the probability of activation at the
anterior side (Fig. 3J) by decreasing the num-
ber of downstream cells (fig. S15), thus allow-
ing cells on the anterior side (G-node and
anterior corners AD and AV) to predomi-
nantly initiate spontaneous activation waves
(60% from the G-node and 97% from all an-
terior sides; Fig. 3J).
However, upon removing the restrictions
on muscle movement, the G-node primarily
acted as a secondary mechanism for control-
ling contractions. Only when the antagonis-
tic muscle contractions were minimal would
the G-node initiate sequential local muscle
activation and contraction, leading to undu-
latory locomotion (fig. S16A and movie S17).
However, in subsequent muscle contractions
the biohybrid fish predominantly exhibited
simultaneous global contractions and oscil-
latory locomotion with minimal body wave
propagations, caused by mechanoelectrical
signaling of the muscular bilayer (Fig. 3E, fig.
S16, B and C, and movie S16). Although G-nodes
are located on both sides of the body, one
dominant G-node controlled initiation of mus-
cle contraction as a secondary pacemaker
(movie S17). Because of the G-node’s role as a
secondary pacing mechanism of antagonistic
contractions, the biohybrid fish equipped with
a G-node had significantly increased sponta-
neous contractile frequencies (Fig. 3K) while
maintaining similar body kinematics (fig.
S16) and a positive frequency–swimming speed
relationship similar to those of externally
stimulated fish (Fig. 3L). As a result, our G-node-
equipped biohybrid fish demonstrated in-
creased maximum swimming speeds of more
than one body length per second (15 mm/s;
movie S11).
Lee et al., Science 375, 639–647 (2022)
11 February 2022
4 of 8
RES EARCH | R E S E A R C H A R T I C L E
A
spontaneous
activities
mechano-
electrical
signaling
B
0 s 0.12
0.2
0.28
0.4
F
100
-streptomycin, -Gd3+
)
%
(
+streptomycin
+Gd 3+
n
o
i
t
c
a
r
t
n
o
c
y
t
i
l
i
b
a
b
o
r
p
l
a
c
i
r
i
p
m
e
i
c
i
t
s
n
o
g
a
t
n
a
f
o
50
0
<65
80–95
110–125
65–80
95–110
peak tail-beat angle (degree)
125–140
C
head
D
l
e
g
n
a
t
a
e
b
-
l
i
a
t
)
e
e
r
g
e
d
(
-90
90
0
time (s)
1
0
0.5
30
)
a
P
k
(
s
s
e
r
t
s
-30
tail
5 mm
)
m
m
(
e
c
n
a
t
s
d
i
8
4
0
0
1
0.5
time (s)
I
E
l
y
d
o
b
g
n
o
a
n
o
i
t
i
s
o
p
5 mm
-1
0 (head)
1
curvature (1/mm)
0.5
1 (tail)
0
0.5
time (s)
1
H
posterior
dorsal
corner
(PD)
anterior dorsal
corner (AD)
posterior
ventral
corner
(PV)
anterior ventral
corner (AV)
0
G-node
activation delay (ms)
051
p = 3.4×10-10
1.5×10-5
2.2×10-4
K
)
z
H
(
y
c
n
e
u
q
e
r
f
t
a
e
b
-
l
i
a
t
s
u
o
e
n
a
t
n
o
p
s
5
4
3
2
1
0
L
15
10
)
s
/
m
m
(
d
e
e
p
s
5
0
single
layer
muscular
bilayer
muscular
bilayer
w/ G-node
J
100
)
%
(
e
c
n
e
r
r
u
c
c
o
80
60
40
20
0
D
P
V
A
D
A
o
/
w
e
d
o
n
-
G
V
P
D
P
V
A
D
A
e
d
o
n
-
G
/
w
e
d
o
n
-
G
p = 2.84×10-16
4.32×10-16
G
10
)
s
/
m
m
(
d
e
e
p
s
8
6
4
2
0
streptomycin
Gd3+
–
–
+
–
–
+
maximum speed
>140
spontaneous
paced
1
2
3
pacing freq (Hz)
4
Fig. 3. Autonomous BCF propulsion. (A to G) Mechano-electrical signaling of
the muscular bilayer. (A) Spontaneous activation of one-side muscle induces
consecutive contraction of the opposite-side muscle through mechano-electrical
signaling between muscular bilayer tissues. (B) Representative time lapse images
of consecutive antagonistic muscle contraction of 49-day-old biohybrid fish.
(C) Midline kinematics (time step: 100 ms). (D) Correlation between spontaneous
muscle activation and the moving distance. (E) Curvature of the midline
during five consecutive left and right muscle strokes. (F) Empirical probability of
antagonistic contraction and (G) moving speed of self-paced biohybrid fish treated
with stretch-activated channel blockers, 250 mM streptomycin (n = 4 biohybrid
fish) and 100 mM Gd3+ (n = 5 biohybrid fish) (box plot: center line, box limits, and
whiskers indicates mean, SEM, and the first and third quartiles, respectively).
The treatment of stretch-activated channel blockers, streptomycin and Gd3+
reduced the chance of antagonistic muscle contraction as well as the swimming
speed of the biohybrid fish. (H to L) Geometrically insulated node (G-node).
Activation pattern of biohybrid fish (H) without G-node and (I) with G-node. (J)
Probability of muscle activation sites. Spontaneous muscle activation from
G-node dominates spontaneous activation from the corners (n = 6 biologically
independent samples without G-node and 5 samples with G-node). (K) Tail-beat
frequency of biohybrid fish equipped with single-layer (n = 9 videos from nine
fish), muscular bilayer (n = 20 videos from 14 fish), and muscular bilayer with
G-node (n = 18 videos from five fish). Significance was determined by the
Tukey-Kramer honestly significant difference test. (L) Positive relationship between
pacing frequency and moving speed of autonomously swimming biohybrid fish
(n = 30 videos from 19 autonomously swimming biohybrid fish and 54 videos
from optogenetically swimming biohybrid fish). Data represent mean ± SEM.
Although these G-node–entrained, mechano-
electrical signaling–sustained, cyclic antagonistic
muscle contractions are autonomous, optoge-
netic stimulation can be used for on-demand
locomotion control. Antagonistic muscle con-
tractions became coupled with optical pacing
within fewer than three sequential light pulses
(movie S18). Further, optogenetic stimulation
can also be used to inhibit autonomous loco-
motion; pausing immediately after a pulsed
stimulation can stop muscle contractions for
an extended period (50 s; movie S19). Pro-
longed continuous optogenetic stimulation
stops muscle contractions and autonomous
locomotion (movie S20). External stimulation
reinitiates autonomous, antagonistic muscle
contractions by activating mechanoelectrical
signaling (movie S21).
Advanced performance of the biohybrid fish
Our autonomously swimming biohybrid fish
(15.0 mm/s; movie S11) outperformed the
locomotory speed of previous biohybrid mus-
cular systems (2, 3, 25–35) [5 to 27 times the
speed of the biohybrid stingray (3) and the bio-
hybrid skeletal muscle biorobot] (32) (Fig. 4A),
highlighting the role of feedback mechanisms
in developing biohybrid systems. Moreover,
when considering the ratio of muscle mass
to the total weight for the biohybrid fish
(1.4%; fig. S7) and biohybrid stingray (9.7%)
(3), the biohybrid fish demonstrated faster
swimming speeds per unit muscle mass by
an order of magnitude (13 times the maximum
swimming speed of the biohybrid stingray) (3)
(Fig. 4B).
The swimming performance of the biohybrid
fish resembles that of wild-type BCF swim-
mers with similar body lengths (juvenile zebra-
fish, juvenile white molly, and Microdevario
kubotai) (Fig. 4, C to F, fig. S8, and movies S4
to S6). Similar to the biohybrid fish, each of
these species moves by shedding a pair of
reverse-sign vortices when their tails reach
maximum lateral excursion (Fig. 4, C to F).
Lee et al., Science 375, 639–647 (2022)
11 February 2022
5 of 8
RES EARCH | R E S E A R C H A R T I C L E
The strength of these vortices between the
biohybrid and wild-type fish were comparable
(Fig. 4, C to F). Rather than forming a contin-
uous chaotic chain of wakes, both the biohybrid
and wild-type fish maintain stable vortex pairs
with minimal vortex interactions (Fig. 4, C to
F). The stable wake pattern is a typical char-
acteristic of juvenile zebrafish locomotion at
relatively high Strouhal numbers (St) and
relatively low Reynolds numbers (Re < 5000)
(36), where viscous forces cannot be neglected
and the lateral velocity of wake flows are rela-
tively high. In this flow regime, the swimming
speed is nearly proportional to the tail-beat
frequency (37). Thus, the juvenile zebrafish,
white molly, and M. kubotai had faster tail-
beat frequencies (16.7, 7.5, and 7.7 Hz, which
were 4.6, 2.1, and 2.1 times as high as that of
the biohybrid fish) and showed proportionally
increased swimming speeds of 59.7, 25.1, and
21.3 mm/s, respectively (4.0, 1.7, and 1.5 times
as high as that of the biohybrid fish). Although
muscle function in wild-type fish encompasses
more than locomotion, when considering
the ratio of total muscle mass to the total
weight of biohybrid fish (1.4%; fig. S7) com-
pared with wild-type fish (80%) (38), the maxi-
mum swimming speed per unit muscle mass
of biohybrid fish exceeded those of wild-type
fish by a factor of 70 to150 (Fig. 4B).
Efficiency of the biohybrid fish
To analyze the efficiency of the biohybrid fish,
we used scaling and dimensional analysis.
Wild-type swimmers achieved energetically
favorable locomotion through convergent
evolution and were found to hew to the two
scaling relationships St ~ Re−1/4 and Re ~ Sw−1/4
in the low Re and high St flow regime (37)
(Fig. 4, G and H). The Strouhal number St =
fA/U (f, tail-beat frequency; A, tail-beat am-
plitude; U, forward speed) represents the ratio
of the lateral oscillation amplitude to swim-
ming distance per lateral tail excursion, the
swimming number Sw = 2pfAL/u (L, char-
acteristic body length of the swimmer; u, fluid
viscosity) represents input kinematics, and
the Reynolds number Re = UL/u compares
inertial to viscous forces and is a function
of swimming speed. Compared with the bio-
hybrid stingray (3), our biohybrid fish oper-
ates much closer to these average scaling
relationships of wild-type swimmers. More-
over, the biohybrid fish swimming at high
tail-beat frequencies (high St and Sw) per-
formed comparably to wild-type swimmers
(Fig. 4, G and H).
The performance of the biohybrid fish is very
sensitive to muscle kinematics and coordi-
nation. Some biohybrid fish accelerated by
increasing tail-beat amplitude (figs. S17 and
S18A and movie S22), which is similar to the
acceleration of wild-type fish (39). This posi-
tive relationship between swimming speed
juvenile zebrafish
juvenile
juvenile
juvenile
juvenile
juvenile
Fig. 4. Comparison of swimming performance between biohybrid and aquatic swimmers. (A and B) Compar-
ison of swimming performance in biohybrid walkers and swimmers. (A) locomotion speed (2, 3, 25–35)
and (B) speed per unit muscle mass. PIV analysis of (C) biohybrid fish [body length (lb): 14 mm];
(D) wild-type juvenile zebrafish (lb: 12 mm); (E) white molly (lb: 19 mm); and (F) and M. kubotai (lb: 20 mm).
(G and H) Scaling analysis of biohybrid fish and wild-type swimmers with (G) Re-St and (H) Sw-Re
(n = 30 movies from 19 biohybrid fish).
and tail-beat amplitude during accelerative
locomotion contrasts with the constant tail-
beat amplitude regardless of swimming speed
during steady locomotion (fig. S9B). Although
St, Sw, and Re numbers increase with its swim-
ming speed (fig. S18, B and C) in the accel-
erative locomotion, the biohybrid fish exhibited
a considerable decrease in propulsive effici-
ency as its speed increases as shown by the
deviation from the optimal St-Re and Re-Sw
relationships of aquatic swimmers (fig. S18,
B and C). The inhibition of muscle coordina-
tion with a stretch-activated ion channel blocker,
Gd3+, also led to a drastic reduction of 80.8%
in Re and 40.6% in Sw and the deviation from
optimal St-Re and Re-Sw relationships of
aquatic swimmers (figs. S12 and S18, D to F,
and movie S12), which demonstrate that mus-
cular coordination is necessary to achieve
effective and efficient swimming.
Long-term performance of the biohybrid fish
Given the autonomous antagonistic muscle
contractions of the biohybrid fish, we ques-
tioned whether this spontaneous activity
would improve its long-term performance.
The biohybrid fish maintained spontaneous
activity for 108 days [16 to 18 times the length
of the biohybrid stingray (6 days) (3) and the
skeletal muscle-based biohybrid actuator
(7 days) (40)], equivalent to 38 million beats
(Fig. 5, A and B, and movies S23 and S24). Fur-
ther, its locomotion could also be controlled
by external optogenetic stimulation (movie
Lee et al., Science 375, 639–647 (2022)
11 February 2022
6 of 8
RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. Long-term swimming performance analysis. (A) Trajectory (grids,
1 cm) and (B) corresponding tail-beat angle of 108-day-old biohybrid fish with
79% antagonistic contractions. (C) Swimming performance for 108 days (n = 4
fish). Biohybrid fish equipped with the muscular bilayer exhibited enhanced
contracting amplitude, maximum swimming speed, and muscle coordination for
the first month and maintained their performance for at least 108 days, whereas
fish made with the single-layer muscle exhibited decreased contracting
amplitude after 28 days. (n = 4 fish; data represent mean ± SEM).
S24) throughout this time. The autonomously
swimming biohybrid fish also increased mus-
cle contraction amplitude, maximum swim-
ming speed, and muscle coordination for the
first month before maintaining its swim-
ming performance over 108 days (Fig. 5C).
By contrast, biohybrid fish equipped with
single-layered muscle showed deteriorating
tail-beat amplitude within the first month (Fig.
5C and movie S25). These data demonstrate
the potential of muscular bilayer systems and
mechanoelectrical signaling as a means to pro-
mote maturation of in vitro muscle tissues.
Discussion
We integrated two functional design features
of the heart—mechanoelectrical signaling and
automaticity—into a biohybrid platform and
recreated an autonomously actuating car-
diac muscular system in the form of a bio-
hybrid fish. This fish is a closed-loop system in
which muscle contraction–induced bending
is used as a feedback input to the endoge-
nous mechanosensors—stretch-activated ion
channels—in the muscles. These channels
respond to this feedback input and induce
muscle activation and contraction, producing
self-sustainable rhythmic BCF propulsion. The
self-driven spontaneous contractions in our
muscular bilayer induced coordinated global
tissue-level contractions with comparable effi-
ciencies to wild-type fish. Alternatively, inte-
grated optogenetic control enabled overriding
of internal control mechanisms to stop and con-
trol asynchronous muscle contractions. There
are few, if any, closed-loop mechanical fish
robots that are free-swimming; fish robots also
typically require numerous actuators and sen-
sors to control fin movements, and these are
difficult to engineer at smaller size scales (mil-
limeters to centimeters scale) (41). However,
integration of the cardiac activation system as
an embedded mechanism of both sensing and
control enabled the generation of fishlike loco-
motion at such smaller scales (42). The use of
biological muscle actuators with intrinsic closed-
loop control simplifies the construction com-
pared with current mechanical robotic systems
and provides control beyond existing biohy-
brid systems.
Additionally, our muscular bilayer con-
struct provides a platform for studying tissue-
level cardiac biophysics. We demonstrate that
dynamic axial stretching can induce excita-
tions and contractions on a beat-by-beat basis
in engineered human stem cell–derived CM
tissues by contributing to antagonistic mus-
cle contractions. We found that antagonistic
contractions are sensitive to streptomycin
and Gd3+, which indicates that mechano-
electrical signaling by means of stretch-activated
ion channels is one of the essential mecha-
nisms that mediate antagonistic contractions.
Notably, in normal myocardium where CMs
are mechanically and electrically coupled,
mechanoelectrical signaling contributes to
synchronizing local ventricular repolarization
and protects against cell-to-cell repolariza-
tions and contractile heterogeneities across
the heart (43). By contrast, in our muscular
bilayer where antagonistic muscle pairs are
mechanically coupled yet electrically decou-
pled across sides, mechanoelectrical signaling
generates stretch-induced depolarizations
on a beat-by-beat basis. The stretch-induced
excitations and contractions were also ob-
served in quiescent single CMs and in a rest-
ing ventricular myocardium (10), but these
observations were restricted to the ectopic
responses of CMs to acute mechanical stimu-
lation, which induced re-entrant arrhythmias.
Our muscular bilayer construct is the first to
demonstrate that the mechanoelectrical sig-
naling of CMs could induce self-sustaining
muscle excitations and contractions for ex-
tended periods (108 days, equivalent to 38 mil-
lion beats). These findings are aligned with
the growing appreciation for cardiac stretch-
activated channels and mechanoelectrical sig-
naling mechanisms as targets of heart rhythm
management (10, 44). The longevity of the
autonomously moving fish system also raises
the question of whether a feedback between
repetitive electrical and mechanical activity
and the regulation of its molecular elements
through altered gene expression or other basic
cellular processes is correlated.
The G-node, an isolated cluster of cells con-
nected through a single conducting exit path-
way, initiated spontaneous activation waves
by reducing the impedance between source
and sink. G-node integration improved locomo-
tion speeds by enhancing the pacing frequency.
This increased frequency in the presence of
the G-node is reminiscent of entrainment in
Lee et al., Science 375, 639–647 (2022)
11 February 2022
7 of 8
RES EARCH | R E S E A R C H A R T I C L E
re-entry cycles in which the focus shortens the
re-entry cycle (45). Another possible under-
lying mechanism of the increased frequency
is that the G-node produced regular contrac-
tions and consequently induced stronger
and more rapid contractions of the muscular
bilayer, which could enhance the dynamics
of antagonistic, asynchronous muscle contrac-
tions. The G-node functionality as a node of
automaticity in the biohybrid fish suggests
that, functionally, a pacemaker may be de-
fined by its geometry and source-sink relation-
ships as well as its ion channel expression.
Taken together, the technology described
here may represent foundational work toward
the goal of creating autonomous systems ca-
pable of homeostatic regulation and adaptive
behavioral control. Our results suggest an op-
portunity to revisit long-standing assumptions
of how the heart works in biomimetic systems,
which may allow a more granular analysis of
structure-function relationships in cardiovas-
cular physiology.
RE FE RENCES AND N OT ES
1. S. Vogel, Vital Circuits: On Pumps, Pipes, and the Workings of
Circulatory Systems (Oxford Univ. Press, 1992).
J. C. Nawroth et al., Nat. Biotechnol. 30, 792–797 (2012).
2.
3. S. J. Park et al., Science 353, 158–162 (2016).
4. L. Ricotti et al., Sci. Robot. 2, eaaq0495 (2017).
5. H. E. ter Keurs, Pflugers Arch. 462, 165–175 (2011).
6. T. A. Quinn, P. Kohl, Physiol. Rev. 101, 37–92 (2021).
7. T. A. Quinn, P. Kohl, U. Ravens, Prog. Biophys. Mol. Biol. 115,
71–75 (2014).
8. B. E. Dabiri, H. Lee, K. K. Parker, Prog. Biophys. Mol. Biol. 110,
196–203 (2012).
9. O. Friedrich, S. Wagner, A. R. Battle, S. Schürmann, B. Martinac,
Prog. Biophys. Mol. Biol. 110, 226–238 (2012).
10. T. A. Quinn, H. Jin, P. Lee, P. Kohl, Circ. Arrhythm.
Electrophysiol. 10, e004777 (2017).
11. S. D. Unudurthi, R. M. Wolf, T. J. Hund, Front. Physiol. 5, 446 (2014).
12. A. G. Kléber, Y. Rudy, Physiol. Rev. 84, 431–488 (2004).
13. S. Rohr, J. P. Kucera, V. G. Fast, A. G. Kléber, Science 275,
841–844 (1997).
14. B. Bullard, A. Pastore, J. Muscle Res. Cell Motil. 32, 303–313 (2011).
15. R. W. Joyner, F. J. van Capelle, Biophys. J. 50, 1157–1164 (1986).
16. Materials and methods are available as supplementary materials.
17. S. J. Park et al., Circulation 140, 390–404 (2019).
18. M. L. McCain, A. Agarwal, H. W. Nesmith, A. P. Nesmith,
K. K. Parker, Biomaterials 35, 5462–5471 (2014).
19. E. S. Boyden, F. Zhang, E. Bamberg, G. Nagel, K. Deisseroth,
Nat. Neurosci. 8, 1263–1268 (2005).
20. N. C. Klapoetke et al., Nat. Methods 11, 338–346 (2014).
21. F. Gannier, E. White, A. Lacampagne, D. Garnier,
J. Y. Le Guennec, Cardiovasc. Res. 28, 1193–1198 (1994).
22. X. C. Yang, F. Sachs, Science 243, 1068–1071 (1989).
23. A. G. Kléber, Q. Jin, Biophys. Rev. 2, 031301 (2021).
24. K. Y. Lee et al., Optical stimulation for independent activation
of muscles: Matlab and Labview, Zenodo (2021); doi:10.5281/
zenodo.5618323.
25. A. W. Feinberg et al., Science 317, 1366–1370 (2007).
26. J. Xi, J. J. Schmidt, C. D. Montemagno, Nat. Mater. 4, 180–184
(2005).
27. C. Cvetkovic et al., Proc. Natl. Acad. Sci. U.S.A. 111, 10125–10130
(2014).
28. V. Chan et al., Sci. Rep. 2, 857 (2012).
29. J. Kim et al., Lab Chip 7, 1504–1508 (2007).
30. Y. Akiyama, T. Hoshino, K. Iwabuchi, K. Morishima, PLOS ONE
7, e38274 (2012).
31. Y. Akiyama et al., Biomed. Microdevices 14, 979–986 (2012).
32. G. J. Pagan-Diaz et al., Adv. Funct. Mater. 28, 1801145 (2018).
33. R. Raman et al., Proc. Natl. Acad. Sci. U.S.A. 113, 3497–3502
(2016).
34. B. J. Williams, S. V. Anand, J. Rajagopalan, M. T. Saif,
Nat. Commun. 5, 3081 (2014).
35. O. Aydin et al., Proc. Natl. Acad. Sci. U.S.A. 116, 19841–19847 (2019).
36. G. K. Taylor, R. L. Nudds, A. L. Thomas, Nature 425, 707–711 (2003).
37. M. Gazzola, M. Argentina, L. Mahadevan, Nat. Phys. 10,
758–761 (2014).
38. H. E. Jackson, P. W. Ingham, Mech. Dev. 130, 447–457 (2013).
39. T. N. Wise, M. A. B. Schwalbe, E. D. Tytell, J. Exp. Biol. 221,
jeb.190892 (2018).
40. Y. Morimoto, H. Onoe, S. Takeuchi, Sci. Robot. 3, eaat4440 (2018).
41. R. Du, Z. Li, K. Youcef-Toumi, P. Valdivia y Alvarado, in Robot Fish:
Bio-inspired Fishlike Underwater Robots, R. Du, Z. Li, K. Youcef-Toumi,
P. Valdivia y Alvarado, Eds. (Springer, 2015), pp. 1–24.
42. G. V. Lauder, J. L. Tangorra, in Robot Fish: Bio-inspired
Fishlike Underwater Robots, R. Du, Z. Li, K. Youcef-Toumi,
P. Valdivia y Alvarado, Eds. (Springer, 2015), pp. 25–49.
43. T. Opthof, V. M. Meijborg, C. N. Belterman, R. Coronel,
Cardiovasc. Res. 108, 181–187 (2015).
44. T. A. Quinn, Cardiovasc. Res. 108, 1–3 (2015).
45. J. Almendral, R. Caulier-Cisterna, J. L. Rojo-Álvarez,
Pacing Clin. Electrophysiol. 36, 508–532 (2013).
ACKN OWLED GMEN TS
We thank M. Rosnach for photography and illustrations.
Funding: this work was funded by the Harvard Paulson School
of Engineering and Applied Sciences, the Wyss Institute for
Biologically Inspired Engineering, National Institutes of Health
National Center for Advancing Translational Sciences grant
UH3TR000522, and National Science Foundation Materials
Research Science and Engineering Center grant DMR-1420570.
K.K.P. was sponsored by National Institutes of Health National Center
for Advancing Translational Sciences grant 1-UG3-HL-141798-01.
S.-J.P. was funded by the Georgia Institute of Technology and
Emory University School of Medicine. G.V.L. was funded by
the Office of Naval Research (T. McKenna, Program Manager,
ONR 341), grant N00014-15-1-2234, and the National Science
Foundation, grant 1830881. H.A.M.A. would like to thank the
American Chemical Society for their generous support through the
Irving S. Sigal Postdoctoral Fellowship. The views and conclusions
contained in this document are those of the authors and should
not be interpreted as representing the official policies, either expressed
or implied, of the National Institutes of Health, the National Science
Foundation, or the US Government. This work was performed in part at
the Harvard Center for Nanoscale Systems, a member of the National
Nanotechnology Infrastructure Network, which is supported by the
National Science Foundation under award ECS-0335765. The Center
for Nanoscale Systems is part of Harvard University. Author
contributions: K.Y.L. and S.-J.P. conceived and designed the study,
developed a geometrically insulated cardiac tissue node-integrated
muscular bilayer construct, designed and performed performance
experiments, analyzed data, organized figures, and wrote the paper.
K.K.P. conceived and designed the study, developed the idea of a
geometrically insulated cardiac tissue node and muscular bilayer, and
supervised the project. A.G.K. and G.V.L. contributed to the concept of
a geometrically insulated cardiac tissue node and muscular bilayer,
respectively. D.M. and G.V.L. contributed to the PIV experiments of both
biohybrid and wild-type fish. S.L.K. designed optogenetic tools and
edited the manuscript. C.M. assisted the fabrication of the biohybrid
fish. S.L.K., J.Z., and H.A.M.A. performed primary neonatal rat ventricular
harvest for the biohybrid fish optimization. All authors contributed to
the preparation of the manuscript; Competing interests: K.Y.L., S.-J.P.,
A.G.K., V.T., and K.K.P. are inventors on a patent filed by Harvard
University, U.S. Provisional Patent Application No. 63/299,920, based
on the results described in this manuscript. The remaining authors
declare no competing interests. Data and materials availability: All
data are available in the main text or the supplementary materials.
Code and scripts are available at Zenodo (24)
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abh0474
Materials and Methods
Figs. S1 to S19
References (46–48)
MDAR Reproducibility Checklist
Movies S1 to S25
View/request a protocol for this paper from Bio-protocol.
16 February 2021; accepted 14 January 2022
10.1126/science.abh0474
Lee et al., Science 375, 639–647 (2022)
11 February 2022
8 of 8
|
10.1126_science.adc9998
|
Corrected 16 March 2023. See full text.
RES EARCH
R E S E A R C H A R T I C L E
◥
BIOELECTRONICS
Metabolite-induced in vivo fabrication of
substrate-free organic bioelectronics
Xenofon Strakosas1,2*†, Hanne Biesmans1†, Tobias Abrahamsson1, Karin Hellman2,
Malin Silverå Ejneby1, Mary J. Donahue1, Peter Ekström2, Fredrik Ek2, Marios Savvakis1, Martin Hjort2,
David Bliman3,4, Mathieu Linares1,5, Caroline Lindholm1, Eleni Stavrinidou1, Jennifer Y. Gerasimov1,
Daniel T. Simon1, Roger Olsson2,3, Magnus Berggren1*
Interfacing electronics with neural tissue is crucial for understanding complex biological functions,
but conventional bioelectronics consist of rigid electrodes fundamentally incompatible with living
systems. The difference between static solid-state electronics and dynamic biological matter makes
seamless integration of the two challenging. To address this incompatibility, we developed a method to
dynamically create soft substrate-free conducting materials within the biological environment. We
demonstrate in vivo electrode formation in zebrafish and leech models, using endogenous metabolites to
trigger enzymatic polymerization of organic precursors within an injectable gel, thereby forming
conducting polymer gels with long-range conductivity. This approach can be used to target specific
biological substructures and is suitable for nerve stimulation, paving the way for fully integrated,
in vivo–fabricated electronics within the nervous system.
I nterfacing neural tissues with advanced
digital and electronic instrumentation is
one approach to probing the complex sig-
naling characteristics of the nervous system.
Even though several functional bioelec-
tronic implants have been demonstrated, the
rapid formation of scar tissue around the im-
plantation site affects the lifetime, precision,
and overall fidelity of the biological–electronic
interface (1). Most of these technologies are
based on thin films, requiring a planar [or in
some cases three-dimensional (3D)–structured]
substrate of varying stiffness. The material in-
terfaces that such technologies necessitate—
e.g., substrate-active material, active material-
encapsulation, etc.—introduce complexity and
lead to the common failure mode of delamina-
tion of one or more layers. Successful long-
term integration of neural stimulation and
recording technologies thus relies on match-
ing the electrical and viscoelastic characteristics
of the device to those of the nervous system
in a seamless manner.
When designing a system complementary
to the characteristics of the nervous system,
bioelectronic electrode technology should ex-
1Laboratory of Organic Electronics, Department of Science
and Technology, Linköping University, 601 74 Norrköping,
Sweden. 2Chemical Biology and Therapeutics, Department of
Experimental Medical Science, Lund University, SE-221 84
Lund, Sweden. 3Department of Chemistry and Molecular
Biology, University of Gothenburg, SE-405 30 Gothenburg,
Sweden. 4IRLAB Therapeutics AB, Arvid Wallgrens Backe 20,
413 46 Gothenburg, Sweden. 5Scientific Visualization Group,
Department of Science and Technology, Linköping University,
601 74 Norrköping, Sweden.
*Corresponding author. Email: magnus.berggren@liu.se (M.B.);
xenofon.strakosas@liu.se (X.S.)
†These authors contributed equally to this work.
tend far beyond the 2D arrangements of lines
and pads organized along hard, planar sub-
strates that form the basis of classic electronic
design. Various strategies have been under-
taken to design electrodes that seamlessly con-
nect to the structures of the nervous system.
With the rise of flexible and soft electronics
based on intrinsically electroactive conducting
polymers capable of mixed ion–electron con-
duction, bioelectronic electrodes with matched
impedance characteristics have been imple-
mented on ultraflexible and ultrasoft substrates
(2). Electrodes based on conductive polymers
(i.e., organic electronics) (3, 4) conformally in-
tegrate with the surface of the brain in animal
models and have been demonstrated for elec-
trocorticography (5, 6). Even though soft and
flexible electrodes allow for conformal contact
with the brain and nerves, the substrate repre-
sents a major limitation for reaching struc-
tures deep within the sensitive neural tissue in
a minimally invasive way (7). To circumvent
this limitation, conducting polymers have been
formed within the target biological system,
introduced as monomers from a solution and
polymerized in vivo to derive “impregnated”
(8) electrodes and even bicontinuous amal-
gamations with neural tissue (9). However,
polymerization requires chemical (10) and/or
electrical (10) energy, which can potentially
be harmful to sensitive neural tissue. To fur-
ther advance the concept of in vivo polymer-
ization, genetic engineering of an animal was
carried out to express enzymes that promote
local polymerization in specific cells in close
proximity to the target tissue in vivo in the
nematode Caenorhabditis elegans (11). Owing
to its reliance on genetic engineering, however,
this technique cannot be ethically extended to
humans to provide an advanced interface with
the brain.
Progress has been made on using endoge-
nous enzymatic activity to catalyze the oxida-
tive polymerization of precursors to produce
mixed ion–electron conduction polymers in vivo.
The biocompatible precursors are based on an
organic trithiophene-based monomer with
a 2,5-bis(2,3-dihydrothieno[3,4-b][1,4]dioxin-
5-yl)thiophene (ETE) backbone, which is
monofunctionalized on the central thiophene
with a 4-ethoxybutane-1-sulfonic acid salt side
chain (ETE-S). The water-soluble trithiophene
monomer has a low oxidation potential (~0.3 V
versus Ag/AgCl), allowing for enzymatic po-
lymerization in vivo (12). Endogenous per-
oxidase enzymes polymerize ETE-S in plants
(13) and in the small freshwater animal Hydra
vulgaris (14). The peroxidases utilize local
H2O2 as a catalyst to produce radical mono-
mers, which further polymerize and aggre-
gate to form conducting polymers integrated
with the tissue.
There is a remaining leap that must be taken
to produce a truly versatile mixed ion–electron
conduction electrode system that would fulfill
all the requirements for a self-organized in vivo
neural interface system that is high-performing,
minimally invasive, and stable. We propose
that such electrode technology would exhibit
the following characteristics: (a) dispensed as
a fluid (15) that is compatible with the envi-
ronment of the target neural tissue and can
diffuse within a desired distance of the injec-
tion site; (b) consists of, and generates, only
nontoxic components to drive polymerization
and cross-linking; forms an electrode that is
(c) homogeneous, (d) long-term stable, and (e)
gelled (soft) with (f) high conductivity and high
volumetric capacitance while (g) conformally
interfacing with neural structures at different
length scales.
We report an approach to producing self-
organized, high-performing electrode structures
within the peripheral nervous system (PNS) and
central nervous system (CNS), which is applica-
ble to a wide range of tissue and animal models.
Our approach relies on the injection of a multi-
component mixture that has been designed to
fulfill all the criteria, (a) to (g), listed above. The
mixture includes an ETE derivative with a
2-ethoxyacetic acid sodium salt side chain [ETE-
COONa: addressing (a), (b), (d), (f)], poly(vinyl
alcohol) (16, 17) [(PVA): (c), (e)], poly-L-lysine (18)
[(PLL: (d), (e), (g)], horseradish peroxidase [HRP:
(b)], oxidase enzymes [ROx: (b)] (12, 19), and 1-
ethyl-3-[3-dimethylaminopropyl]carbodiimide/
N-hydroxysulfosuccinimide [EDC/sulfo-NHS: (d)].
The polymerization of ETE-COONa is medi-
ated by the ROx–HRP enzyme cascade where-
by the ROx consumes endogenous metabolites
within the physiological target to produce H2O2
locally, which then serves as an electron acceptor
Strakosas et al., Science 379, 795–802 (2023)
24 February 2023
1 of 8
RES EARCH | R E S E A R C H A R T I C L E
Corrected 16 March 2023. See full text.
Fig. 1. In vivo polymeriza-
tion of conducting polymer
gels induced by endoge-
nous metabolites. (A) Sche-
matic representation showing
the gel components. Each
component contributes to the
process of polymerization or
cross-linking. (B) Polymeriza-
tion occurs dynamically in
the CNS after the introduc-
tion of the gel. (C) Schematic
of the in vivo polymerization
mechanism. (D) Cocktail gel
(yellow-brown color) injected
in agarose gel containing
5 mM glucose and (E) subse-
quently polymerized (turning
dark blue) after 10 min.
Scale bars, 1 mm. (F) UV-vis
spectra from cocktail gels
injected into agarose 1 min
after injection (black line) and
120 min after injection (red
line) (n = 1 representative
UV-vis measurement).
in the oxidation of ETE-COONa by HRP. We
show that by taking advantage of the enzymatic
polymerization and cross-linking around ner-
vous tissue, it is possible to fabricate soft
electrodes in vivo in a broad variety of model
organisms.
Design, operating principle, and polymerization
of gel electrodes.
Unlike in plants (13) and hydra (14), where the
endogenous environment promotes spontane-
ous ETE-S polymerization, the native environ-
ment in zebrafish brains does not efficiently
polymerize ETE-S. We introduced neat ETE-S
solution (2 mg ml−1 ETE-S in water) in ex-
tracted brains of zebrafish (Danio rerio) and
allowed the polymerization to proceed over
3 hours. Fluorescence microscopy images from
zebrafish brain slices showed that the in-
jected ETE-S can be excited with blue light (ex-
citation: 485 nm; emission: 525 nm) (fig. S1A).
This fluorescence is attributed to aggregated
but nonpolymerized ETE-S (fig. S1B). Be-
cause ETE-S is enzymatically polymerized
by peroxidases in the presence of H2O2, in vivo
polymerization in nerve tissue is limited ei-
ther by the absence of peroxidases or H2O2, or
both.
To overcome this limitation, we devised a
strategy to polymerize ETE analogs in various
tissues and animal models. The concept al-
lows for local production of H2O2 by oxidase
enzymes and subsequent polymerization of
ETE-derivatives by peroxidase enzymes (12).
The enzymes are introduced in the tissue, in
which endogenous metabolites that are ubiq-
uitous in nervous tissue fuel the enzymatic
polymerization cascade. To improve the sta-
bility of the conducting polymer, we used a
cocktail gel containing ETE-COONa (rather
than ETE-S), oxidase enzymes [ROx: glucose
oxidase (GOx) or lactate oxidase (LOx)], and
HRP embedded in a PVA:PLL polymer matrix
(Fig. 1A). EDC/sulfo-NHS was added in the gel
to activate the carboxylic acid toward amide
bond formation. Specifically, EDC/sulfo-NHS
reacts with the carboxyl groups of ETE-COONa
to generate an intermediate ester (ETE-NHS;
Fig. 1C, upper right) that reacts with the pri-
mary amine groups of PLL, hydroxyl groups of
PVA, as well as other amine and alcohol groups
that may exist in tissue (20). The gels are then
introduced into the tissue to polymerize in situ
and form soft electronic conductors (Fig. 1B).
Metabolites from the tissue diffuse into the
injected gels and ROx enzymes catalyze their
oxidation, producing H2O2 as a byproduct in
the presence of O2. Then HRP locally converts
H2O2 to H2O and in the process, ETE mono-
mers are converted into radicals that undergo
further polymerization. The polymerized ETE-
NHS intermediates cross-link with the PLL to
form a stable pETE-PLL–conducting polymer
gel electrode (Fig. 1C). The resulting electrode
reaches a shape and configuration dictated
by the diffusion and chemical kinetics of the
Strakosas et al., Science 379, 795–802 (2023)
24 February 2023
2 of 8
RES EARCH | R E S E A R C H A R T I C L E
components introduced into the target biolog-
ical system, as well as the native microenviron-
ment within that system.
The polymerization of ETE can be visualized
optically or by fluorescence imaging. Cocktail
gels with glucose oxidase (GOx) were injected
in 0.6% agarose gels containing 5 mM glucose.
Agarose has been widely used to mimic the
mechanical properties of the brain, and the
GOx gel polymerizes fully at glucose concen-
trations >100 mM within a time span of 45 min
(fig. S2). The color of the injected solution
transitioned from yellow to dark blue within
seconds (Fig. 1, D and E, and movie S1). The
dark blue color is distinctive for p-type semicon-
ducting and conducting polymers. Ultraviolet-
visible (UV-vis) absorption spectroscopy further
validated the formation of a doped conduct-
ing polymer. In solution, ETE monomers ab-
sorb light with a peak at 350 nm. At higher
concentrations, the monomer will aggregate
and shift the peak to 400 to 450 nm. Un-
doped polymer exhibits absorption between
500 and 600 nm, and absorbances >600 nm
correspond to doped polymer. UV-vis spectra
of the injected gels in agarose showed that
the absorption peak at 350 nm decreased
over time, and a broad absorption band with a
maximum at 850 nm (and extending beyond
1000 nm) increased over time. This broad peak
is in part composed of polaron transition bands
(13, 21), indicating the formation of a doped
conducting polymer (Fig. 1F). Rheology mea-
surements of the gels before and after polym-
erization were performed and show that the
gels have low viscosity. After polymerization,
the increase in viscosity implies the formation
of a soft polymer network (fig. S3).
We additionally injected gels into agarose
using an Au-coated microcapillary and moni-
tored the ETE monomer fluorescence (excita-
tion: 350 nm; emission: 450 nm) over time (fig.
S4, A and B). Time lapse imaging in fluorescence
mode showed a depletion of nonpolymerized
monomer at the injection site. During injec-
tion of the gel into the agarose with 5 mM glu-
cose, fluorescence increased rapidly, followed
by a 99% decay in intensity within 1 min indi-
cating initial diffusion and subsequent polym-
erization of ETE (fig. S4, C and D, and movie
S2). In agarose without glucose, the fluores-
cence intensity dropped by 20% after injection
of the gel, likely owing to diffusion and subse-
quent dilution of the monomer (fig. S4, E and
F, and movie S3). In glucose-containing agar-
ose, the impedance measured using the Au-
coated capillary dropped significantly after
injection of the polymer (fig. S4, G and H),
indicating effective in situ formation of con-
ducting material.
Electrical properties of gels
All cocktail gels were prepared by mixing
the individual components in MES buffer at
Corrected 16 March 2023. See full text.
pH 6.9. During mixing, however, strong elec-
trostatic interactions between the negatively
charged ETE-COONa or ETE-NHS, the me-
dium, and the positively charged lysines of
PLL, HRP, or both resulted in the formation
of particulates, which produced a cloudy sus-
pension (fig. S5, G and H). The effect was more
pronounced with increased concentrations of
ETE and PLL. Density functional theory and
molecular dynamics calculations (fig. S5) con-
firm the aggregation. In the absence of PLL,
ETE-NHS molecules aggregate through p-p
stacking (fig. S5, A and B). In the presence of
PLL, however, the sulfonate group of ETE-NHS
intermediate interacts mainly Coulombically
with the charged group of the lysine, resulting
in aggregation between a PLL chain and a
manifold of ETE-NHS monomers (fig. S5, B
to D). Aggregation between ETE-NHS mole-
cules still occurs to some extent in the absence
of PLL, but to a lesser degree (fig. S5, E and F).
To reduce the aggregation and improve the
A
V
I
Ag/AgCl
PBS
PVA:PLL:ETE
PaC
Au
Glass
ETE-NHS
ETE-NHS Sonication
ETE-S
ETE-S Sonication
)
3
–
m
c
F
(
e
c
n
a
t
i
c
a
p
a
C
C
)
A
µ
(
t
n
e
r
r
u
C
60
40
20
0
-20
-40
B
)
1
–
m
c
S
(
y
t
i
v
i
t
c
u
d
n
o
C
ETE (mg ml–1)
-0.8
-0.4
0
0.4
Voltage (V)
Fig. 2. Electrical and electrochemical stability. (A) Schematic of the electrochemical platform used to
characterize the electrical, electrochemical, and stability properties of the polymerized gels. (B) Conductivity
and specific capacitance as a function of monomer concentration (n = 3 independent measurements).
Datapoints represent the mean ±SD. (C) Cyclic voltammetry of ETE-COONa with sulfo-NHS (ETE-NHS) versus
ETE-S gel films before and after the platforms were sonicated for 10 min in deionized water (n = 3
independent current-voltage measurements).
A
B
C
Fig. 3. Cell biocompatibility. (A) Confocal fluorescent images of cultured PC12 cells in control media
without ETE or enzymes. WGA (red) was used to stain the cell membranes. Scale bar, 20 mm. (B) Confocal
fluorescent images of cultured PC12 cells in media incubated with cocktail gels containing ETE-COONa
with NHS (ETE-NHS) and LOx for 24 hours, showing unpolymerized monomer uptake by the cells. Red structures
are cell membranes, and blue structures represent ETE-monomers. Scale bar, 20 mm. (C) Live/Dead assay
of cells incubated with media preconditioned with gels containing ETE and various combinations of cocktail
enzymes. Data are presented as means ± SD. Kruskas-Wallis was used to test significance. *P = 0.0332, **P =
0.0021 after comparison of all means (n = 6 replicates for every group).
Strakosas et al., Science 379, 795–802 (2023)
24 February 2023
3 of 8
RES EARCH | R E S E A R C H A R T I C L E
Corrected 16 March 2023. See full text.
A
C
E
B
Injected gel with LOx:HRP
Injected gel with LOx:HRP
Injected gel with LOx:HRP
Polymerized gel
t = 0 min
Injected gel with GOx:HRP
t = 10 min
Injected gel with GOx:HRP
t = 20 min
Injected gel with GOx:HRP
injected gel
D
Brain Slice
Polymer
Glass
V
Au
I
Brain Slice
Polymer
50
F
)
A
n
(
t
n
e
r
r
u
C
25
0
-25
Au electrodes
-50
-0.4
-0.2
0
Voltage (V)
0.2
0.4
Tissue
Polymer in tissue
G
PVA:PLL:GOx:HRP
t = 10 min
H
PVA:PLL:GOx:HRP
t = 60 min
polymer formation
polymer formation
I
PVA:PLL:LOx:HRP
t = 60 min
J
)
A
µ
(
polymer formation
0.8
0.6
0.4
0.2
0
-0.2
t
n
e
r
r
u
C
-0.4
-0.6
-0.8
Electrolyte
Electronic heart
-0.1
0
Voltage (V)
0.1
Fig. 4. In vivo polymerization in zebrafish. (A) Photograph of an anesthetized
albino zebrafish with gel polymerized in the fin. (B) Time-lapse microscope
images from a zebrafish fin during polymerization of gels. (C) Schematic showing
injection of gel into zebrafish brain. (D) Schematic of a zebrafish brain slice
integrated with Au metal electrodes for electrical measurements.
(E) Microscope image of brain slice integrated with metal electrodes for
electrical measurements. Scale bar, 200 mm. (F) Current–voltage characteristics from
brain slice samples on Au electrodes (electrode spacing, 15 mm) (n = 3 independent
current-voltage measurements). (G and H) Microscope images of extracted zebrafish
heart immersed in cocktail gels with GOx as oxidase enzyme. (I) Microscope image
of a heart that is immersed in cocktail gels with LOx as the oxidase enzyme. (J) Electrical
measurements from a zebrafish heart on Au electrodes (20-mm spacing) compared
to phosphate-buffered saline electrolyte control. The heart was previously immersed in
gels with LOx for 120 min (n = 3 independent current-voltage measurements).
homogeneity of the suspension, we added PVA
with the intent of balancing the charge on the
PLL (17) (fig. S5, G and H). The addition of PVA
notably enhanced the solubility of the ETE and
PLL mixture (17), resulting in a solution that
was visibly clear for concentrations of ETE-
NHS up to 8 mg ml−1 (fig. S5, G and H).
The electrical characteristics of the conduct-
ing gels are highly dependent on gel formula-
tion. We studied the electrical characteristics
of the gels for different ETE-NHS concentra-
tions using Au microelectrode arrays (MEAs)
patterned with a double layer of parylene C
(PaC) (22, 23) (Fig. 2A). Gel cocktails were
drop-cast on top of the MEAs and allowed to
dehydrate for 10 min. Upon the addition of
5 mM glucose, the gels polymerized (fig. S6A).
For further characterization steps, the area
of the formed conducting polymer gel was
defined through PaC peel-off (of the top PaC
layer) (22) after the gels were polymerized
and dried (fig. S6B). During polymerization, a
small voltage of 0.1 V, which does not induce
polymerization, was applied between two ad-
jacent Au electrodes (fig. S6C). An increase in
current exhibiting sigmoidal behavior was
measured with a delayed onset of typically
500 s after addition of glucose (fig. S6D). The
time at which current onset occurred de-
pended on the concentrations of enzymes, met-
abolites, and supporting biopolymers. However,
the maximum current level was primarily de-
pendent on the concentration of ETE-monomer.
Both conductivity and capacitance increased
while increasing ETE-NHS concentration up
to 8 mg ml−1, and peak values of conductivity
s = 0.25 S cm−1 and specific capacitance Cv =
25 F cm−3 (Fig. 2B) were measured. The mea-
sured conductivity values, however, did not
represent the maximum conductivity of the
gels, as they could be further doped electro-
chemically. To illustrate this concept, we used
the gels as channel materials in organic elec-
trochemical transistors (OECTs) (fig. S6E).
We found that the drain current (Id) at zero
gate voltage was low, and that Id increased
for negative gate voltages. The increase in the
current indicates that the gels were not intrin-
sically fully doped (fig. S6F). We believe that
this was the case because after polymerization,
the carboxyl/sulfo-NHS groups of ETE-NHS
were bound to the amine groups of PLL and
thus could not self-dope the polymer.
Stability and biocompatiblity
Apart from the electrical properties, long-term
structural stability is an important feature of
the formed gels when interfacing with biolog-
ical structures. We designed the gels such that
the formed polymer would be cross-linked. To
test the stability, glucose-induced gels were
polymerized on top of MEAs. After the for-
mation and characterization of the conduct-
ing gels, the coated MEAs were sonicated in
DI water for 10 min to induce harsh mechan-
ical stresses. Cyclic voltammetry before and
after sonication showed that ETE-NHS–based
gel films did not lose electrical or electrochem-
ical characteristics after this treatment (Fig. 2C,
solid lines). By contrast, when ETE-S was
used as the monomer in the gels, the ETE-S–
based films lost most of their capacitance after
Strakosas et al., Science 379, 795–802 (2023)
24 February 2023
4 of 8
RES EARCH | R E S E A R C H A R T I C L E
Corrected 16 March 2023. See full text.
Fig. 5. In vivo polymerization
around the nervous system
of medicinal leeches and the
dynamic change of electro-
chemical properties on flexi-
ble probes. (A) Leech with
polymerized gels, induced by
lactate, around its central nerve.
(B) Close-up of the polymerized
gel from (A). (C) Microscope
image depicting the flexible PaC-
based probe with eight pat-
terned electrodes. The Au line
interconnects are insulated with
PaC. Dashed white lines show
the edges of the probe. The PaC
substrate is extended beyond
the Au electrodes. Scale bar,
500 mm. (D to F) Change in the
color of the conducting gel from
clear to dark blue on top of
the flexible probes over 1 hour.
Scale bars, 2.5 mm. (G), Sche-
matic of circuit connections on
the Au electrodes for measure-
ment of EIS and current between
Au electrodes (after connection
via the polymerized gel). (H) EIS
measurements of all Au electrodes
2 min (black) and 50 min
(red) after gel deposition (Z is
solid line, phase is dashed line;
n = 1 representative impedance
measurement). (I) Current–
voltage characteristics
between two adjacent electro-
des after 2 min (black) and
50 min (red) of gel deposition
(n = 1 representative current-
voltage measurement).
A
B
C
Nerve
PaC
Polymer
Exposed Au
D
t = 0 min
E
t = 10 min
F
t = 60 min
Nerve
Gel
Nerve
Gel
Nerve
Gel
Electrodes
Probe
Electrodes
Electrodes
G
V
Needle-RE
ETE-NHS
gel
Flexible
Probe
V
H
106
Leech
Incision
Nerve
)
(
)
Z
(
g
o
l
105
2 min
50 min
104
100
101
102
103
104
Frequency (Hz)
0
10
20
30
40
50
60
70
I
)
g
e
d
(
e
s
a
h
P
)
A
µ
(
t
n
e
r
r
u
C
80
60
40
20
0
-20
-40
-60
2 min
50 min
-0.2
0
0.2
Voltage (V)
sonication (Fig. 2C, dotted lines). The stabil-
ity of the ETE-NHS–based films and degra-
dation of ETE-S–based films could also be
visually observed before and after sonication
(fig. S7, A to D). In addition to these mechan-
ical fatigue tests, the devices were imposed to
repeated voltage cycling (1000 cycles) with-
out significant loss of capacitance (fig. S7E).
Lastly, ETE-NHS patterned polymer gels were
incubated with PC-12 cells for 72 hours with-
out loss of electrical performance (fig. S7F).
These complementary findings indicate that
ETE-NHS does indeed covalently bind to PLL
and cross-link the bulk of the gel films, allow-
ing for long-term stability.
Finally, gel formulations were optimized for
biocompatibility. The gel electrode system can
potentially induce toxicity in sensitive tissues
either through exposure to individual reac-
tants or as a result of the chemistry associated
with the actual polymerization mechanism.
Although we expect most activated ETE-NHS
monomers to bind to PLL, unbound mono-
mers may diffuse away from the gel to the
surrounding tissue with potentially toxic ef-
fects. Moreover, ROx enzymes can continu-
ously turn over metabolites and produce high
amounts of H2O2 that can be cytotoxic (24).
We prepared three different gel formulations:
(i) ETE-NHS without enzymes, (ii) ETE-NHS +
LOx, and (iii) ETE-NHS + LOx + HRP. Organic
electrochemical transistor-based H2O2 sensors
(with gates modified with platinum nanopar-
ticles to catalyze H2O2) (24, 25) were used to
verify and compare levels of H2O2 production
with and without HRP (fig. S8, A to D). Con-
ditioned media incubated with gels showed
increasing H2O2 with increasing incubation
time until all present metabolites were con-
sumed. We saw that a ratio of 27:1 units of
HRP:LOx produced an optimal amount of
H2O2 (i.e., less than 100 mM) for this specific
system, so this ratio was used for further cell
experiments. Fluorescent images and UV-vis
spectra show that only the gels with a combi-
nation of ETE-NHS + LOx + HRP were nonflu-
orescent, indicating that the ETE polymerized
in conditioned cell media (fig. S9, A to C). For
nonpolymerized gels, ETE monomers diffused
through the media and subsequently attached
to, or passed through, the cell membrane when
the conditioned medium was used to incubate
PC-12 cells (Fig. 3, A and B). The diffused
monomers were found to be nontoxic to cells
(Fig. 3C, first bin). The LOx activity, however,
was toxic to cells owing to the excessive prod-
uction of H2O2 (Fig. 3C, second bin). The gels
with optimized HRP:LOx ratios activities
were shown to be significantly better for cell
viability survival than the gels without HRP
(Fig. 3C, second and third bins).
In vivo polymerization in zebrafish brain,
fin, and heart
Gels were injected in live zebrafish to validate
the enzyme-triggered in vivo polymerization.
Strakosas et al., Science 379, 795–802 (2023)
24 February 2023
5 of 8
RES EARCH | R E S E A R C H A R T I C L E
Anesthetized albino zebrafish were used as a
model to visualize polymerization owing to
their versatility and general transparency. Cock-
tail gels (~100 nl) injected into the tailfins
polymerized in vivo, resulting in a distinct dark
color along the whole length of the fin cavities
(Fig. 4A). Glucose and lactate were both shown
to act as catalysts for polymerization, with lac-
tate inducing a relatively faster polymerization,
presumably because of its higher concentrations
in zebrafish tissues (Fig. 4B and movie S4) (26).
The versatility of this approach is highlighted
in gels injected into the zebrafish brain. Cocktail
gels with LOx (HRP:LOx, 27:1 U/ml; ~10 nl total
volume) as the oxidase enzyme were injected
into the brains of anesthetized zebrafish (Fig.
4C). After injection, zebrafish were left to swim
for 3 hours. The brains were then extracted and
dissected, and brain slices were prepared. The
dark blue polymer could also be seen optical-
ly on fixed brain slices (fig. S10A). By contrast,
a yellow color was observed on brain slices
where only neat ETE-S solution was injected
(fig. S10B). UV-vis spectra from brain slices
that optically exhibited signs of polymeriza-
tion showed an absorption spectrum with a
peak at 600 nm and no peak at 350 to 400 nm
(Fig. 1F and fig. S10C), indicating depletion
of the trimers and polymer formation with
semiconducting behavior. The brain slices
were then placed on top of MEAs to charac-
terize their electrical behavior (Fig. 4D). Mea-
surements were performed in regions where
dark spots were observed along the brain slices
(Fig. 4E). In dehydrated brain slices, a linear
(ohmic) current was observed when a small
voltage between −0.5 and 0.5 V was applied
(n = 3 brains). The magnitude of the current
was higher than in control tissue samples
(Fig. 4F). When cocktail gels with GOx as the
oxidative enzyme were injected into the brains
of zebrafish, however, no polymerization oc-
curred. Optical inspection of the brain slices
showed no sign of a dark blue region. The poor
performance of GOx-containing gels was ex-
pected as zebrafish brains, much like their
human counterparts, are known to have low
glucose and high lactate concentrations (26, 27).
UV-vis spectra from brain slices that optically
exhibited signs of polymerization showed a wide
absorption spectrum from 550 to 950 nm
with a peak at 600 nm. These gels do not have
peaks at 350 to 400 nm, in contrast to an in-
jected ETE-S solution (Fig. 1F and fig. S10C).
To better visualize the in vivo polymerization
induced by lactate versus glucose, we extracted
whole brains, dissected them in half, and im-
mersed the brains in cocktail gels (fig. S10D). In
gels with LOx, the interface between the solu-
tion and the dissected part of the brain turned
dark blue (fig. S10E). In gels with GOx, no
change in color was observed (fig. S10F). The
brains were subsequently placed on top of pat-
terned metal electrodes to examine the electri-
Corrected 16 March 2023. See full text.
cal behavior of the dark-colored regions. The
brains were placed with the dissected part,
corresponding to the interior of the brain,
facing the electrodes (fig. S10, G and H). In
brains that were soaked in LOx gels, a linear
current was observed when a voltage sweep
was applied (fig. S10I, red line). The magni-
tude of the current could reach microamperes
with linear behavior for both wet and dry
samples. No current was observed in brains
that were immersed in GOx cocktail gels (fig.
S10I, black line).
Moreover, LOx-based gels appeared to be
nontoxic. Gels based on LOx were injected
into the brains of anesthetized zebrafish. After
anesthesia, the zebrafish were left to swim for
72 hours with the injected gels (n = 9 fish). The
fish suffered no mortalities, and we did not
observe any changes to the zebrafish swimming
behavior or general appearance (see materials
and methods). To investigate the impact of
the gels on the brain tissue, we extracted the
brains after 48 hours and serially sectioned
them (14-mm thickness). Alternate sections
were stained with methylene blue/thionine
(MB&T) or hematoxylin and Eosin (H&E).
Polymer was observed in the granular cell layer
of the cerebellum (corpus cerebelli) without
any signs of structural tissue damage due to
the injection or the formed polymer (fig. S11).
Zebrafish hearts were also extracted and
immersed in cocktail gels. In gels with either
GOx or LOx, dark blue lines were observed
on the surface of the hearts, indicating that
both glucose and lactate induced polymeri-
zation in the zebrafish heart. The glucose-
induced polymerization was overall slower
compared to lactate. The polymerization in
the heart did not occur everywhere, but rather
threaded structures along the surface of the
heart were observed, with the structures’ size
increasing over time (Fig. 4,G and H). We
speculate that the cocktail gel polymerizes in
regions in which exchange of glucose or lactate
between the heart tissue and the surrounding
electrolyte occurs, e.g., around coronary arteries.
This is expected, as local tissue glucose levels
often covary with local blood flow (28). The
polymerization and formation of the threaded
structures were more prominent in heart sam-
ples that were incubated in gels with LOx (Fig.
4I). The hearts were removed from the LOx
gels and then integrated with MEAs. A linear
current response to the application of a linear
voltage sweep, following Ohm’s law, was ap-
parent for hearts incubated in LOx gels but
not evident for the electrolyte-only controls
(Fig. 4J). In addition to electrical function-
alization, the selective polymerization, in com-
bination with local coloration from monomer
to polymer transition, may be of utility for local
staining in regions of interest.
These findings highlight that the formation
of electronic conductors fueled by endogenous
metabolites can develop soft electronics in var-
ious biological tissues and environments. In-
deed, we injected gels with GOx and LOx in
tenderized beef, pork, and freshly killed chicken
muscle, observing polymerization in all tissues
explored (fig. S12 and movie S5). In a plant-
based tofu sample, polymerization is ineffi-
cient or does not occur owing to the lack or
low concentration of specific metabolites (fig.
S12). Owing to the abundance of metabolites
and neurotransmitters in animal models and
their corresponding enzymes, in vivo poly-
merization and formation of conductors are
not limited to a specific tissue or model, and
large local variations in metabolite concen-
trations between and within animal tissues
(e.g., glucose levels have been reported to vary
from 0.4 mM in the lower respiratory tract
(29) to up to 500 mM in the lumen of the small
intestine (30)) should allow for structure-
specific polymerization as long as a suitable
oxidase can be used.
Ex vivo polymerization around the nervous
system of medicinal leeches as an in situ
electrode area extension
For neuroscience applications, conducting poly-
mers have been extensively used as coatings on
implantable metal electrodes to improve record-
ings and stimulation in the CNS or PNS because
they reduce the electrochemical impedance of
metal electrodes by increasing their electro-
active surface area. This benefit is more promi-
nent for micrometer-sized electrodes, for which
impedance becomes a limiting factor. More-
over, conducting polymers are softer than con-
ventional electrode materials and are thus less
prone to triggering an immune response in
the tissue.
As a proof of concept, we first attempted
to induce ex vivo polymerization of gels at
the interface between nervous tissue and
Au electrodes to reduce their impedance. We
used medicinal leeches (Hirudo medicinalis)
as models owing to their simple, easily acces-
sible, and well-characterized nervous system.
Drop-cast gels with LOx as the primary enzyme
polymerized around the connecting nerve of
the leech (Fig. 5, A and B), and in the leech
muscles, in the span of 5 to 15 min (fig. S13),
whereas gels with GOx did not polymerize
(fig. S13). As in zebrafish brains, lactate con-
centrations in leeches are higher than those
of glucose (31).
Patterned Au electrodes on a flexible PaC-
based probe were used to interface with the
leech nerve via gels polymerized around the
nerve. A printed circuit board was applied as
an interface between the electrode probe and
the recording system (fig. S14E). Each probe
comprised eight electrodes in total, each elec-
trode having an area of 450 by 200 mm2 and
with 50-mm spacing between adjacent electro-
des (Fig. 5C). The Au electrodes were placed
Strakosas et al., Science 379, 795–802 (2023)
24 February 2023
6 of 8
RES EARCH | R E S E A R C H A R T I C L E
1 to 2 mm away from the nerve (with the probe
substrate placed under the nerve), and 2 ml of
Lox-based ETE cocktail gel was drop-cast on
top of the electrodes and around the nerve.
The color of the dropcast solution changed
from white to dark blue over time, and the
polymerized gel surrounded the nerve while
also covering the Au electrodes (Fig. 5, D to
F). When the presence of the probe prevented
metabolites from reaching the gel, blood was
manually transferred on top of the gel. Dur-
ing polymerization, the impedance of the Au
electrodes was monitored by using electrochem-
ical impedance spectroscopy (EIS) versus a
reference metal electrode in the leech muscle
(Fig. 5G). This impedance decreased signifi-
cantly with a phase shift toward lower frequen-
cies (Fig. 5H), indicating increased capacitance
of the electrode. In this case, the reduction in
impedance was caused by the presence of the
conducting polymer itself along with the con-
tribution of multiple metal electrodes (of the
eight in total on the probe) being intercon-
nected through the in situ–formed conducting
polymer. The connection between Au electrodes
was also verified by applying a voltage sweep
between two adjacent electrodes, which re-
sulted in a linear current 50 min after deposi-
tion of the cocktail gel (Fig. 5I, red line). No
current was observed 2 min after deposition
of the gels (Fig. 5I, black line). To verify that
the polymer itself can reduce the impedance
of individual electrodes, we also patterned
gels on top of individual Au electrodes on the
flexible probe by using a parylene peel-off tech-
nique (fig. S14, A and B) and induced polymer-
ization with a lactate-containing solution (fig.
S14C), which decreased the impedance of the
gel patterned Au electrode at low frequencies
(fig. S14D). These impedance measurements
agree with studies in which conducting polymers
increase the capacitance of Au electrodes (22).
The patterned polymer electrodes were then
used to investigate whether the polymerized
gel could act as an extended electrode to stim-
ulate the connecting nerve and induce muscle
contractions in the semi-intact leech model
(fig S14E). When a control Au electrode on the
flexible probe was placed 1 to 2 mm away from
the nerve, stimulation with 250 cathodic cur-
rent pulses (100 to 300 mA, 1-ms duration, 30 Hz)
resulted in clear muscle movements, with a
response frequency of 50 to 60% (table S1).
Equivalent movement response occurred with
the polymerized patterned gel electrodes (movie
S6). There was no obvious difference in the
stimulation threshold when the polymerized
patterned gel was in contact with the nerve
from underneath (fig. S14E and table S1). To
test more precisely if it is possible to stim-
ulate the nerve with the patterned gel as an
extended electrode, a nerve model suitable for
more fine and robust electrophysiological mea-
surements would be preferable. However, al-
Corrected 16 March 2023. See full text.
though, the current threshold did not change,
the voltage output for patterned gels was ~0.5 V
and capacitive in nature (fig. S14E), whereas
the required voltage for the Au electrode alone
was ~2 V. Capacitive charge injection is in
general preferable to faradaic charge injection
because it prevents redox reactions that can
create or deplete chemical species in the bi-
ological environment (32). Overall, in the leech
experiments, we demonstrated that the con-
cept of in vivo–polymerized gels is still under
exploration and needs optimization before
stimulating nerves (e.g., gel area, speed of
polymerization, nerve model, etc.). However,
we demonstrate that polymerized gels can in-
deed improve the properties and performance
of current electrodes by, for example, lowering
the electrode impedance and minimizing or
eliminating faradaic charge injection during
stimulation. Because the relatively large Au
electrodes and gels were not suitable for low-
amplitude and high-frequency recording, it
was not possible to use them for electrophysio-
logical recordings in the leech system.
A strategy for enzymatic polymerization
fueled by endogenous metabolites for in vivo
formation of substrate-free organic electronics,
in both vertebrate and invertebrate models,
was reported. Gels containing enzymes and
small electroactive monomers were injected
into the biological tissue, and endogenous
metabolites induced polymerization of the
monomers. This resulted in organic electronic
gels without the requirement of a rigid—and
thus inherently biologically and rheologically
incompatible—substrate material. Such in vivo–
manufactured gel electrodes were demon-
strated in agarose gels, cell culture, zebrafish
and leech model systems, and mammalian
muscle tissue, and as the channel material in
OECTs, demonstrating the broad versatility of
the endogenously fueled polymerization meth-
od. From spectroscopic, electrical, mechanical,
and neurophysiological investigations, we con-
clude that the developed electrode technology
forms a biological-electronic amalgamation
with precisely the performance and compati-
bility characteristics required for seamless in-
tegration with living systems, going far beyond
the current state of the art in thin-film and/or
substrate-based bioelectronics. The gel formula-
tion was optimized for conducting polymer gels
with high stability and biocompatibility. Owing
to the plethora of metabolites, our method is
not limited to specific biological models but is
rather universal and could be tailored to match
local fluxes of endogenous substances matched
to their corresponding oxidase enzymes
(metabolites, neurotransmitters, other biomark-
ers). Furthermore, the endogenous compound–
fueled polymerization method does not neces-
sitate genetic manipulation of target cells or
tissue, making it more easily translatable. In
addition, selective polymerization combined
with the color change ensuing from monomer
to polymer transition may be of utility for local
staining in regions of interest. We believe that
these locally and in vivo–formed organic elec-
tronic systems promise a paradigm shift in the
development and interface of electronics with
biology, seamlessly blurring the distinction
between the biological and technological or
electronic materials and systems.
REFERENCES AND NOTES
1.
J. W. Salatino, K. A. Ludwig, T. D. Y. Kozai, E. K. Purcell,
Nat. Biomed. Eng. 1, 862–877 (2017).
2. M. Aizawa, Proc. Annu. Int. Conf. IEE Eng. Med. Biol. Soc. 13,
1792–1793 (1991).
3. C. Boehler, S. Carli, L. Fadiga, T. Stieglitz, M. Asplund,
Nat. Protoc. 15, 3557–3578 (2020).
4. Y. Liu, V. R. Feig, Z. Bao, Adv. Healthc. Mater. 10, e2001916
(2021).
5. D. Khodagholy et al., Adv. Mater. 23, H268–H272 (2011).
6. E. Castagnola et al., IEEE Trans. Neural Syst. Rehabil. Eng. 23,
342–350 (2015).
7. E. Lopez-Sola, R. Sanchez-Todo, È. Lleal, J. Neural Eng. 17,
041002 (2020).
8. E. Khor, H. C. Li, A. Wee, Biomaterials 16, 657–661
(1995).
9. S. M. Richardson-Burns et al., Biomaterials 28, 1539–1552
(2007).
10. L. Ouyang, C. L. Shaw, C. C. Kuo, A. L. Griffin, D. C. Martin,
J. Neural Eng. 11, 026005 (2014).
11. J. Liu et al., Science 367, 1372–1376 (2020).
12. G. Dufil et al., J. Mater. Chem. B Mater. Biol. Med. 8, 4221–4227
(2020).
13. E. Stavrinidou et al., Proc. Natl. Acad. Sci. U.S.A. 114,
2807–2812 (2017).
14. G. Tommasini et al., Bioact. Mater. 10, 107–116 (2021).
15. S. Spieth et al., Biomed. Microdevices 14, 799–809 (2012).
16. B. E. B. Jensen et al., Nanoscale 5, 6758–6766 (2013).
17. C. Brough et al., AAPS PharmSciTech 17, 167–179
(2016).
18. L. Richert et al., Biomacromolecules 5, 284–294 (2004).
19. K. E. Hammel et al., J. Biol. Chem. 268, 12274–12281
(1993).
20. J. Deng et al., Nat. Mater. 20, 229–236 (2021).
21. A. V. Volkov et al., Adv. Electron. Mater. 3, 1700096
(2017).
22. M. Sessolo et al., Adv. Mater. 25, 2135–2139 (2013).
23. M. J. Donahue et al., Adv. Mater. 30, 1705031 (2018).
24. X. Strakosas et al., Adv. Sci. (Weinh.) 9, e2101711
(2022).
25. X. Strakosas et al., J. Appl. Polym. Sci. 134, 44483 (2017).
26. Y. C. Tseng et al., Front. Zool. 11, 53 (2014).
27. W. M. Abi-Saab et al., J. Cereb. Blood Flow Metab. 22, 271–279
(2002).
28. K. Takeuchi et al., Circulation 98 (Suppl), II234–II239,
discussion II240–II241 (1998).
29. J. P. Garnett, E. H. Baker, D. L. Baines, Eur. Respir. J. 40,
1269–1276 (2012).
30. R. P. Ferraris, S. Yasharpour, K. C. K. Lloyd, R. Mirzayan,
J. M. Diamond, Am. J. Physiol. 259, G822–G837 (1990).
31. U. Hoeger, A. Wenning, U. Greisinger, J. Exp. Biol. 147, 43–51
(1989).
32. S. F. Cogan, Annu. Rev. Biomed. Eng. 10, 275–309 (2008).
33. X. Strakosas et al., Metabolite-induced in vivo fabrication of
substrate-free organic bioelectronics, Zenodo (2022);
https://doi.org/10.5281/zenodo.7078627.
AC KNOWL ED GME NTS
We thank T.-P. Ruoko, E. Głowacki, C. Musumeci, T. Arbring Sjöström,
S. Rossi, C. Li, D. Priyadarshini, and A. Halder for invaluable
insight and guidance, and R. Henriksen for donating fresh chicken
tissue. This work was carried out within the “e-NeuroPharma”
projects supported by the European Research Council (AdG 2018
Magnus Berggren, 834677), the Swedish Research Council
(2018-06197), and the Swedish Foundation for Strategic Research
(RMX18-0083). Additional funding was provided by the Knut and
Alice Wallenberg Foundation and the Önnesjö Foundation. M.L.
thanks the Swedish National Infrastructure for Computing (SNIC)
for providing computer resources. Part of the study was
Strakosas et al., Science 379, 795–802 (2023)
24 February 2023
7 of 8
RES EARCH | R E S E A R C H A R T I C L E
Corrected 16 March 2023. See full text.
accomplished within MultiPark and NanoLund Strategic Research
Areas at Lund University. Funding for M.H. was provided by the
Swedish Research Council (2021-05231). Funding: European Research
Council (AdG 2018 Magnus Berggren, 834677), Swedish
Research Council (2018-06197), Swedish Foundation for Strategic
Research (RMX18-0083), and Swedish Research Council
(2021-05231). Additional funding was provided by the Knut and
Alice Wallenberg Foundation and the Önnesjö Foundation. Author
contributions: Conceptualization: X.S., H.B., T.A., D.T.S., R.O.,
M.B. Methodology: X.S., H.B., T.A., E.S., D.T.S., R.O., M.B. Monomer
synthesis: T.A., D.B. Zebrafish experiments: X.S., K.H., P.E., M.H.,
F.E. Leech experiments: X.S., M.J.D., M.S.E., H.B. Cell work:
H.B., X.S., C.L. Electrical characterization: X.S., M.J.D., M.S., M.J.
Simulations: M.L. Funding acquisition: R.O., M.B. Supervision: J.G.,
C.L., D.T.S., R.O., M.B. Writing – original draft: X.S., H.B., M.B.
Writing – review and editing: All authors reviewed and edited the
draft. Competing interests: The authors declare that they have
no competing interests. Data and materials availability: The
data and codes that support the findings of this study are available
at Zenodo (33). License information: Copyright © 2023 the
authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.sciencemag.org/about/
science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adc9998
Materials and Methods
Supplementary Text
Figs. S1 to S20
Table S1
References (34–45)
MDAR Reproducibility Checklist
Movies S1 to S6
Submitted 13 May 2022; accepted 22 December 2022
10.1126/science.adc9998
Strakosas et al., Science 379, 795–802 (2023)
24 February 2023
8 of 8
|
10.1126_science.add1969
|
RES EARCH
QUANTUM SIMULATION
Observation of universal Hall response in strongly
interacting Fermions
T.-W. Zhou1,2, G. Cappellini2,3, D. Tusi3, L. Franchi1, J. Parravicini1,2,3, C. Repellin4, S. Greschner5,
M. Inguscio6,2,3, T. Giamarchi5, M. Filippone7, J. Catani2,3, L. Fallani1,2,3*
The Hall effect, which originates from the motion of charged particles in magnetic fields, has
deep consequences for the description of materials, extending far beyond condensed matter.
Understanding such an effect in interacting systems represents a fundamental challenge, even
for small magnetic fields. In this work, we used an atomic quantum simulator in which we tracked the
motion of ultracold fermions in two-leg ribbons threaded by artificial magnetic fields. Through
controllable quench dynamics, we measured the Hall response for a range of synthetic tunneling and
atomic interaction strengths. We unveil a universal interaction-independent behavior above an
interaction threshold, in agreement with theoretical analyses. The ability to reach hard-to-compute
regimes demonstrates the power of quantum simulation to describe strongly correlated topological
states of matter.
S ince its first observation in 1879 (1), the
Hall effect has been an extraordinary
tool for understanding solid-state sys-
tems (2). This phenomenon is a macro-
scopic manifestation of the motion of
charge carriers in materials subjected to a
magnetic field B, generating an electric field
Ey perpendicular to the longitudinal current
Jx flowing in the system. At a small magnetic
field, the Hall coefficient RH = Ey/(BJx) per-
mits the extraction of the effective charge q
and carrier density n, because RH ≃ (cid:1)1=nq
in conventional conductors. The Hall effect
has widespread applications in metrology
and materials science, such as sensitive mea-
surements of magnetic fields and resistance
standards based on its quantized behavior
at large B (3). The modern understanding
of the Hall effect establishes it as a mani-
festation of robust geometric properties of
quantum systems: Fermi-surface curvature
of metals under weak magnetic fields (4, 5),
Berry curvature of anomalous Hall systems (6),
and topological invariants of band insulators
(7). Studies of Hall responses are ubiquitous in
fields that address topological quantum mat-
ter (8) and synthetic realizations thereof (9, 10).
However, when interactions are present
among the carriers, understanding the Hall
coefficient becomes a theoretical challenge.
At large magnetic fields, interactions lead to
the fractional quantum Hall effect (11), where
1Department of Physics and Astronomy, University of
Florence, 50019 Sesto Fiorentino, Italy. 2Istituto Nazionale di
Ottica del Consiglio Nazionale delle Ricerche (CNR-INO),
Sezione di Sesto Fiorentino, 50019 Sesto Fiorentino, Italy.
3European Laboratory for Non-Linear Spectroscopy (LENS),
50019 Sesto Fiorentino, Italy. 4Université Grenoble Alpes,
CNRS, LPMMC, 38000 Grenoble, France. 5Department of
Quantum Matter Physics, University of Geneva, 1211 Geneva,
Switzerland. 6Department of Engineering, Campus Bio-Medico
University of Rome, 00128 Rome, Italy. 7Université Grenoble
Alpes, CEA, IRIG-MEM-L_SIM, 38000 Grenoble, France.
*Corresponding author. Email: fallani@lens.unifi.it
the quantization of RHB to fractions of h/e2
(where h is Planck’s constant) reveals the emer-
gence of elementary excitations with fractional
charge and anyonic statistics (12, 13). For small
fields, the connection of RH with carrier den-
sities and topological invariants is lost, lead-
ing to numerous theoretical attempts (14–21)
to understand the effects of many-body corre-
lations on this quantity. This complicates the
interpretation of measured anomalous tem-
perature dependence and sign changes of RH in
the normal phase of cuprates (22, 23), dis-
ordered superconducting films (24), and organ-
ic compounds (25, 26). Numerical progress
has recently allowed a reliable calculation
of the Hall coefficient (27) in a quasi–one-
dimensional (1D) geometry and predicted a
threshold of interactions above which the Hall
coefficient becomes interaction independent
and thus universal.
In this context, ultracold atoms in optical
lattices provide an opportunity to gain insight
into the fundamental aspects of interacting
Hall systems, owing to their flexibility and con-
trollability. A notable recent advance was the
realization of artificial magnetic fields in op-
tical lattices through various schemes, including
laser-induced tunneling, Floquet engineering,
and synthetic dimensions (9, 28–31). These
schemes have been exploited to explore single-
particle (32–35) and few-body (36, 37) phenomena,
whereas the observation of strongly correlated
many-body effects triggered by interactions
has remained elusive.
In this work, we report on the measurement
of the Hall response in a quantum simulator
with strongly interacting ultracold fermions.
By controlling the repulsion between particles,
we obtained experimental evidence of the uni-
versal response that is predicted at a large
interaction strength. We used a synthetic di-
mension to engineer a two-leg ladder whose
plaquettes are threaded by a synthetic mag-
Fig. 1. Experimental scheme. A synthetic ladder is
realized by trapping fermionic 173Yb atoms in a
1D optical lattice and coupling their nuclear spins
mF = −1/2 and mF = −5/2 via two-photon Raman
transitions. The position-dependent phase of the
Raman coupling simulates a magnetic field B
with Aharonov-Bohm phase ϕ per unit cell. An
atomic current is activated by tilting the ladder with
an optical gradient, equivalent to a constant electric
field Ex. The radius difference between the green
and blue spheres illustrates the leg population
imbalance (Hall polarization) induced by the Hall
drift. The time-dependent longitudinal current Jx(t)
and the Hall polarization Py(t) are measured with
time-of-flight imaging and optical Stern-Gerlach
detection, respectively (typical acquisitions are
shown below the ladder).
netic flux ϕ (Fig. 1). We monitored the real-time
dynamics of the system after the instantaneous
quench of a linear potential, which tilts the
lattice along ^x and mimics the action of a lon-
gitudinal electric field Ex. We observe that the
combined action of Ex and ϕ triggers a lon-
gitudinal current Jx, accompanied by the Hall
polarization of the system along the transverse
direction Py. Even though the dynamics of Jx
and Py strongly depend on microscopic ladder
parameters, we observe that a proxy of the
Hall coefficient, the Hall imbalance (27)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
Py
(cid:1)
(cid:1)
Jx
DH ¼
ð1Þ
converges toward an interaction-independent
value for large atomic repulsions. Our obser-
vations quantitatively agree with theoretical
calculations and confirm the predictions re-
ported in (27). The results showcase the im-
portance of interactions in Hall systems, paving
the way toward the investigation of strongly
correlated effects in topological phases of syn-
thetic quantum matter.
Making and probing a synthetic ladder
Our experiment exploits an ultracold Fermi
gas of 173Yb atoms initially polarized in the
Zhou et al., Science 381, 427–430 (2023)
28 July 2023
1 of 4
RES EARCH | R E S E A R C H A R T I C L E
jF ¼ 5=2; mF ¼ (cid:1)5=2i hyperfine state. The
atoms are trapped in a 1D optical lattice, which
allows real tunneling between different sites
along direction ^x. An additional 2D lattice (not
shown in Fig. 1) freezes the atomic motion
along the orthogonal real-space directions,
forming a 2D array of fermionic tubes. By
adiabatically activating the coherent Raman
couplings between nuclear spin states jmF ¼
(cid:1)1=2i and jmF ¼ (cid:1)5=2i (denoted m = 1, 2
respectively), our system realizes a two-leg
ladder (32), in which the nuclear spins act
as different sites along a synthetic dimension
^y (see Fig. 1). The system is described by a
two-leg version of the interacting Harper-
Hofstadter Hamiltonian
H ¼ (cid:1)tx
X
j;m
h
a†
j;majþ1;m þ h:c:
i
X
h
(cid:1)ty
j
eiϕja†
j;1
aj;2 þ h:c:
i
þ U
X
j
nj;1nj;2
ð2Þ
where aj;mða†
j;mÞ is the fermionic annihilation
(creation) operator on site (j, m) in the real
and synthetic (m = 1, 2) dimensions, nj;m ¼
a†
j;maj;m, and h.c. is the Hermitian conjugate
operator. Here, tx is the nearest-neighbor
tunneling amplitude, and U > 0 is the “on-
rung” interaction energy between two atoms
with different nuclear spins in the same real-
lattice site. The coupling between two spin
states tyeiϕj is interpreted as a tunneling along
the synthetic dimension, whereby the posi-
tion-dependent phase simulates the effect of a
static magnetic flux ϕ threading the ladder;
in our experiments, ϕj
j ¼ 0:32p . A residual
harmonic confining potential results in an
j2nj;m , with
additional term Hconf: ¼ Vx
X
j;m
the confinement strength Vx = 0.01tx. The
atomic repulsion U is controlled, independently
from tx and ty, by changing the radial confine-
ment of fermionic tubes via the 2D lattice depth;
to keep Vx constant while changing U, we added
a weak double-well potential along direction ^x,
compensating the trapping frequency by ad-
justing the potential barrier between the two
wells (38).
To generate a current along ^x, we switched on
jnj;m ,
an optical gradient Hquench ¼ (cid:1)Ex
X
j;m
tilting the ladder along the real-lattice direction,
with Ex = 0.5tx. After time t, we measured the
particle current Jx in the real dimension and the
spin polarization Py in the synthetic dimension.
To perform these measurements, the Raman cou-
pling was abruptly switched off to freeze the
population along the synthetic dimension. The
lattice momentum distribution in the real di-
mension n(k, t), normalized to the total atom
number, was then measured with a band-
mapping technique, where the lattice momenta
k are expressed in units of the real-lattice wave
number kL = p/d and d = 380 nm is the lattice
spacing. We thus access the current Jx, given by
Jx tð Þ ¼ ∫1
(cid:1)1sin pkð
ð
Þn k; t
Þdk
ð3Þ
In the synthetic dimension, the spin distrib-
ution is measured by performing an optical
Stern-Gerlach detection (39). This method,
based on the spin-dependent force exerted
by a near-resonant laser beam, allows a spatial
separation of the two spin components and
the separate count of the atom number Nm in
both of them. The spin polarization coincides
with the transverse (Hall) polarization Py of
the system, which we define as
Py tð Þ ¼
N1 tð Þ (cid:1) N2 tð Þ
N1 tð Þ þ N2 tð Þ
(cid:1)
N1 0ð Þ (cid:1) N2 0ð Þ
N1 0ð Þ þ N2 0ð Þ
ð4Þ
This definition evaluates the difference in frac-
tional spin population with respect to the ini-
tial value, with the populations N1(0) and N2(0)
measured right before the application of the
optical gradient. The definition in Eq. 4 accounts
for the small initial population difference caused
by residual off-resonant coupling to the nuclear
spin statejmF ¼ þ3=2i (38); this initial difference
can safely be neglected owing to the averaging
procedure discussed in the next section. We de-
termined the Hall imbalance DH from the ratio
between the measured Py and Jx, following Eq. 1.
Measuring the Hall effect
Figure 2 shows the measured current, polar-
ization, and Hall imbalance as a function of
time t (defined in units of ħ/tx, where ħ is the
reduced Planck’s constant) for a particular
choice of experimental parameters ty = 3.39tx
and U = 6.56tx. We performed identical mea-
surements with both ϕ = +0.32p and a re-
versed direction of the synthetic magnetic field
ϕ = −0.32p and observed a change of the sign
in Py(t) (38). This behavior confirms the in-
terpretation of our data in terms of the emer-
gence of a Hall response. We averaged these
two independent measurements of {Jx(t), Py
(t)} for ϕ = +0.32p and {Jx(t), −Py(t)} for ϕ =
−0.32p to improve the signal-to-noise ratio
and minimize the effect of the residual off-
resonant coupling to the third state (38). We
observed that the Hall imbalance DH (Fig. 2,
bottom) rapidly approaches a stationary re-
gime, with small amplitude deviations around
a limiting value, whereas the dynamical evo-
lution of Jx and Py remains transient. This fast
convergence of DH is reproduced by the theo-
retical model described in the next paragraph
and conveniently allows us to measure the sta-
tionary Hall response using quench dynamics.
According to the theoretical predictions re-
ported in (27), the stationary Hall imbalance
for strong interactions (U ≫ tx ) is expected to
reach the U-independent universal value
DH ¼ 2
tx
ty
(cid:3) (cid:4)
ϕ
(cid:1)
(cid:1)
(cid:1)
tan
2
(cid:1)
(cid:1)
(cid:1)
ð5Þ
Our simulator yields results consistent with
this universal value (Fig. 2, bottom) despite
important differences with the setup used in
(27) (parabolic confinement, nonlinear drive,
tubes with different particle occupations, and
finite temperatures T ≃ tx) and without using
any fitting procedure. To explain this robustness,
we provide a theoretical analysis. First, we per-
formed extensive density matrix renormaliza-
tion group (DMRG) (40) simulations at zero
Fig. 2. Time evolution of the particle current Jx,
Hall polarization Py, and Hall imbalance DH. The
experimental data are measured at dimensionless
time t (in units of ħ/tx) for ty = 3.39tx and U = 6.56tx,
after applying an instantaneous tilt Ex = 0.5tx.
The values of Jx and Py (top plot; green and red,
respectively) are evaluated by averaging two individual
sets of measurements for ϕ = +0.32p and ϕ = −0.32p,
each comprising 10 to 15 images at every time
step; the error bars represent the standard error of
the mean and are obtained with a statistical
Bootstrap method. The values of DH (bottom plot;
blue) are computed from the data in the top plot
according to Eq. 1, and the error bars represent the
standard error of the mean and are obtained with
standard uncertainty propagation. The shaded areas
are theoretical predictions accounting for the
distribution of atom numbers in the tubes and
experimental temperature uncertainty 1 ≤ T/tx ≤ 2.
They result from a MFA (see main text), where the
renormalized tunneling t(cid:3)
y
parameter ~ty, is introduced to allow a meaningful comparison between the MFA and experiment. The gray
dashed line indicates the universal relation Eq. 5.
¼ 5tx is evaluated through comparison with zero-temperature DMRG. The
Zhou et al., Science 381, 427–430 (2023)
28 July 2023
2 of 4
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Time-averaged
Hall imbalance as a
function of synthetic
tunneling. (A to C)
Single-particle energy
spectrum e(k) calcu-
lated as a function
of the quasimomentum
k for different values of
ty/tx [labeled as A, B,
and C in (D)]. The
interband gap increases
with ty/tx, eventually
leading to two separate
bands (C). The color
scale represents the
population of the m = 1
state. (D) The experi-
mental data (blue
circles) are measured
at U = 6.56tx and
ϕj
j ¼ 0:32p, with the
averaging procedure
and error analysis
detailed in Fig. 2. The
horizontal and vertical error bars show the experimental uncertainty in ty and the uncertainty resulting from
the time average, respectively. The red solid line is the DMRG simulation at zero temperature for a fixed
atomic interaction U = 6.56tx and a number of rungs L = 200, accounting for different tube occupations. The
yellow solid line is the MFA at zero temperature with renormalized t(cid:3)
¼ ty þ 0:1U (see main text for details).
y
The shaded area illustrates the MFA at finite temperatures 0.5 ≤ T/tx ≤ 2, and the gray dashed line indicates
the universal relation from Eq. 5.
Fig. 4. Time-averaged Hall imbalance
as a function of atomic interaction.
The experimental data (blue circles)
are measured at ty = 1.15tx and
j ¼ 0:32p, with the averaging
ϕj
procedure and error analysis detailed
in Fig. 2. The horizontal and vertical
error bars show the experimental
uncertainty in U and the uncertainty
resulting from the time average,
respectively. The red solid line is the
DMRG simulation at zero temperature
for a fixed synthetic tunneling ty =
1.15tx and L = 200 rungs, accounting
for different tube occupations. The yellow solid line is the MFA at zero temperature, with the substitution
ty → t(cid:3)
¼ ty þ 0:30 (cid:4) U. The shaded area illustrates MFA results for finite temperatures 0.5 ≤ T/tx ≤ 2. The
y
gray dashed line indicates the universal relation from Eq. 5, and the dot-dashed line depicts the result for
noninteracting fermions at zero temperature.
temperature (finite-temperature DMRG would
be prohibitively costly). To give a semiquanti-
tative account of the effects of finite tempera-
tures in the nonuniversal regime (intermediate
U and ty), we resorted to a mean-field approxi-
mation (MFA) of interactions, which results in
an effective increase of the transverse coupling
ty → t(cid:3)
y. For each value of U, we first find the t(cid:3)
y
that best reproduces the zero-temperature
DMRG real-time simulations of the current
Jx and polarization Py. We find a quantitative
agreement between MFA and DMRG, if the
MFA polarization is multiplied by t(cid:3)
y=ty ; no
rescaling is required for Jx. We discuss the clear
limitations of MFA to describe the dynamics
of such strongly correlated low-dimensional
systems (38). Nonetheless, this approach gives
a reasonable explanation for why, at finite
temperatures, larger values of the transverse
hopping ty or the interaction U are required
to observe the universal Hall response (Eq. 5),
as observed and explained below.
Testing universality
To pinpoint the onset of the universal regime,
we measured the dependence of the Hall im-
balance’s stationary value on the system pa-
rameters. We considered the averaged Hall
ratio hDHi ¼ hPy tð Þ=Jx tð Þit in the time inter-
val t ∈ [1, 5]. Figure 3D shows the measured
hDHi for a fixed interaction strength U = 6.56tx
and different values of the tunneling ratio
ty/tx, which is controlled by changing the
Raman beam power. The averaged Hall im-
balance hDHi is small at small synthetic tun-
neling ( ty ≪ tx ) and reaches the universal
value (Eq. 5) for ty=tx ≳ 2. Provided that the
system is below half-filling, this transition
also exists in noninteracting systems, where
a large transverse hopping ty opens a large
gap between the two bands of the system,
stabilizing a single-band metal whose Hall im-
balance has the universal value (Eq. 5) (27, 38)
(see single-particle energy bands in Fig. 3, A to
C). Because finite temperatures tend to pro-
mote particles to the upper band, we expect
the finite temperature (T ≃ tx ) in our setup
to push the transition to the universal regime
toward larger values of ty/tx. This effect is
visible in Fig. 3: Whereas zero-temperature
DMRG predicts a transition to the universal
regime at smaller values of ty/tx , the finite-
temperature MFA yields a better quantitative
agreement with the experiment.
Þ
ð
Finally, we demonstrate the interaction-
driven origin of the universal Hall response
of Eq. 5. Figure 4 illustrates the behavior of
the Hall imbalance hDHi upon changing the
interaction strength U/tx at a fixed, nearly
isotropic tunneling ty = 1.15tx. We observe that
hDHi quickly deviates from the noninteracting
value and approaches the U-independent uni-
j ≃ 1:1 at large
Þ tan ϕ=2
ð
versal value 2 tx=ty
j
U/tx. In the spirit of the MFA, this behavior
can be partially explained by the two-band sce-
nario discussed before: Interactions renor-
malize ty toward an interaction-dependent
value t(cid:3)
y > ty, enlarging the gap between the
bands. In the large-U limit, the bands are
well separated and the highest band becomes
empty, similarly to what happens in the large-
ty limit (Fig. 3, A to C). Increasing U thus leads
to a robust single-band metallic state charac-
terized by a constant, universal value of DH
(27). As discussed above, the MFA accounts for
finite-temperature effects and permits a quan-
titative comparison between experiment and
theory. Similar to the transition from weak
to large transverse tunneling described in the
previous paragraph, the finite temperature
pushes the transition toward larger interaction
strengths than those predicted by the zero-
temperature DMRG, as observed in the exper-
imental data.
Zhou et al., Science 381, 427–430 (2023)
28 July 2023
3 of 4
RES EARCH | R E S E A R C H A R T I C L E
Despite the effectiveness of the MFA pic-
ture, we stress the essential and nonperturba-
tive role played by strong interactions to reach
the universal regime in Fig. 4, which funda-
mentally differentiates it from the large-ty limit
(Fig. 3). Indeed, although strong interactions
can be modeled by the MFA using a renor-
malized t(cid:3)
y , the universal regime is reached
anyway for DH ¼ 2 tx=ty
j, and not
(cid:3)
for DH ¼ 2 tx=t(cid:3)
tan ϕ=2
ð
j
y
j. We emphasize
Þ tan ϕ=2
ð
j
(cid:4)
ð
Þ
Þ
that the observed effect is truly a many-body
effect, as captured by the DMRG calculations.
The MFA allows us to make progress at finite
temperatures because it reproduces the sta-
bilization of the single-band metal at large U
by increasing ty, where the Hall imbalance
approaches a universal value when ty > T (38).
Nonetheless, MFA requires its results to be
rescaled to give a quantitative account of the
universal regime and has clear limitations with
respect to reproducing the fine dynamics of the
polarization at large U (38). The MFA should
thus be complemented with more complete,
but much more difficult, exact finite-temperature
studies.
Conclusions
In this experiment, we have shown distinctive
many-body effects triggered by strong interac-
tions in the Hall response of a controllable
quantum simulator of a two-leg ladder threaded
by a synthetic magnetic flux. Beyond the clear
potential of such experiments to measure Hall
voltages (41) and clarify the exotic Hall response
of strongly correlated solid-state conductors,
our work paves the way to the investigation
of the exotic transport properties of strongly
correlated topological phases of matter. This
cold-atom experiment enters unknown ter-
ritory for theory because it features strong
correlations and finite temperatures and yet
shows full control of the simulation param-
eters. An interesting perspective resides in
investigating interacting ladders with a larger
number of nuclear spin states, a regime no-
toriously difficult to access with present com-
putational techniques.
RE FERENCES AND NOTES
33. M. Aidelsburger et al., Nat. Phys. 11, 162–166
1. E. H. Hall, Am. J. Math. 2, 287 (1879).
2. R. S. Popovic, Hall Effect Devices (CRC Press, ed. 2, 2003).
3. K. Klitzing, G. Dorda, M. Pepper, Phys. Rev. Lett. 45, 494–497
(1980).
4. N. P. Ong, Phys. Rev. B 43, 193–201
(1991).
5. M. Tsuji, J. Phys. Soc. Jpn. 13, 979–986
(1958).
6. D. Xiao, M.-C. Chang, Q. Niu, Rev. Mod. Phys. 82, 1959–2007
(2010).
7. D. J. Thouless, M. Kohmoto, M. P. Nightingale, M. den Nijs,
Phys. Rev. Lett. 49, 405–408 (1982).
8. M. Z. Hasan, C. L. Kane, Rev. Mod. Phys. 82, 3045–3067
(2010).
9. N. Goldman, J. C. Budich, P. Zoller, Nat. Phys. 12, 639–645
(2016).
10. T. Ozawa et al., Rev. Mod. Phys. 91, 015006
(2019).
11. D. C. Tsui, H. L. Stormer, A. C. Gossard, Phys. Rev. Lett. 48,
1559–1562 (1982).
12. R. B. Laughlin, Phys. Rev. Lett. 50, 1395–1398
(1983).
13. D. Yoshioka, The Quantum Hall Effect, Series
in Solid-state Science, vol. 133 (Springer, ed. 1,
2002).
14. A. Kapitulnik, S. A. Kivelson, B. Spivak, Rev. Mod. Phys. 91,
011002 (2019).
15. W. F. Brinkman, T. M. Rice, Phys. Rev. B 4, 1566–1571
(1971).
16. G. León, C. Berthod, T. Giamarchi, Phys. Rev. B 75, 195123
(2007).
17. A. Auerbach, Phys. Rev. Lett. 121, 066601 (2018).
18. X. Zotos, F. Naef, M. Long, P. Prelovšek, Phys. Rev. Lett. 85,
377 (2000).
19. E. Lange, Phys. Rev. B 55, 3907–3928
(1997).
20. B. S. Shastry, B. I. Shraiman, R. R. P. Singh, Phys. Rev. Lett. 70,
2004–2007 (1993).
21. E. Berg, S. D. Huber, N. H. Lindner, Phys. Rev. B 91, 024507
(2015).
22. S. J. Hagen, C. J. Lobb, R. L. Greene, M. G. Forrester,
J. H. Kang, Phys. Rev. B 41, 11630–11633
(190).
23. S. Badoux et al., Nature 531, 210–214 (2016).
24. A. W. Smith, T. W. Clinton, C. C. Tsuei, C. J. Lobb,
Phys. Rev. B 49, 12927–12930 (1994).
25. G. Mihály, I. Kézsmárki, F. Zámborszky, L. Forró,
Phys. Rev. Lett. 84, 2670–2673 (2000).
26. J. Moser et al., Phys. Rev. Lett. 84, 2674–2677
(2000).
27. S. Greschner, M. Filippone, T. Giamarchi, Phys. Rev. Lett. 122,
083402 (2019).
(2015).
34. D. Genkina et al., New J. Phys. 21, 053021 (2019).
35. T. Chalopin et al., Nat. Phys. 16, 1017–1021
(2020).
36. M. E. Tai et al., Nature 546, 519–523 (2017).
37. J. Léonard et al., Nature 546, 10.1038/s41586-023-06122-4
(2023).
38. See supplementary materials.
39. S. Taie et al., Phys. Rev. Lett. 105, 190401 (2010).
40. S. R. White, Phys. Rev. Lett. 69, 2863–2866 (1992).
41. M. Buser, S. Greschner, U. Schollwöck, T. Giamarchi,
Phys. Rev. Lett. 126, 030501 (2021).
42. J. Hauschild, F. Pollmann, SciPost Phys. Lect. Notes 2018, 5
(2018).
43. TeNPy code; https://github.com/tenpy/tenpy.
44. T.-W. Zhou et al., Datasets for “Observation of universal Hall
response in strongly interacting Fermions”. Zenodo
(2023); https://doi.org/10.5281/zenodo.7485725.
AC KNOWLED GME NTS
We thank T. Beller for very helpful comments on the manuscript
and C. Berthod and N. Cooper for useful discussions. Funding:
We acknowledge financial support from the Topology and
Symmetries in Synthetic Fermionic Systems (TOPSIM) European
Research Council (ERC) Consolidator Grant 682629, the
TOPSPACE MIUR FARE project, Quantum Technologies For LAttice
Gauge (QTFLAG) QuantERA ERA-NET Cofund in Quantum
Technologies, and MIUR PRIN project 2017E44HRF. This work was
supported in part by the Swiss National Science Foundation
under Division II grant 200020-188687. M.F. acknowledges support
from Swiss National Science Foundation/Schweizerischer
Nationalfonds (FNS/SNF) Ambizione grant PZ00P2_174038 and
the EPiQ ANR-22-PETQ-0007 part of Plan France 2030. Author
contributions: L.Fa., J.C., G.C., M.I., M.F., and T.G. conceived
the experiments. T.-W.Z., D.T., L.Fr., G.C., and J.P. carried out the
experimental work. T.-W.Z., D.T., and L.Fr. analyzed the
experimental results. S.G., C.R., M.F., and T.G. performed
theoretical work. All authors contributed extensively to the
discussion of the results and to the writing of the manuscript.
Competing interests: The authors declare no competing interests.
Data and materials availability: DMRG and time-dependent
variational principle calculations were performed using the TeNPy
Library (version 0.7.2) (42, 43). All of the experimental and
theoretical data presented in the main figures are available for
download from Zenodo, an open-access repository (44). License
information: Copyright © 2023 the authors, some rights
reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government
works. https://www.science.org/about/science-licenses-
journal-article-reuse
28. J. Dalibard, F. Gerbier, G. Juzeliūnas, P. Öhberg, Rev. Mod.
SUPPLEMENTARY MATERIALS
Phys. 83, 1523 (2011).
29. N. Goldman, G. Juzeliūnas, P. Öhberg, I. B. Spielman,
Rep. Prog. Phys. 77, 126401 (2014).
30. N. R. Cooper, J. Dalibard, I. B. Spielman, Rev. Mod. Phys. 91,
015005 (2019).
31. T. Ozawa, H. M. Price, Nat. Rev. Phys. 1, 349–357
(2019).
32. M. Mancini et al., Science 349, 1510–1513 (2015).
science.org/doi/10.1126/science.add1969
Supplementary Text
Figs. S1 to S10
References (45–53)
Submitted 26 May 2022; accepted 14 June 2023
10.1126/science.add1969
Zhou et al., Science 381, 427–430 (2023)
28 July 2023
4 of 4
|
10.1126_science.add1219
|
RES EARCH
CYTOKINE SIGNALING
Structural basis of g chain family receptor sharing
at the membrane level
Tiantian Cai1†‡, Rachel Lenoir Capello1†, Xiong Pi1,2, Hao Wu1,2, James J. Chou1*§
Common g chain (gc) cytokine receptors, including interleukin-2 (IL-2), IL-4, IL-7, IL-9, IL-15, and IL-21 receptors,
are activated upon engagement with a common gc receptor (CD132) by concomitant binding of their
ectodomains to an interleukin. In this work, we find that direct interactions between the transmembrane
domains (TMDs) of both the gc and the interleukin receptors (ILRs) are also required for receptor activation.
Moreover, the same gc TMD can specifically recognize multiple ILR TMDs of diverse sequences within the
family. Heterodimer structures of gc TMD bound to IL-7 and IL-9 receptor TMDs—determined in a lipid
bilayer–like environment by nuclear magnetic resonance spectroscopy—reveal a conserved knob-into-hole
mechanism of recognition that mediates receptor sharing within the membrane. Thus, signaling in the gc
receptor family requires specific heterotypic interactions of the TMDs.
M embers of the common g chain (gc)
family of cytokine receptors, upon
binding to their corresponding inter-
leukin, pair specifically with a com-
mon, shared gc receptor (CD132) to
allow signal transduction (1, 2). To date, the
family consists of six members: the interleukin-2
receptor (IL-2R), IL-4R, IL-7R, IL-9R, IL-15R,
and IL-21R. These receptors play crucial roles in
the development and proliferation of multiple
lymphocyte lineages of both the innate and
adaptive immune systems (1, 3) and are clinically
important targets for immunotherapy (2, 4, 5).
Interleukin receptors (ILRs) and the gc are type
I transmembrane proteins comprising a ligand-
binding ectodomain, a transmembrane domain
(TMD), and an intracellular domain (ICD) con-
taining a Janus kinase (JAK)–binding site.
The mechanism by which interleukins me-
diate specific pairing of the gc with ILRs has
been revealed by crystallographic studies of the
IL-2–bound and IL-4–bound receptor com-
plexes (6–8). In both cases, the interleukin
binding to an ILR generates a composite sur-
face for subsequent interaction with the gc.
This interaction interface is small (~1000 Å2)
and characterized by relatively flat surface com-
plementarity (7), constituting the structural
basis for the degenerate cytokine recognition
by the gc.
One aspect of ILRs not well understood is
the TMD and its possible role in receptor sig-
naling. This question was inspired by previous
studies on other receptors that have revealed
1Department of Biological Chemistry and Molecular Pharmacology,
Harvard Medical School, Boston, MA 02115, USA. 2Program in
Cellular and Molecular Medicine, Boston Children’s Hospital,
Boston, MA 02115, USA.
*Corresponding author. Email: james_chou@hms.harvard.edu
†These authors contributed equally to this work.
‡Present address: High Magnetic Field Laboratory, Hefei Institutes
of Physical Science, Chinese Academy of Sciences, Science Island,
Hefei, 230031, China.
§Present address: Interdisciplinary Research Center on Biology and
Chemistry, Shanghai Institute of Organic Chemistry, Chinese
Academy of Sciences, Shanghai, 201203, China.
much more active roles of TMDs in mediating
receptor oligomerization and activation than
conventionally appreciated. Examples include
type 1 cytokine receptors, such as erythropoi-
etin receptor (9), thrombopoietin receptor (10),
growth hormone receptor (11, 12), receptors in
the tumor necrosis factor receptor superfamily
(13, 14), and receptor tyrosine kinases (15, 16).
An early cryo–electron microscopy (cryo-EM)
study of a full-length transmembrane form of
a quaternary cytokine-receptor complex com-
prising gp130, LIF-R, the cytokine ciliary neuro-
trophic factor (CNTF), and its alpha receptor
(CNTF-Ra) has suggested that the TMDs of these
receptor chains are associated (17). Furthermore,
mutations within TMDs of the gc family recep-
tors have been recorded in clinical studies. For
example, a single mutation (V253G) in the IL-7R
TMD is associated with acute lymphoblastic
leukemias (ALLs) (18). The structural basis of
these disease mutations is unknown owing to
a lack of structural information for the trans-
membrane region of the gc family receptors.
In this study, we report that heterotypic in-
teractions between the TMDs of the gc and its
family members are an essential component of
receptor signaling, which raises the question
how a single transmembrane helix (TMH) can
have the structural sophistication to specifi-
cally recognize a variety of TMHs with highly
divergent sequences. We determined high-
resolution structures of the gc TMD in com-
plex with IL-7R and IL-9R TMDs in bicelles
that mimic a lipid bilayer. The two heterodimer
structures and functional mutagenesis reveal
a common knob-into-hole mechanism that
underlies degenerate ILR pairing with the
gc within the membrane.
Removal of an ILR (IL-7R or IL-9R) and
gc ectodomains activates JAK-STAT signaling
A large body of structural and functional studies
on members of the gc family receptors have
described receptor activation (4, 6–8, 19).
Briefly, interleukins bind to the ectodomains
of their a or b chain receptors to form a com-
plex that subsequently recruits the gc, thus
positioning the ICDs in the correct arrange-
ment to allow reciprocal phosphorylation of
JAK1 and JAK3 and to activate downstream
signal transducers and activators of transcrip-
tion (STAT) signaling (Fig. 1A). However, the
ectodomain and ICD are separated by the
TMD, which may also contribute to driving a
signaling-compatible configuration. The pre-
ligand conformation of receptor ectodomains
could physically prevent ligand-independent,
TMD-driven signaling (14, 20). To test whether
heterotypic association between the ILR TMD
and the gc TMD can be achieved by the remo-
val of ectodomains, wild-type (WT) or cleav-
able human IL-7R (hIL-7R) and gc were
designed (Fig. 1B) and coexpressed in BaF3
cells. After incubation with TEV protease to
shed the ectodomains, activation was eval-
uated with STAT5 phosphorylation.
Both WT receptors (hIL-7R and hgc) and
cleavable receptors (TEV–hIL-7R and TEV-hgc)
were well expressed on the cell membrane with
enhanced green fluorescent protein (EGFP)
and mCherry fused to the C termini of hIL-7R
and hgc, respectively (fig. S1A). Colocalization
of hIL-7R and hgc as well as sparse bright
puncta were observed in the absence of hIL-7,
consistent with previous accounts of both
homotypic preassociation of hIL-7R (21) and
heterotypic preassociation with hgc (22, 23).
More puncta appeared on the cell surface after
ligand incubation as a result of greater recep-
tor clustering induced by ligand binding. Pre-
dictably, this was accompanied by receptor
activation as indicated by STAT5 phosphoryl-
ation (fig. S1B). The TEV-cleavable hIL-7R and
hgc exhibited analogous ligand response to
that of the WT receptors (fig. S1B) but could be
activated by incubating BaF3 cells with TEV
protease in the absence of ligand (Fig. 1C). The
TEV protease dose-response profile revealed
that ~40% maximal activation could be reached
compared with IL-7–induced activation (Fig.
1D) and that the increase of receptor activa-
tion correlated with increased shedding of
receptor ectodomains from the cell surface
(fig. S1, C to H). More cell-surface puncta were
observed after treatment with 50 mg/ml of TEV
protease, which suggests the ability of TMH-
ICD to cluster, though substantially less than
in the context of ectodomain-ligand engage-
ment (fig. S1I). By contrast, no activation was
detected after adding TEV protease to cells
expressing WT receptors (Fig. 1, C to D, and fig.
S1I). Similar receptor activity induced by pro-
teolytic removal of ectodomain was observed
for cleavable hIL-9R and hgc (fig. S2), which
suggests that heterotypic interaction of TMDs
is a general phenomenon for gc receptor fam-
ily members.
To independently examine the function-
al consequence of ectodomain removal, we
Cai et al., Science 381, 569–576 (2023)
4 August 2023
1 of 8
RES EARCH | R E S E A R C H A R T I C L E
designed two constructs with the ectodo-
main (ECD) deleted (hIL-7R–DECD and hgc-
DECD) and expressed each of them alone or
paired with a WT or ectodomain-deleted
receptor (Fig. 1E). The membrane integra-
tions of hIL-7R–DECD and hgc-DECD were
confirmed by confocal microscopy (fig. S1J).
Among receptor-expressing cells, those co-
expressing hIL-7R–DECD and hgc-DECD ex-
hibited the strongest signaling, reaching ~40%
of ligand-induced STAT5 phosphorylation
by the WT receptors (Fig. 1, F and G). This result
A
B
P
i
l
c
c
c
c
IL
D2
D2
ns
ns
ns
0
0
IL
IL
D1
D1
D1
D1
D2
D1
D2
D1
D1
D2
D2
0.0
0.4
0.2
0.6
E
site 1
ICD
C
D
TMD
site 2b
Jak1
Jak3
Jak3
Jak1
site 2a
ECD
Jak3
Jak3
Jak1
Jak1
Rest
STAT
Actin
Actin
Jak1
P
IL-7R
STAT5
STAT5
WT c
Jak3
P
Jak1
P
pSTAT5
pSTAT5
a
m
r
o
N
hIL-7
hIL-7
/ -chain
/ -chain
/ -chain
TEV
protease
TEV
protease
WT IL-7R
Intermediate
TEV
cleavage
site
5
T
A
T
S
/
5
T
A
T
S
p
d
e
z
WT IL-7R / WT c
TEV protease
WT IL-7R / WT c
TEV-IL-7R / TEV- c
10 20 30 40 50 60 100 200
?
Jak3
P
STAT
P
Activation
TEV-IL-7R / TEV- c
TEV protease
10 20 30 40 50 60 100 200
Fig. 1. Removal of the
ectodomains of IL-7R and gc
activates IL-7R signaling.
(A) The prevailing model
of cytokine-mediated receptor
sharing for receptors in the
common gc receptor family.
Receptors in this family
are type I membrane proteins
comprising a ligand binding
ectodomain (ECD), a TMD, and
an ICD with docking sites for
the JAK-STAT signaling mole-
cules. (B) The autoinhibition
hypothesis postulating that
preligand association of ECDs
hinders clustering of TMD-ICD
for activation and that proteo-
lytic removal of the ECDs
should allow TMD-driven
receptor pairing and cytokine-
independent signaling.
(C) IL-7R pathway activation
detected by immunoblotting
of STAT5 phosphorylation
(pSTAT5) in BaF3 cells coex-
pressing cleavable receptors
(TEV–IL-7R and TEV-gc) (top)
or WT receptors (WT IL-7R and
WT gc) (bottom) after treat-
ment with 0 to 200 mg/ml of
TEV protease or 50 ng/ml
of hIL-7. (D) Quantification of
protease-induced pSTAT5
signals in (C) as pSTAT5/
STAT5 intensity ratios, which
are further normalized relative
to that induced by hIL-7.
The data are shown as
means ± SEMs calculated from
all technical replicates from
three independent experiments
(with one to two technical
replicates per experiment).
Statistical significance by
Student’s t test. ns, not
significant; **P ≤ 0.01;
***P ≤ 0.001. (E) Schematic
illustration of all possible
combinations of full-length
receptors (IL-7R, gc) and
ECD-deleted receptors
(IL-7R–DECD, gc-DECD)
expressed in BaF3 cells for
testing ligand-independent signaling. (F) IL-7R pathway activation reported by phosphorylation of STAT5, JAK1, and JAK3 in BaF3 cells expressing full-length and
ECD-deleted receptor combinations illustrated in (E). (G) Quantification of pSTAT5 signals in (F) as pSTAT5/STAT5 intensity ratios, which are further normalized
relative to that induced by hIL-7. The data are shown as means ± SEMs calculated from all technical replicates from three independent experiments (with one to two
technical replicates per experiment). Statistical significance by Student’s t test. ns, not significant; *P ≤ 0.05; **P ≤ 0.01.
ECD
T c
/ WT IL-7
/ W
WT IL-7R / WT
ECD
ECD
WT IL-7
c-
ECD
ECD
c-
c-
ECD
/
ECD
c-
WT IL-7R
WT IL-7R / WT
ECD /
IL-7R-
/ WT
ECD /
R / WT c
TEV protease(µg/ml)
IL-7R- ECD
+ c- ECD
ECD
c-
WT IL-7R
+ c- ECD
5
T
A
T
S
5
T
A
T
S
p
d
e
z
WT IL-7R
+ WT c
IL-7R-
L-7R-
IL-7R-
IL-7R-
IL-7R-
ECD
IL-7R- ECD
IL-7R- ECD
ECD
pSTAT5
STAT5
a
m
r
o
N
c- ECD
Jak1 Jak3
+ WT c
pJak1
pJak3
hIL-7
Actin
Jak1
Jak3
Jak1
-
c
G
Jak1
Jak3
Jak3
Jak3
Jak1
R
F
200
100
c
0.6
0.0
0.1
0.3
0.2
0.4
0.5
D2
c
c
10
20
40
30
60
50
I
ns
ns
+
_
_
_
_
_
_
i
l
0
/
Cai et al., Science 381, 569–576 (2023)
4 August 2023
2 of 8
RES EARCH | R E S E A R C H A R T I C L E
was consistent with the activation achieved by
proteolytic removal of ectodomains (Fig. 1, C
and D). The cells expressing hIL-7R–DECD or
hgc-DECD (with a WT receptor) also showed
weaker but measurable ligand-independent
signaling (Fig. 1, F and G), which suggests that
both hIL-7R and hgc ectodomains are involved
in receptor autoinhibition in the absence of
ligand. They would prevent the TMD-ICD re-
gions of hIL-7R and hgc from forming signaling-
competent complexes. Thus, TMDs appear
to play an active role in the heterotypic asso-
ciation between hIL-7R and hgc that drives
signaling.
TMD interactions match the
receptor-sharing network
We next investigated whether the TMD of
the gc (designated gcTMD) could directly as-
sociate with that of ILR members (designated
IL-xRTMD) in the cell membrane and whether
the heterodimerization of TMDs also exhib-
ited properties of receptor sharing as do the
ectodomains. We selected the TMDs of gc,
IL-4R, IL-7R, and IL-9R from both human
and mouse to test oligomerization (Fig. 2A).
We also selected the TMD of IL-5R from the
b chain (bc) family to examine possible non-
specific association with gcTMD. The TMD
sequences of these receptors (fig. S3) did not
show any conserved small-xxx-small motif
that mediates TMH dimerization in growth
factor receptors (16, 24–26). Nor did they
show charged residues that mediate intra-
membrane assembly of activating immuno-
receptors (27–29). Thus, pairing of the gc TMD
with different ILR TMDs was mediated pos-
sibly by a previously unrecognized type of
interaction.
TMD interactions were examined using a
bacterial adenylate cyclase two-hybrid (BACTH)
assay (30). Briefly, two complementary domains
(T18 and T25) of adenylate cyclase (AC) were
A
C
m c
mIL-4R
T25
mIL-7R
mIL-9R mIL-5R
c
hIL-4R
T25
hIL-7R
hIL-9R hIL-5R
h c
hIL-4R
8
1
T
hIL-7R
hIL-9R
hIL-5R
Empty
vector
IbaG-T18+
IbaG-T25
I
IL-13R
IL-9R
IL-4R
c
IL-2R
IL-2R
IL-7R
IL-21R
IL-15R
IL-3R
c
TSLPR
GM-
CSFR
IL-5R
B
TMD1 TMD2
TMD1
TMD2
Bacteria
Inner membrane
T18 T25
T18
T25
ATP
cAMP
ATP
cAMP
lacZ expression
No lacZ expression
X-Gal fermentation
No X-Gal fermentation
m c
mIL-4R
8
1
T
mIL-7R
mIL-9R
mIL-5R
Empty
vector
IbaG-T18+
IbaG-T25
D
)
y
t
i
v
i
t
c
A
e
v
i
t
l
a
e
R
(
n
o
i
t
c
a
r
e
t
n
I
D
M
T
2
1
0
Blue colonies
White colonies
IbaG/IbaG
Empty vector
c
c/hIL-2R
c/h
hIL-2R
h
h
h
c
/h
c/hIL-2R
hIL-2R
c
c
c
/h
c/hIL-4R
hIL-4R/h
c/hIL-7R
hIL-7R/h
c/hIL-9R
hIL-9R/h
h
h
c
c
c
c/hIL-15R
hIL-15R/h
c/hIL-21R
hIL-21R/h
c/hIL-5R
hIL-5R/h
h
h
h
h
c
Fig. 2. Specific heterodimerization of gcTMD with the TMDs of its family
members. (A) Schematic diagram showing receptor-sharing network for the gc
(blue) and bc (yellow) receptor families. Some complexes require a third chain for
cytokine recognition and signal transduction, such as IL-2Ra and IL15Ra. IL-4Ra
and IL-7Ra can also form heterodimer complexes with IL-13Ra and TSLPR,
respectively. (B) Schematic illustration of the BACTH system (30) for analyzing
TMD interactions. ATP, adenosine 5′-triphosphate; cAMP, adenosine 3′,5′-
monophosphate; lacZ, b-galactosidase; X-Gal, X-galactosidase. (C) BACTH analysis
of TMD interactions for representative gc family receptors and the bc family
member IL-5R from both mice and humans. Blue colonies indicate TMD-TMD
association in the bacteria inner membrane. (D) Quantification of BACTH colony
colors, which are normalized relative to the IbaG/IbaG positive control. The data are
shown as means ± SEMs calculated from technical replicates (typically three to
four) from one representative of three independent experiments.
Cai et al., Science 381, 569–576 (2023)
4 August 2023
3 of 8
RES EARCH | R E S E A R C H A R T I C L E
fused, separately, to the TMDs under inves-
tigation (Fig. 2B). The TMD fusion proteins
were expressed in an AC-deficient (CyaA) strain
that yielded white bacteria colonies by default.
Stable association of two TMDs reconstitutes
AC activity, resulting in blue colonies. For
both mouse and human sequences, gcTMD
can stably heterodimerize with the TMDs of
IL-4R, IL-7R, and IL-9R—the corresponding
colonies showed the strongest blue colors
(Fig. 2C and fig. S4). By contrast, no asso-
ciation between the TMDs of gc and IL-5R
was detected (Fig. 2C and fig. S4), which in-
dicates no cross-family TMD association. Ad-
ditional BACTH results for human gc, IL-2Ra,
IL-2Rb, IL-15R, and IL-21R from the gc
family and for human bc and IL-5R from the
bc family (fig. S5) demonstrated that hetero-
typic TMD associations are consistent with the
gc family receptor–sharing network (Fig. 2D).
A homotypic TMD interaction was also ob-
served for IL-2Ra and IL-2Rb, IL-4R, IL-7R, IL-
9R, and IL-21R, though on average it was
weaker than the heterotypic interaction in-
volving gcTMD (Fig. 2C and fig. S5). Although
the function of homotypic TMD interactions is
unknown, it may contribute to preassembly
of ILRs before ligand engagement. A previous
study had suggested that the IL-2Ra and IL-
2Rb are already colocalized in resting T cells in
the absence of IL-2 (31).
Structures of gcTMD in complex with
IL-7RTMD and with IL-9RTMD in bicelles
The specific recognition of the TMDs of gc
family ILRs by gcTMD was unexpected be-
cause their sequences are highly divergent (fig.
S3). To understand the structural basis of re-
ceptor sharing at the membrane level, we de-
termined high-resolution nuclear magnetic
resonance (NMR) structures of two represen-
tative heterodimer complexes in bicelles
mimicking a lipid bilayer: gcTMD bound to
IL-7RTMD (figs. S6 and S7) and gcTMD bound
to IL-9RTMD (figs. S6 and S8). In addition,
an intermolecular nuclear Overhauser effect
(NOE) difference experiment revealed gcTMD
residues in the interaction interface. I274,
L277, and I278 in particular experienced inter-
molecular NOE both in complex with IL-7RTMD
and with IL-9RTMD (Fig. 3A).
The two transmembrane heterodimer struc-
tures were unexpectedly homologous. Foremost,
this structural similarity was characterized by
Equilibrium
Excited
NOE
c TMD
15N, 2H
15N
H
H
IL-xR TMD
13C
13C
c TMD / IL-7R TMD
c TMD / IL-9R TMD
G272
G269
107
110
G272
G269
107
110
V280
E262
8.5
I266
I275
F279
Y281
W283
F282
S257
T270
I278
I274
T276
L284
F259
A263
L277
V271
V268
115
V280
E285
120
E262
8.5
I266
I275
F279
Y281
W283
F282
S257
T270
I278
T276
L284
F259
A263
I274
L277
V271
V268
115
1
5
N
(
p
p
m
)
120
E285
L273
L258
125
L273
L258
125
R286
R286
7.5
8.5
8.0
1H (ppm)
8.5
8.0
1H (ppm)
7.5
c ECD
c ECD
IL-7R ECD
IL-9R ECD
extracellular
30 Å
intracellular
D
c
IL-9R
L255
I258
V262
L283
I274
c
IL-7R
L259
I274
A
B
C
Fig. 3. NMR structures of gcTMD
in complex with IL-7RTMD and
with IL-9RTMD show a common
knob-into-hole mechanism of
recognition. (A) (Left) Detecting
residue-specific interchain NOEs
using the sample of (15N, 2H)
gcTMD mixed with (13C) IL-7RTMD
or IL-9RTMD at a 1:1 molar ratio.
The experiment involves recording
two interleaved 1H-15N TROSY-
HSQC spectra: one at equilibrium
and the other with the 13C-attached
aliphatic protons inverted during
200 ms of NOE mixing. (Right)
Overlaying the difference between
two interleaved 1H-15N TROSY-
HSQC spectra (red) onto the
reference spectrum (blue) reveals
gcTMD residues in close contact
with IL-7/9RTMD. The spectra
were recorded at 303 K and 1H
frequency of 900 MHz. ppm, parts
per million. (B) Ribbon representa-
tion of the structures of gcTMD
in complex with IL-7RTMD (left) and
with IL-9RTMD (right) in DMPC-
DH6PC bicelles with q = 0.4.
(C) A close-up view of gcTMD I274
(sphere) fitting into the hydrophobic
pocket of IL-7RTMD (yellow) formed
by the F1-xx-F2F3-xx-F motif.
(D) The same close-up view as in
(C) of gcTMD I274 fitting into the
pocket of IL-9RTMD (pink). Single-
letter abbreviations for the amino
acid residues are as follows: A,
Ala; C, Cys; D, Asp; E, Glu; F, Phe;
G, Gly; H, His; I, Ile; K, Lys; L,
Leu; M, Met; N, Asn; P, Pro; Q,
Gln; R, Arg; S, Ser; T, Thr; V, Val;
W, Trp; and Y, Tyr.
Cai et al., Science 381, 569–576 (2023)
4 August 2023
I279
L282
F286
4 of 8
RES EARCH | R E S E A R C H A R T I C L E
an ~20° helical packing angle and the involve-
ment of the same face of gcTMD in packing
against IL-7RTMD or IL-9RTMD (IL-7/9RTMD)
(Fig. 3B). Closer inspection of the two struc-
tures identified I274 of gcTMD as the key
residue that filled a hydrophobic hole of IL-
7/9RTMD comprising four residues (Fig. 3, C
and D), reminiscent of a knob-into-hole mech-
anism. These four residues had a sequence
arrangement of F1-xx-F2F3-xx-F4, where Fi
represents hydrophobic residues such as iso-
leucine, leucine, valine, and phenylalanine that
constitute the pocket. Other gcTMD residues,
I266, P267, T270, and L277 also made close
van der Waals (VDW) contacts with receptor
TMDs but did not appear to be locked into a
hole (Fig. 4, A and B). For the receptor TMD,
S252, L255, L259, and V262 of IL-7RTMD and
I279, L282, L283, and F286 of IL-9RTMD were
in close VDW contact with gcTMD, but none
of them were wrapped by a defined pocket
(Fig. 4, A and B).
Mutating P267, T270, I274, or L277 of gcTMD
to tyrosine disrupted heterodimerization with
IL-4RTMD, IL-7RTMD, and IL-9RTMD (Fig.
4C and fig. S9, A to H). A similar pattern was
observed for IL-21RTMD (fig. S9I). Mutating
S252 or L255 of IL-7RTMD to tyrosine also abol-
ished heterodimerization (Fig. 4D and fig. S9J).
Thus, these NMR structures are consistent with
those expressed in the cell membrane and re-
veal that intramembrane recognition by the gc
of different family members is based largely on
the same knob-into-hole mechanism (Fig. 4E).
This relatively small binding area—501.4 Å2
between gcTMD and IL-7RTMD or 479.0 Å2 be-
tween gcTMD and IL-9RTMD—was consistent
with the small number of residues involved in
the knob-into-hole interaction. Confinement
of interaction to a small area could explain
why IL-7RTMD and IL-9RTMD could both be
recognized by gcTMD despite only having 35%
identity. This interaction is reminiscent of the
small interaction interface between the gc
ectodomain and cytokines (7), which suggests
that degenerate recognition by the gc is pos-
sible because of a small but structurally con-
sistent recognition surface pattern.
Essential role of the knob-into-hole interaction
in the membrane for IL-7R signaling
To determine whether the specific knob-into-
hole mechanism of TMD heterodimerization
is relevant to receptor signaling, we performed
IL-7R
IL-9R
IL-4R
A
B
c TMD
IL-7R TMD
c TMD
IL-7R TMD
C
T270
G272
L273
S252
L255
P267
I274
L277
Y281
S249
V253
V257
c TMD
IL-9R TMD
c TMD
IL-9R TMD
c
WT
P267Y
T270Y
G272Y
L273Y
I274Y
L277Y
T270
G272
L273
E
P267
I274
L277
Y281
I279
L283
L282
F286
D
c
WT
IL-7R S249Y
S252Y
V253G
V253Y
L255Y
V257Y
IL-xR TMD
c TMD
IL-xR TMD
c TMD
Fig. 4. Validation of the surface complementarity and the common mecha-
nism of gcTMD sharing by mutagenesis. (A) The contours of the gcTMD–IL-
7RTMD heterodimer interface shown with the IL-7RTMD strand surface rendered
(left) and the gcTMD surface rendered (right). The gcTMD and IL-7RTMD residues in
sphere are residues tested by mutagenesis in (C), with red and white indicating
dimer disruption and no effect, respectively. (B) Same as in (A) for the gcTMD–IL-
9RTMD heterodimer. (C) BACTH analysis of the effect of gcTMD mutations on
gcTMD association with IL-7RTMD, IL-9RTMD, or IL-4RTMD. Corresponding mutations
for human sequences are shown in fig. S9. Results are from one representative of
three independent experiments with three replicates per experiment. (D) BACTH
analysis of the effect of IL-7RTMD mutations on association with gcTMD. Results are
from one representative of three independent experiments with three replicates per
experiment. (E) Schematic illustration of the knob-into-hole mechanism of recognition
that mediates receptor sharing within the membrane.
Cai et al., Science 381, 569–576 (2023)
4 August 2023
5 of 8
RES EARCH | R E S E A R C H A R T I C L E
A
C
E
Bright field Nucleus
hIL-7R
h c
Merge
c
G272
(G271 in h c)
I274
(I273 in h c)
IL-7R
V253
L255
B
WT hIL-7R
+ WT h c
- hIL-7
WT hIL-7R
+ WT h c
+ hIL-7
D
Bright field
Nucleus
hIL-7R
h c
Merge
Bright field Nucleus
hIL-7R
h c
Merge
WT hIL-7R
+ G271Y h c
- hIL-7
WT hIL-7R
+ G271Y h c
+ hIL-7
WT hIL-7R
+ I273Y h c
- hIL-7
WT hIL-7R
+ I273Y h c
+ hIL-7
V253Y hIL-7R
+ WT h c
- hIL-7
V253Y hIL-7R
+ WT h c
+ hIL-7
L255Y hIL-7R
+ WT h c
- hIL-7
L255Y hIL-7R
+ WT h c
+ hIL-7
F
WT hIL-7R
+ WT h c
L255Y hIL-7R
+ WT h c
V253Y hIL-7R
+ WT h c
WT hIL-7R
+ I273Y h c
WT hIL-7R
+ G271Y h c
hIL-7
-
+
-
+
-
+
-
+
-
+
pSTAT5
STAT5
Actin
5
T
A
T
S
/
5
T
A
T
S
p
d
e
z
i
l
a
m
r
o
N
1.0
0.5
0.0
ns
ns
PBS
hIL7
WT hIL-7R
+ WT h c
L255Y hIL-7R
+ WT h c
V253Y hIL-7R
+ WT h c
WT hIL-7R
+ I273Y h c
WT hIL-7R
+ G271Y h c
Fig. 5. Specific TMD heterodimerization is required for ligand-induced
IL-7R signaling. (A) Cylinder representation of the gcTMD–IL-7RTMD
heterodimer structure showing the positions of the TMD residues of gc and
IL-7R (sphere) that were tested for ligand-induced receptor signaling.
Specifically, I274 of gcTMD and L255 of IL-7R are important for the knob-
into-hole mechanism, whereas G272 of gcTMD and V253 of IL-7R are not
involved in heterodimerization. (B to D) Coexpression and distribution
of WT hIL-7R (green) and hgc (magenta), of WT hIL-7R (green) and hgc
mutants (magenta), and of WT hgc (magenta) and hIL-7R mutants (green)
on the surface of BaF3 without and with treatment with hIL-7. The nuclei
of 1 × 106 cells were stained with 4′,6-diamidino-2-phenylindole (DAPI)
(blue). Three independent experiments were performed with two
replicates per experiment. The images represent one of two biological
replicates with similar results. Scale bars, 5 mm. (E) Phosphorylation of
STAT5 (pSTAT5) in BaF3 cells coexpressing WT or mutant hIL-7R with
WT or mutant hgc after treatment with 50 ng/ml of hIL-7. The pSTAT5
signals were detected by immunoblotting (top) and compared with
immunoblot signals of STAT5 (middle). (F) Quantification of pSTAT5
signals in (E) as pSTAT5/STAT5 intensity ratios, normalized relative to
that of the WT receptors. The data are shown as means ± SEMs calculated
from all technical replicates from three independent experiments
(with one to two technical replicates per experiment). Statistical
significance by Student’s t test. ns, not significant; ****P ≤ 0.0001.
PBS, phosphate-buffered saline.
Cai et al., Science 381, 569–576 (2023)
4 August 2023
6 of 8
RES EARCH | R E S E A R C H A R T I C L E
structure-guided mutagenesis and tested these
mutants using functional assays similar to
those used for receptor activation by TEV pro-
tease (Fig. 1, C and D). On the basis of the
structure of gcTMD in complex with IL-7RTMD,
the knob residue of gcTMD was I274 in mice
and I273 in humans, and a key constituent of
the hole in IL-7RTMD is L255 (Fig. 5A). These
residues were mutated to tyrosine to disrupt
the knob-into-hole interaction. In addition, the
glycine on the opposite side of the gcTMD
(G271 in humans) and V253 on the opposite
side of the IL-7RTMD were mutated to tyro-
sine as no-effect mutations.
Full-length hgc and hIL-7R as well as their
single mutants were expressed in BaF3 cells.
The surface expressions of the mutants were
all within ~28% of that of the WT (fig. S10) and
exhibited similar levels of hIL-7 binding (fig.
S11). Sparse puncta were observed before lig-
and addition for all variants, which suggests
that the introduction of the TMD mutations
did not affect preligand association (Fig. 5, B
to D). Many more puncta appeared after ligand
addition for cells coexpressing WT hIL-7R and
G271Y hgc and for cells coexpressing V253Y
hIL-7R and WT hgc (Fig. 5, C and D), consist-
ent with IL-7–induced receptor clustering. By
contrast, cells expressing either I273Y hgc or
L255Y hIL-7R showed no detectable increase
of puncta after ligand addition, which indicates
that preventing TMD heterodimerization re-
duced ligand-induced receptor clustering.
The imaging results were independently
confirmed by immunoblot analysis of STAT5
phosphorylation. Disrupting either the knob
with the I273Y mutation or the hole with the
L255Y mutation completely abolished signali-
ng, whereas G271Y in hgc or V253Y in hIL-7R
showed no detectable differences in signaling
compared with the WT receptors (Fig. 5, E
and F). Similarly, a mutation of IL-9R that
disrupts TMD heterodimerization (L282Y) also
abolished signaling, whereas a mutation away
from the heterodimer interface (T284Y) showed
no effect on signaling (fig. S12). Thus, the het-
erotypic association of receptor TMDs me-
diated by the knob-into-hole mechanism is
essential for ligand-induced signaling of IL-
7R and IL-9R. Moreover, this mechanism is
likely applicable to other members of the gc
family.
Although the V253Y mutation of IL-7RTMD
did not alter signaling, consistent with the
structure, the gain-of-function mutation V253G
(18) could mediate a homotypic IL-7RTMD in-
teraction that enhanced receptor clustering
and signaling, as glycine is often implicated
in TMH dimerization. In addition to knob-
into-hole interaction mutations, we tested hgc
TMD mutations found in patients with severe
combined immunodeficiency diseases (SCIDs),
including I265N, G268R, M270R, G271E, C278W,
and V279M (32–34) (fig. S13). These results
are consistent with the finding that TMD het-
erodimerization is required for signaling.
The mutations C278W and V279M did not
significantly reduce TMD heterodimerization
between gc and IL-7R– or IL-7–induced signal-
ing (fig. S13, F and H). However, these mutations
may affect gc intramembrane interaction in
the more complicated trimeric complex com-
prising IL-2Ra or IL-15Ra, IL-2Rb, and gc TMDs
(Fig. 2A).
Discussion
The gc TMD is capable of specific recognition
of multiple ILR TMDs that have low sequence
homology (fig. S3). A notable structural ques-
tion is how specificity and promiscuity are
concomitantly achieved. The complementary
surface representations of IL-7/9RTMD and
gcTMD (Fig. 3) show that IL-7RTMD and IL-
9RTMD both have a deep pocket fitting the
I274 knob of gcTMD. In this context, the I274
knob of gcTMD can be perceived as the key
that fits the small hydrophobic hole (or lock)
presented by the different ILR TMDs, which
would explain the promiscuity. The structural
features governing specificity, however, may
go beyond the knob-into-hole module. The
vertical positions in the lipid bilayer of the
two TMDs are important for aligning the I274
knob with the hydrophobic hole for binding.
Other supporting interactions may further en-
hance specificity, such as the interaction of
L277 and Y281 of gcTMD with the comple-
mentary surface of the receptor TMD near its
C-terminal end (Fig. 4, A and B). The three
gcTMD residues with 100% conservation across
species (I274, L277, and Y281) (fig. S3) are pre-
cisely those residues making close VDW con-
tacts with the receptor TMDs in our structures.
Specific TMD pairing required for signaling
of the gc family receptors may inform the
generation of TMDs that can modulate recep-
tor activity. Although the design of transmem-
brane proteins remains challenging, several
studies have demonstrated its feasibility for
TMD oligomerization (35–37). The use of a
designed transmembrane peptide to modu-
late immunoreceptor activity by competing
with the TMD heterodimerization between
the integrin a and b subunits could push the
equilibrium from an inactive to an active
state (38). More recently, designed TMD oligo-
merization has been applied to mediate chi-
meric antigen receptor (CAR) oligomerization to
enhance the therapeutic window of CAR T
immunotherapies (39). The distinct features
of the knob-into-hole interaction that medi-
ate gc sharing and the structural differences
among the ILR members of the family could
be exploited to fashion decoy gcTMDs for the
selective interference of cytokine signaling.
The partial signaling after removal of recep-
tor ectodomains strongly suggests that they
are arranged in an autoinhibitory state in the
absence of ligand. This autoinhibition likely
involves homotypic and heterotypic associa-
tions of the ectodomains of ILRs and gc, which
is consistent with our observation that the co-
expression of IL-7R and gc resulted in their
colocalization on the cell surface (fig. S1, A
and I). The preligand association of ILRs and
gc may also organize the receptors for a biased
response to ligand activation. Previous studies
have shown that preassociation of the gc with
the IL-7R reduces the amount of free gc for
other members of the family, which results in
asymmetrical cross-talk toward stimulation
with different cytokines (22, 23). To date, no
structural information has been made availa-
ble for heterotypic preligand association in
the gc receptor family. A crystal structure of
homotypic association, however, has been re-
ported for the IL-7R ectodomain (21), which
shows a head-to-head antiparallel configura-
tion proposed to be autoinhibitory, as it keeps
the TMD-ICDs apart (19).
It is unclear whether the TMD homotypic
interaction competes or synergizes with the
heterotypic TMD interaction in receptor activa-
tion. Our intermolecular NOE data of IL-7RTMD
in bicelles suggest that TMD self-association is
driven by the packing of hydrophobic residues
(F251, L255, and I258) from the opposite
TMDs (fig. S14, A and B). The L255Y mutation
prevented hIL-7RTMD self-association in the
BACTH assay (fig. S14C), but L255 is an integ-
ral part of the hydrophobic pocket in the
knob-into-hole mechanism (Fig. 3C). Thus, at
least in the case of IL-7RTMD, the homotypic
association is expected to compete with the
heterotypic interaction with the gc TMD. The
functional relevance of this result remains to
be studied.
We have demonstrated that TMDs of the gc
family of cytokine receptors can conserve spec-
ificity without sacrificing promiscuity using a
knob-into-hole recognition pattern that medi-
ates receptor sharing in the membrane and is
important for receptor activation. Our finding
provides a conceptual framework for con-
structing TMHs to modulate gc family recep-
tor activities to potentially address the gain- or
loss-of-function disease mutations associated
with these TMDs.
REFERENCES AND NOTES
1. W. J. Leonard, Nat. Rev. Immunol. 1, 200–208
(2001).
2. Y. Rochman, R. Spolski, W. J. Leonard, Nat. Rev. Immunol. 9,
480–490 (2009).
3. R. Spolski, W. J. Leonard, Annu. Rev. Immunol. 26, 57–79
4.
(2008).
J.-X. Lin, W. J. Leonard, Cold Spring Harb. Perspect. Biol. 10,
a028449 (2018).
5. W. J. Leonard, J. X. Lin, J. J. O’Shea, Immunity 50, 832–850
(2019).
6. X. Wang, M. Rickert, K. C. Garcia, Science 310, 1159–1163
(2005).
7. X. Wang, P. Lupardus, S. L. Laporte, K. C. Garcia, Annu. Rev.
Immunol. 27, 29–60 (2009).
8. S. L. LaPorte et al., Cell 132, 259–272 (2008).
Cai et al., Science 381, 569–576 (2023)
4 August 2023
7 of 8
RES EARCH | R E S E A R C H A R T I C L E
9. K. F. Kubatzky et al., Curr. Biol. 11, 110–115 (2001).
10. E. E. Matthews et al., FASEB J. 25, 2234–2244 (2011).
11. N. Yang, X. Wang, J. Jiang, S. J. Frank, Mol. Endocrinol. 21,
1642–1655 (2007).
29. L. K. Fong, M. J. Chalkley, S. K. Tan, M. Grabe, W. F. DeGrado,
Proc. Natl. Acad. Sci. U.S.A. 118, e2026318118 (2021).
30. A. Battesti, E. Bouveret, Methods 58, 325–334 (2012).
31. S. Damjanovich et al., Proc. Natl. Acad. Sci. U.S.A. 94,
12. E. V. Bocharov et al., Biochim. Biophys. Acta Gen. Subj. 1862,
13134–13139 (1997).
1410–1420 (2018).
13. Q. Fu et al., Mol. Cell 61, 602–613 (2016).
14. L. Pan et al., Cell 176, 1477–1489.e14 (2019).
15. A. Arkhipov et al., Cell 152, 557–569 (2013).
16. N. F. Endres et al., Cell 152, 543–556 (2013).
17. G. Skiniotis, P. J. Lupardus, M. Martick, T. Walz, K. C. Garcia,
Mol. Cell 31, 737–748 (2008).
18. C. Shochat et al., Blood 124, 106–110 (2014).
19. S. T. Walsh, Immunol. Rev. 250, 303–316 (2012).
20. G. Du et al., Cell Res. 33, 131–146 (2023).
21. C. A. McElroy et al., Proc. Natl. Acad. Sci. U.S.A. 109,
2503–2508 (2012).
22. P. Gonnord et al., Sci. Signal. 11, eaal1253 (2018).
23. A. T. Waickman et al., iScience 23, 101421 (2020).
24. K. R. MacKenzie, J. H. Prestegard, D. M. Engelman, Science
276, 131–133 (1997).
25. E. V. Bocharov et al., J. Biol. Chem. 283, 6950–6956 (2008).
26. R. Trenker, M. E. Call, M. J. Call, J. Am. Chem. Soc. 137,
15676–15679 (2015).
27. M. E. Call, K. W. Wucherpfennig, Nat. Rev. Immunol. 7, 841–850
(2007).
28. M. E. Call, K. W. Wucherpfennig, J. J. Chou, Nat. Immunol. 11,
1023–1029 (2010).
32. J. M. Puck et al., Blood 89, 1968–1977 (1997).
33. J. E. Niemela, J. M. Puck, R. E. Fischer, T. A. Fleisher, A. P. Hsu,
Clin. Immunol. 95, 33–38 (2000).
34. C. M. Yao et al., J. Clin. Immunol. 33, 526–539 (2013).
35. I. V. Korendovych, W. F. DeGrado, Q. Rev. Biophys. 53, e3
(2020).
36. P. Lu et al., Science 359, 1042–1046 (2018).
37. P. Barth, A. Senes, Nat. Struct. Mol. Biol. 23, 475–480 (2016).
38. H. Yin et al., Science 315, 1817–1822 (2007).
39. A. Elazar et al., eLife 11, e75660 (2022).
ACKN OWLED GMEN TS
We thank Q. Fu and W. Chen for their help with protein
biochemistry, A. Zoued and R. Ian for sharing materials for the
BACTH assay, and J. Li for their technical support of virus
packaging. We thank C. Xie and S. Blacklow for their insightful
discussions. Funding: This work was supported by NIH grants
GM140887 (to J.J.C.) and AI150709 (to J.J.C. and H.W.). The NMR
data were collected at the MIT-Harvard Center for Magnetic
Resonance (supported by NIH grants P41 GM132079 and S10
OD023513). Author contributions: T.C., R.L.C., H.W., and J.J.C.
conceived the study. T.C. and R.L.C. prepared samples and
performed NMR analyses. T.C. and R.L.C. performed the BACTH
assays. T.C. performed functional cell assays. X.P. assisted with
cell assays. J.J.C., T.C., and R.L.C. wrote the paper, and the other
authors helped with editing the paper. Competing interests:
The authors declare no competing interests. Data and materials
availability: The atomic structure coordinate and structural
constraints have been deposited in the Protein Data Bank (PDB)
with accession numbers 8DDC (gcTMD–IL-7RTMD) and 8DDD
(gcTMD–IL-9RTMD). The chemical shift values have been
deposited in the Biological Magnetic Resonance Data Bank (BMRB)
with accession numbers 31026 (gcTMD–IL-7RTMD) and 31027
(gcTMD–IL-9RTMD). All data are available in the main text or the
supplementary materials. License information: Copyright © 2023
the authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.science.org/about/science-
licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add1219
Materials and Methods
Figs. S1 to S14
Tables S1 and S2
References (40–53)
MDAR Reproducibility Checklist
Submitted 22 May 2022; resubmitted 25 October 2022
Accepted 23 June 2023
10.1126/science.add1219
Cai et al., Science 381, 569–576 (2023)
4 August 2023
8 of 8
|
10.1126_science.add2844
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
BITCOIN
Are cryptocurrencies currencies?
Bitcoin as legal tender in El Salvador
Fernando Alvarez*, David Argente*, Diana Van Patten*
INTRODUCTION: The introduction of digital cur-
rencies is perhaps the most important develop-
ment in monetary economics in the past decade.
However, a currency’s defining role is to serve
as a medium of exchange, and cryptocurrencies
have yet to be widely adopted as such. This
study leverages a unique quasinatural exper-
iment that can shed light on the reasons behind
this lack of adoption. El Salvador became the
first country to make bitcoin legal tender; not
only must bitcoin be accepted as a means of
payment for taxes and debts, but also busi-
nesses are required to accept bitcoin as a
medium of exchange. The government also
launched an app called “Chivo Wallet,” which
allows users to digitally trade both bitcoins
and US dollars (USD, the official currency in
El Salvador) without paying transaction fees,
and provided major adoption incentives such as
a large bonus for downloaders. Moreover, the
pandemic provided an additional incentive
to adopt touchless payments; if bitcoin has a
chance to be used as a medium of exchange,
then this setting gave the cryptocurrency a prime
opportunity. Furthermore, the study of Chivo
Wallet, a digital currency backed by a central
bank, is informative to the debate surround-
ing central bank digital currencies (CBDCs).
RATIONALE: We conducted a nationally repre-
sentative face-to-face survey involving 1800
households in El Salvador and complemented
its results with an analysis using all transactions
identified as involving Chivo Wallet leverag-
ing data from the blockchain. We explored
whether Chivo Wallet and bitcoin were adopted
after the government’s “Big Push,” what fac-
tors deterred adoption by individuals and firms,
and what insights can be obtained from block-
chain data. We also analyzed the broader les-
sons learned from this example.
RESULTS: We found that bitcoin was not widely
used as a medium of exchange and usage of
Chivo Wallet was low. Most downloads took
place just as the app was launched. Since then,
adoption and remittances using Chivo Wallet
have been decreasing over time. These results
suggest that it is unlikely that the usage of
bitcoin and Chivo Wallet will increase. Privacy
and transparency concerns appear to be key
barriers to adoption. We also documented that
this technology involves a large initial adop-
tion cost, has benefits that significantly increase
as more people use it, and faces resistance from
firms in terms of its adoption. These findings
are relevant for countries studying the viability
of CBDCs and of crypto as a currency. Further,
our survey sheds light on how it is the already
wealthy and banked who use crypto, which
stands in stark contrast with recurrent hy-
potheses claiming that the use of crypto may
particularly help the poor and unbanked. An
analysis relying on all blockchain transaction–
level data from Chivo allowed us to validate
and better understand our survey results and
provided new insights on the dynamics of the
use of Chivo Wallet.
CONCLUSION: Despite bitcoin’s legal tender
status and the large incentives to promote
Chivo Wallet in El Salvador, the cryptocurrency
was not adopted at large as a medium of ex-
change, and digital payments were scarce and
concentrated. These findings are informative
about the intrinsic value of cryptocurrencies as
means of payments and about the scope of
CBDCs in developing countries.▪
The list of author affiliations is available in the full article online.
*Corresponding author. Email: f-alvarez1@uchicago.edu (F.A.);
david.argente@yale.edu (D.A.); diana.vanpatten@yale.edu
(D.V.P.)
Cite this article as F. Alvarez et al., Science 382, eadd2844
(2023). DOI: 10.1126/science.add2844
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.add2844
A Timing of adoption: Monthly downloads as a share of total downloads
C Awareness and individual use of Chivo Wallet
Adoption of Chivo
Wallet in El Salvador.
(A) Dynamics of Chivo
Wallet downloads and
(B) regional variations in
adoption across El Salvadoran
regions by shares of
unbanked. (C and D) Sum-
mary of the survey’s results
on adoption by individuals
and firms, respectively.
s
r
e
t
p
o
d
a
f
o
e
r
a
h
S
0.4
0.3
0.2
0.1
0
2021m9
2021m10
2021m11
2021m12
2022m1
2022m2
B Regional variation in adoption by share of unbanked population
Know about Chivo
Try to use Chivo
Able to use Chivo
Use Chivo after $30
Use Chivo after $30 in bitcoin
Remittances via Chivo
Pay taxes with Chivo
Remittances via Chivo in bitcoin
0
0.2
0.4
0.6
0.8
Share of population
D Acceptance and use of bitcoin among firms
Firms accepting bitcoin
Firms with sales in bitcoin
Panama
La Libertad
San Salvador
La Paz
Sonsonate
San Vicente
Santa Ana
Ahuachapán
)
.
p
o
p
f
o
e
r
a
h
s
(
o
v
i
h
C
g
n
d
a
o
n
w
o
d
y
r
T
i
l
0.8
0.7
0.6
0.5
0.4
0.3
Cuscatlán
Cabañas
Chalatenango
Usulután
Total sales in bitcoin
San Miguel
Morazán
CAF
La Unión
Total sales kept in bitcoin
0.5
0.6
0.7
Unbanked
0.8
0.9
0
5
10
15
Percent
20
Alvarez et al., Science 382, 1375 (2023)
22 December 2023
1 of 1
RES EARCH
R E S E A R C H A R T I C L E
◥
BITCOIN
Are cryptocurrencies currencies?
Bitcoin as legal tender in El Salvador
Fernando Alvarez1*, David Argente2*, Diana Van Patten2*
A currency’s essential feature is to be a medium of exchange. This study explores the potential of
cryptocurrencies to be used in daily transactions in El Salvador, the first country to make bitcoin legal
tender. The government’s “big push” introduced “Chivo Wallet,” a digital wallet sharing features with
Central Bank Digital Currencies (CBDCs), with perks to use it for trading bitcoins and US dollars.
Through a nationally representative, face-to-face survey of 1800 households and blockchain data
encompassing all Chivo Wallet transactions, we document a pattern of low and decreasing usage of
digital payments and bitcoin. Privacy and security concerns are key adoption barriers, which speaks
to a policy debate on crypto and CBDCs with anonymity at its core. Additionally, we estimate
Chivo Wallet’s adoption cost and complementarities among adopters.
In its first form money is simply any
commodity … which any person will
readily receive, and which, therefore,
every person desires to have by him in
greater or less quantity, in order that
he may have the means of procuring
necessaries of life at any time.
– William Stanley Jevons
T he introduction of digital currencies in
general, and of cryptocurrencies in par-
ticular, is perhaps the most important
development in monetary economics in
the past decade. Cryptocurrencies such
as bitcoin differ markedly from traditional
banks. Bitcoin relies on cryptography for secu-
rity and operates on a decentralized network
with verifiable transactions, contrasting with
centralized banks governed by regulations. For
the unbanked and those reliant on remittances,
bitcoin presents a potential solution by en-
abling financial transactions, bridging the gap
left by traditional banking systems. However,
a currency’s key and defining role is to serve
as a medium of exchange (1, 2), and crypto-
currencies have yet to be widely adopted for
this purpose (3).
This study leverages a unique quasinatural
experiment that can shed light on the reasons
behind bitcoin’s lack of adoption. On 7 Sep-
tember 2021, El Salvador became the first
country to make bitcoin legal tender through
the “Bitcoin Law.” A legal tender refers to a
form of payment that is recognized by law as
valid for settling financial obligations within a
particular jurisdiction. According to the Bit-
coin Law, not only must bitcoin be accepted as
1Kenneth C. Griffin Department of Economics, University of
Chicago, Chicago, IL 60637, USA. 2School of Management, Yale
University, New Haven, CT 06511, USA.
*Corresponding author. Email: f-alvarez1@uchicago.edu (F.A.);
david.argente@yale.edu (D.A.); diana.vanpatten@yale.edu (D.V.P.)
a means of payment for taxes and outstanding
debts, but also all businesses are required to
accept bitcoin as a medium of exchange for all
transactions (4). The Salvadoran government
also launched an app called “Chivo Wallet,” a
custodial wallet app that allows users to dig-
itally trade both bitcoins and US dollars (USD)
without paying any transaction fees. The gov-
ernment also provided major adoption incen-
tives, such as a large bonus for downloaders
that could potentially solve the coordination
failure, and also subsidized fees. Moreover,
the COVID-19 pandemic provided an addi-
tional incentive to adopt touchless payment
methods. If bitcoin has a chance to be used in
transactions as a medium of exchange, then
this setting gave the cryptocurrency a prime
opportunity.
Furthermore, central banks are considering
alternatives to enter the era of digital payments.
Nine of 10 central banks are exploring central
bank digital currencies (CBDCs), and more than
half are developing them or running concrete
experiments (5). A retail CBDC, a digital cur-
rency backed by a central bank with legal ten-
der status, shares many features with a fast
payment system such as Chivo Wallet. More-
over, because Chivo Wallet allows for pay-
ments both in bitcoins and in USD, an analysis
of its implementation is informative to the
debate surrounding CBDCs, and a comparison
between bitcoin and USD usage within the
app is informative about the use of crypto in
particular.
Situating the current study and
key contributions
Unique monetary episodes can provide valu-
able insights into the workings of the economy
and inform future policy-making. Sargent’s
seminal work on hyperinflations is a prime
example of this research tradition (6). Our
study of the Salvadorean experience follows in
this tradition by studying an unprecedented
monetary experiment in which bitcoin be-
came legal tender and digital currency started
being traded through Chivo Wallet. Our ex-
amination shows that the designation of bitcoin
as legal tender does not imply that it becomes
a general medium of exchange as defined by
previous work, i.e., an object “which is habit-
ually, and without hesitation, taken by anybody
in exchange for any commodity” (7). Important
references in the literature argue that “accept-
ability” makes an object more likely to become a
medium of exchange and can be influenced by
government policies (2, 8, 9), and that the state
can give a currency value by allowing the public
to use it to pay taxes (1, 10–13). We contribute to
this long-standing work by documenting that
accepting a digital currency to pay for taxes
is not a sufficient condition for it to become
widely accepted.
Our work also contributes to the study of
cryptocurrencies. Empirically, the literature
has focused on the risks faced by individuals
(14, 15), arbitrage opportunities and price
manipulation (16, 17), bitcoin network par-
ticipants (18, 19), bitcoin’s price fluctuations
(3, 62), the determinants of asset pricing (20),
and developing the notion that bitcoin seems
to function more like a speculative investment
than a bona fide currency (3). Our results pro-
vide insights on the characteristics of adopters
and the bottlenecks of adoption in a setting
where incentives to adopt are high, fees are
subsidized, and we have measurable variation
in determinants such as income and finan-
cial literacy. This study is also related to the
growing theoretical literature on cryptocur-
rencies, which has built models stressing the
network effects of its adoption (18, 21, 22)
and the cost of its production (23, 24). Com-
plementary to these studies, our work quanti-
fies the fixed costs of adoption along with the
network effects.
Further, through the study of Chivo Wallet
payments in USD, we address the literature on
CBDCs, in which empirical evidence is scarce
(25) (26). As in the case of Chivo Wallet, recent
policy briefs argue that CBDCs should not be
bearer instruments (27). This is the case, for
instance, for China’s CBDC (28), and is also the
case of Chivo Wallet. Moreover, whereas Chivo
Wallet is not backed by a central bank, it is
backed by the government and is not required
to be linked to a bank account, just as would
be the case with a CBDC. Our work highlights
several challenges to the implementation of
CBDCs, such as the role of privacy and trans-
parency concerns, while suggesting there is a
role for policies that incentivize adoption given
the presence of strong complementarities among
adopters. More broadly, our study relates to
work on the adoption of payment methods be-
yond cash (29–32).
Alvarez et al., Science 382, eadd2844 (2023)
22 December 2023
1 of 9
RES EARCH | R E S E A R C H A R T I C L E
Research questions
RQ 1: Were Chivo Wallet and bitcoin actually adopted
after the government’s “big push”?
RQ 2: What factors deterred the adoption of Chivo
Wallet and bitcoin by individuals?
RQ 3: What other broader lessons can be drawn
from this experiment?
RQ 4: What were the drivers of adoption by firms?
RQ 5: What insights can be obtained from block-
chain data?
Methods
The context
El Salvador has been the stage for several mon-
etary experiments. In 2001, the USD became
legal tender and the country’s only official
currency (33). Later, on 7 September 2021,
El Salvador became the first country to make
bitcoin legal tender through the Bitcoin Law.
Although there might be many reasons be-
hind the decision, when the policy was an-
nounced, the president stated that it would
generate jobs, provide financial inclusion, and
facilitate sending remittances (34). In addition,
bitcoin was seen as a mechanism to introduce
people into financial services. For context, only
one-third of the Salvadorian adult population
had a bank account at a financial institution in
2017 (35). Moreover, our survey’s background
questions show that in El Salvador, most trans-
actions are paid in cash; in fact, >50% of people
only use cash, >70% of adults are unbanked, and
almost 90% of them do not use mobile bank-
ing, as reported in fig. S7 (36). We found that
64.6% of Salvadorans have access to a mobile
phone with internet, a prerequisite to adopt
Chivo Wallet (37).
The Bitcoin Law
The first article of the Bitcoin Law describes
its main objective and endows bitcoin with a
legal tender status (38). It also makes bitcoin
a unit of account within the country and, ac-
cording to the theory of chartalism, endows it
with value by accepting it as a means of pay-
ment for tax purposes. The Bitcoin Law also
goes beyond the usual provisions of a legal
tender, making bitcoin a medium of exchange
of mandatory acceptance nationwide. Article
7 reads: “Every economic agent must accept
bitcoin as payment when offered to him by
whoever acquires a good or service.” Another
relevant article of the law is related to how bit-
coin usage will be implemented in the country.
In particular, Article 8 mandates the govern-
ment to provide the means to conduct trans-
actions using bitcoin. How was the adoption
of bitcoin facilitated and promoted by the
state? The government’s answer was “Chivo
Wallet” (39).
The Chivo Wallet app
Just as bitcoin became legal tender, the gov-
ernment launched Chivo Wallet and an edu-
cational campaign on how to use it. This digital
wallet allows users to convert bitcoins into USD
and vice versa without a fee and to send or
receive either currency (40). As shown in fig.
S4, payments are made through the application
by entering the recipient’s identification num-
ber or phone number and the payment amount
(41). The app can also be used to pay at local
establishments, is compatible with other bit-
coin on-chain and Lightning wallets, and con-
nects with El Salvador’s banking system to
deposit or withdraw USD from a bank account
(42). Chivo Wallet can be used by registered
Salvadorans even if they reside abroad to fa-
cilitate sending remittances potentially faster
and at a lower cost than alternative services.
Chivo Wallet also has a version intended for
firms, which allows them to charge their cli-
ents and pay taxes. It does not provide users
with the key to their bitcoin, which makes it
a “custodial” wallet in which transactions are
not anonymous; users are required to enter
their personal information after downloading
the app, just as in the case of several CBDCs
(27, 28).
Adoption incentives
Usage of bitcoin in El Salvador is related to
Chivo Wallet’s adoption, and as an adoption
incentive, citizens who downloaded the app
could receive a $30 bitcoin bonus from the
government, which is a substantial amount
in this Central American country with a GDP
per capita of $4131 (43). These $30 bonuses
were automatically deposited in their wallets;
however, the money could not be withdrawn
as cash before first being transferred to an-
other Chivo Wallet because the bonus was in-
tended to promote bitcoin usage. As another
government incentive, users could get a signif-
icant discount on gasoline if they paid using
Chivo Wallet (44). Moreover, transactions in
bitcoin usually involve substantial fees. For in-
stance, bitcoin ATM fees can range from 5%
to over 20%, with an average of about 8.5%,
and transactions in bitcoin reached a fee of
>$60 USD per transaction in April 2021 and
an average value of $1.8 USD in February 2022.
Transactions in bitcoin and conversions from
bitcoins to USD using Chivo Wallet and cash
withdrawals at Chivo Wallet ATMs do not
incur any fees. This can be interpreted as an
additional government subsidy. In El Salvador,
payments of public salaries and pensions re-
main in USD. Allowing for these payments
to be in bitcoins could have provided another
adoption incentive (45).
Bitcoin in other countries
The lack of access to banking services and
infrastructure increases the potential of digi-
tal payments to promote financial inclusion.
Consistent with this, most of the top 20 coun-
tries in the 2021 Global Crypto Adoption Index
are emerging economies. The Central African
Republic (CAF) was the second country, after
El Salvador, to make bitcoin legal tender in
April 2022, the same month in which Panama
approved its own Crypto Law (46). High-income
countries have not been absent from the crypto
stage. For instance, an Arizona senator proposed
a bill to make bitcoin legal tender in that state
in January 2022 (47).
Measures
In the midst of a growing interest to promote
digital currencies among monetary authori-
ties, El Salvador offers a rare opportunity to
learn about the potential of cryptocurrencies
to become a widely used payment method.
However, access to data poses a challenge be-
cause El Salvador’s government reveals only
selected information (48). To overcome this
challenge, we conducted surveys to generate
data that would be otherwise unobtainable.
This allowed us to measure the adoption of
respondents based on their characteristics,
focusing not only on downloads but also on
usage.
The survey was face-to-face, nationally rep-
resentative, and spanned 1800 households
during February 2022 (49), leading to results
with a 95% confidence interval and a 1.94%
margin of error. Respondents were all >18 years
old, as this is a prerequisite to be eligible to use
Chivo Wallet. The national survey was con-
ducted in partnership with CID-Gallup (50).
Interviewers were trained a week in advance
to conduct the survey, and we implemented a
pilot interviewing 50 people to ensure that
survey questions were clear. Our sample val-
idation can be found in table S2; the sample
almost exactly matches total population shares
in terms of gender, age, districts, and educa-
tion levels.
The sample is also representative in terms
of bank account ownership (51). The national-
scale and face-to-face nature of the survey
posed a challenge compared with an internet
or phone survey. However, both features are
important in our setting. First, understanding
adoption patterns requires a sample that in-
cludes small cities and rural areas; focusing
on main population centers may bias results.
Second, because bitcoin’s adoption through
Chivo Wallet requires access to both a cell
phone and an internet connection, a survey by
phone or internet, which relies on respon-
dents having access to either communication
method would mechanically underestimate
adoption costs. Finally, the face-to-face for-
mat of our survey preserves data quality while
allowing us to conduct a longer survey with
more detailed questions than would be fea-
sible through the phone or internet (52). The
survey measures sociodemographic varia-
bles, knowledge about Chivo Wallet, down-
loads of the app, and usage both in bitcoins
Alvarez et al., Science 382, eadd2844 (2023)
22 December 2023
2 of 9
RES EARCH | R E S E A R C H A R T I C L E
and USD. We include details on the specific
questions in the supplementary materials,
section D.
We complement our survey results with an
analysis using all transactions identified as
involving Chivo Wallet, leveraging data from
the blockchain, a distributed public ledger. We
not only studied overall volumes transacted
through Chivo Wallet, but separately analyzed
the patterns of deposits and withdrawals, and
identified consistencies between the survey
outcomes and the blockchain results.
Results
RQ 1: Were Chivo Wallet and bitcoin actually
adopted after the government’s “big push”?
Awareness
We found that 68% of potential users knew
about the app’s existence. Most of those who
were aware of the app learned about it through
social media, followed by television and radio,
news, and friends and family, as summarized
by fig. S8. Almost 78% of those who were aware
of the app had tried to download it. Most
downloads happened just as Chivo Wallet was
launched. Figure 1A shows that 40% of all
downloads occurred in September 2021 and
there were virtually no downloads in 2022. The
latter suggests that our survey was already
capturing the most relevant share of adopters
of this digital wallet.
In terms of heterogeneity, we found that
banked, educated, and young men were more
likely to know about Chivo Wallet (table S3),
as were people who owned a cell phone with
internet. Moreover, conditional on knowing
about Chivo Wallet, these characteristics also
make a person more likely to try to adopt it, as
documented in Table 1 (53). People with these
demographics also tended to download the app
on their own without help (table S4). These
findings suggest that the introduction of Chivo
Wallet mainly provided an additional means
of payment among those already banked in-
stead of stimulating more financial inclusion
among the unbanked.
Not all users agreed with the widespread
use of Chivo Wallet. Individuals who agreed
tended to own a mobile phone with internet,
and were younger and male. Columns 1 to 3 in
table S5 show that people who agreed with the
use of Chivo Wallet were 0.3 percentage points
more likely to download the app, and columns
4 to 6 show that individuals who were less
likely to agree also tended to be those who
needed help installing the app.
Reasons to download Chivo Wallet
The key incentive for downloading the app was
the $30 bonus, which is equivalent to 0.7% of
annual income per capita. Other reasons deemed
as the most important were the contactless
nature of the payment method in the midst
of the pandemic and the potential to receive
remittances; fig. S11 summarizes all reasons
regarded as most important.
Chivo Wallet usage by households
Most respondents spent their $30 bonus to
pay for expenses in bitcoins, and almost 20%
of those who downloaded the app had not yet
used their bonus (54). However, most users
did not keep using Chivo Wallet after spend-
ing their bonus. Table 2 presents descriptive
statistics on Chivo Wallet’s usage among those
who downloaded it and who reported using
the app after spending the bonus. A salient fea-
ture of people who downloaded Chivo Wallet
and kept using it after spending their bonus
is that they were more likely to be young, edu-
cated, male, banked, and much more likely (26%)
to be using other digital wallets in addition to
Chivo Wallet to conduct transfers (55). Dis-
tance to a Chivo Wallet ATM and facing issues
with the app, however, were not good predic-
tors of whether the user remained active, sug-
gesting that these were not the binding barriers
to sustain usage (56).
More than half of these “active users” had not
made a cash withdrawal from a Chivo Wallet
ATM, although the mean number of withdraw-
als was 2.59, given the presence of extreme
values in the right tail (57). The number of
payments and transfers received or sent was
also largely driven by very active users in the
right tail. Deposits in USD is the only statistic
in which users in the 25th percentile have a
nonzero value. We can conclude that active
Chivo Wallet users transact in USD more than
bitcoins, because the average amount of pay-
ments and transfers, sent or received, was
consistently larger in USD.
Regional variation
Figure 1B shows important regional variation
in the probability of downloading Chivo Wallet
depending on the share of unbanked popula-
tion in each department. It also benchmarks the
CAF, the second country to make bitcoin legal
tender, and Panama, which enacted a crypto
law in April 2022, with respect to departments
in El Salvador given our estimates and their
share of unbanked. Figures S18 and S19 also
show regional differences in adoption and
Fig. 1. Chivo Wallet’s adoption. (A) Timing of adoption: Monthly downloads. (B) Regional variation in adoption as a share of total downloads by share of unbanked
population. (A) shows the month in which each user in our sample downloaded Chivo Wallet as a share of total downloads. (B) shows the relationship between
the share of people who have tried to use Chivo Wallet and the fraction of people who do not have access to a bank account in El Salvador, by department. (B) also
includes, for comparison, the shares of unbanked in Panama and the CAF.
Alvarez et al., Science 382, eadd2844 (2023)
22 December 2023
3 of 9
RES EARCH | R E S E A R C H A R T I C L E
Table 1. Adoption of Chivo Wallet. The dependent variable was “have you tried to download
Chivo Wallet?”
(1)
(2)
(3)
(4)
0.1085***
(0.036)
0.0757**
Cell phone with internet
.....................................................................................................................................................................................................................
(0.035)
.....................................................................................................................................................................................................................
–0.0815***
Unbanked
.....................................................................................................................................................................................................................
(0.026)
.....................................................................................................................................................................................................................
Years of schooling
.....................................................................................................................................................................................................................
0.0676**
Middle school
.....................................................................................................................................................................................................................
(0.024)
.....................................................................................................................................................................................................................
0.0832**
High school+
.....................................................................................................................................................................................................................
(0.036)
0.0849***
(0.023)
0.1168***
(0.029)
.....................................................................................................................................................................................................................
–0.1132***
(0.023)
.....................................................................................................................................................................................................................
Age 35 to 44
.....................................................................................................................................................................................................................
Age.....................................................................................................................................................................................................................
–0.0241*
Age 25 to 34
.....................................................................................................................................................................................................................
(0.013)
–0.0473
(0.032)
.....................................................................................................................................................................................................................
–0.0888*
Age 45 to 54
.....................................................................................................................................................................................................................
(0.041)
.....................................................................................................................................................................................................................
–0.1238***
Age 55+
.....................................................................................................................................................................................................................
(0.028)
–0.0236
(0.014)
–0.0480
(0.032)
–0.0969*
(0.045)
–0.1349***
(0.029)
.....................................................................................................................................................................................................................
Gender
.....................................................................................................................................................................................................................
–0.0089
Female
.....................................................................................................................................................................................................................
(0.020)
.....................................................................................................................................................................................................................
–0.0528**
Single
.....................................................................................................................................................................................................................
(0.023)
1224
0.055
Yes
.....................................................................................................................................................................................................................
Observations
.....................................................................................................................................................................................................................
R2
.....................................................................................................................................................................................................................
Department
.....................................................................................................................................................................................................................
The sample only includes respondents who knew about the existence of Chivo Wallet. Results in this table rely on a linear
probability model. Results are robust to other specifications, in particular, columns (1) and (3) of table S12 show the marginal
effects under a logit model. The regression includes department fixed effects, and each of the controls is obtained from
survey questions on whether a person owns a financial instrument (unbanked), years of schooling, age, gender, and marital
status. Standard errors are clustered by department.
–0.0292
(0.021)
–0.0567**
(0.023)
1224
0.041
Yes
1224
0.023
Yes
1224
0.019
Yes
awareness about Chivo Wallet depending both
on the average income and the share of un-
banked per “department,” with departments
being similar to counties. Regions with higher
levels of development tended to be more active
using Chivo Wallet. The share of users who
continued using the application after spend-
ing the $30 USD bonus in departments such
as San Salvador and La Libertad, which have
the highest income per capita in the country,
was twice as large as in departments with
low income per capita, such as Usulután and
Chalatenango (58). Similarly, departments
with a larger share of unbanked population
had as little as half the adoption levels as de-
partments in which most of the population
had access to banking services.
Along similar lines, assuming that the im-
plementation of a digital wallet is similar in
other contexts, our estimates allow us to ex-
plore how other features of adoption would
manifest in other countries, which could prove
valuable to policy makers. Given our estimates,
in the CAF, only 37 to 45% of the population
would have been aware of the app’s existence,
8 to 14% would continue using the app given
similar adoption incentives as in El Salvador,
and <2% would use the app for remittances. In
the case of Panama, income per capita is higher
than in El Salvador, as is access to banking
services. We estimate that >95% of the adult
population in Panama would be aware of the
technology, between 30 and 56% would con-
tinue using it after spending the adoption
incentives, and 10 to 30% would use it for re-
mittances. The last two estimates are cut in
half when considering payments in bitcoins in
either country.
Role of taxes and remittances
By law, bitcoin can be used to pay taxes.
Chartalism implies that endowing a currency
with the power to pay taxes gives it value as
a means of exchange. However, only 5% of
Salvadorans have paid taxes using Chivo Wallet.
Moreover, in El Salvador, some households
receive >60% of their income from remit-
tances, as summarized in fig. S15. Chivo Wallet
is not widely used to receive remittances from
abroad; only 3% (8%) of people have received
remittances in bitcoins (USD) using the app.
This finding aligns with reports from the
Central Reserve Bank of El Salvador, which
found that only 1.45% of remittances were
received through digital wallets in March
2022, and provides external validation to our
survey (59).
RQ 2: What factors deterred the adoption of
Chivo Wallet and bitcoin by individuals?
Chivo Wallet adoption deterrents
More than 21% of respondents knew about
Chivo Wallet but did not try to download it.
The reasons not to download it are summar-
ized in fig. S13A. The most important reason
was that users preferred to use cash. The sec-
ond most relevant reason not to download
Chivo Wallet were trust issues: Respondents
did not trust the system or bitcoin itself (60).
Privacy and security are at the heart of the
debate around CBDCs and bitcoin. Concerns
regarding lack of anonymity and secure trans-
actions could then explain, at least partially,
the main two reasons not to download the
app, because cash is an anonymous payment
method (61). The next most frequent reason
mentioned was not owning a phone with in-
ternet, followed by the technology being com-
plicated. In sixth place, Salvadorans mentioned
technical difficulties using the app; fig. S14
summarizes the main reported problems.
Bitcoin adoption deterrents
Figure S13B reports the main reasons why in-
dividuals do not use bitcoin. For >50% of re-
spondents, the main reason not to use bitcoin
was that they did not understand it nor trust
it. Although the volatility of bitcoin has poten-
tial as a deterrent (62), trust and transparency
seem to be more salient than uncertainty, be-
cause bitcoin’s volatility was mentioned by <10%
of respondents. If volatility were the main de-
terrent from using Chivo Wallet, then we
should then see people downloading the app
and transacting in USD, which are very stable;
however, this was not the case, as explained in
the previous paragraph.
Taking stock
Figure 2A summarizes results from our first
two research questions. We documented that
more than two-thirds of Salvadorans were
aware that Chivo Wallet exists. However, not
all of those who knew about the app have tried
to download it; just over half of all respond-
ents did so. The main reason not to download
Chivo Wallet was that individuals prefer to pay
in cash, followed by mistrust; these motifs
may be related to privacy concerns. The main
reason that they downloaded the app was to
use the $30 bonus offered by the government,
but less than half of those who were able to
download Chivo Wallet, 20% of adult citizens,
continued to use it after spending the bonus
and they mostly used it to transact in USD, not
in bitcoins.
Alvarez et al., Science 382, eadd2844 (2023)
22 December 2023
4 of 9
RES EARCH | R E S E A R C H A R T I C L E
Table 2. Descriptive statistics: Active Chivo Wallet users.
(1) Mean
(2) SD
(3) 10th
(4) 25th
(5) Median
(6) 75th
(7) 90th
ATM withdrawals
............................................................................................................................................................................................................................................................................................................................................
Average amount of ATM withdrawals (in USD)
............................................................................................................................................................................................................................................................................................................................................
Payments/transfers sent in bitcoins
............................................................................................................................................................................................................................................................................................................................................
Payments/transfers sent in USD
............................................................................................................................................................................................................................................................................................................................................
Average amount of payments/transfers sent in bitcoin (in USD)
............................................................................................................................................................................................................................................................................................................................................
Average amount of payments/transfers sent in dollars (in USD)
............................................................................................................................................................................................................................................................................................................................................
Payments/transfers received in bitcoins
............................................................................................................................................................................................................................................................................................................................................
Payments/transfers received in USD
............................................................................................................................................................................................................................................................................................................................................
Average amount of payments/transfers received in bitcoin (in USD)
............................................................................................................................................................................................................................................................................................................................................
Average amount of payments/transfers received in dollars (in USD)
............................................................................................................................................................................................................................................................................................................................................
Deposits in bitcoins
............................................................................................................................................................................................................................................................................................................................................
Deposits in USD
............................................................................................................................................................................................................................................................................................................................................
The table shows distribution of responses to the questions: (i) How many times per month do you withdraw money from Chivo Wallet ATMs?; (ii) What is the average amount of your ATM
withdrawals?; (iii) How many payments or transfers do you perform per month using Chivo Wallet in bitcoins or in USD?; (iv) What is the average amount of your payments or transfers in bitcoins
or in USD?; (v) How many payments or transfers did you receive per month using Chivo Wallet in bitcoins or in USD?; (vi) What is the average amount of your payments or transfers you received
in bitcoins or in USD?; and (vii) How many times have you deposited money to your Chivo Wallet in bitcoins or in USD? We divided the number of deposits by the months a person was active in
Chivo Wallet to convert them to a monthly variable and rounded the values to the closest integer. The sample includes those who kept using Chivo Wallet after spending their $30 bonus (20.6% of
respondents). We dropped observations above the 99th percentile to avoid extreme outliers.
8.7
65.6
7.8
24.8
38.2
47.1
7
18
77
78.9
3.9
13.8
2.5
54.9
2.3
9.2
32.5
39.6
2.1
6.2
51.3
55.3
1.31
4.4
2
60
2
5
42.5
50
1
2
55
70
1
2
4
120
5
20
80
100
4
15
100
120
2.5
10
0
30
0
1
20
20
0
0
25
30
0
1
0
20
0
0
10
12
0
0
10
15
0
0
0
10
0
0
3
7
0
0
2
5
0
0
Fig. 2. Taking stock. (A) Awareness and individual use. (B) Acceptance and use of bitcoin with Chivo Wallet among firms. (A) shows shares with respect to the entire
sample, so it is subject to a 1.94% margin of error. In (B), the top two bars show percentages with respect to all surveyed owners and employees who knew about
payment methods at the firm. The bottom two bars show percentages with respect to total sales.
Moreover, most individuals who used Chivo
Wallet after spending the bonus did not engage
with the app intensively; the median user re-
ported no ATM withdrawals and no payments,
sent or received, in bitcoin in a given month. To
put this in perspective, the median number of
daily transactions per person across means of
payments was between 1.3 and 1.4 in several
countries (63), and Chivo Wallet’s developer
indicates that there are 0.001 to 0.003 daily
transactions per adult (64). Further, we did
not find evidence of Chivo Wallet being used
to pay for taxes or to send remittances at a
substantial scale. Figure S20 replicates Fig. 2A
for the share of the banked and unbanked
population, respectively.
Overall, we have documented that bitcoin is
not being widely used as a medium of ex-
change and that Chivo Wallet’s usage is low in
El Salvador. The latter stands despite the “big
push” exerted by the government, which in-
volved endowing bitcoin with legal tender status
through the Bitcoin Law, the $30 bonus, gas
discounts, and no fees, and despite the pan-
demic’s incentive to use touchless payment
methods.
RQ 3: What other broader lessons can be drawn
from this experiment?
Complementarities
For some technologies, the benefit of adopt-
ing increases as more people adopt (65). Ar-
guably, such complementarities, also called
network externalities, are an inherent feature
of digital payment methods and give a po-
tential role for policy to improve allocations
(66). Thus, we can draw broad lessons appli-
cable to other payment technologies from the
analysis of Chivo Wallet. We found evidence
of complementarities, both in the decision
to adopt the app and on how intensively peo-
ple used it, as reported in the supplementary
materials, section C.
Adoption and variable costs
We leveraged the familiarity with the $30 in-
troductory bonus and asked two questions
to estimate the distribution of (self-reported)
Alvarez et al., Science 382, eadd2844 (2023)
22 December 2023
5 of 9
RES EARCH | R E S E A R C H A R T I C L E
adoption costs. The first question was: “How
large does the bonus need to be to convince
you to download Chivo Wallet?,” which was
directed to people who had not downloaded
the app, but knew about it (14.5% of respon-
dents). The second question was: “What is the
minimum bonus that would have convinced
you to download Chivo Wallet?,” which was
directed to people who had downloaded the
app (53.5% of respondents). Table S7 displays
our results. Although the mean reported val-
ue was $30, the median user would have ac-
cepted $20 USD, and there were people in
the 10th percentile who would have adopted
it even without a bonus. The adoption cost was
larger for individuals with certain demograph-
ics: Unbanked respondents reported $6.9 USD
higher cost than those who were banked, peo-
ple without a cellphone with internet reported
a $8.6 USD higher cost than those with one,
it was $2.9 USD costlier for households with
only elementary education to adopt compared
with those with education beyond elemen-
tary, and finally, women reported a $8.9 USD
higher cost than men.
Chivo Wallet allows users to withdraw cash
from Chivo Wallet ATMs and convert bitcoin
into USD without a fee. However, outside of
Chivo Wallet, most providers charge significant
fees. Table S8 shows the maximum reported
willingness to pay to withdraw $100 USD at a
Chivo Wallet ATM was $3.3 USD on average.
This amount is less than half of the mean fee
to purchase cash at bitcoin ATMs outside of
El Salvador. Moreover, the median respondent
was willing to pay only $1 USD. These findings
suggest that Chivo Wallet users would not en-
gage in cash withdrawals if they faced non-
subsidized fees. Table S8 also reports that the
average willingness to pay to convert bitcoins
into USD was $2.9 USD, and the median user
would be willing to pay only $0.05 USD. These
amounts are much smaller than any transac-
tion cost of exchanges, indicating that Chivo
Wallet would not be used in the absence of the
subsidies.
Impact on usage of other payment methods
If users adopt Chivo Wallet, then they might
substitute it for other payment methods such
as cash and cards. We found some evidence
consistent with this substitution. We docu-
mented that 10% of users who downloaded
Chivo Wallet decreased their use of cash and
11% reduced their use of debit cards (67). The
government also offered a discount of ~8% per
gallon for gas purchases with Chivo Wallet,
which allowed us to measure the elasticity of
substitution, which measures how easily people
switch between Chivo Wallet and other pay-
ment methods, as detailed in the supplementary
materials, section C. Although the sample size is
small, the estimated elasticities of substitution
are positive and large, which suggests that the
welfare costs of policies disincentivizing other
payment methods (such as cash) are lower if
digital payments are available.
RQ 4: What were the drivers of adoption
by firms?
The Bitcoin Law states that all economic agents
must accept bitcoin, but this does not neces-
sarily translate into all firms effectively doing
so (68). To study the extent to which firms ac-
cepted bitcoin, we relied on a subset of re-
spondents who identified themselves as owners
of firms or as employees who knew about the
payment methods accepted by their employer,
who then answered a series of questions about
their business. Results are summarized in Fig. 2B.
First, we documented that whereas almost
all firms accepted cash, slightly over 20% ac-
cepted bitcoin (69). Among those that did
accept bitcoin, 75% started accepting it just
as the law was enacted. Only 11.4% of firms
had positive sales in bitcoin. This estimate
aligns with results from two independent sur-
veys targeting firms of all sizes and across
sectors (70). Further, our survey indicates that
81% of firms accepting bitcoin have not seen
a change in their sales since starting to accept
it, and whereas the median firm made no sales
in bitcoin, 4.9% of all sales were paid in bitcoin
through Chivo Wallet, mainly to large firms.
These estimates align with those by two in-
dependent local surveys (71).
Second, we documented that firms accept-
ing bitcoin were mostly large and in the fifth
quintile of the firm size distribution (72). These
large firms were also more likely to accept cards.
Third, most firms reporting sales in bitcoin
converted them into USD: 71% converted sales
into USD and then withdrew them as cash,
17% converted sales into USD and kept them
in Chivo Wallet, and only 12% of firms stored
their sales in bitcoin within Chivo Wallet. Fi-
nally, we found that 11% of firms have increased
prices since bitcoin became legal tender, which
is consistent with the hypothesis that firms
might be transferring costs related to the crypto-
currency (e.g., volatility) to customers (73).
RQ 5: What insights can be obtained from
blockchain data?
So far, our conclusions discussed have been
drawn from the survey data that we collected.
This section leverages that all bitcoin transac-
tions are recorded on the blockchain, a dis-
tributed public ledger, to analyze Chivo Wallet’s
activity based on actual transaction data. This
exercise allowed us to validate and better un-
derstand our survey results. The analysis using
transaction-level data for all of Chivo Wallet’s
transactions on the blockchain also provides
new insights into how bitcoin transactions are
carried out in El Salvador and by whom. Data
sources are detailed in the supplementary ma-
terials, section E.
It is important to understand which Chivo
Wallet transactions would surface in the block-
chain and which ones would not. As of today,
verifying a bitcoin transaction on the block-
chain is both costly and takes several hours
(74). Given this constraint, many wallets that
use bitcoin for relatively small payments do
not verify all transactions on the blockchain.
Instead, they are custodial wallets and rely
on a clearing house. Chivo Wallet is no excep-
tion; therefore, transactions from one Chivo
Wallet to another one, in general, would not
register on the blockchain. Transactions be-
tween different addresses owned by Chivo
Wallet as an entity do appear on the blockchain,
and we label them as internal transactions (75).
Transactions from Chivo Wallet to external
crypto wallets also surface in the public ledger.
These would include, for example, payments
from tourists visiting El Salvador and paying
in bitcoin from their foreign wallets.
According to our data, as of 3 November 2022,
Chivo Wallet was associated with 142,148 ad-
dresses, which were involved in 425,514 trans-
actions and a total of 3,424 bitcoins deposited
into Chivo Wallet. These are all the transac-
tions that can be identified as involving Chivo
Wallet either as a buyer or a seller of bitcoin.
Figure 3 summarizes some of the observed
dynamics. As shown in Fig. 3A, the total trans-
actions in bitcoins, expressed in USD, reached
their peak between October and December
2021 and decreased significantly thereafter.
The latter is consistent with the results of our
survey, which document high activity within
the first months of Chivo Wallet’s operation
and a sharp decrease thereafter.
Figure 3A shows all activity, whereas Fig. 3B
considers only external transactions and de-
composes them as total deposits into and
withdrawals from Chivo Wallet (76). First, the
co-movement between both types of external
transactions was substantial. Second, an anal-
ysis of the average size of each type of transac-
tion, reported in fig. S24, shows that deposits
were composed by many small and relatively
frequent transactions; for example, these could
be transactions from tourists visiting El Salvador
to use bitcoin or residents from El Salvador
who had bitcoin in other wallets (77). Their
active behavior resembles the one by the right
tail of Chivo Wallet users who were extremely
active, as documented in table S6. The magni-
tude of inflows of bitcoin in the survey and on
the blockchain data also align. According to
our survey, between $221,000 and $334,000
USD flowed into Chivo Wallet per day, where-
as according to blockchain data, this amount
was ~$245,000 USD per day (78). Third, a joint
analysis of Fig. 3B and fig. S24 shows that with-
drawals (i.e., sales of bitcoin by Chivo Wallet)
tended to be large and happen rarely, and in
synchrony, with the pace of accumulated de-
posits. This pattern suggests that withdrawals
Alvarez et al., Science 382, eadd2844 (2023)
22 December 2023
6 of 9
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Chivo Wallet’s blockchain transactions. (A) Total transactions, both internal and external. (B) External deposits and withdrawals (in USD). (A) shows the
total number of transactions in Chivo Wallet, including internal transactions and external withdrawals and deposits in USD. We converted bitcoin’s value into USD
because otherwise the patterns would also reflect bitcoin’s price changes, which were substantial in this period. (B) shows the dynamics of external withdrawals and
deposits. The vertical dashed lines date moments when El Salvador’s government announced a bitcoin purchase.
occured as part of Chivo Wallet’s bitcoin in-
ventory management, such that Chivo Wallet
accumulated balances of bitcoin to lower the
transaction cost of selling them. This behavior
is consistent with the almost zero net accu-
mulation of bitcoin within the wallet shown in
fig. S25. This behavior resembles the one dis-
played by firms in El Salvador, which tended
to convert all of the bitcoin they received into
USD almost immediately. Finally, the data
verify that trading volumes in Chivo Wallet are
uncorrelated with bitcoin prices; thus, Chivo
Wallet trading volumes seem to be driven by
idiosyncratic reasons rather than by bitcoin
market prices; section E of the supplemen-
tary materials provides more details on this
relationship.
Discussion
Following the tradition of studying unique
monetary episodes to inform policymaking, our
analysis of the Salvadorean experience with
bitcoin as legal tender offers valuable insights
into the complexities of the adoption of crypto-
currencies as a medium of exchange and the
implementation of CBDCs. El Salvador’s govern-
ment provided a “big push” to incentivize the
use of digital payments and bitcoin, including
a large sign-up bonus and subsidized fees.
Bitcoin is not only endowed with legal tender
status, allowing the currency to be used to pay
taxes and debts, but also must be accepted by
any economic agent by law. Monetary theories
such as chartalism suggest that these condi-
tions should be sufficient for bitcoin to be-
come a medium of exchange.
However, our results show that, despite all
incentives and the enhanced attractiveness of
contactless payments in the midst of the pan-
demic, bitcoin is being not widely used as a
medium of exchange and usage of Chivo Wallet
is low. Most downloads took place just as Chivo
Wallet was launched. Since then, adoption and
levels of remittances through Chivo Wallet have
been decreasing over time. These results sug-
gest that it is unlikely that usage of bitcoin and
Chivo Wallet will increase. Our empirical results
challenge the implications of chartalism.
Privacy and transparency concerns appear
to be key barriers to adoption; unexpectedly,
these are the two concerns that decentralized
currencies such as crypto aim to address. More-
over, we document that this payment technol-
ogy involves a large initial adoption cost, has
benefits that significantly increase as more
people use it (i.e., complementarities), and faces
resistance from firms in terms of its adoption.
Our findings lay out the challenges faced by
digital payments and cryptocurrencies to be-
come widely accepted, and are relevant for
countries studying the viability of CBDCs and
of crypto as a currency. Moreover, our survey
work using a representative sample sheds light
on how it is the already wealthy and banked
who use crypto, which stands in stark contrast
with recurrent hypotheses claiming that the
use of crypto may help the poor and unbanked
the most.
There is substantial heterogeneity across dem-
ographic groups in the likelihood of adopting
and using bitcoin as a means of payment. The
reasons that young, educated men are more
likely to use bitcoin for transactions remain
an open question. One hypothesis is that this
group has higher financial literacy. We found
that, even conditional on access to financial
services and education, young men were still
more likely to use bitcoin. However, financial
literacy encompasses several other areas of
knowledge that are not captured by these con-
trols. An alternative hypothesis is that young,
educated men have a higher propensity to
adopt new technologies in general. The litera-
ture on payment methods has documented
that young individuals have a greater propen-
sity to adopt means of payment beyond cash,
such as cards (87). Nevertheless, further re-
search is necessary to causally identify the
factors contributing to the observed hetero-
geneity across demographic groups. An anal-
ysis relying on all blockchain transaction-level
data from Chivo Wallet allowed us to validate
and better understand our survey results and
is a unique opportunity to provide new in-
sights on the dynamics of Chivo Wallet’s ac-
tivity. The latter is valuable because the app
is a unique exchange in that it can also be used
as means of payment by law.
Furthermore, the results carry policy impli-
cations for other countries. A study of this
experience is informative in drawing broader
lessons on the likelihood of success of CBDCs
and cryptocurrencies in contexts outside of El
Salvador. Assuming that the implementation
of a digital wallet is similar in other contexts,
our estimates allow us to explore what would
be the adoption of the technology in other
countries, which can prove valuable to policy
makers. Two interesting cases are the CAF,
which recently made bitcoin legal tender and
has a stable local currency, as well as Panama,
which also enacted a Crypto Law and where
the USD is the official currency, as in El
Salvador. El Salvador falls in between these
Alvarez et al., Science 382, eadd2844 (2023)
22 December 2023
7 of 9
RES EARCH | R E S E A R C H A R T I C L E
countries in terms of both income per capita
and access to banking services (79). The in-
troduction of a cryptocurrency could lead to
outcomes different from the ones we docu-
mented in countries where the local currency
is unstable and there are restrictions on capi-
tal mobility, such as Argentina and Turkey.
Thus, an analysis of these contexts may offer
fertile ground for future research to explore.
Overall, we conclude that despite bitcoin’s
legal tender status and the large incentives to
promote Chivo Wallet, the cryptocurrency is
not adopted at large by the population as a
medium of exchange and digital payments are
scarce and concentrated. These findings are
informative about the intrinsic value of crypto-
currencies as means of payments, as viewed
in the larger context of monetary models in
economics and about the scope of CBDCs in
developing countries.
RE FE RENCES AND N OT ES
1. W. Jevons, Money and the Mechanism of Exchange, The
International Scientifie Series (D. Appleton, 1875).
2. N. Kiyotaki, R. Wright, Acceptability, means of payment, and
media of exchange. Fed. Reserve Bank Minneap. Q. Rev. 16, 18
(1992). doi: 10.21034/qr.1632
3. D. Yermack, Handbook of Digital Currency (Elsevier, 2015),
pp. 31–43. doi: 10.1016/B978-0-12-802117-0.00002-3
4. Although a means of payment is an object used to make
purchases and settle debts, the concept of medium of
exchange is broader; it is an object that is taken in exchange
for something to then be exchanged for something else,
without the “prerequisite” of an existing debt (8).
5. A. Kosse, I. Mattei, “Gaining momentum: Results of the 2021
BIS survey on central bank digital currencies” (Bank for
International Settlements, 2022); https://www.bis.org/publ/
bppdf/bispap125.htm.
6. T. J. Sargent, in: Inflation: Causes and Effects, R. E. Hall, Ed.
(Univ. of Chicago Press, 1982), pp. 41–98.
7. K. Wicksell, “Money and credit,” in Lectures on Political
Economy, K. Wicksell, Ed. (Routledge, 1906).
8. N. Kiyotaki, R. Wright, On money as a medium of exchange.
J. Polit. Econ. 97, 927–954 (1989). doi: 10.1086/261634
9. S. R. Aiyagari, N. Wallace, Government transaction policy, the
medium of exchange, and welfare. J. Econ. Theory 74, 1–18
(1997). doi: 10.1006/jeth.1996.2250
10. A. Smith, An Inquiry into the Nature and Causes of the Wealth
of Nations (McMaster University Archive for the History of
Economic Thought, 1776).
11. E. Cannan, P. H. Wicksteed, The common sense of political
economy, including a study of the human basis of economic
law. Econ. J. (Lond.) 20, 394 (1910). doi: 10.2307/2221031
12. R. M. Starr, The price of money in a pure exchange monetary
economy with taxation. Econometrica 42, 45 (1974).
doi: 10.2307/1913684
13. For instance, Adam Smith describes that “A prince who should
enact that a certain proportion of his taxes should be paid in
paper money of a certain kind might thereby give a certain
value to this paper money” (10).
14. D. Duffie, “Digital currencies and fast payment systems:
Disruption is coming,” paper presented at the Asian Monetary
Policy Forum, 31 May 2019, Sentosa, Singapore.
15. N. Borri, Conditional tail-risk in cryptocurrency markets.
16.
J. Empir. Finance 50, 1–19 (2019). doi: 10.1016/
j.jempfin.2018.11.002
I. Makarov, A. Schoar, Trading and arbitrage in cryptocurrency
markets. J. Financ. Econ. 135, 293–319 (2020). doi: 10.1016/
j.jfineco.2019.07.001
17. J. M. Griffin, A. Shams, Is bitcoin really untethered? J. Finance
75, 1913–1964 (2020). doi: 10.1111/jofi.12903
18. L. W. Cong, Y. Li, N. Wang, Token-based platform finance.
J. Financ. Econ. 144, 972–991 (2022). doi: 10.1016/
j.jfineco.2021.10.002
I. Makarov, A. Schoar, “Blockchain analysis of the bitcoin
market” (Tech. Rep., National Bureau of Economic Research,
2021); http://dx.doi.org/10.2139/ssrn.3942181.
19.
20. Y. Liu, A. Tsyvinski, Risks and returns of cryptocurrency. Rev.
Financ. Stud. 34, 2689–2727 (2020). doi: 10.1093/rfs/hhaa113
21. E. Pagnotta, A. Buraschi, “An equilibrium valuation of bitcoin and
decentralized network assets” (Tech. Rep., Social Science
Research Network Electronic Journal, 2018); http://dx.doi.org/
10.2139/ssrn.3142022.
22. B. Biais, A. Menkveld, C. Casamatta, C. Bisi’ere, M. Bouvard,
“Equilibrium bitcoin pricing” (Tech. Rep., Social Science
Research Network Electronic Journal, 2019); http://dx.doi.org/
10.2139/ssrn.3261063.
23. L. W. Cong, Z. He, J. Li, Decentralized mining in centralized
pools. Rev. Financ. Stud. 34, 1191–1235 (2021).
doi: 10.1093/rfs/hhaa040
24. M. Sockin, W. Xiong, “A model of cryptocurrencies” (Working
Paper 26816, National Bureau of Economic Research, 2020);
https://www.nber.org/papers/w26816.
25. R. Auer et al., Central bank digital currencies: Motives,
economic implications, and the research frontier. Annu. Rev.
Econ. 14, 697–721 (2022). doi: 10.1146/annurev-economics-
051420-020324
26. Detailed literature reviews on CBDCs can be found in (80)
and (25).
27. D. Duffie, K. Mathieson, D. Pilav, “Central bank digital currency:
Principles for technical implementation” (Tech. Rep., Social
Science Research Network Electronic Journal, 2021);
http://dx.doi.org/10.2139/ssrn.3837669.
28. D. Duffie, “Can China conquer crypto?” Foreign Affairs,
22 April 2022.
29. T. Suri, Mobile money. Annu. Rev. Econ. 9, 497–520 (2017).
doi: 10.1146/annurev-economics-063016-103638
30. R. Borzekowski, K. K. Elizabeth, A. Shaista, Consumers’
use of debit cards: Patterns, preferences, and price response.
J. Money Credit Bank. 40, 149–172 (2008). doi: 10.1111/
j.1538-4616.2008.00107.x
41. Users can withdraw USD from their wallet either by doing a
transfer from their bank account or by withdrawing cash from
a Chivo Wallet ATM without a fee. As of September 2021, there
were 200 Chivo Wallet ATMs in El Salvador (see figs. S5 and
S6), and 51 in the US. Similarly, users can load money into
their wallets through an official website using a credit or debit
card or with cash through Chivo Wallet ATMs. Although funds
remain in Chivo Wallet, they represent a claim to either USD
or bitcoins, which is not uncommon in payment platforms.
In other words, both USD and bitcoins are a parallel digital
asset with a fixed exchange rate. In Chivo Wallet, the price of
bitcoin is adjusted in real time to its market price. For instance,
a customer could pay a firm or another user the USD price
of an item in bitcoins, and the app would use the real-time
exchange rate to charge her.
42. The Lightning Network is a protocol that uses temporary
payment channelsoperating off-chain. After a channel is
closed, payments are validated on the blockchain.
43. World Bank, “GDP per capita (current US$)” (2020);
https://data.worldbank.org/indicator/NY.GDP.PCAP.CD.
44. Major gas stations dropped the gallon price by $0.20 for
customers who paid with Chivo Wallet between September
and October, and another drop of $0.30 per gallon was
announced in November.
45. For instance, during the 2001–2002 Argentinean crisis, several
provinces introduced low- denomination bonds (“quasi
money”) and used them to pay wages and other inputs (86). It
is worth noting that El Salvador is relatively small and is
therefore a bitcoin price-taker; indeed, fig. S3 shows there
were no large changes in the (global) price of bitcoin after the
Bitcoin Law or Chivo Wallet’s launching. Thus, the experiment
speaks about whether bitcoin is used as a means of payment
given the above-described incentives despite the fact that it
has a given resale value.
31. B. Yang, A. T. Ching, Dynamics of consumer adoption of
46. Panama and CAF are benchmarked against regions in El
financial innovation: The case of ATM cards. Manage. Sci. 60,
903–922 (2014). doi: 10.1287/mnsc.2013.1792
32. The app that we studied differs in important aspects from
other mobile payment technologies. First, it was launched and
sponsored by the central government and allows for payments
both in a cryptocurrency and in the local currency (USD);
thus, it shares features with CBDCs. Second, the app was
launched nationwide along with generous incentives to adopt
and no fees, which allows us to provide statistics on the
distribution of adoption costs while isolating the fees’ impact.
Our work also relates to recent work studying the degree of
substitutability between payment methods (81–84). We
quantified the degree of substitutability between mobile
payments and other payment methods and found it to be
larger than the substitutability between cash and cards.
33. The former currency is no longer circulated; therefore, prices,
accounts, and transactions were converted into USD (85).
34. For instance, in terms of job generation, as a way to encourage
investments from bitcoin entrepreneurs, the government
offered permanent residency to anyone who spends three
bitcoin in the country and explained that, since bitcoin is legal
tender, foreigners would not have to pay capital gains tax
in El Salvador on profits made if bitcoin’s value goes up.
Moreover, remittances make up 22% of El Salvador’s GDP, and
bitcoin could potentially be a channel to send these
remittances while paying lower fees.
35. Consejo Nacional de Inclusión y Educación Financiera, “Politica
nacional de inclusion financiera para El Salvador (PNIF-SLV)”
(BCR, 2021); https://www.bcr.gob.sv/bcrsite/uploaded/
content/category/387473516.pdf.
36. More details on financial inclusion in El Salvador are provided
in table S1.
37. We collected data on access to a cell phone with internet
ourselves, because information on cell phone and internet
access was only available for each one separately in household
surveys. Fig. S9 and Fig. 10 provide details on these measures
separately and other demographics relying on survey data.
38. Article 1 reads: “The purpose of this law is to regulate bitcoin
as unrestricted legal tender with liberating power, unlimited
in any transaction, and to any title that public or private
natural or legal persons require carrying out.”
39. In El Salvador, “chivo” is a slang term meaning “cool.”
40. El Salvador established a trust fund, which is known to have a
limit of $150 million, to allow for the automatic conversion
of bitcoin into USD without fees. Official details on the trust
fund or Salvadoran bitcoin purchases have not been disclosed.
Hitherto, the only sources of information have been the
president’s Twitter posts, which indicate that the country had
acquired approximately ,800 bitcoin as of April 2022.
Salvador in Fig. 1B.
47. Bill SB 1341 was introduced by state Sen. Wendy Rogers.
48. Most information comes from the president’s Twitter account.
We tried, unsuccessfully, to contact multiple government
entities, including Chivo Wallet customer service, El Salvador’s
Superintendency of the Financial System, Central Bank, and
Casa Presidencial, to receive more quantitative information.
49. In terms of the timing of the survey, information was collected
across several weeks always including weekends; weekends
were important in reaching a representative sample of profiles.
50. CID-Gallup has been conducting surveys in Latin America for
>40 years. It has an office in El Salvador that periodically
conducts large-scale surveys.
51. Total population shares match the General Directorate of
Statistics and Censuses’ 2021 projections.
52. Approximate survey length was 27 min. To obtain candid
responses, respondents were guaranteed confidentiality and
notified that the survey aimed to inform academic research.
53. Table 1 relies on a linear probability model. Results are robust
to other specifications, in particular, columns (1) and (3) of
table S12 show the marginal effects under a logit model.
54. According to Chivo Wallet’s regulations, users must spend
their bonus in bitcoins to incentivize its usage. Some people
found ways to circumvent this restriction; for instance, sending
the bonus to a family member and asking her to withdraw
the money from a Chivo Wallet ATM.
55. These findings regarding the prominence of young adoption is
consistent with (87).
56. Table S6 shows no evidence of technical issues with the app
being a concern by constructing a dummy equal to one if the
user faced problems using the app.
57. Figs. S5 and S6 show Chivo Wallet ATM locations. Fig. S12
displays mean distances to a Chivo Wallet ATM across
population shares.
58. In general, extending income adoption relations requires
caution, as countries with higher income, such as Panama,
may have higher adoption of digital payments (e.g., card or
mobile) and, thus, lower incentives to adopt a Chivo Wallet
type of service (88). However, the adoption of digital payments
in Panama was similar to that in El Salvador; in Panama,
13.3% of people over 15 years of age report having borrowed
from a financial institution or used a credit card, whereas
11.5% is the corresponding percentage in El Salvador.
Moreover, in both countries, 6.5% of people over 15 years of
age have made a payment using their mobile phone or
internet according to the World Bank’s G20 Financial
Inclusion Indicators.
59. Fig. S2 reports official monthly data on remittances in bitcoins,
and fig. S17 summarizes our results.
Alvarez et al., Science 382, eadd2844 (2023)
22 December 2023
8 of 9
RES EARCH | R E S E A R C H A R T I C L E
60. Mistrust is also the main reason not to agree with the use of
Chivo Wallet (fig. S16).
61. Note that, in the US, apps to trade bitcoin are required to
gather information on the identity of the trader, so bitcoin is
not associated with anonymity in the US, just as in
El Salvador’s case.
62. D. G. Baur, K. Hong, A. D. Lee, Bitcoin: Medium of exchange or
speculative assets? J. Int. Financ. Mark. Inst. Money 54,
177–189 (2018). doi: 10.1016/j.intfin.2017.12.004
63. J. Bagnall et al., “Consumer cash usage: A cross-country
comparison with payment diary survey data” (Tech. Rep.,
Social Science Research Network Electronic Journal, 2014);
http://dx.doi.org/10.2139/ssrn.2436365.
64. Source: “Chivo Wallet registra un promedio de 6,000 trans-
acciones por d´ıa, segu´n experto argentino,” Diario El Mundo,
November 2021. Estimate obtained based on an adult
population of 4.3 million.
65. F. Alvarez, D. Argente, F. Lippi, E. M´endez, D. Van Patten,
“Strategic complementarities in a dynamic model of
technology adoption: P2P digital payments” (Working Paper
31280, National Bureau of Economic Research, 2023);
https://doi.org/10.3386/w31280.
66. The diffusion of many technologies is also shaped by learning;
this mechanism, however, does not necessarily create an
externality or room for policy interventions to improve outcomes.
67. More details are reported in fig. S21.
68. Businesses that refuse to accept bitcoin are operating in
violation of local regulations and exposed to sanctions
under the Consumer Protection Law; our survey points to
enforcement on firm adoption being imperfect.
69. The share that accepts cards is only a little over 25%.
Even among firms that accept bitcoin, prices were quoted in USD
and the Chivo Wallet app provided real-time bitcoin equivalents.
70. First, a survey ran by the Salvadoran Foundation for
Economic and Social Development (FU- SADES) toward the
end of 2021 indicates that 10% of businesses have made
sales in bitcoin (“Institutional Position N.106,” FUSADES,
December 2021). Second, the Chamber of Commerce and
Industry of El Salvador (Camarasal) conducted a survey in
February 2022 reporting that 13.9% of businesses have
made sales in bitcoin (“First Business Survey 2022,” Camarasal,
March 2022).
71. The Chamber of Commerce and Industry of El Salvador reports
a similar estimate of firms that have not changed their sales,
and (91.7%) the Salvadoran Foundation for Economic and
Social Development estimates that the share of sales paid in
bitcoin is between 1 and 5%.
72. Table S9 shows results robust to controlling for the sector of
the firm. Findings are very similar if only including responses
from the firm’s owner or from an employee who reports
to work in sales.
73. Fig. S22 shows (i) a summary of the results on prices from
the consumer’s perspective (21% have encountered higher
prices at some businesses) and (ii) the full distribution
of shares of sales in bitcoin across firms. Fig. S23 summarizes
findings on firms.
74. Although it can be verified faster, this extra speed incurs an
additional cost.
75. Although one entity can own several addresses, these are not
transactions between Chivo Wallets owned by individuals.
76. Thus, this figure considers transactions that involve an address
that can be identified as Chivo Wallet and another address.
77. The fees paid for these deposits tended to be higher closer to
Chivo Wallet’s launch (see fig. S26), which would be consistent
with more urgency from bitcoiners trying to pay for goods
and services when Chivo Wallet’s hype was at its peak.
Throughout the period, fees for deposits into Chivo Wallet
tended to be higher than those paid for withdrawals, which
also points to more urgency on the deposits’ front compared
with withdrawals. The data indicate that Chivo Wallet mostly
transacted with well-known exchanges; the main one being
Binance (12% of all the volume transacted), followed by Bitso,
OKX, and Coinbase.
78. Flows from the blockchain data have a standard deviation of
184,300. To calculate these flows using our survey, we focused
on inflows of bitcoin into Chivo Wallet from other wallets,
because these are the transactions recorded on the block-
chain. Thus, our population of interest consists of individuals
who have deposited bitcoin into Chivo Wallet and have
transferred bitcoins to wallets other than Chivo Wallet, ~2% of
the adult population of El Salvador. For this sample, we computed
total deposits per day as the difference between the total amount
sent per day and the total amount received per day in the app,
including transactions in both USD and bitcoins, because
convertibility across currencies is free within the app. To estimate
the total deposits in bitcoins per day, we multiplied total
deposits times the share of deposits in bitcoins (17.3%).
79. The CAF has an income per capita of ~$418 USD and Panama
of approximately $12,172 USD, and as in El Salvador, the
alternative to bitcoin is a stable currency. Approximately 13.7%
of the population in the CAF has access to a bank account,
whereas in Panama this number is ~46.5% (Fig. 1B).
80. F. Carapella, J. Flemming, “Central bank digital currency:
A literature review” (Tech. Rep., Social Science Research Network
Electronic Journal, 2020); https://doi.org/10.17016/2380-7172.2790.
81. A. Deviatov, N. Wallace, Optimal inflation in a model of inside
money. Rev. Econ. Dyn. 17, 287–293 (2014). doi: 10.1016/
j.red.2013.06.003
82. F. Alvarez, F. Lippi, Cash burns: An inventory model with a
cash-credit choice. J. Monet. Econ. 90, 99–112 (2017).
doi: 10.1016/j.jmoneco.2017.07.001
83. F. Alvarez, D. Argente, On the effects of the availability of
means of payments: The case of Uber. Q. J. Econ. 137,
1737–1789 (2022). doi: 10.1093/qje/qjac008
84. F. Alvarez, D. Argente, “Consumer surplus of alternative
payment methods: Paying Uber with cash” (Working Paper
28133, National Bureau of Economic Research, 2022);
http://dx.doi.org/10.2139/ssrn.3462480.
85. A. J. Swiston, “Official dollarization as a monetary regime: Its
effects on El Salvador” (Working Paper 2011/129, International
Monetary Fund, 2008); https://ssrn.com/abstract=1864432.
86. G. B. Gorton, E. W. Tallman, Fighting Financial Crises: Learning
from the Past (Univ. of Chicago Press, 2018). doi: 10.7208/
chicago/9780226479651.001.0001
87. M. Brown, N. Hentschel, H. Mettler, H. Stix, The convenience of
electronic payments and consumer cash demand: Causal
evidence from the staggered introduction of contactless debit
cards” (Research Paper 2020/02, University of St. Gallen
School of Finance, 2022); http://dx.doi.org/10.2139/
ssrn.3582388.
88. Z. Wang, P. Han, “Technology adoption and leapfrogging:
Racing for mobile payments” (Working Paper 21-5, Federal
Reserve Bank of Richmond, 2021); http://dx.doi.org/10.21144/
wp21-05.
89. S. Meiklejohn et al., “A fistful of bitcoins: Characterizing
payments among men with no names,” in Proceedings
of the 2013 Conference on Internet Measurement Conference
(IMC ’13), Barcelona, Spain, 23 to 25 October 2013
(Association for Computing Machinery, 2013), pp. 127–140;
https://doi.org/10.1145/2504730.2504747.
90. D. Ermilov, M. Panov, Y. Yanovich, “Automatic bitcoin address
clustering,” in 2017 16th IEEE International Conference on
Machine Learning and Applications (ICMLA), Cancún,
Mexico, 18 to 21 December 2017 (IEEE, 2017), pp. 461–466;
https://doi.org/10.1109/ICMLA.2017.0-118.
91. S. Foley, J. R. Karlsen, T. J. Putnins, Sex, drugs, and bitcoin:
How much illegal activity is financed through
cryptocurrencies? Rev. Financ. Stud. 32, 1798–1853 (2019).
doi: 10.1093/rfs/hhz015
92. Data and code for: F. Alvarez, D. Argente, D. Van Patten, Are
cryptocurrencies currencies? Bitcoin as legal tender in
El Salvador, Dryad (2023); https://doi.org/10.5061/
dryad.z8w9ghxjm.
AC KNOWLED GME NTS
We thank M. Bassetto, M. Brown, D. Duffie, M. Golosov, K. Huynh,
C. Monnet, A. Neumeyer, D. Niepelt, A. Schoar, H. Uhlig, N. Wallace,
and Z. Wang for helpful comments and discussions. Funding:
The authors used funding from their own universities, which at the
time the project was conducted were: University of Chicago (F.A.),
The Pennsylvania State University (D.A.), and Yale University
(D.V.P.). Author contributions: Conceptualization: F.A., D.A., D.V.P.;
Funding acquisition: F.A., D.A., D.V.P.; Investigation: F.A., D.A.,
D.V.P.; Methodology: F.A., D.A., D.V.P.; Writing – original draft:
F.A., D.A., D.V.P.; Writing – review and editing: F.A., D.A., D.V.P.
Competing interests: The authors declare no competing interests.
Data and materials availability: Data and code for this study
have been deposited at Dryad (92). License information:
Copyright © 2023 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science.
No claim to original US government works. https://www.science.
org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add2844
Supplementary Text
Figs. S1 to S27
Tables S1 to S13
References (93–96)
Submitted 1 June 2022; resubmitted 23 April 2023
Accepted 7 November 2023
10.1126/science.add2844
Alvarez et al., Science 382, eadd2844 (2023)
22 December 2023
9 of 9
|
10.1126_science.add4445
|
tables S2 and S3)—two with low activity rates
and three with high activity.
Slow descent to ~110 m (duration
21.27 ± 22.7 min)
Beginning their dives at depths varying from
the surface to 60 m, the sharks descended to
~110 m with a gradually decreasing pitch angle
(i.e., pointing increasingly downward) from
0° to −15° with an average activity intensity [as
measured by overall dynamic body acceleration
(23)] of 0.7 ± 0.2 m/s2, an average tailbeat fre-
quency of 0.4 ± 0.1 Hz, and an average tailbeat
sway amplitude of 0.3 ± 0.2 m/s2.
Fast descent to bottom (duration 6.45 ± 1.82 min)
At ~110 m, the sharks abruptly oriented at a
steep pitch angle between −70° and −80° and
increased swimming intensity (with an aver-
age overall dynamic body acceleration of 1.3 ±
0.4 m/s2). At depths between 300 and 500 m
and while still at a steep dive angle, they con-
ducted sporadic bursts of intense swimming
activity lasting between 5 and 20 s, with over-
all dynamic body acceleration increasing by
more than an order of magnitude (with peaks
averaging 22.2 ± 12.0 m/s2 and peaking as high
as 49.0 m/s2). The frequency of these burst
events increased as the sharks approached the
bottom of their dives.
Bottom time (duration 4.3 ± 3.02 min)
The sharks reached an average maximum depth
of 635 ± 90 m with a range of 418 to 825 m. The
bottom portions of the dives exhibited V-, U-,
vU-, Uv-, and W-shaped depth profiles that are
similar to those observed in deep-diving ma-
rine mammals such as elephant seals and other
fish species that bounce dive, such as sicklefin
devil rays (Mobula tarapacana) (11, 24, 25).
Tagged sharks would either hold a steady depth
within a 30-m range (V- and U-shaped dives)
or move between depths varying by >100 m
(vU-, Uv-, and W-shaped dives). Intense swim-
ming activity occurred either in frequent bursts
or was sustained throughout the duration of the
bottom time (average overall dynamic body
RES EARCH
THERMAL REGULATION
“Breath holding” as a thermoregulation
strategy in the deep-diving scalloped
hammerhead shark
Mark Royer1*, Carl Meyer1, John Royer2, Kelsey Maloney1, Edward Cardona1, Chloé Blandino1,
Guilherme Fernandes da Silva3,4, Kate Whittingham5, Kim N. Holland1
Fish moving between different thermal environments experience heat exchange via conduction
through the body wall and convection from blood flow across the gills. We report a strategy of
preventing convective heat loss at the gills during excursions into deep, cold water by the tropical
scalloped hammerhead shark (Sphryna lewini). Adult scalloped hammerhead sharks dive rapidly and
repeatedly from warm (~26°C) surface waters to depths exceeding 800 meters with temperatures
as low as 5°C. Biologgers attached to adult sharks show that warm muscle temperatures were
maintained throughout the deepest portion of each dive. Substantive cooling only occurred during
the latter stages of the ascent phase and, once initiated, was rapid. Heat transfer coefficient
modeling indicated that convective heat transfer was suspended, probably by suppressing gill
function during deep dives. This previously unobserved strategy has broad similarities to marine
mammal “breath hold” diving.
experienced during deep dives, we equipped
adult individuals with instrument packages
that measured depth, ambient water tempera-
ture, activity rates, and body orientation using
triaxial accelerometers and measured core
muscle temperature with a thermistor probe
inserted ~8 cm into the dorsal musculature
near the dorsal fin (table S1). Our specific ob-
jectives were to (i) characterize swimming per-
formance during repetitive deep dives and (ii)
measure core body temperature during dives
to determine whether core body temperature
is a result of simple thermal inertia or active
thermoregulation [distinct from simply select-
ing optimal thermal environments to achieve
optimal body temperature (20–22)].
Deep-diving behavior
We defined deep dives as those exceeding 400 m
with starting and ending depths <100 m. All
deep dives (n = 106) were conducted at night
(Fig. 1 and figs. S2 to S4) and consisted of five
distinctive phases (Fig. 2, figs. S5 and S6, and
A s fish move from warm to cold water,
metabolic heat generated in muscle tis-
sue is carried away by the blood and is
rapidly lost to the environment at the
gills (1, 2). Regionally endothermic fishes
(those that warm part but not all of their
body), such as lamnid sharks and tunas (fam-
ilies Lamnidae and Scombridae), retain heat by
using vascular countercurrent heat exchangers
to keep specific organs or tissues warm during
deep dives into cold water (3–7). Although large
body sizes can passively buffer the rate of change
of body temperature, most fishes, including
the scalloped hammerhead shark (Sphyrna
lewini), are ectotherms that lack morpholog-
ical and vascular adaptations to actively con-
serve heat (fig. S1).
Scalloped hammerhead sharks occupy warm
surface waters in tropical and warm temperate
oceans but make repeated nocturnal dives to
depths exceeding 800 m, where temperatures
are as low as 5°C (8–11). These dives provide
access to the slow-moving mesopelagic and
bathypelagic prey that dominate their diets
(12–18). Such deep diving is risky because body
cooling can reduce visual acuity, cardiac func-
tion, and swimming muscle power, which could
be detrimental for obligate ram ventilators—
fish that rely on their forward movement to
force water across their gills for respiration—
such as scalloped hammerhead sharks (4, 19).
To determine how scalloped hammerhead
sharks function in frigid water temperatures
1Hawai‘i Institute of Marine Biology, University of Hawai‘i at
Mānoa, Kāne‘ohe, HI 96744, USA. 2School of Physics and
Astronomy, The University of Edinburgh, Edinburgh EH9 3FD, UK.
3Department of Marine Sciences, Federal University of São Paulo,
11070-102 Santos, Brazil. 4Ocean and Resources Engineering,
University of Hawai‘i at Mānoa, Honolulu, HI 96822, USA. 5Biology
Department, Whitman College, Walla Walla, WA 99362, USA.
*Corresponding author. Email: royerm@hawaii.edu
Table 1. Results from representative whole-body heat coefficient model 4 for scalloped
hammerheads HH7, HH8, and HH9. Parameter estimates include ±95% confidence intervals. k1,
whole-body heat transfer coefficient during normal swimming; k2, whole-body heat transfer coefficient
during the high-activity phase of a deep dive; T
(metabolic) heat production during normal swimming; T
of internal (metabolic) heat production during the high-activity phase of a deep dive.
o1, rate of temperature change because of internal
o2, rate of temperature change because
(cid:1)
(cid:1)
Shark
k1 normal
k2 diving
T(cid:1)
o1 normal
T(cid:1)
o2 diving
0.0522 ± 0.0424
HH7
.....................................................................................................................................................................................................................
0.1057 ± 0.0437
HH8
.....................................................................................................................................................................................................................
0.0427 ± 0.0090
HH9
.....................................................................................................................................................................................................................
0.0050 ± 0.0031
0.0158 ± 0.0060
0.0020 ± 0.0006
0.0040 ± 0.0024
0.0339 ± 0.0022
0.0052 ± 0.0003
0.0175 ± 0.0026
0.1129 ± 0.0046
0.0311 ± 0.0007
Royer et al., Science 380, 651–655 (2023)
12 May 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
acceleration values of 1.9 ± 0.5 m/s2 with bursts
to 17.0 m/s2). Tailbeat frequency during this
stage of the dive was difficult to distinguish
because of the erratic nature of the movement,
as indicated by rapid fluctuations in pitch and
roll angles.
A
Fast ascent to inflection point (duration 6.45 ±
1.9 min)
Ascents began with a steep, abrupt increase in
pitch angle to 70° to 80° and sudden increases
in swimming intensity (tailbeat frequency =
1.2 ± 0.6 Hz; tailbeat sway amplitude up to
4.5 m/s2, averaging 2.0 ± 0.6 m/s2; and over-
all dynamic body acceleration averaging 2.5 ±
0.3 m/s2, with average peaks at 10.5 ± 6.6 m/s2
and reaching as high as 41.1 m/s2). Sharks sus-
tained this intense swimming effort and steep
pitch angle until they reached depths of be-
tween 350 and 250 m (average 290 ± 19 m).
At this point, there was a consistent inflection
in the profile characterized by an abrupt de-
crease in ascent rate associated with a change
to a shallow pitch angle and reduction in swim-
ming intensity (indicated by drops in overall
dynamic body acceleration, tailbeat frequency,
and tailbeat sway amplitude).
Slow ascent to upper 50 m (18.12 ± 3.55 min)
After the inflection point, the sharks continued
their ascent at a slower rate to roughly 50-m
depth. Swimming activity was greatly reduced
compared with the fast ascent phase (overall
dynamic body acceleration averaging 1.2 m/s2,
tailbeat frequency 0.5 Hz, and tailbeat sway
amplitude 0.4 ± 0.2 m/s2). During interdive
intervals, sharks stayed within the top 50 m
of the water column and resumed their char-
acteristic nighttime side-swimming behavior
of maintaining a rolled angle of ~60° during
steady swimming (26) (fig. S5). Interdive in-
tervals lasted an average of 43 ± 28 min and
ranged from 18 min to >3 hours.
Overall, dives lasted an average of 56.0 ±
22.0 min (from the start of the slow descent
to the end of the slow ascent). Bottom times
(end of the fast descent to the start of the fast
ascent) lasted an average of 4.0 ± 3.0 min. The
average high-activity phase of the dives (from
the start of the fast descent to the end of the
fast ascent at the inflection point) lasted an
average of 17.0 ± 3.5 min (table S3). Between
deep dives, sharks spent most (98%) of their time
between 175 m and the surface, with ambient
water temperatures averaging 25.6°C (±1.6°C)
and shark intramuscular temperatures averag-
ing 26°C (±0.4°C) with a narrow variation (mini-
mum 21.3°C, maximum 28.2°C) (table S4).
Body temperature profiles
At the bottom of deep dives, sharks experienced
ambient temperatures of 5° to 11°C [mean 6.8°C
(±1.2°C)]. Thus, during the fast descent phase
of a dive, they experienced an ~20°C drop in
B
Fig. 1. Scalloped hammerhead shark diving behavior and temperature. (A) Eleven days of depth
(top), ambient (blue), and intramuscular (red) temperature profiles from scalloped hammerhead shark HH7. All
deep dives were conducted at night (gray shading). (B) Depth (top), ambient (blue), and intramuscular (red)
temperature profiles from scalloped hammerhead shark HH7 during six deep dives conducted in a single evening.
water temperature over a period of 5 to 7 min
(tables S3 and S4). However, the body temper-
ature of the sharks did not change appreciably
until rapidly cooling as the shark approached
the surface waters during the final ascent stage
of the dive (Fig. 2B and figs. S7 to S9).
During their fast descents into cold water,
shark body temperatures initially cooled very
slightly (0.1°C) but then either held constant
or warmed again slightly (0.1° to 0.25°C) dur-
ing the rest of the descent and through at
least half of their bottom time. After an average
of 4 min of bottom time and during the initial
fast ascent, body temperatures began to drop
slowly (0.03° ± 0.02°C/min). Near the inflection
point of the ascent (290-m average depth),
heat loss increased by an order of magnitude
to 0.23° ± 0.15°C/min. Body temperatures con-
tinued to drop at this faster rate until the final
phase of the dive, when sharks leveled off in
the surface mixed layer (Fig. 2 and figs. S7 to
S9). Thus, sharks experienced an average drop
in body temperature of 2.8°C from the bottom
of their dive to the start of their surface in-
terdive interval, with the greatest heat loss oc-
curring around the ascent inflection point
(290-m average depth) that is concurrent with
an abrupt decline in swimming intensity (Fig.
2, figs. S7 to S9, and tables S2 and S3).
Whole-body heat transfer coefficient (k) modeling
Heat exchange is proportional to the differ-
ence between body temperature and ambient
water temperature
Royer et al., Science 380, 651–655 (2023)
12 May 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Scalloped hammerhead shark swimming behavior and body temperature during deep dives.
(A) Depth profiles (black), body temperature (red), ambient water temperature (blue dashed), and swimming
activity [tailbeat sway acceleration (green)] during two successive deep dives by shark HH7. Dotted lines
indicate the inflection point of the ascents when the swimming activity and pitch angle abruptly decrease.
(B) Distinct phases of a deep dive using a representative dive. Shown are depth (black), intramuscular
temperature (red), and tailbeat sway acceleration (green).
dTb tð Þ
dt
¼ k Ta tð Þ (cid:3) Tb tð Þ
½
(cid:1)
(cid:4) þ T
o
ð1Þ
where k is the whole-body thermal rate co-
efficient (degrees Celsius per minute per
degree Celsius), Ta(t) is the ambient water
temperature (degrees Celsius) as a function of
time t, Tb(t) is the core muscle temperature
(degrees Celsius) as a function of time t, and
o is the rate of temperature change because
T
of metabolic heat production (degrees Celsius
per minute) by the swimming muscles (27–32).
The whole-body heat transfer coefficient (k)
accounts for both convective heat transfer (i.e.,
(cid:1)
heat exchange between blood flowing through
the gills and ambient seawater) and conductive
heat exchange across the body wall (33–35).
Predicted values of the whole-body heat trans-
fer coefficient (k) and rate of temperature
change because of metabolic heat production
(cid:1)
were modeled to match the observed rates
T
of body warming and cooling (27, 28).
(cid:3)
o
(cid:1)
Modeling (k) for three sharks revealed that
more than one k value is needed to accurately
describe the rate of heat transfer required to
produce body temperatures observed during deep
dives. Model 4, which used two k and two T
o
(heat production) values based on the different
(cid:1)
(cid:1)
(cid:1)
phases of the deep dives, was the most accurate
in replicating the observed body temperatures
(Table 1, table S5, Fig. 3, and figs. S10 and S11).
Values for k2 estimated for the intense swim-
ming phases (fast descent, bottom time, and
fast ascent) of deep dives were an order of
magnitude less than k1 values estimated for
the slow descent and slow ascent phases of the
dive (Table 1). This indicates a substantially
slower rate of heat transfer during the intense
swimming phases of deep dives and suggests
the existence of a mechanism enabling them
to reduce heat loss and conserve body temper-
ature until they reach the inflection point
during ascent. The modeled metabolic heat
production during the intense swimming
o2, was greater than the
phases of deep dives, T
heat production rate during slow swimming,
T
o1, by a factor of ~3 to 13 (Table 1).
Scalloped hammerhead k2:k1 ratios were ex-
ceptional compared with those of other ecto-
thermic sharks and fishes of similar body
sizes [e.g., the blue shark (Prionace glauca)]
and were comparable to those of regionally
endothermic species, such as the bigeye tuna
(Thunnus obesus) and swordfish (Xiphias
gladius), both of which have specialized ana-
tomical adaptations for conserving body heat
(32) (fig. S12). The large differences in the
heating and cooling coefficients of the scal-
loped hammerhead sharks in our study exceed
those of the ectothermic ocean sunfish (Mola
mola), which exhibits similarly large differ-
ences in these coefficients. This is surprising
because the disk-like anatomy of the ocean
sunfish is well suited for surface basking for
rewarming, whereas scalloped hammerhead
sharks have a more typical fusiform shape
similar to that of blue sharks or makos (Isurus
oxyrynchus). It is worth noting that although
the coefficient differences and ratios of the
scalloped hammerhead sharks are compara-
ble to those of regionally endothermic spe-
cies, scalloped hammerhead sharks lack the
specialized anatomy characteristic of these spe-
cies (32) (figs. S1 and S12).
The whole-body heat transfer coefficient was
also estimated for two dead adult scalloped
hammerheads that were obtained as freshly
dead incidental fisheries mortalities (table S6).
The single–k value model 1 was used because
heat flux can only be attributed to conductive
heat transfer across the body wall (there is no
blood flow to the gills to support convective
heat transfer), and physiological processes are
absent. The whole-body heat transfer coef-
ficients (k) from the dead sharks (0.0039°
and 0.0038°C min−1 °C−1) were similar (between
0.0011° and 0.0019°C min−1 °C−1 difference
for comparably sized sharks HH7 and HH9,
0.0121°C min−1 °C−1 for the smaller HH8) to
the k2 values estimated for the tagged scal-
loped hammerheads during the intense phases
of their deep dives using model 4 (Table 1 and
Royer et al., Science 380, 651–655 (2023)
12 May 2023
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Body temperature modeling across repetitive deep dives. Observed (yellow) and modeled
intramuscular temperature during six deep dives from scalloped hammerhead shark HH7 with depth (black).
Model 1 (blue dashed) assumes that the rate of heat exchange between the shark and the environment
(k) and the rate of temperature change because of metabolic heat production
is constant and not
altered by the shark. Model 4 (cyan solid) assumes k1 and T
and T
estimated on the basis of the coefficient of determination (R2) between the observed and predicted body
temperature. Models were run for each night that consisted of more than one deep dive for each shark. Start and stop
times for each model duration were arbitrarily set to include all deep dives within a night duration and to include the
durations where body temperature reaches the ambient temperature after the conclusion of a series of deep dives.
o2 during the high activity phases of a deep dive. The optimized parameters for each model were
o1 when the shark is swimming normally and k2
T
(cid:1)
(cid:3)
o
(cid:1)
(cid:1)
(cid:1)
table S6). The similarity of the heat transfer
coefficients of the dead sharks and the k2
values of similarly sized live sharks indicates
that scalloped hammerheads do not lose body
heat through convective heat transfer at the
gills during the intense swimming phases of
their dives.
Discussion
If convective heat transfer occurs at the gills
throughout the dive, body cooling should rou-
tinely begin and persist during the deeper
stages of the dive; yet, this is not the case.
Instead, these sharks maintained an elevated
body temperature (up to 20°C above ambient)
during most of each dive and only rapidly lost
heat near the inflection point in the ascent.
Sharp changes in body temperature at differ-
ent points of the dive cannot be explained by
simple passive buffering, and the apparent
lack of convective heat transfer suggests that
scalloped hammerhead sharks have an active
mechanism to prevent heat loss from their
gills during excursions into cold water. Possible
mechanisms include shunting blood away from
the gills or reducing the flow of water across
the gills by reducing ram-ventilation by clos-
ing the mouth, the gill slits, or both. Either
mechanism will inhibit the shark’s ability to
absorb oxygen from the environment. Essen-
tially, the sharks seem to be holding their
breath. We believe that the breath holding
mechanism is the closing of the gill slits, and
video evidence supports this contention. For
example, remotely operated vehicle (ROV) foot-
age (36) of an adult scalloped hammerhead
shark swimming at a depth of 1043 m off
Tanzania appeared to have its gill slits closed,
whereas videos of scalloped hammerhead
sharks in surface waters shows their gill slits
open (figs. S13 and S14). No anatomical evi-
dence of true vascular shunts at the gills has
been observed in sharks (37); therefore, the
most parsimonious explanation for the ob-
served heat retention is closing of the gill
slits. Further research is needed to confirm
the gill closing hypothesis, for instance by at-
taching cameras to the pectoral fins of sharks
that would capture the opening or closing
of the gills. The faster decline in body tem-
perature at or near the ascent inflection
point (250- to 300-m depth) probably reflects
reopening of the gill slits and the resump-
tion of convective heat transfer. The interdive
period might facilitate repayment of oxygen
debt incurred from the recruitment of the
white anaerobic musculature during the sprint-
ing portion of each dive. This highly active
swimming combined with apparent breath
holding should be reflected in the activity
levels of key muscle enzymes involved with lo-
comotion and energy mobilization and should
be examined further.
The physiological advantages conferred by
maintaining a warm body temperature at
depth enable scalloped hammerhead sharks to
exploit resources in the cold mesopelagic
depths lying below warm surface waters (38).
These advantages are similar to those ascribed
to so-called high-performance, regionally endo-
thermic fishes, such as tuna and lamnid sharks,
and include faster swimming capability be-
cause of enhanced muscle power output and
enhanced enzyme and cardiac performance
and neural processing (39). These benefits
are consistent with our observations of burst
swimming and high activity rates at the bot-
tom of dives that may be feeding events. Feed-
ing events are probably very brief. Hence, even
if the sharks fleetingly open their gills while
feeding at depth, it would be unlikely to be
reflected in the core muscle temperature. For
example, recorded feeding events at depth in
blue sharks (Prionace glauca) (32) and ocean
sunfish (Mola mola) (30) did not result in de-
tectable changes in muscle temperature.
There are distinct physiological differences
between scalloped hammerhead sharks and
high-performance fishes. The latter maintain
warm muscles and sensory organs during dives,
but they are nevertheless affected by reduc-
tions in heart function because the heart is
supplied by cold blood flowing from the gills
(40–45). By reducing convective heat loss at
the gills, scalloped hammerhead sharks main-
tain both muscle and heart temperature, there-
by possibly preserving cardiac function during
deep dives (45).
REFERENCES AND NOTES
1. E. D. Stevens, A. M. Sutterlin, J. Exp. Biol. 65, 131–145 (1976).
2. F. G. Carey, Q. H. Gibson, Physiol. Zool. 60, 138–148 (1987).
3. R. W. Brill, H. Dewar, J. B. Graham, Environ. Biol. Fishes 40,
109–124 (1994).
4. B. A. Block, J. R. Finnerty, Environ. Biol. Fishes 40, 283–302
(1994).
5. C. A. Sepulveda, S. Kohin, C. Chan, R. Vetter, J. B. Graham,
6.
Mar. Biol. 145, 191–199 (2004).
J. C. Patterson, C. A. Sepulveda, D. Bernal, J. Morphol. 272,
1353–1364 (2011).
7. D. Bernal, J. K. Carlson, K. J. Goldman, C. G. Lowe, in Biology of
Sharks and Their Relatives, J. C. Carrier, J. A. Musick,
M. R. Heithaus, Eds. (CRC Press, ed. 2, 2012), pp. 211–237.
8. S. J. Jorgensen, A. P. Klimley, A. F. Muhlia-Melo, J. Fish Biol.
74, 1682–1687 (2009).
9. S. Bessudo et al., Bol. Investig. Mar. Costeras 40, 91–106 (2011).
10. E. R. Hoffmayer, J. S. Franks, W. B. Driggers, P. W. Howey,
Bull. Mar. Sci. 89, 551–557 (2013).
11. J. L. Y. Spaet, C. H. Lam, C. D. Braun, M. L. Berumen,
Anim. Biotelem. 5, 20 (2017).
12. B. A. Seibel, E. V. Thuesen, J. J. Childress, L. A. Gorodezky,
Biol. Bull. 192, 262–278 (1997).
13. J. C. Drazen, B. A. Seibel, Limnol. Oceanogr. 52, 2306–2316
(2007).
14. T. A. Clarke, Pac. Sci. 25, 133–144 (1971).
15. A. P. Klimley, D. R. Nelson, Behav. Ecol. Sociobiol. 15, 45–54
(1984).
16. J. D. Stevens, J. M. Lyle, Mar. Freshw. Res. 40, 46–129
(1989).
17. M. J. Smale, G. Cliff, S. Afr. J. Mar. Sci. 7615, 241–253
(1998).
18. T. Vaske Júnior, C. M. Vooren, R. P. Lessa, Braz. J. Oceanogr. 57,
97–104 (2009).
19. A. F. Bennett, Am. J. Physiol. 259, R253–R258 (1990).
20. W. H. Neill, Amer. Zool. 19, 305–317 (1979).
21. S. A. Matern, J. J. Cech, T. E. Hopkins, Environ. Biol. Fishes 58,
173–182 (2000).
22. D. W. Sims et al., J. Anim. Ecol. 75, 176–190 (2006).
23. A. C. Gleiss, R. P. Wilson, E. L. C. Shepard, Methods Ecol. Evol.
2, 23–33 (2011).
24. B. J. Le Boeuf, Y. Naito, T. Asaga, D. Crocker, D. P. Costa, Can.
J. Zool. 70, 786–795 (1992).
25. C. D. Braun et al., Annu. Rev. Mar. Sci. 14, 129–159 (2022).
26. M. Royer et al., Anim. Biotelem. 8, 11 (2020).
Royer et al., Science 380, 651–655 (2023)
12 May 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
27. K. N. Holland, R. W. Brill, R. K. C. Chang, J. R. Sibert,
D. A. Fournier, Nature 358, 410–412 (1992).
28. K. N. Holland, J. R. Sibert, Environ. Biol. Fishes 40, 319–327 (1994).
29. H. Dewar, J. B. Graham, R. W. Brill, J. Exp. Biol. 192, 33–44
47. M. Royer et al., “Breath-holding” as a thermoregulation
strategy in the scalloped hammerhead, a deep-diving tropical
ectothermic shark, Zenodo (2023); https://doi.org/10.5281/
zenodo.7697788.
(1994).
30. I. Nakamura, Y. Goto, K. Sato, J. Anim. Ecol. 84, 590–603 (2015).
31. A. Stoehr, J. St. Martin, S. Aalbers, C. Sepulveda, D. Bernal,
ICES J. Mar. Sci. 75, 858–870 (2018).
32. Y. Y. Watanabe, I. Nakamura, W. C. Chiang, Mar. Biol. 168, 161
(2021).
33. W. H. Neill, E. D. Stevens, Science 184, 1008–1010 (1974).
34. R. G. Fechhelm, W. H. Neill, Physiol. Zool. 55, 229–239 (1982).
35. M. Pépino, K. Goyer, P. Magnan, J. Exp. Biol. 218, 3461–3467
(2015).
36. A. B. M. Moore, A. R. Gates, Mar. Biodivers. Rec. 8, 1–4 (2015).
37. N. C. Wegner, in Physiology of Elasmobranch Fishes: Structure
and Interaction with Environment, R. E. Shadwick, A. P. Farrell,
C. J. Brauner, Eds., vol. 34 of Fish Physiology (Academic Press,
2015), pp. 101–151.
38. J. B. Graham, K. A. Dickson, Fish Physiol. 19, 121–165 (2001).
39. K. A. Dickson, J. B. Graham, Physiol. Biochem. Zool. 77,
998–1018 (2004).
40. K. E. Korsmeyer, N. C. Lai, R. E. Shadwick, J. B. Graham,
J. Exp. Biol. 200, 1975–1986 (1997).
41. R. W. Brill, T. E. Lowe, K. L. Cousins, in Cardiovascular Function
in Fishes, K. Gamperl, D. MacKinlay, A. Farrel, Eds. (American
Fisheries Society, 1998), pp. 57–62.
42. J. M. Blank et al., J. Exp. Biol. 207, 881–890 (2004).
43. K. C. Weng et al., Science 310, 104–106 (2005).
44. H. A. Shiels, G. L. J. Galli, B. A. Block, Proc. R. Soc. B 282,
20141989 (2015).
45. N. C. Wegner, O. E. Snodgrass, H. Dewar, J. R. Hyde, Science
348, 786–789 (2015).
46. M. Royer et al., “Breath-holding” as a thermoregulation
strategy in the scalloped hammerhead, a deep-diving tropical
ectothermic shark, dataset, Dryad (2023); https://doi.org/
10.5061/dryad.p8cz8w9v0.
ACKN OWLED GMEN TS
We thank K. Carlson, K. Haiat, N. Hu, B. Rackliffe, J. Lazor, P. Mino,
T. Clegg, M. Maxwell, and S. Rucker for assisting with fieldwork
and J. Muir for providing crucial boat support for tag recovery. The
lead author thanks C. Meyer, B. Bowen, A. Seale, and M. Yoshizawa
for providing feedback on this project as dissertation committee
members and K. Holland for being the chair of this committee. We
thank L. Robinson, the Robinson family, and the people of Ni‘ihau
for finding our tag package and J. TenBruggencate for coordinating
communication and shipping of the recovered tag to us. We thank
the anonymous reviewers for their insightful feedback. This is
contribution no. 1918 from the Hawai‘i Institute of Marine Biology
and no. 11660 from the School of Ocean and Earth Science and
Technology at the University of Hawai‘i. Data for this study are
available on Dryad (46). The Matlab code used to for thermal
coefficient analysis is available on Zenodo (47). Funding: This
study received support from the Jessie D. Kay Memorial Research
Grant, UH Mānoa Department of Biology (M.R.); a Hawai‘i Institute
of Marine Biology Lord Endowed Scholarship (M.R.); a Robinson
Family Ocean Studies Assistantship (M.R.); and Pacific Islands
Ocean Observation Systems (K.N.H. and C.M.). Ethics
statement: Scalloped hammerhead shark handling and tagging
procedures were approved by the ethics committee at the
University of Hawai‘i (Institutional Animal Care and Use Committee
Protocol no. 05-053). Author contributions: M.R. designed the
study. M.R., K.M., E.C., K.W., C.B., and G.F.d.S. designed, constructed,
and tested biologging packages. M.R. led fieldwork operations and
attached biologging packages to all sharks. K.M., E.C., K.W., C.B., and
G.F.d.S. assisted in fieldwork operations, including field preparation
and planning, fishing effort, shark tagging, and package recovery.
J.R. developed the Matlab code used for thermal coefficient modeling
and analysis. C.M. provided detailed instruction for analyzing
accelerometry data. M.R., K.M., E.C., G.F.d.S., and K.W. analyzed
telemetry data. M.R. and J.R. conducted analyses on thermal
coefficient modeling. M.R. drafted the figures. M.R. drafted the main
body, and K.N.H. and C.M. substantially revised it. M.R. wrote the
supplementary materials section. C.M. and K.N.H. provided
substantial contributions to all written material and revisions
during initial draft writing. All authors gave substantive feedback
and suggested edits during the final drafting of the manuscript.
M.R., C.M., and K.N.H. secured funding. Competing interests: The
authors declare that there are no competing interests. Data and
materials availability: The following data are available on Dryad (46):
results of thermal coefficient modeling (Matlab code) for each night
of deep diving for each shark; deep-diving telemetry metrics for
each individual dive from each shark and summary statistics for each
metric; and depth, ambient temperature, body temperature, and
triaxial acceleration telemetry for each shark for each night of deep
diving analyzed. Thermal coefficient modeling was performed using
Matlab code that is available through Zenodo (47). Telemetry data
were analyzed with Igor Pro 9 (Wavemetrics, Portland, OR, USA).
Correspondence should be addressed to M.R. License information:
Copyright © 2023 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science. No
claim to original US government works. https://www.science.org/
about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add4445
Materials and Methods
Figs. S1 to S14
Tables S1 to S6
References (48–55)
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
Submitted 5 July 2022; accepted 15 March 2023
10.1126/science.add4445
Royer et al., Science 380, 651–655 (2023)
12 May 2023
5 of 5
|
10.1126_science.add2915
|
RES EARCH
AVIAN ECOLOGY
Agriculture and hot temperatures interactively erode
the nest success of habitat generalist birds across
the United States
Katherine S. Lauck1*†, Alison Ke1†, Elissa M. Olimpi1,2, Daniel Paredes1,3, Kees Hood1,
Thomas Phillips1, William R. L. Anderegg4,5, Daniel S. Karp1
Habitat conversion and climate change are fundamental drivers of biodiversity loss worldwide but are
often analyzed in isolation. We used a continental-scale, decades-long database of more than 150,000 bird
nesting attempts to explore how extreme heat affects avian reproduction in forests, grasslands, and
agricultural and developed areas across the US. We found that in forests, extreme heat increased nest
success, but birds nesting in agricultural settings were less likely to successfully fledge young when
temperatures reached anomalously high levels. Species that build exposed cup nests and species of
higher conservation concern were particularly vulnerable to maximum temperature anomalies in
agricultural settings. Finally, future projections suggested that ongoing climate change may exacerbate
the negative effects of habitat conversion on avian nesting success, thereby compromising conservation
efforts in human-dominated landscapes.
H abitat conversion is the primary driver
of terrestrial biodiversity loss, and cli-
mate change is projected to cause wide-
spread extirpations (1, 2). However, the
effects of habitat conversion and climate
change are often analyzed in isolation even
though the fate of many species will ultimately
depend on how they interact (3). For example,
many forms of habitat conversion (e.g., agri-
cultural or urban expansion) remove insulating
tree canopies or other complex microhabitats,
thereby exposing organisms to warmer maxi-
mum and/or cooler minimum temperatures
[i.e., reducing thermal buffering (4)]. Indeed,
temperatures in agricultural settings regularly
attain levels >10°C higher than in nearby natu-
ral habitats (5). Other stressors related to human
land use may increase the sensitivity of bio-
diversity to heat; for example, pesticide use
and low vegetation complexity may reduce in-
sect prey availability, limiting food and water
available for thermoregulation (6). In addition,
trees may protect understory species from
heavy rains and retain moisture, buffering
against drought (4). Thus, as temperatures
warm and precipitation regimes shift, climate
change may cause cities and farms to become
even less hospitable, undermining efforts to
safeguard biodiversity in human-dominated
landscapes (7).
1Department of Wildlife, Fish, and Conservation Biology,
University of California, Davis, Davis, CA, USA. 2Department
of Fish and Wildlife Conservation, Virginia Tech, Blacksburg,
VA, USA. 3Environmental Analysis Group, Department of
Plant Biology, Ecology and Earth Science, University of
Extremadura, Extremadura, Spain. 4School of Biological
Sciences, University of Utah, Salt Lake City, UT, USA. 5Wilkes
Center for Climate Science and Policy, University of Utah,
Salt Lake City, UT, USA.
*Corresponding author. Email: kslauck@ucdavis.edu
†These authors contributed equally to this work.
Birds may be particularly sensitive to tem-
perature extremes, because species with altri-
cial young are ectothermic for the first few
weeks of life, and extreme temperatures can
divert energy from growth to thermoregula-
tion (8). Although the effects of cold snaps on
avian reproduction are well documented (9, 10),
recent work suggests that high temperatures
can also reduce avian survival (11–13) and even
cause community collapse (13). Temperature
extremes may limit species persistence more
than long-term increases in average tempera-
tures (14), and variations in microclimate buf-
fering can thus influence bird distributions (15).
Unfortunately, understanding the interactive
effects of climate and land-use change on spe-
cies persistence requires demographic data that
are difficult to obtain over large spatiotemporal
scales (3). Thus, our knowledge of how climate
and land-use change interactively affect species
is mostly restricted to changes in species dis-
tributions and/or abundances (4, 16). Project
NestWatch, a citizen-science nest-monitoring
program organized by the Cornell Laboratory
of Ornithology, offers a rare opportunity to
explore how temperature extremes and land
use interact to affect avian nesting success at
a national scale. Using their data, we analyzed
152,863 nesting attempts by 58 bird species
across 23 years (1998 to 2020) and 37,869 sites
in four land-use types within the conterminous
US: forests, open natural habitats, agricultural
settings, and developed areas [table S1 (17)].
We used GridMET (18) to measure tempera-
ture extremes during each nesting attempt by
calculating temperature anomalies, which we
defined as the average maximum (or minimum)
temperatures during the 45 days after each
nesting attempt’s date of first egg relative to
conditions during the same date range over
a historical reference period (1980 to 2000).
Our work was guided by four questions.
First, how do the effects of temperature on
nesting success vary across land-cover types?
We hypothesized that maximum temperature
anomalies would reduce success in open, human-
dominated habitats with fewer thermally buf-
fered areas. We also quantified precipitation
anomalies during nesting, hypothesizing that
extreme events would also reduce success in
open, human-dominated habitats. Second, are
some species more vulnerable to the inter-
active effects of habitat conversion and climate
change than others? We predicted that species
that build exposed nests (as opposed to cavity
nests) and species of higher conservation con-
cern would be the most sensitive to climate and
land-use interactions. Third, are the effects of
temperature and land cover consistent across
species’ ranges? We predicted that maximum
temperature anomalies would disproportionately
affect nesting success in agricultural areas
within warmer regions. Finally, looking for-
ward across the 21st century, how will nesting
success likely change across space, time, and
alternative climate change scenarios? We hypo-
thesized that declines would be pronounced
in agricultural settings, grassland, and devel-
oped areas, but not in forests.
The interactive effects of climate and land
cover could arise from individuals of the same
species varying in their temperature responses
across land-cover types (e.g., if some provide
more thermal buffering than others) or from
shifts in species composition (e.g., if agriculture-
associated species are more thermally sensitive
than forest-associated species). To determine
whether thermal buffering in some land-cover
types increases avian resilience against heat
waves, we focused on habitat generalist species
that could conceivably exhibit different re-
sponses to heat waves in different land-cover
types. Additionally, we modeled climate and
land-use interactions using both generalized
linear mixed models (GLMMs) that integrate
within- and across-species effects and Bayesian
models that only consider variation within
species (17).
Climate and land-cover interactions
Interactions between maximum temperature
anomalies and land-cover type
Across both modeling frameworks, we found
that the effects of maximum temperature ex-
tremes on avian nesting success differed among
land-use types (P < 0.001 for GLMM; Bayesian
confidence interval: agricultural settings, –0.21
to –0.01; forests, 0.03 to 0.15; Fig. 1 and tables
S2 to S5). In agricultural settings, the probabil-
ity of successfully fledging at least one offspring
declined by 6% (from 75 to 69%) between nests
that experienced cooler versus warmer maxi-
mum temperature anomalies (i.e., 2 SDs lower
versus 2 SDs higher than mean historical tem-
peratures). Only considering within-species
Lauck et al., Science 382, 290–294 (2023)
20 October 2023
1 of 5
An Erratum was posted on 14 December 2023. See Erratum. RES EARCH | R E S E A R C H A R T I C L E
effects produced a similar decline (i.e., an 8%
decline of the same temperature range in the
Bayesian model). In developed areas, the prob-
ability of success declined by only 2% from
cooler to warmer maximum temperature anom-
alies (not significant in the Bayesian model).
One reason for the more muted decline might
be that nests in “developed areas” were often
in residential areas that can have high tree
cover, not in city centers that may be subject to
particularly extreme heat island effects. By
contrast, only 1.5% of agricultural sites were in
plantations with tree canopies (table S6). As a
result, developed areas had higher canopy cover
than both agricultural and natural open sites
(mean cover: developed, 21%; agricultural, 11%;
natural open, 17.6%). Future work could pro-
fitably focus on using temperature loggers to
link nesting success to local microclimates
rather than the coarser air temperature mea-
surements analyzed here.
Although temperature increases in forests
have been previously shown to reduce nest
productivity through increased predation (19),
we found that reproductive success increased
in forests by 5% across the same temperature
range (6% in the Bayesian model). Why might
this be? First, because tree canopies may keep
nests cool, adults nesting in forests could be
released from thermoregulatory care when
temperatures increase, allowing them to spend
more time foraging (20). Increased tempera-
tures in cold microclimates might also reduce
the reproductive cost of time spent off the nest
(21). Maximum temperature anomalies could
also drive phenological shifts that improve
forest-nesting birds’ access to food resources
[e.g., by increasing insect abundance earlier in
the season (22)]. Critically, the positive relation-
ship between nesting success and maximum
temperature anomalies in forests does not sug-
gest that climate change is benefiting forest birds,
because warming temperatures may be decreas-
ing other demographic parameters (e.g., adult
and/or juvenile survival).
Finally, in natural open habitats, nest success
exhibited a nonlinear relationship with maxi-
mum temperature anomalies, peaking at an
intermediate value (in the Bayesian model,
no relationship with temperature was observed).
One explanation is that grassland species could
be adapted to temperature regimes in open-
canopied environments, but too much devia-
tion from historical norms (in either direction)
may decrease success. Indeed, most nests in
natural open habitats were in grasslands or
prairies (~85% of attempts; table S6).
Our finding that the effects of temperature
extremes on nesting success vary across
land-use types persisted when additionally
accounting for latitude, elevation, nest preda-
tion, spatial autocorrelation, extreme observa-
tions, the scale of landscape cover composition,
and nests in more thermally buffered agri-
Fig. 1. Maximum temperature extremes reduce avian nesting success in agricultural settings but
increase it in forests. For nests in each land-use type, panels present the relationship between maximum
temperature anomalies 45 days after lay date as z-scores relative to historical temperatures (17) and the
predicted proportion of nests with at least one offspring surviving to fledging. Solid lines indicate model
predictions; shaded regions represent 95% confidence regions. Histograms in the bottom panels depict the
distribution of maximum temperature anomalies across all nesting attempts. (A) Results from GLMMs
combining within- and among-species variations in response to temperature. Asterisks indicate the level of
significance in each land use. *P < 0.05; **P < 0.01. The quadratic effect of maximum temperature
anomaly is only significant in natural open land cover. (B) Results from a Bayesian analysis that only assesses
the effects of temperatures within species (i.e., factoring out changes in species composition among
land-cover types). Asterisks indicate that the 95% Bayesian confidence interval did not overlap zero.
cultural habitats [table S7 and fig. S1 (17)].
Birds were also not more sensitive to temper-
ature anomalies in agricultural settings be-
cause of differences in air temperature among
land-use types (fig. S2).
Interactions among minimum temperature
anomaly, precipitation, and land cover
Unlike maximum temperatures, the effects of
other climate variables exhibited much less
variation among land-cover types (tables S2,
S3, and S8). First, the linear effect of minimum
temperature anomalies did not vary among
land-cover types (P = 0.28). However, the qua-
dratic effect did so marginally (P = 0.05), with
nesting success exhibiting a more convex rela-
tionship with minimum temperature anomaly
in forests and a more linear relationship in the
other land-cover types. Second, precipitation
over the prior year also exhibited no variation
in linear effects on nesting success across land-
cover types (P = 0.19). However, the quadratic
effect again varied (table S3 and S8), exhibiting
a more convex relationship in forests and a more
linear effect in the other land-cover types. Final-
ly, the linear effect of precipitation anomalies
Lauck et al., Science 382, 290–294 (2023)
20 October 2023
2 of 5
An Erratum was posted on 14 December 2023. See Erratum. RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Interactive effects of
temperature extremes and
land use are exacerbated for
species of higher conservation
concern and for cup-nesting
species in agricultural settings.
(A) Maximum temperature
anomalies 45 days after lay
date on the proportion of nests
with at least one fledging
(i.e., nest success) in each
land-use type and for species of
varying levels of conservation
concern. Colored lines depict
predictions for species of
varying levels of conservation
concern, represented by: house
sparrow (Passer domesticus;
score level 4, lowest concern),
barn swallow (Hirundo rustica,
level 8), chestnut-backed
chickadee (Poecile rufescens,
level 12), and oak titmouse
(Baeolophhus inornatus; score
level 15, highest concern).
(B) Maximum temperature
anomaly effects on birds
nesting in cavities (blue) versus
cup nests (red). Asterisks
indicate the level of significance
of the interaction between
maximum temperature anomalies
and species’ conservation scores
or nest types in each land use.
+P < 0.1, *P < 0.05, **P < 0.01.
during nesting varied by land-cover type, but
with smaller effect sizes than maximum temper-
ature anomalies (P < 0.001; fig. S3 and tables S3,
S8, and S9). Specifically, higher precipitation
during nesting was associated with higher suc-
cess in agricultural settings, decreased success
in forests, and small effects in other land-cover
types. This suggests that birds in agricultural
settings may suffer even more when precipita-
tion is low and temperatures are high.
Identifying species sensitive to climate
and land-cover interactions
The negative effects of maximum temperature
anomalies on avian reproduction in agricul-
tural settings and the positive effects in forests
were generally consistent across species, but
were not consistent across species in other
land-cover types (fig. S4 and table S5). Mul-
tiple mechanisms may explain these trends.
First, temperature extremes in agricultural
settings could induce physiological stress and
require nestlings to expend energy on thermo-
regulation rather than growth or to suffer water
costs and dehydration (8). Second, maximum
temperature anomalies may affect birds in-
directly, for example, by reducing arthropod
prey abundance or adult foraging efficiency
in agricultural settings but not in natural habi-
tats with microclimate refugia (23, 24). Indeed,
arthropods are particularly sensitive to climate
warming in agricultural settings but are more
buffered in natural habitats (25).
Nest type
Identifying traits that explain variations in
species’ responses to maximum temperature
anomalies might provide insight into the under-
lying mechanisms. We found that cup nests in
agricultural settings experienced particularly
severe declines in nest success with higher
maximum temperatures (P = 0.005; Fig. 2 and
tables S10 to S12). Cup nests are less thermally
buffered than cavity nests, suggesting that
maximum temperature anomalies may indeed
reduce nesting success directly [i.e., through
avian physiology (26)]. We also found similar re-
sults when comparing nests in human-constructed
nest boxes with other nests (fig. S5 and tables S10
to S12); however, because 99% of cavity nests
were also in nest boxes, these two compari-
sons are functionally equivalent. This finding
suggests that the effect sizes of the tempera-
ture and land-use interactions reported above
may be conservative given the dominance of
artificial nest box observations in the database
(table S1).
Conservation concern
Species of higher conservation concern [as
defined by (27)] were also more vulnerable to
maximum temperature anomalies in agricul-
tural settings and more successful in forests
(P = 0.027; Fig. 2 and tables S10 to S12). This
again suggests that the effect sizes reported
above may be conservative for rarer, less fre-
quently sampled species. For example, in agri-
cultural settings, GLMMs predicted that hotter
maximum temperature anomalies would de-
crease nest success by 15% (from 90 to 75%;
maximum temperatures ± 2 SDs) for species of
highest conservation concern but increase suc-
cess by 1% for species of low concern. Because
species of conservation concern are often already
sensitive to anthropogenic land uses (28), they
Lauck et al., Science 382, 290–294 (2023)
20 October 2023
3 of 5
An Erratum was posted on 14 December 2023. See Erratum. RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Climate change is expected to decrease nesting success in agricultural settings but increase
it in forests. Density diagrams (left) depict how the probability of successfully fledging at least one offspring
is predicted to change by 2100 under a high-emissions scenario (RCP 8.5) across sites for each land-use
type. Maps depict the spatial distribution of predicted changes; dark blue, red, and gray points indicate sites
where nesting success is expected to increase by 5% or more, decrease by 5% or more, or remain largely
unchanged, respectively. Points and histograms represent average predictions across five climate models (17).
may simply be unable to cope when tempera-
tures increase in marginal environments.
Thermal limits
Avian reproduction is expected to be most sen-
sitive to temperature extremes in regions that
experience temperatures near species’ thermal
limits (29). To examine spatial variation in spe-
cies’ responses to maximum temperature ano-
malies and land use, we calculated site tem-
perature baselines in four ways, including
approaches that quantify temperatures in ab-
solute terms and relative to temperature var-
iability across each species’ geographic range
(17). Although nest success was more sensitive
to maximum temperature anomalies in warmer
regions (P < 0.001 for three of four measures;
fig. S6 and tables S13 and S14), the interactive
effect of maximum temperature anomalies and
land-use types was consistent across regions: In
both hot and cold regions, maximum temper-
ature anomalies were the most harmful in agri-
cultural settings (P ≥ 0.6; tables S13 and S15).
Future climate change scenarios
Our results suggest that bird species nesting in
agricultural settings may be more vulnerable
to climate change than those in forests. Lever-
aging five global climate models and multiple
climate change scenarios [representative con-
centration pathways (RCPs)], we used our
models to explore how nest success would
change if each nesting attempt in our dataset
instead occurred with the climate conditions
of 2040 to 2059 or 2080 to 2099 (Fig. 3, fig.
S7A, and table S16). These analyses implicitly
assume that nesting phenology is fixed in time,
even though some birds can shift to earlier
breeding times to track climate niches (29).
Therefore, these results should be considered
a sensitivity analysis rather than explicit pre-
dictions. Climatic uncertainty for agricultural
settings was slightly higher than for the other
land-cover types, especially in the southeastern
US (figs. S7B and S8), but we suspect that
climatic uncertainty may have constrained
the projected effect sizes. Statistical uncertainty
exceeded climatic uncertainty (figs. S7C and
S8). Our models suggest that nesting success
in agricultural settings would decline by an
additional 4.41% by 2100 if greenhouse gas
emissions maintain their current rate of in-
crease (RCP 8.5), whereas success in forests
would increase by 1.78%. By contrast, if emis-
sions were reduced (RCP 4.5), then nesting
success in agricultural settings would decline
by only 1.14% and success in forests would in-
crease by 1.07% (fig. S7A and table S16). Thus,
if emissions are curtailed, then birds nesting
in human-dominated areas would likely be
more successful over the long term.
Conclusions
Our results highlight the vulnerability of birds
nesting in agricultural settings to temperature
extremes and may offer insight into mecha-
nisms underlying North American bird declines
(30). They also align with recent findings from
Europe suggesting that climate change may
be causing larger population declines in gen-
eralist, farmland-associated birds compared
with specialist, woodland-associated species
(31). Maintaining forest patches in anthropo-
genic landscapes may thus increase avian re-
silience to extreme climatic events (32). An
important caveat is that the species studied
here are habitat generalists. Enhancing forest
cover in naturally open land covers could harm
grassland-obligate birds, which might be able
to leverage habitat heterogeneity in natural
grasslands to find thermally buffered areas (33).
Nonetheless, for other species, erecting
Lauck et al., Science 382, 290–294 (2023)
20 October 2023
4 of 5
An Erratum was posted on 14 December 2023. See Erratum. RES EARCH | R E S E A R C H A R T I C L E
sun-shielded or insulated nest boxes in shaded
locations (34), conserving forest patches, and/or
planting scattered trees in human-dominated
landscapes may help species cope with climate
change–driven temperature extremes by provid-
ing thermal buffering (35), especially for the
species of conservation concern studied here.
RE FE RENCES AND N OT ES
1. R. Dirzo, P. H. Raven, Annu. Rev. Environ. Resour. 28, 137–167
(2003).
2. F. E. B. Spooner, R. G. Pearson, R. Freeman, Glob. Chang. Biol.
24, 4521–4531 (2018).
3. T. H. Oliver, M. D. Morecroft,Wiley Interdiscip. Rev. Clim. Change
4.
5, 317–335 (2014).
J. J. Williams, T. Newbold, Divers. Distrib. 26, 76–92
(2020).
5. R. A. Senior, J. K. Hill, P. González Del Pliego, L. K. Goode,
D. P. Edwards, Ecol. Evol. 7, 7897–7908 (2017).
6. D. R. Garrett, F. Pelletier, D. Garant, M. Bélisle, Ecol. Monogr.
92, e1518 (2022).
7. C. Kremen, A. M. Merenlender, Science 362, eaau6020
8.
(2018).
J. C. Wingfield et al., Philos. Trans. R. Soc. Lond. B Biol. Sci.
372, 20160140 (2017).
9. D. W. Winkler, M. K. Luo, E. Rakhimberdiev, Oecologia 173,
129–138 (2013).
10. J. R. Shipley et al., Proc. Natl. Acad. Sci. U.S.A. 117,
25590–25594 (2020).
11. B. Zuckerberg, C. A. Ribic, L. A. McCauley, Conserv. Biol. 32,
872–882 (2018).
12. A. R. Bourne, S. J. Cunningham, C. N. Spottiswoode,
A. R. Ridley, Proc. Biol. Sci. 287, 20201140 (2020).
13. E. A. Riddell, K. J. Iknayan, B. O. Wolf, B. Sinervo, S. R. Beissinger,
Proc. Natl. Acad. Sci. U.S.A. 116, 21609–21615 (2019).
14. N. E. Zimmermann et al., Proc. Natl. Acad. Sci. U.S.A. 106,
19723–19728 (2009).
15. S. J. K. Frey, A. S. Hadley, M. G. Betts, Divers. Distrib. 22,
944–959 (2016).
16. J. M. Northrup, J. W. Rivers, Z. Yang, M. G. Betts, Glob. Chang.
Biol. 25, 1561–1575 (2019).
17. Materials and methods are available as supplementary
materials.
18. J. T. Abatzoglou, Int. J. Climatol. 33, 121–131
37. Code for: K. S. Lauck et al., Agriculture and hot temperatures
interactively erode the nest success of habitat generalist
birds across the United States, Zenodo (2023); https://
doi.org/10.5281/zenodo.8226606.
(2013).
AC KNOWLED GME NTS
19. W. A. Cox, F. R. Thompson 3rd, J. L. Reidy, J. Faaborg,
Glob. Chang. Biol. 19, 1064–1074 (2013).
20. V. G. Rohwer, J. R. Purcell, PLOS ONE 14, e0219907
(2019).
21. B. H. Coe, M. L. Beck, S. Y. Chin, C. M. B. Jachowski,
W. A. Hopkins, J. Avian Biol. 46, 385–394 (2015).
22. E. K. Bowers et al., Ecology 97, 2880–2891 (2016).
23. A. J. Suggitt et al., Oikos 120, 1–8 (2011).
24. C. Funghi, L. S. C. McCowan, W. Schuett, S. C. Griffith,
Anim. Behav. 149, 33–43 (2019).
25. C. L. Outhwaite, P. McCann, T. Newbold, Nature 605, 97–102
(2022).
26. M. C. Mainwaring, I. R. Hartley, M. M. Lambrechts,
D. C. Deeming, Ecol. Evol. 4, 3909–3928 (2014).
27. North American Bird Conservation Initiative, “The state of
North America’s birds 2016,” in Environment and Climate
Change (Cornell Univ., 2016); www.stateofthebirds.org.
28. L. Sykes, L. Santini, A. Etard, T. Newbold, Conserv. Biol. 34,
688–696 (2020).
29. J. B. Socolar, P. N. Epanchin, S. R. Beissinger,
M. W. Tingley, Proc. Natl. Acad. Sci. U.S.A. 114,
12976–12981 (2017).
30. K. V. Rosenberg et al., Science 366, 120–124 (2019).
31. S. Rigal et al., Proc. Natl. Acad. Sci. U.S.A. 120, e2216573120
(2023).
32. D. G. Nimmo, A. Haslem, J. Q. Radford, M. Hall, A. F. Bennett,
J. Appl. Ecol. 53, 449–458 (2016).
33. J. S. Bernath-Plaisted, C. A. Ribic, W. B. Hills,
P. A. Townsend, B. Zuckerberg, Environ. Res. Lett. 18,
064023 (2023).
34. E. R. Larson, J. R. Eastwood, K. L. Buchanan, A. T. D. Bennett,
M. L. Berg, Ecol. Manage. Restor. 19, 39–48 (2018).
35. T. Vanwalleghem, R. K. Meentemeyer, Ecosystems 12,
1158–1172 (2009).
36. Data for: K. S. Lauck et al., Agriculture and hot temperatures
We thank the volunteers who contributed data to the Cornell
Laboratory of Ornithology’s Project NestWatch program, as well as
program staff; R. Bailey for help in preparing and sharing the
database; W. Brooks of the UC Davis DataLab for invaluable advice
regarding statistical approaches for spatial autocorrelation; and
our anonymous reviewers for their rigorous and constructive
feedback. Funding: This work was supported by a UC Davis
Graduate Group in Ecology Fellowship (K.S.L.); a National Science
Foundation Graduate Research Fellowship (A.K.); and an
Achievement Rewards for College Scientists Fellowship (A.K.).
Author contributions: Conceptualization: K.S.L., A.K., E.M.O., D.P.,
K.H., T.P., W.R.L.A., D.S.K.; Data curation: K.S.L., A.K., E.M.O., D.P.,
W.R.L.A., D.S.K.; Formal analysis: K.S.L., A.K., E.M.O., D.P.; Funding
acquisition: K.S.L., A.K., D.S.K.; Investigation: K.S.L., A.K., E.M.O.,
D.P., K.H., T.P., W.R.L.A., D.S.K.; Methodology: K.S.L., A.K., E.M.O.,
D.P., K.H., T.P., W.R.L.A., D.S.K.; Supervision: D.S.K.; Visualization:
K.S.L., A.K., E.M.O., D.P.; Writing – original draft: K.S.L., A.K.,
D.S.K.; Writing – review and editing: K.S.L., A.K., E.M.O., D.P., K.H.,
T.P., W.R.L.A., D.S.K. Competing interests: The authors declare
no competing interests. Data and materials availability: All data
are available in the manuscript, the supplementary materials, or
Dryad (36). Code are available on Zenodo (37). Data and codebase
are available under the GNU General Purpose License 3.0. License
information: Copyright © 2023 the authors, some rights reserved;
exclusive licensee American Association for the Advancement of
Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add2915
Materials and Methods
Figs. S1 to S8
Tables S1 to S16
References (38–49)
interactively erode the nest success of habitat generalist birds across
the United States, Dryad (2023); https://doi.org/10.25338/B8ZD1P.
Submitted 1 June 2022; accepted 11 August 2023
10.1126/science.add2915
Lauck et al., Science 382, 290–294 (2023)
20 October 2023
5 of 5
An Erratum was posted on 14 December 2023. See Erratum.
|
10.1126_science.add3672
|
RES EARCH
SUPERCONDUCTIVITY
Correlating the charge-transfer gap to the maximum
transition temperature in Bi2Sr2Can-1CunO2n+4+d
Zechao Wang1,2†, Changwei Zou3†, Chengtian Lin4, Xiangyu Luo5, Hongtao Yan5, Chaohui Yin5,
Yong Xu3,6,7, Xingjiang Zhou5, Yayu Wang3,6*, Jing Zhu1,2*
As the number of CuO2 layers, n, in each unit cell of a cuprate family increases, the maximum transition
temperature (Tc,max) exhibits a universal bell-shaped curve with a peak at n = 3. The microscopic
mechanism of this trend remains elusive. In this study, we used advanced electron microscopy to image
the atomic structure of cuprates in the Bi2Sr2Can-1CunO2n+4+d family with 1 ≤ n ≤ 9; the evolution of the
charge-transfer gap size (D) with n can be measured simultaneously. We determined that the n
dependence of D follows an inverted bell-shaped curve with the minimum D value at n = 3. The
correlation between D, n, and Tc,max may clarify the origin of superconductivity in cuprates.
p
ffiffiffiffiffi
M
I dentifying universal trends of the super-
conducting transition temperature (Tc)
and its correlations to other physical param-
eters may provide crucial clues for elucidat-
ing the origin of superconductivity (1). For
example, the isotope effect Tcº1=
, where
M is the isotopic mass, has inspired the phonon-
mediated pairing picture in establishing the
microscopic theory of conventional supercon-
ductivity in metals and alloys (2, 3). Since the
discovery of high-Tc superconductivity in cop-
per oxide materials (referred to as cuprates),
there have been considerable efforts in under-
standing what controls Tc. There are two well-
established trends about Tc in cuprates: One is
the dome-shaped doping dependence of Tc for a
specific cuprate compound, where the maxi-
mum Tc (Tc,max) is located at hole concentration
p ∼ 0.16 (1). The other is the variation of Tc,max
with the number of CuO2 planes (n) per unit
cell in a homologous series, which reaches the
highest when n = 3 (4). Understanding the
physical origin behind these empirical rules
could lead to the eventual solution for the
pairing-mechanism problem in cuprates (1, 4).
Whereas the second trend remains univer-
sal for many cuprate families (Fig. 1A), the first
is apparently violated in the trilayer cuprate
Bi2Sr2Ca2Cu3O10+d (Bi-2223), which has the high-
est Tc,max of ~115 K in the Bi2Sr2Can-1CunO2n+4+d
family (Bi-family) compounds. Recent scanning
tunneling microscopy (STM) studies (5) have
1National Center for Electron Microscopy in Beijing, School of
Materials Science and Engineering, Key Laboratory of
Advanced Materials (MOE), The State Key Laboratory of New
Ceramics and Fine Processing, Tsinghua University, Beijing,
P.R. China. 2Ji Hua Laboratory, Foshan, Guangdong, P.R.
China. 3State Key Laboratory of Low-Dimensional Quantum
Physics and Department of Physics, Tsinghua University,
Beijing, P.R. China. 4Max Planck Institute for Solid State
Research, Stuttgart, Germany. 5Beijing National Laboratory
for Condensed Matter Physics, Institute of Physics, Chinese
Academy of Sciences, Beijing, P.R. China. 6New Cornerstone
Science Laboratory, Frontier Science Center for Quantum
Information, Beijing, P.R. China. 7RIKEN Center for Emergent
Matter Science (CEMS), Wako, Japan.
†These authors contributed equally to this work.
*Corresponding author. Email: jzhu@mail.tsinghua.edu.cn (J.Z.);
yayuwang@mail.tsinghua.edu.cn. (Y.W.)
shown that the superconducting gap of Bi-2223
decreases monotonously with p in the over-
doped regime. Unexpectedly, the Tc in this
doping range exhibits a plateau rather than
a continuous drop, most likely owing to the
existence of two types of inequivalent CuO2
planes (6, 7). By contrast, the Bi-2223 system
provides an ideal platform for investigating
the evolution of Tc,max with n and the physi-
cal parameters that control Tc,max. A promising
starting point is the Mott-insulating “parent”
state (p ~ 0) of cuprates, where the electronic
structure can be described by a single param-
eter referred to as the charge-transfer gap
(CTG) D, the size of which varies from 1 to 3 eV (8).
A previous STM work (9) in Bi2Sr2Can-1CunO2n+4+d
and Can+1CunOn+2Cl2 with n = 1 and 2 suggests
that Tc,max is anticorrelated with the D, together
with theoretical calculations (10–12). However,
a measurement covering the whole homologous
cuprate series has not yet been achieved because
of two main obstacles. First, it is difficult to syn-
thesize high-quality single crystals for n ≥ 4 in
any cuprate family, and it is even more difficult
to reach the p ~ 0 limit. Second, most of the
spectral information about cuprates cannot
distinguish the distinct CuO2 planes for n ≥ 3.
The development of advanced electron micros-
copy enables one to achieve through-depth
(defined in this paper as along the crystallo-
graphic c axis) imaging and carry out high-spatial
(subangstrom)–resolution electron energy–loss
spectrometry (EELS) experiments. Thus, it
allows a local probe of the intrinsic structural
and electronic properties in confined unit cells
(13). Here, we used scanning transmission
electron microscopy (STEM) and EELS tech-
niques to directly image the layer-by-layer lat-
tice configuration and electronic structure of the
Bi-family cuprates, covering an unprecedented
layer structure range with 1 ≤ n ≤ 9.
Imaging the atomic structure of the Bi-family
compounds with 1 ≤ n ≤ 9.
The evolution of Tc,max with n for the Bi-family
cuprates is indicated with a bold orange line in
Fig. 1A, which displays the universal bell-shaped
trend (14–19). Owing to the high spatial reso-
lution of the STEM technique, the layered
atomic structure can be clearly visualized, as
shown in Fig. 1B for an optimally doped Bi-
2223 sample. Each Cu atom in the inner CuO2
plane (IP) forms an in-plane CuO4 plaquette,
whereas in the outer CuO2 planes (OP), the
configuration is a CuO5 pyramid with one apical
oxygen (Fig. 1B, right). Some previous studies
(20, 21) have suggested that the different envi-
ronments of IP and OP may have a profound
influence on their superconducting properties,
but their layer-resolved electronic properties
remain to be unveiled.
Figure 1, C to K, displays a series of high-
quality cross-sectional STEM images, in which
Bi2Sr2Can-1CunO2n+4+d with 1 ≤ n ≤ 9 are all
identified. We obtained the data for n ≥ 3 by
carefully searching within a dozen Bi-2223 sam-
ples, which include very scarce regions that have
four or more CuO2 planes per unit cell. The
details about the search and discovery of the
n > 3 phases are shown in fig. S5. Because of
the quasi-two-dimensional nature of Bi-family
cuprates, it has been shown recently that all
the essential electronic properties are contained
within a unit cell, BiO/SrO charge reservoir
layers plus the CuO2 layers (22). The method-
ology used here opens a gate for studying
multilayer cuprates.
Measuring the charge-transfer gap of the
Bi-family compounds.
The electronic band structure of cuprates is
well characterized by the in-plane Cu 3dx2(cid:2)y2
and doped O 2px/2py orbitals (23). Within the
Zaanen-Sawatzky-Allen scheme (24), the un-
doped cuprate is a charge-transfer–type Mott
insulator, in which the lowest-energy excita-
tion is from the O charge-transfer band (CTB)
to the unoccupied upper Hubbard band (UHB)
of the Cu sites, which is schematically illus-
trated in Fig. 2A. Figure 2B shows the STM
spectra in three insulating Bi-family cuprates
with n = 1, 2, and 3, in which the D value be-
tween the edges of CTB and UHB can be
clearly extracted. The results here expand on
those of previous studies (9) to n = 3 and
reveal the same trend of a monotonic decrease
of D from 1.5 eV in Bi-2201 to 1.0 eV in Bi-2212
and 0.7 eV in Bi-2223.
Upon hole doping, an additional feature
emerges from the CTB that is known as the
Zhang-Rice singlet (ZRS) band (Fig. 2C). This
feature can be attributed to the hybridized
state of a doped hole involving one Cu-3d and
the four nearest O-2p orbitals (25). In this
circumstance, the main features in the STM
spectra are the spectral weight transfer from
high energy to low energy and the gradual for-
mation of the pseudogap (26); the D value
becomes less well defined. Nevertheless, the
energy difference between the centers of the ZRS
Wang et al., Science 381, 227–231 (2023)
14 July 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
Figure 1
Wang et al., Science 381, 227–231 (2023)
14 July 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. The Tc,max and atomically resolved crystal structure of Bi2Sr2Can-1CunO2n+4+d cuprates with 1 ≤ n ≤ 9. (A) Relationship between Tc,max and n for the
homologous series of n-layered cuprates (14–19) including Tl-12(n-1)n, Hg-12(n-1)n, Cu-12(n-1)n, and Bi-22(n-1)n. The bold orange line indicates the Tc,max as a
function of n in the Bi-family. The n = 9 phase has never been discovered, so the Tc,max value was obtained with linear extrapolation (as indicated with the black
dotted line) and represented by a hollow circle. (B) A zoomed-in area of the n = 3 sample (Bi-2223) with the schematic crystal structure shown at right. OP, outer
CuO2 plane; IP, inner CuO2 plane. (C to K) Cross-sectional layered structure in the Bi-family cuprates with 1 ≤ n ≤ 9, where the c axis direction is along the vertical
blue arrow. The yellow arrows indicate the positions of CuO2 planes in the crystal (fig. S4). Scale bar, 5 Å. All the data were taken at T = 300 K.
Fig. 2. Probing D with STM and STEM-EELS
techniques. (A) Schematic band structure of the
undoped cuprates. LHB, lower Hubbard band; CTB,
charge transfer band; UHB, upper Hubbard band.
(B) The Mott-insulator type dI/dV (differential
conductance) spectra of undoped Bi-2201 (n = 1),
Bi-2212 (n = 2), and Bi-2223 (n = 3) samples taken
with STM. The D here characterizes the distance
between the edges of CTB and UHB, as shown in (A).
(C) Schematic band structure of hole-doped cup-
rates. ZRS, Zhang-Rice singlet. The red and blue
arrows indicate the EELS excitation process from
core levels to the unoccupied UHB and ZRS,
respectively (the principle is shown in fig. S6).
(D) STEM-EELS of the O K edge from n = 1 to 3.
The D here characterizes the distance between the
centers of ZRS and UHB, as shown in (C). The
Gaussian functions were used to simulate the
ZRS (red dotted lines) and the UHB (purple, IP;
green, OP). (Left) The STEM images show the real-
space areas where the EELS are taken. All the
STEM-EELS data were taken at T = 300 K. a.u.,
arbitrary units.
and UHB is still a valid parameter that charac-
terizes the charge-transfer energy (27), which
varies little with doping (28, 29). The state-of-
the-art STEM-EELS is an ideal technique for
measuring such peak-to-peak CTG, where the
spectral peaks shown in Fig. 2D indicate the
large probability of exciting core-level electrons
to the unoccupied ZRS and UHB in Fig. 2C (27),
in analogy to the x-ray absorption spectroscopy
(XAS) at the O K edge (excitation energy of an
O 1s core electron to an empty state). In par-
ticular, the ZRS peak centered around 528.5 eV
(Fig. 2D, red dashed line) corresponds well with
the measurements of the bulk-sensitive XAS
technique (27). The ZRS peak is robust in cup-
rates against doping, temperature, and materials.
Similar to that of STM, the D obtained with
EELS is mainly controlled by the unoccupied
UHB, which displays a red-shift trend from
n = 1 to 3, as revealed in Fig. 2D. We extracted
the D size from a reliable fitting procedure as
described in (30), which renders D = 2.5 eV for
Bi-2201 and D = 2.0 eV for Bi-2212, respectively.
Owing to the high atomic-plane resolution, we
distinguished the CTG corresponding to the OP
(DOP = 1.7 eV) and IP (DIP = 1.5 eV) in trilayer
Bi-2223. The spectral weight ratio between the
ZRS and UHB is smaller in the IP, which is sug-
gestive of the lower hole concentration (27).
Correlating observables with the number of
CuO2 layers.
To investigate how the D size evolves with the
number of CuO2 layers per unit cell, we have
further determined the STEM-EELS of the O K
edge with 1 ≤ n ≤ 9 (Fig. 3, A and B). The
definition and description of different CuO2
planes where we took the STEM-EELS meas-
urements are shown in fig. S7. We repeated
the measurements 10 times on different areas
with the same n to improve statistical robust-
ness (full dataset is provided in figs. S8 to S17)
(30). For clarity, we only show the averaged
spectra for the OP layer; those for the IP layers
are shown in fig. S19. The overall trend is
shown in Fig. 3, C and D, which clearly dem-
onstrate that the n dependence of D is rep-
resented by an inverted bell-shaped curve with
the minimum D = 1.8 eV at n = 3. There is an
apparent anticorrelation between the n de-
pendence of D and Tc,max. For all the phases in
which n ≥ 3, the D value of the IP is smaller
than the D value of OP, but exhibits the same
dependence on n (Fig. 3D).
In the scenario of single-band Hubbard model,
we may define an effective superexchange energy
eff/D, where teff is the effective hop-
Jeff ~ 4t2
pd/D , and tpd
ping term that is proportional to t2
is the hopping integral between neighboring O
2p and Cu 3d orbitals (8). Because the Jeff
between local Cu moments has been proposed
to be responsible for the spin singlet pairing
(31), Jeff ~ 1/D3 thus correlates the CTG and the
Cooper pairing strength. For n = 1 and 2, only
the OP is present, and the increase of JOP is
consistent with that derived through STM (9)
and inelastic photon scattering (32). For the
whole range with 1 ≤ n ≤ 9, Fig. 3E shows that
the n dependence of Jeff for the OP layer
exhibits a bell shape highly analogous to that
of Tc,max. It reaches a peak at n = 3 and de-
creases for 4 ≤ n ≤ 9. The close correlation
between Jeff and Tc,max indicates that the D,
and thus the underlying strong correlation
effects, determine Tc,max, at least in the Bi-family
cuprates with 1 ≤ n ≤ 9.
The combined STEM-EELS and STM results
here touch on the central issue regarding the
bell-shaped evolution of Tc,max with n, which is
universal in all cuprate families. From the
theoretical point of view, a previous study (4)
Wang et al., Science 381, 227–231 (2023)
14 July 2023
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
has proposed that the enigmatic balance be-
tween Josephson tunneling and competing
electronic orders is the underlying physical
mechanism. Our results reveal that it is the
CTG, which displays an inverted bell-shaped
trend compared with the Tc,max that plays a
deterministic role in this empirical rule. This
conclusion is supported by the consistency of
D characterized by energy scales between band
centers (EELS and XAS) and edges (STM). We
summarize the CTG sizes taken with different
experimental probes in Fig. 4. The difference
between STM and EELS shows a rigid shift of
~1 eV, which is caused by the finite bandwidths
and provides useful information about the
Fig. 3. The evolution of D with n in Bi-family cuprates. (A) STEM-EELS with a
higher energy range of the O K edge with 1 ≤ n ≤ 9. For each curve, we averaged
the data measured on the OP layer of 10 different areas with the same n (full
datasets are provided in figs. S8 to S17). (B) Zoomed-in STEM-EELS of (A). The
Gaussian functions were used to determine the position of ZRS and UHB (fig.
S18). (C and D) The statistical results of D for the OP and IP with 1 ≤ n ≤ 9, which
was determined with 10 measurements taken under the same experimental
conditions. All the data were taken at T = 300 K. Each dot and error bar in (D)
are the calculated average and standard deviation of the result in (C) for the
same n. (E) The effective superexchange Jeff ~ 1/D3 as a function of n, assuming
the same teff for the OP and IP of different n in a homologous series. a.u.,
arbitrary units.
Fig. 4. The correlation between D, n, and Tc,max. (A and B) Summarized plots of (A) D versus n and (B) D versus Tc,max for the Bi-family cuprates. Our STM and
EELS (on the OP) data, as well as previous XAS data, are shown together. The STM results are generally ~1 eV smaller than the EELS and XAS values, owing to the
finite bandwidth or hopping integral t. The wide gray-red and blue-green strips are guides to the eye. See table S1 for more details.
Wang et al., Science 381, 227–231 (2023)
14 July 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
effective hopping integral in the Bi-family
cuprates.
Discussion and outlook
Our study shows that the CuO2 planes with
different apical environments have very dif-
ferent D, which is seen not only in the evo-
lution with n but also in that the IP always
has a smaller D size than the OP of the same
compound. The Jeff of IP is thus larger than
that of OP, as shown in Fig. 3E, which is con-
sistent with previous studies of the pairing
strength (21, 33). Therefore, the ultimate reason
why D, and hence Tc,max, varies substantially for
different n in a homologous series is most likely
related to the orbitals out of the CuO2 plane,
indicating an interplay between the under-
lying electronic structure and orbital param-
eter symmetry in cuprates (4, 34, 35).
RE FE RENCE AND NOTES
1. B. Keimer, S. A. Kivelson, M. R. Norman, S. Uchida, J. Zaanen,
Nature 518, 179–186 (2015).
2. C. A. Reynolds, B. Serin, W. H. Wright, L. B. Nesbitt, Phys. Rev.
78, 487 (1950).
3. E. Maxwell, Phys. Rev. 78, 477 (1950).
4. S. Chakravarty, H.-Y. Kee, K. Völker, Nature 428, 53–55
(2004).
5. C. W. Zou et al., Nat. Phys. 18, 551–557 (2022).
6. T. Fujii, T. Watanabe, A. Matsuda, Physica C Supercond
357–360, 173–176 (2001).
7. Z. Hao et al., Phys. Rev. Lett. 125, 237005 (2020).
8. P. A. Lee, N. Nagaosa, X. G. Wen, Rev. Mod. Phys. 78, 17–85
(2006).
9. W. Ruan et al., Sci. Bull. 61, 1826–1832 (2016).
10. C. Weber, C. Yee, K. Haule, G. Kotliar, Europhys. Lett. 100,
AC KNOWLED GME NTS
37001 (2012).
11. N. Kowalski, S. S. Dash, P. Sémon, D. Sénéchal, A.-M. Tremblay,
Proc. Natl. Acad. Sci. U.S.A. 118, e2106476118 (2021).
12. Z. H. Cui et al., Phys. Rev. Res. 2, 043259 (2020).
13. D. A. Muller et al., Science 319, 1073–1076 (2008).
14. H. Mukuda et al., J. Phys. Soc. Jpn. 81, 011008 (2012).
15. C. W. Chu, L. Z. Deng, B. Lv, Physica C 514, 290–313
(2015).
16. Bismuth-Based High-Temperature Superconductors, H. Maeda,
Ed. (CRC Press, 1996).
17. H. Wang et al., Appl. Phys. Lett. 57, 710–711 (1990).
18. H. Narita, T. Hatano, K. Nakamura, J. Appl. Phys. 72,
19.
5778–5785 (1992).
I. Bozovic, J. N. Eckstein, G. F. Virshup, Physica C 235, 178–181
(1994).
20. C. Zou et al., Phys. Rev. Lett. 124, 047003 (2020).
21. S. Kunisada et al., Science 369, 833–838 (2020).
22. Y. Yu et al., Nature 575, 156–163 (2019).
23. V. J. Emery, Phys. Rev. Lett. 58, 2794–2797 (1987).
24. J. Zaanen, G. A. Sawatzky, J. W. Allen, Phys. Rev. Lett. 55,
418–421 (1985).
25. F. C. Zhang, T. M. Rice, Phys. Rev. B Condens. Matter 37,
3759–3761 (1988).
26. P. Cai et al., Nat. Phys. 12, 1047–1051 (2016).
27. C. T. Chen et al., Phys. Rev. Lett. 66, 104–107 (1991).
28. Y. Zhong et al., Phys. Rev. Lett. 125, 077002 (2020).
29. C. Hu et al., Nat. Commun. 12, 1356 (2021).
30. Materials and methods and additional data are available as
supplementary materials.
31. P. W. Anderson, Science 235, 1196–1198 (1987).
32. L. Wang et al., Nat. Commun. 13, 3163 (2022).
33. H. Kotegawa et al., Phys. Rev. B Condens. Matter 64, 064515
(2001).
34. E. Pavarini, I. Dasgupta, T. Saha-Dasgupta, O. Jepsen,
O. K. Andersen, Phys. Rev. Lett. 87, 047003 (2001).
35. T. Xiang, J. M. Wheatley, Phys. Rev. Lett. 77, 4632–4635
(1996).
36. Z. C. Wang et al., Source Data for: Correlating the charge
transfer gap to the maximum transition temperature in
Bi2Sr2Can-1CunO2n+4+d, version 1, Zenodo (2023); https://doi.
org/10.5281/zenodo.7608360.
We gratefully acknowledge Q. K. Xue for helpful discussions and
assistance. Funding: This work was financially supported by the
Chinese National Natural Science Foundation (Basic Science
Center Project of NSFC, grant no. 52388201), the Basic and
Applied Basic Research Major Programme of Guangdong Province,
China (grant no. 2021B0301030003), and Jihua Laboratory
(project no. X210141TL210). Y.W. is supported by the Innovation
Program for Quantum Science and Technology (grant no.
2021ZD0302502) and the New Cornerstone Science Foundation
through the New Cornerstone Investigator Program and the
XPLORER PRIZE. X.Z. is supported by the Chinese National
Natural Science Foundation of China (grant no. 11888101), the
National Key Research and Development Program (grant no.
2021YFA1401800), and the Strategic Priority Research Program
(B) of the Chinese Academy of Sciences (grant no. XDB25000000).
This work used the resources of the National Center for Electron
Microscopy in Beijing. Author contributions: J.Z. and Z.W. initiated
the idea and designed the studies. Z.W. performed the electron
microscopy experiments and data analysis with the help of J.Z. C.L.
grew the crystals of Bi-2223. C.Z., X.L., H.Y., C.Y. and X.Z. processed
the samples. C.Z. and Y.W. performed the STM experiment. All
authors contributed to the scientific discussions. Y.X. offered some
theoretical guidance on the DFT calculation. Z.W., C.Z., Y.W., and
J.Z. wrote the paper with contributions from all authors. Competing
interests: The authors declare that they have no competing interests.
Data and materials availability: All data needed to evaluate the
conclusions in the paper are deposited in a public database of Zenodo
(36). License information: Copyright © 2023 the authors, some
rights reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government works.
https://www.science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add3672
Materials and Methods
Figs. S1 to S20
Table S1
References (37–47)
Submitted 7 June 2022; accepted 6 June 2023
10.1126/science.add3672
Wang et al., Science 381, 227–231 (2023)
14 July 2023
5 of 5
|
10.1126_science.adc9150
|
Corrected 23 January 2023. See full text.
RES EARCH
R E S E A R C H A R T I C L E
◥
NANOFLUIDICS
Neuromorphic functions with
a polyelectrolyte-confined fluidic memristor
Tianyi Xiong1,2, Changwei Li1,3, Xiulan He1, Boyang Xie1,2, Jianwei Zong1,2, Yanan Jiang4, Wenjie Ma1,
Fei Wu1, Junjie Fei3, Ping Yu1,2*, Lanqun Mao1,4*
Reproducing ion channel–based neural functions with artificial fluidic systems has long been an
aspirational goal for both neuromorphic computing and biomedical applications. In this study,
neuromorphic functions were successfully accomplished with a polyelectrolyte-confined fluidic
memristor (PFM), in which confined polyelectrolyte–ion interactions contributed to hysteretic ion
transport, resulting in ion memory effects. Various electric pulse patterns were emulated by PFM with
ultralow energy consumption. The fluidic property of PFM enabled the mimicking of chemical-regulated
electric pulses. More importantly, chemical-electric signal transduction was implemented with a
single PFM. With its structural similarity to ion channels, PFM is versatile and easily interfaces with
biological systems, paving a way to building neuromorphic devices with advanced functions by
introducing rich chemical designs.
T he development of artificial systems with
brainlike functions (i.e., neuromorphic
devices) is rapidly expanding because of
their promising applications in neuro-
morphic computing (1, 2), bioinspired
sensorimotor implementation (3, 4), brain–
machine interfaces (5, 6), and neuroprosthet-
ics (7, 8). So far, neuromorphic functions with
diverse patterns have been achieved and
incorporated into applications in various ways,
mainly with history-dependent solid-state
resistance-switchable devices, including two-
terminal memristors (9–11) and three-terminal
transistors (12, 13). However, most of the
neuromorphic functions achieved thus far are
based on the emulation of the electric pulse
pattern using solid-state devices. An analog to
the biological synapse—especially the emula-
tion of a chemical synapse in a solution-based
context—remains very challenging with these
solid-state devices. In this regard, a fluidic-
based memristor is highly desirable to achieve
neuromorphic functions in an aqueous envi-
ronment, because of its superior compatibil-
ity with biological systems and the larger
number of functions endowed to the neuro-
morphic devices by introducing diverse chem-
istries (14).
Previous attempts have revealed that ion-
based micro- or nanofluidic devices with
1Beijing National Laboratory for Molecular Sciences, Key
Laboratory of Analytical Chemistry for Living Biosystems,
Institute of Chemistry, Chinese Academy of Sciences (CAS),
Beijing 100190, China. 2School of Chemical Sciences, University
of Chinese Academy of Sciences, Beijing 100049, China.
3Key Laboratory of Environmentally Friendly Chemistry and
Applications of Ministry of Education, College of Chemistry,
Xiangtan University, Xiangtan 411105, China. 4College of
Chemistry, Beijing Normal University, Beijing 100875, China.
*Corresponding author. Email: yuping@iccas.ac.cn (P.Y.);
lqmao@bnu.edu.cn (L.M.)
advanced functionalities [e.g., the ion diode
(15), ion transistor (16), or ion switch (17)]
are achievable by confining an electrolyte
into micro- or nanochannels. Several studies
reported that these confined systems feature
memresistance and memcapacitance (18–20).
Moreover, long-term plasticity was obtained
with nanochannels by introducing an ionic
liquid–electrolyte interface (21). Despite these
efforts, realizing neuromorphic functions in
aqueous media is still a long-standing chal-
lenge, mainly because the strong shielding
effect in an aqueous environment greatly
hinders interionic interactions, thereby lim-
iting the formation of memory in fluidic-based
systems. In 2021, a milestone theoretical
model predicted that ion memory functions
could be accomplished with two-dimensional
extremely confined channels (22), which has
been experimentally realized by the same
group (23).
Here, we report a polyelectrolyte-confined
fluidic memristor (PFM) that can successfully
accomplish various neuromorphic functions
for mimicking not only electric pulse patterns
but also chemical-electric signal transduction.
Inspired by biological ion channels that func-
tion as natural memristors by controlling ion
flux with spatial confinement and molecular
recognition (24) (Fig. 1A), we designed and
fabricated a polyimidazolium brush (PimB)–
confined fluidic channel (Fig. 1B). We se-
lected polyimidazolium because of its high
charge density, rich chemistry, and versatile
ability to recognize different anions (25). Typ-
ically, PimBs were grown onto the inner wall
of the glass micro- or nanopipette by surface-
initiated atomic transfer radical polymeriza-
tion (26) (fig. S1, A and B). In this way, the
fluid was confined by PimBs, in which the
establishment of anion concentration equilib-
rium and charge balance between the inside
and outside of PimBs under the stimulation of
electric fields or chemicals would be hysteretic,
resulting in history-dependent ion memory.
Polyelectrolyte-confined fluidic memristor
The device was constructed by a PimB-confined
conically fluidic channel, an electrolyte, and,
to complete the electric connection, two Ag/
AgCl electrodes (fig. S1C). A triangle wave volt-
age was applied to the device to investigate the
current-voltage (I-V) relationship. The recti-
fied I-V curve was recorded with a modified
micropipette (Fig. 1C, red) because of the geo-
metric asymmetry of the conical channel and
the anion selectivity of the positively charged
PimB (26). Meanwhile, the pinched I-V curve
collected under this periodic voltage with a
nonzero cross-point voltage (Vcp) satisfied
the history-dependent memristor nature ac-
cording to Chua’s theory (27). This offset (Vcp)
originates from the influence of surface charge
in this asymmetric channel (28), which is
usually observed for biological memristors,
such as the K+ ion channel (29). In contrast,
the bare micropipette only yields the linear
ohmic I-V curve (Fig. 1C, blue), demonstrating
the essential role of PimB in this pinched
hysteretic loop.
We further investigated the dependence of
current on voltage scanning frequency (i.e.,
scan rates, n). The I-V curve experiences a
transition from the hysteretic and rectified
form at lower scan rates to the linear-like form
at higher scan rates (Fig. 1D and fig. S2). The
area (S) inside the hysteresis loop shrinks as
the scan rate increases and degenerates to
zero at infinite scan rate, as suggested by the
fitting of the S-n relationship (Fig. 1E and
supplementary text). The pinched hysteresis
loop, reduced loop area with increasing fre-
quency, and the linear I-V relationship under
infinite frequency satisfy three fingerprints of
a memristor (29).
To explore the origin of memristive be-
havior of the PFM, we conducted dynamic
monitoring of ion conductivity under dif-
ferent constant bias voltages (Fig. 1F). At a
voltage equal to Vcp (53 mV), PFM maintains
a static ion conductivity over time (Fig. 1F,
yellow). At a voltage higher than Vcp (200 mV),
the conductivity gradually increases to reach a
plateau in ~2 s (Fig. 1F, red), whereas it de-
creases to leveling in 1 s at a voltage lower than
Vcp (−200 mV) (Fig. 1F, blue). These results
confirm that the ion conductivity change of
PFM is a time-dependent process.
The conductivity change is closely asso-
ciated with ion dynamic distribution inside
the channel. We carried out finite element
modeling (FEM) to describe the changes of
ion distribution over time under different bias
voltages (supplementary text and fig. S3A).
Xiong et al., Science 379, 156–161 (2023)
13 January 2023
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
Corrected 23 January 2023. See full text.
Fig. 1. Conductivity changes
of PFM. (A and B) Schematic
illustration of the neural
functions realized by interac-
tion-gated ion current in
biological neurons (A) and a
PimB-confined fluidic system
(B). (C) I-V curves of PFM (red)
and a bare micropipette (blue)
in 10 mM KCl aqueous solution
under a triangle wave with a
scan rate of 50 mV/s. The
hysteresis loop area is shaded
in purple. The arrows show the
scan direction. (D) I-V curves of
PFM in 10 mM KCl under
triangle waves with a fast (10 V/s,
red) and a slow (10 mV/s,
blue) scan rate. (E) Plot of the
hysteresis loop area (S) with scan
rate (n). (Inset) Zoom-in of the
plot at low-scan rate. (F) Time-
dependent conductivity changes
of PFM in 10 mM KCl under a
constant bias voltage equal to Vcp
(53 mV, yellow), higher than Vcp
(+200 mV, red), and lower
than Vcp (−200 mV, blue).
(G) Simulated I-t curves of PFM
in 10 mM KCl under +1 V (top)
and −1 V (bottom) bias voltages.
(H) Schematic illustration of
time-dependent ion redistribution
processes in PFM at potentials
higher than Vcp (from I to III) and
lower than Vcp (from I to V).
/
/
Without a bias voltage, cation (K+) and anion
(Cl−) concentrations in the bulk layer are bal-
anced to maintain charge neutrality, which
are equal to the bulk electrolyte concentration
(fig. S3B). While in the positively charged
PimB layer, Cl− ions overwhelm K+ ions with
a high concentration (~104.5 mM) owing to
the strong electrostatic attraction between
anions and the imidazolium moieties (fig.
S3B). This high surface charge density of the
PimB confers large anion storage, and the
anions in PimB have relatively slow diffusion
dynamics, necessitating prolonged enrichment
of Cl− in the PimB layer from the rest state to
reach steady ion distribution at +1 V (Fig. 1H
and fig. S3C). Therefore, the ion current under-
goes a gradual increase to its steady state (Fig.
1G, red). Similarly, the slowed depletion of Cl−
in the PimB layer at −1 V (Fig. 1H and fig. S3D)
results in the gradual decrease of ion current
(Fig. 1G, blue). These simulated consequences
match well the temporal profiles of ion con-
ductivity at constant bias potentials (Fig. 1F),
suggesting that the relatively slow diffusion
dynamics of anions into and out of the PimBs
leads to time-dependent ion memory and,
consequently, the memristive behavior of
the device.
Mimicking short-term plasticity patterns
with PFM
To mimic short-term plasticity (STP) electric
pulse patterns, we applied paired voltage pulses
to the PFM and recorded current spikes in ac-
cordance with pulsed stimulations. As shown,
two continuous pulses of +2 or −2 V induce a
current increase (DI = 8.9 nA), called a paired-
pulse facilitation (PPF; Fig. 2A), or a signifi-
cant current decline (DI = −48.4 nA), called a
paired-pulse depression (PPD; Fig. 2B), validat-
ing the capability of PFM in emulating the STP
electric pulses. FEM simulation of ion dynam-
ics under the same pulsed voltage waveforms
reproduces the similar trends of current change
(Fig. 2C), further demonstrating that STP elec-
tric pulses originated from time-dependent ion
redistribution in PFM. The applied voltage
drives the ion concentration polarization in
the PimB layer that gives rise to the observed
current spike. Upon removal of the external
electric field (i.e., at the pulse interval), slow
anion diffusion dynamics in the PimB layer
would briefly hold the ion concentration po-
larization state owing to the strong interaction
between Pim and anions. Hysteretic ion redis-
tribution during the pulse interval continues to
influence the anion enrichment or depletion in
Xiong et al., Science 379, 156–161 (2023)
13 January 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
Corrected 23 January 2023. See full text.
Fig. 2. STP electric pulses of PFM. PPF (A) and PPD (B) of PFM in
10 mM NaCl. The upper plots show voltage pulse waveforms. (C) The
simulated I-t responses with FEM under the same voltage waveforms in (A)
(red) and (B) (blue). (D) Plots of current changes (DI) with the paired
pulse intervals (Dt) under positive (V = +2V, tp = 10 ms, red) and negative
(V = −2V, tp = 10 ms, blue) bias voltages. Error bars present standard
deviation of three measurements with the same device. (E) Current
responses under voltage pulse train (V = −2 V, tp = 5 ms) with varying
frequency from 1 to 100 Hz. (F) Plot of conductivity versus the applied voltage
pulse frequency.
the PimB layer when a voltage pulse was ap-
plied again, leading to the conductivity (or ion
current) level up or down of the next spike,
thus enabling the mimicking of STP electric
pulses with the PFM. This conductivity change
is strongly related to the pulse interval (Dt)
following a biexponential relationship
(cid:1) (cid:3)
(cid:1) (cid:3)
DI ¼ A1e
(cid:2) Dt
t1
(cid:2) Dt
t2
þ A2e
ð1Þ
where t1 and t2 are two time constants that
might be related to the ion redistribution in
the bulk and PimB layer, respectively; and A1
and A2 are the weights of these two dynamic
processes. Shortening of the voltage pulse in-
terval (Dt) leads to increased current change
(DI), whereas a prolonged voltage pulse interval
results in reduced DI. This decline dynamics
emulates the change of synaptic weight in
shaping biological STPs under repeated stim-
uli of varying frequencies (30). We then cal-
culated the retention time (tr) of PFM, defined
as the pulse interval where DI drops to 5% of
Þ. For a cer-
its maximum DIDt¼tr
tain voltage pulse stimulation (e.g., V = ±2 V,
tp = 10 ms), tr was calculated to be ~500 ms
(Fig. 2D and table S1; tr = 484.36 ms for facil-
¼ 5%DIDt¼0
ð
itation and tr = 501.22 ms for depression) in
10 mM NaCl aqueous solution. These values
of PFM are comparable to those of STP in bio-
logical systems (table S1), where facilitation
and depression of the synaptic weight occurs
on a 102-ms scale (30).
We further validated the ability of the PFM
to emulate dynamic filtering functions in sen-
sory neurons. By applying a negative voltage
pulse train with the pulse frequency ranging
from 1 to 100 Hz (V = −2 V, tp = 5 ms; Fig. 2E),
frequency-dependent conductivity was ob-
tained (Fig. 2F). The minor difference in the
frequency of the voltage pulse train (e.g., 19,
20, or 21 Hz) could be differentiated by the
conductivity differences with this filtering
function (fig. S4). In addition, programming
voltage pulses (Vset = +2 V, Vreset = −0.6 V)
causes an analog ion conductivity switch
among 100 continuous states. The device
maintains good performance after 30,000 set-
reset tests (fig. S5), demonstrating the endur-
ance of the PFM. Energy consumption (W) of
the PFM was calculated on the basis of the
integration of I-t responses under single volt-
age pulses with W ¼ ∫VIdt, which was closely
related to the orifice size of the pipette and the
applied voltage (fig. S6, A and B). For a 150-nm-
diameter nanopipette-based device, energy
consumption under voltage stimulation of
−100 mV (tp = Dt = 10 ms) is 0.66 pJ per spike
(fig. S6C), which is close to the biological volt-
age (−70 mV) and energy consumption (31).
In contrast to most of the reported memristors
that require high voltages, our PFM can oper-
ate at the voltage and energy consumption as
low as those biological systems (table S2), dem-
onstrating its potential for application in bio-
inspired sensorimotor implementation and
neuroprosthetics.
Chemical-regulated STP electric pulses
Neurons work in a complex chemical envi-
ronment wherein ions and molecules lay the
foundations for all neuroactivities. Changes in
the chemical environment fundamentally con-
tribute to manifold behaviors of neurons,
including synaptic scaling induced by N-methyl-
D-aspartate (32) and enhanced transmission
induced by neurotrophins (33). This chemical
modulation effect was emulated with the PFM
by tuning the Pim–anion interactions. The I-V
curves collected in different electrolyte solu-
tion (NaCl, NaBF4, or NaClO4) illustrate the
dependence of hysteresis loop area (S) on the
species of anions (Fig. 3A). We then investigated
Xiong et al., Science 379, 156–161 (2023)
13 January 2023
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
Corrected 23 January 2023. See full text.
w/o ATP
w/o ATP
w/o ATP
w/o ATP
Fig. 3. Chemical-regulated STP electric pulses. (A) I-V curves of PFM in
10 mM NaCl, NaBF4, or NaClO4 solution with a scan rate of 50 mV/s. (Inset) The
hysteresis loop area (S) from the I-V curves. (B and C) PPF (B) and PPD (C)
of PFM in different electrolyte solutions. (D) Plots of current changes with
pulse intervals in different electrolyte solution (V = −2 V, tp = 10 ms). (Inset)
Schematic illustration of free anions and PimA pairs. (E) Plots of normalized
current changes with pulse intervals in KCl solutions with different concentrations
(V = −2 V, tp = 10 ms). (Inset) Zoom-in of the plot at short pulse intervals.
(F) Current spikes (top) and conductivity changes (bottom) of PFM in phosphate-
buffered saline (pH 7.4) with (blue) and without (red) 1 mM ATP under a 10-pulse train
(V = +0.5 V, Dt = tp = 10 ms). (G) Plots of current changes (DI = I10 − I1) under a
10-pulse train (V = +0.5 V, tp = 10 ms) with pulse intervals. (Inset) Values of retention
time with (blue) and without (red) 1 mM ATP in solution. Error bars in (D), (E), and
(G) present standard deviation of three measurements with the same device.
the chemical regulation of STP by different
anions (V = ±2 V, Dt = tp = 10 ms). Compared
with that in NaCl aqueous solution (Fig. 3, B
and C, red), the PFM in NaBF4 solution ex-
hibits stronger PPF and slightly attenuated
PPD (Fig. 3, B and C, blue). Both effects were
attenuated in NaClO4 solution (Fig. 3, B and
C, yellow). Moreover, tr is related to the chem-
ical environment, as revealed by the biexpo-
nential dynamic curve with different anion
species (Fig. 3D and fig. S7A). The value of
tr decreases in the order of tr;ClO(cid:2)
<
tr;Cl(cid:2) both for positive and negative voltage
pulse stimulation (table S1). In addition, the
STP performance of the PFM could be regu-
lated by the ionic strength. High-concentration
electrolyte accelerates the disappearance of
STP (the lower tr) under either positive or
negative voltage pulse stimulation (Fig. 3E,
fig. S7B, and table S1).
< tr;BF(cid:2)
4
4
To provide in-depth insights into the mech-
anism of this dependence between ion species
or concentration and STP performance of the
PFM, we used the Dukhin number at the
pipette tip (Du0) to evaluate the surface con-
ductivity changes in the system. Larger Du0
indicates that there are more anions partici-
pating in the ion redistribution to intensify
memristive effect, and vice versa. The influence
of Pim–anion interactions on Du0 at steady
state without bias voltages is described by the
following equation
Du0 ¼ KCA
(cid:5)(cid:2)1
(cid:4)
(cid:2)1 1 þ
k2
k1
CA
ð2Þ
where, K is a structural constant of the PFM,
CA is the anion concentration of electrolyte,
and k1 and k2 are the dissociation and asso-
ciation kinetic constants, respectively (see sup-
plementary text). Here, the value of k2/k1 is
closely related to the hydration energy of anions
(25). For anions with a larger hydration energy,
the value of k2/k1 would be lower, yielding a
larger value of Du0. The Cl− bears the highest
hydration energy ðDH 0
<
Hyd;Cl(cid:2) Þ and thus the largest value of Du0.
DH 0
That is, more free Cl− counterions in the
PimB layer participate in the redistribution
process (Fig. 3D, inset, top), resulting in longer
< tr;Cl(cid:2) Þ. For
retention time tr;ClO(cid:2)
ions with lower hydration energy, such as
−, the increased k2/k1 value
BF4
results in a smaller Du0. This means that more
− and ClO4
< tr;BF(cid:2)
Hyd;ClO(cid:2)
4
< DH 0
Hyd;BF(cid:2)
4
(cid:6)
4
4
− and ClO4
− counterions in the PimB layer
BF4
exist in the form of imidazolium–anion (PimA)
pairs than do Cl− counterions, resulting in a
decrease of mobile anions participating in the
ion redistribution in the PFM (Fig. 3D, inset,
bottom, and supplementary text), and conse-
quently the shorter retention time. Increasing
the ion concentration also leads to the Du0
decrease according to Eq. 2, explaining the
change of retention time with ion strength.
These chemical-regulated STP changes hold
promise for the possibility of realizing neuro-
morphic functions with the synergism of mul-
tiple ion species, which is almost impossible
for solid-state systems.
We further exploited bioactive molecules
to modulate STP patterns in a complex en-
vironment stimulated by mild voltage pulses.
In a physiological electrolyte (i.e., phosphate-
buffered saline solution, pH 7.4), the PFM
maintains its STP characteristic under the
stimulation of 10 voltage pulses (V = +0.5 V,
Dt = tp = 10 ms) (Fig. 3F, red), validating the
PFM in biological environment. More impor-
tantly, when 1 mM adenosine triphosphate
(ATP) was added into the solution, reduced
conductivity changes were observed (Fig. 3F),
Xiong et al., Science 379, 156–161 (2023)
13 January 2023
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
Corrected 23 January 2023. See full text.
Fig. 4. Chemical-electric signal transduction of PFM. (A) Schematic illustration of the chemical-electric
transduction of PFM (top) and biological synapses (bottom). (B) Electric pulse response of PFM under the stimulation
−. (C) Zoom-in of a single electric pulse. (D) Schematic illustration
of ClO4
of the ion distribution changes in the PFM when chemical stimuli were delivered into the system.
−. The blue arrows indicate the delivery of ClO4
along with a change in retention time from
156 ms to 138 ms (Fig. 3G). This phenomenon
can also be well understood by Eq. 2 owing to
the strong interaction of ATP with Pim (34).
As indicated, the PFM may allow direct inter-
facing and communicating with biological
systems, given that its neuroplastic behaviors
are controllable by bioactive molecules.
Chemical-electric signal transduction
with PFM
For biological systems, signal transduction at
chemical synapses is mediated by the release
and recognition of neurotransmitters (35)
(Fig. 4A). Such a process is almost impossible
to emulate with solid-state memristors, which
hardly respond to external chemical stimula-
tions. For fluidic-based devices with easily
adaptable configurations, however, this chemical-
electric signal transduction may be realized
by tuning the behaviors of multiple ion spe-
cies in the PimB-confined channel.
To achieve this target, a capillary controlled
by a microinjector pump was inserted into the
inner solution of the PFM, serving as the pre-
synaptic neuron for the delivery of chemical
− was chosen as an “arti-
stimuli. Herein, ClO4
ficial neurotransmitter” for demonstration (Fig.
−)
4A, top). When the chemical stimulus (ClO4
was back-injected into the micropipette, a re-
sponsive electric pulse was observed (Fig.
4B), analogous to neural spikes induced by a
neurotransmitter. Under a negative bias volt-
age (−1 V), the PFM holds at a low conductive
state owing to the low anion concentration
in the PimB layer, emulating the resting state
− anions were re-
of neurons. When the ClO4
leased from the capillary and transported
to the sensitive tip region driven by electro-
phoresis and convective flow, formation of
− pairs due to the stronger interac-
Pim-ClO4
− would decrease
tion between Pim and ClO4
the effective surface charge density, further
hindering the depletion of anions in PimB
layers, resulting in the increase of ion cur-
rent under negative voltage (25, 26). This
phenomenon is similar to the opening of
postsynaptic ion channels activated by neu-
rotransmitters. Then, electrophoresis and
electroosmotic flow drive Cl− ions moving
toward the tip, and the subsequent dissoci-
− pairs brings the current
ation of Pim-ClO4
back to the initial state, emulating the clearance
of transmitters (Fig. 4, C and D). In comparison,
no current spike occurs after injection of pure
NaCl solution with the same stimulation time
(fig. S8A). The PFM shows the capability of
individually accomplishing transduction from
chemical stimuli of certain species to electric
pulse signals.
Moreover, the ion current pulses induced by
chemical stimulation show typical spiking la-
tency behavior. A time lapse was observed be-
tween the stimulation and the occurrence of
neuromorphic spikes (Fig. 4, B and C). When
stimulation intensity increases, the time lapse
decreases accordingly (fig. S8, B and C). The
spiking latency behavior in real neurons plays
a key role in encoding input strength with
spike timing (36), this result thus provides
the possibility for encoding chemical stimula-
tions based on PFM.
Discussion
In this work, we have experimentally demon-
strated a fluidic memristor with neuromorphic
functions by using polyelectrolyte-confined
fluidic structure, which features the typical
fingerprints of a memristor. The time-dependent
ion redistribution controlled by Pim–anion
interactions under spatial confinement con-
tributes to the ion memory. The as-fabricated
PFM features the powerful ability to mimic
STP electric pulse patterns with retention time
and energy consumption comparable to those
of the ion channels in biological systems. More
importantly, the fluidic-based ion redistribu-
tion dynamics can endow the PFMs with neu-
romorphic function versatility that is hardly
achievable with solid-state devices, offering
the opportunity to introduce specific chemical
regulation pathways to neuromorphic func-
tions. Even more impressively, the emulation
of chemical-electric signal transduction can be
accomplished with this device. Compared with
neuromorphic devices based on other mecha-
nisms, our fluidic-based device offers not only
performances comparable to biological systems
but also more advanced neuromorphic func-
tionalities, especially chemical-related func-
tions (table S2).
Although the as-presented PFM features a
series of advantages—diversity in neuromor-
phic functions, the possibility of regulation
and coexistence of multiple ion carriers, and con-
venient interfacing with biological systems—
big challenges remain on the way toward real-
izing broader applications for PFMs. For ex-
ample, realizing long-term plasticity functions
is a key goal for fluidic-based systems, where
the introduction of much stronger (even irre-
versible) interfacial recognition interactions
(e.g., aptamers toward substrates) would be
potentially helpful for prolonging ion memory.
The scale-up of fluidic memristors for in-
memory computing is another challenge,
for which porous micro- or nanofluidic arrays
might offer a solution.
Xiong et al., Science 379, 156–161 (2023)
13 January 2023
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
Corrected 23 January 2023. See full text.
RE FE RENCES AND N OT ES
1. Z. Wang et al., Nat. Rev. Mater. 5, 173–195 (2020).
2. M. A. Zidan, J. P. Strachan, W. D. Lu, Nat. Electron. 1, 22–29
(2018).
3. Y. Lee et al., Sci. Adv. 4, eaat7387 (2018).
4. X. Zhang et al., Nat. Commun. 11, 51 (2020).
5. S. T. Keene et al., Nat. Mater. 19, 969–973 (2020).
6. T. Wang et al., Nat. Electron. 5, 586–595 (2022).
7. Y. Kim et al., Science 360, 998–1003 (2018).
8. S. Kim et al., Sci. Adv. 7, eabe3996 (2021).
9. D. B. Strukov, G. S. Snider, D. R. Stewart, R. S. Williams, Nature
453, 80–83 (2008).
10. Z. Wang et al., Nat. Mater. 16, 101–108 (2017).
11. Q. Xia, J. J. Yang, Nat. Mater. 18, 309–323 (2019).
12. Y. van de Burgt et al., Nat. Mater. 16, 414–418 (2017).
13. E. J. Fuller et al., Science 364, 570–574 (2019).
14. C. E. G. Hoskin, V. R. Schild, J. Vinals, H. Bayley, Nat. Chem. 14,
650–657 (2022).
15. C. Wei, A. J. Bard, S. W. Feldberg, Anal. Chem. 69, 4627–4633
(1997).
16. R. A. Lucas, C.-Y. Lin, L. A. Baker, Z. S. Siwy, Nat. Commun. 11,
1568 (2020).
17. M. R. Powell, L. Cleary, M. Davenport, K. J. Shea, Z. S. Siwy,
Nat. Nanotechnol. 6, 798–802 (2011).
18. D. Wang et al., J. Am. Chem. Soc. 134, 3651–3654 (2012).
19. Q. Sheng, Y. Xie, J. Li, X. Wang, J. Xue, Chem. Commun. 53,
6125–6127 (2017).
20. J. S. Najem et al., ACS Nano 12, 4702–4711 (2018).
21. P. Zhang et al., Nano Lett. 19, 4279–4286 (2019).
22. P. Robin, N. Kavokine, L. Bocquet, Science 373, 687–691 (2021).
23. P. Robin et al., arXiv:2205.07653 [cond-mat.soft] (2022).
24. V. M. Ho, J.-A. Lee, K. C. Martin, Science 334, 623–628 (2011).
25. X. He et al., Angew. Chem. Int. Ed. 57, 4590–4593 (2018).
26. X. He et al., J. Am. Chem. Soc. 139, 1396–1399 (2017).
27. L. Chua, IEEE Trans. Circuit Theory 18, 507–519 (1971).
28. D. Wang et al., ChemElectroChem 5, 3089–3095 (2018).
29. L. Chua, G. Ch. Sirakoulis, A. Adamatzky, Eds., Handbook of
Memristor Networks (Springer, 2019).
30. R. S. Zucker, W. G. Regehr, Annu. Rev. Physiol. 64, 355–405
(2002).
31. S. B. Laughlin, R. R. de Ruyter van Steveninck, J. C. Anderson,
Nat. Neurosci. 1, 36–41 (1998).
32. D. Stellwagen, R. C. Malenka, Nature 440, 1054–1059 (2006).
33. H. Thoenen, Science 270, 593–598 (1995).
34. K. Zhang et al., Sci. China Chem. 63, 1004–1011 (2020).
35. A. E. Pereda, Nat. Rev. Neurosci. 15, 250–263 (2014).
36. E. M. Izhikevich, IEEE Trans. Neural Netw. 15, 1063–1070
(2004).
37. T. Xiong et al., Data for: Neuromorphic functions with a
polyelectrolyte-confined fluidic memristor, Dryad (2022);
https://doi.org/10.5061/dryad.d2547d85r.
ACKN OWLED GMEN TS
We are grateful to the anonymous reviewers for their thoughtful
comments and advice on improving this manuscript. Funding:
Funding was provided by National Natural Science Foundation
of China grants 21790390, 21790391, and 22134002 (L.M.);
22125406, 21790053, and 22074149 (P.Y.); and 22222411 (F.W.).
Funding was also provided by the Natural Science Foundation
of Beijing (JQ19009), the Strategic Priority Research Program of
the Chinese Academy of Sciences (XDB30000000), and the
CAS Project for Young Scientist in Basic Research (YSBR-050).
Author contributions: Conceptualization: P.Y. and L.M.
Supervision: P.Y. and L.M. Funding acquisition: F.W., P.Y., and
L.M. Methodology: T.X., F.W., W.M., Y.J., P.Y., and L.M.
Investigation: T.X., C.L., X.H., B.X., and J.Z. Visualization: T.X.,
C.L., X.H., B.X., J.Z., F.W., W.M., Y.J., and J.F. Writing – original
draft: T.X. and P.Y. Writing – review: T.X., F.W., P.Y., and L.M.
Competing interests: The authors declare no competing interests.
Data and materials availability: All data needed to evaluate
the conclusions in the paper are present in the main text or the
supplementary materials. The tabulated numerical data underlying
the figures and the model for FEM simulation are deposited in
Dryad (37). License information: Copyright © 2023 the authors,
some rights reserved; exclusive licensee American Association
for the Advancement of Science. No claim to original US
government works. https://www.science.org/about/science-
licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adc9150
Materials and Methods
Supplementary Text
Figs. S1 to S8
Tables S1 and S2
References (38–47)
Submitted 9 May 2022; accepted 10 November 2022
10.1126/science.adc9150
Xiong et al., Science 379, 156–161 (2023)
13 January 2023
6 of 6
|
10.1126_science.add4679
|
and are of a continuous-time nature. Stratego
is a game for which little progress has been
achieved by the AI research community be-
cause of the many complex aspects of its struc-
ture. Successes in the game have been limited,
with artificial agents only able to play at a
level comparable to a human amateur [see,
e.g., (13–17)].
This work introduces a novel game-theoretic
method that allows an AI for learning to play
Stratego in self-play in a model-free manner
without human demonstration and from
scratch. This new method resulted in a bot
called DeepNash that beat previous state-of-
the-art AI agents and achieved human expert–
level performance in the most complex variant
of the game, Stratego Classic. At the core of
DeepNash is a principled, novel, model-free
reinforcement learning (RL) algorithm called
Regularized Nash Dynamics (R-NaD). Our
method resulting in DeepNash combines
R-NaD with a deep neural network architec-
ture to learn a strategy that plays at a highly
competitive level by aiming to find a Nash
equilibrium (18) (i.e., an unexploitable strategy
in zero-sum two-player games). In earlier work,
it was formally shown that an R-NaD ap-
proach converges to a Nash equilibrium in
several classes of matrix games, including
two-player zero-sum games (19). The present
work suggests that R-NaD at scale converges
empirically to an approximate Nash equilib-
rium in Stratego. Figure 2 illustrates a high-level
overview of this approach, which underlies
DeepNash.
The performance of DeepNash was system-
atically evaluated against various state-of-the-
art Stratego bots and human expert players on
the Gravon games platform. DeepNash con-
vincingly won against all current state-of-the-
art bots that have been developed to play
Stratego, producing a win rate of >97%, and
it achieved a highly competitive level of play
against human expert Stratego players on
Gravon, where it ranked among the top three
players, both on the year-to-date 2022 (deter-
mined on 22 April 2022) and on all-time
RES EARCH
MACHINE LEARNING
Mastering the game of Stratego with model-free
multiagent reinforcement learning
Julien Perolat*†, Bart De Vylder*†, Daniel Hennes, Eugene Tarassov, Florian Strub,
Vincent de Boer‡, Paul Muller, Jerome T. Connor, Neil Burch, Thomas Anthony,
Stephen McAleer, Romuald Elie, Sarah H. Cen, Zhe Wang, Audrunas Gruslys,
Aleksandra Malysheva, Mina Khan, Sherjil Ozair, Finbarr Timbers, Toby Pohlen, Tom Eccles,
Mark Rowland, Marc Lanctot, Jean-Baptiste Lespiau, Bilal Piot, Shayegan Omidshafiei,
Edward Lockhart, Laurent Sifre, Nathalie Beauguerlange, Remi Munos, David Silver,
Satinder Singh, Demis Hassabis, Karl Tuyls*†
We introduce DeepNash, an autonomous agent that plays the imperfect information game Stratego at
a human expert level. Stratego is one of the few iconic board games that artificial intelligence (AI)
has not yet mastered. It is a game characterized by a twin challenge: It requires long-term strategic
thinking as in chess, but it also requires dealing with imperfect information as in poker. The technique
underpinning DeepNash uses a game-theoretic, model-free deep reinforcement learning method, without
search, that learns to master Stratego through self-play from scratch. DeepNash beat existing state-
of-the-art AI methods in Stratego and achieved a year-to-date (2022) and all-time top-three ranking on
the Gravon games platform, competing with human expert players.
state-of-the-art imperfect information search
techniques that break down the game into
independent situations (5, 6).
For these reasons, Stratego is a major chal-
lenge for the AI community and provides a
hard benchmark for studying strategic inter-
actions at an unparalleled scale. As in most
board games, Stratego tests the ability to make
relatively slow, deliberative, and logical deci-
sions sequentially. Additionally, in most im-
perfect information games, other tactics that
better reflect decision-making processes in
the real world need to be deployed. As von
Neumann described it, “real life consists of
bluffing, of little tactics of deception, of ask-
ing yourself what is the other man going to
think I mean to do” (8). Most recent successes
in large imperfect information games have
been achieved in real-time strategy games
such as StarCraft, Dota, and Capture the Flag
(9–11) or in racing simulation video games
such as Gran Turismo (12), in which most de-
cisions must be made quickly and instinctively
P rogress in artificial intelligence (AI) has
been measured through the mastery of
board games since the inception of the
field. Board games allow us to gauge and
evaluate how humans and machines de-
velop and execute strategies in a controlled
environment. The ability to plan ahead has
been at the heart of successes in AI for decades
in perfect information games such as chess,
checkers, shogi, and Go, as well as in imper-
fect information games such as poker and
Scotland Yard (1–6). For many years, the Strat-
ego (7) board game has constituted one of the
next frontiers of AI research (for a visualiza-
tion of the game phases and game mechanics,
see Fig. 1). The game poses two key challenges.
First, the game tree of Stratego has 10535 pos-
sible states, which is larger than both no-limit
Texas Hold’em poker, a well-researched im-
perfect information game with 10164 states,
and the game of Go, which has 10360 states.
Second, acting in a given situation in Stratego
requires reasoning >1066 possible pairs of
private deployments at the start of the game,
whereas in Texas Hold’em poker, players are
dealt one of 103 different two-card hands for
106 possible private configurations with two
players. Perfect information games such as Go
and chess do not have a private deployment
phase, thus avoiding the complexity that this
challenge poses in Stratego. Currently, it is not
possible to use state-of-the-art model-based
perfect information planning techniques nor
DeepMind Technologies Ltd., London, UK.
*Corresponding author. Email: perolat@deepmind.com (J.P.);
bartdv@deepmind.com (B.D.V.); karltuyls@deepmind.com (K.T.)
†These authors contributed equally to this work and are
co-lead authors.
‡Independent consultant.
Table 1. Evaluation of DeepNash against existing Stratego bots. The numbers are reported from
DeepNash’s point of view. More games (800) were played against bots that could be run automatically.
Opponent
No. of games
Wins
Draws
Losses
0.0%
Probe
.....................................................................................................................................................................................................................
0.0%
Master of the Flag
.....................................................................................................................................................................................................................
1.1%
Demon of Ignorance
.....................................................................................................................................................................................................................
0.3%
Asmodeus
.....................................................................................................................................................................................................................
1.8%
Celsius
.....................................................................................................................................................................................................................
2.1%
Celsius1.1
.....................................................................................................................................................................................................................
PeternLewis
0.1%
.....................................................................................................................................................................................................................
0.0%
Vixen
.....................................................................................................................................................................................................................
100.0%
100.0%
97.1%
99.7%
98.2%
97.9%
99.9%
100.0%
0.0%
0.0%
1.8%
0.0%
0.0%
0.0%
0.0%
0.0%
30
30
800
800
800
800
800
800
Perolat et al., Science 378, 990–996 (2022)
2 December 2022
1 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Stratego is a two-player board game in which players try to capture the opponent’s flag. Initially, the players secretly deploy 40 pieces of diverse
strengths on the board. Then, they take turns moving pieces, possibly encountering an opponent piece that reveals both piece identities, and then the weaker piece is
removed. Two lakes (indicated in blue) cannot be crossed by any piece. The complete rules are defined by the International Stratego Federation.
Fig. 2. Overview of R-NaD. (A) Overview of the R-NaD approach at scale
underlying DeepNash, which allows for learning to play the imperfect
information game Stratego. (B to D) R-NaD learns a policy represented
by a deep neural network (B) through self-play from scratch (C) and
aims at converging to a Nash equilibrium (D). The approach relies on
two core ideas to reach convergence: replicator dynamics and reward
transformation. Their equations are shown for illustrative purposes in their
simplest form.
leaderboards, with a win rate of 84%. There-
fore, to the best of our knowledge, this is the
first time an AI algorithm was able to learn
to play Stratego at a human expert level. It
is worth mentioning that this performance
was achieved without deploying any search
method, which was a key ingredient of many
milestone achievements in AI for board games
in the past.
Methods
R-NaD at scale takes an end-to-end learning
approach to solving Stratego by incorporating
the learning of the deployment phase, i.e., put-
ting the pieces tactically on the board at the
start of a game (Fig. 1), into the learning of the
game-play phase using an integrated deep RL
and game-theoretic approach. As with much
work in two-player zero-sum games, the pur-
pose is to learn an approximate Nash equi-
librium through self-play. In the context of
two-player zero-sum games, a Nash equilib-
rium guarantees that the agent will perform
well, even against a worst-case opponent. Such
robustness typically allows an algorithm to
perform well against humans [see, e.g., (3–5)].
In perfect information games, search tech-
niques aided by RL have provided state-of-the-
art superhuman bots in Go and chess (2, 20).
However, searching for a Nash equilibrium in
imperfect information games requires esti-
mating private information of the opponent
from public states (3, 6). Given the vast num-
ber of such possible private configurations in a
public state, Stratego is computationally too
challenging for all existing search techniques
because the search space becomes intractable.
This work therefore chose a different route,
without search, and proposed a new method
that combines model-free RL in self-play with
a game-theoretic algorithmic idea, R-NaD. The
model-free part implies that R-NaD does not
build an explicit opponent model–tracking
belief space (calculating a likelihood of the
opponent’s state), and the game-theoretic part
is based on the idea that by modifying the dy-
namical system underpinning the reinforcement-
learning algorithm, one can steer the learning
behavior of the agent in the direction of the
Nash equilibrium. The main advantage of this
combined approach is that one does not need
to explicitly model private states from public
ones. A complex challenge, on the other hand,
is to scale up this model-free RL approach with
Perolat et al., Science 378, 990–996 (2022)
2 December 2022
2 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Illustrating R-NaD using the matching pennies game. (A) Payoff table. (B) Algorithmic stages. (C) Dynamics and Lyapunov function.
R-NaD to make self-play competitive against
human expert players in Stratego, which had
not been achieved to date. This combined ap-
proach is illustrated in Fig. 2.
The following subsections present the learn-
ing algorithm behind DeepNash, referring to
the supplementary materials for more techni-
cal details where relevant.
Learning approach
The approach underpinning DeepNash aims
to learn a Nash equilibrium in Stratego through
self-play and model-free RL. The idea of com-
bining the two has been tried before, but it has
been empirically challenging to stabilize such
learning algorithms when scaling up to com-
plex games such as Capture the Flag, Dota, and
StarCraft (9–11). Some empirical work man-
ages to stabilize the learning either by training
against past versions of the agent (9–11) or by
adding reward shaping (10, 11) or expert data
(9) to the training algorithm. Although these
approaches help, they lack theoretical founda-
tions, remain difficult to tune, and are rather
domain dependent. Furthermore, in a game
such as Stratego, it is difficult to define a loss
for which minimization would converge to a
Nash equilibrium without introducing pro-
hibitive computational obstacles at large scale.
For instance, minimizing the exploitability
(21), a well-known quantity that measures the
distance to a Nash equilibrium, requires esti-
mating an agent’s best response during train-
ing, which is computationally intractable in
Stratego. However, it is possible to define a learn-
ing update rule that induces a dynamical sys-
tem for which there exists a so-called Lyapunov
function. This function can be shown to de-
crease during learning and thus guarantees
convergence to a fixed point. This is the cen-
tral idea behind the R-NaD algorithm and is
the successful recipe for DeepNash, which scales
this approach using a deep neural network.
R-NaD algorithm
The R-NaD learning algorithm used in DeepNash
is an actor-critic method based on the idea of
regularization for convergence purposes (19),
which is briefly explained first in the context
of zero-sum two-player normal form games
(illustrated on the matching pennies game).
A normal form game is an abstraction of a
decision-making situation involving more
than one agent. Each agent (indexed by i ∈
{1,2}) needs to simultaneously take an action
ai (in a set of possible actions Ai) according to
a policy pi(.) (i.e., a distribution over possible
actions Ai), after which it receives a game re-
ward [ri (a1, a2)], and then the game is re-
peated. For convenience, the opponent of
player i is indexed by –i. R-NaD relies on three
key stages (Fig. 3B), described below.
Perolat et al., Science 378, 990–996 (2022)
2 December 2022
In the first stage, the reward is transformed
based on a regularization policy preg, which
induces a modified game with rewards
(cid:3)
(cid:1)
ri pip(cid:1)iaia(cid:1)1
(cid:3)
(cid:1)
¼ ri aia(cid:1)1
(cid:1) ηlog
þ ηlog
!
!
(cid:3)
(cid:1) (cid:3)
pi ai
pi
reg aið
Þ
(cid:1)
p(cid:1)i a(cid:1)i
p(cid:1)i
ð
reg a(cid:1)i
ð1Þ
Þ
where η > 0 is a regularization parameter and
i is the player index (i ∈ {1,2}). Note that this
transformed reward is policy dependent.
Second, in the dynamics stage, the system
evolves according to the replicator dynam-
ics system (22, 23) on this modified game. The
basic replicator dynamics equations are a de-
scriptive learning process from evolutionary
game theory, equivalent to RL algorithms (23),
which are also equivalent to an instance of
follow-the-regularized-leader (24) and are de-
fined as follows:
(cid:1) (cid:3)
pi
τ ai
d
dτ
"
(cid:1) (cid:3)
τ ai
¼ pi
(cid:1) (cid:3)
pτ ai
Qi
(cid:1)
X
bi
#
(cid:1) (cid:3)
pi
τ bi
(cid:1) (cid:3)
pτ bi
Qi
ð2Þ
3 of 7
RES EARCH | R E S E A R C H A R T I C L E
(cid:1) (cid:3)
ai
(cid:1) (cid:3)
p ai
(cid:4)
(cid:3)
where Qi
is the quality or fitness of an ac-
(cid:5)
(cid:1)
pτ
¼ Ea(cid:1)i∼p(cid:1)i ri pi; p(cid:1)i; ai; a(cid:1)i
tion: i.e.,Qi
.
These dynamics reinforce the probability of
taking actions with high fitness (relative to
other actions). Because of the reward transfor-
mation, this system has a unique fixed point,
p fix, and convergence to it is guaranteed in
zero-sum two-player games [see (19)], which
can be proven by the Lyapunov function:
Hpfix
pð Þ ¼
X2
X
i¼1
ai∈Ai
(cid:1) (cid:3)
pi
fix ai
log
(cid:6)
(cid:7)
fix aið Þ
pi
Þ
pi aið
(19).
However, this fixed point is not yet a Nash
equilibrium of the original game.
In the final update stage, the fixed point
obtained is used as the regularization policy
for the next iteration. These three stages are
applied repeatedly, generating a sequence of
fixed points that can be proven to converge to a
Nash equilibrium of the original (unmodified)
game (19) in zero-sum two-player games, but
not in all general-sum games. Figure 3C illus-
trates the R-NaD algorithm on the two-player
matching pennies game (the payoff table is
shown in Fig. 3A). The first iteration starts from
Þ and the
(cid:3) ¼ 0:999; 0:001
½
0;reg H; T
pi
(cid:3) ¼
½
0;fix H; T
replicator dynamics converge to p0
0:896; 0:104
(cid:3) and p1
½
(cid:3).
(cid:3) ¼ 0:263; 0:737
½
0;fix H; T
The right figure shows the evolution of the lo-
garithm of the Lyapunov function and illus-
trates that it decreases while learning. Three
iterations of R-NaD are shown.
(cid:3); η ¼ 0:2
ð
½
½
R-NaD at scale
Our method consists of three components:
(i) a core training component R-NaD, i.e., the
model-free RL algorithm presented above,
which is implemented using a deep convolu-
tional network; (ii) a component that fine-
tunes the learned policy to reduce the residual
probabilities of taking highly improbable ac-
tions; and (iii) a test-time postprocessing com-
ponent that uses game-specific knowledge
[whereas (i) and (ii) are game agnostic] to
filter out low-probability actions and obvious
mistakes.
The following section starts by concisely
laying out some essential background informa-
tion on imperfect information games neces-
sary to understand how R-NaD is scaled to a
deep learning model. Then, the implementation
of the three algorithmic stages of R-NaD are
summarized. A detailed description of R-NaD
is provided in the supplementary materials.
Imperfect information games
In a two-player zero-sum imperfect informa-
tion game, two players (player i = 1 or i = 2)
sequentially interact in turns. At turn t, the
(cid:3)
players receive a reward signal r1
, and the
current player i = Ψt observes the game state
through an observation ot and selects an ac-
tion at according to a parameterized policy
function p(.|ot). In model-free RL, the trajecto-
t r2
t
(cid:1)
Fig. 4. Illustration of DeepNash’s assessment of the relative value of material versus information.
Shown is an illustration of DeepNash’s assessment of the relative value of material versus information in two
human (red) versus DeepNash (blue) matches. (A) Although blue is behind a 7 and 8, no pieces are revealed
and only two are moved. As a result, DeepNash assesses its chance of winning to still be ~70% (blue indeed won
this match). (B) Blue to move. DeepNash’s policy supports three moves at this state, with the indicated
probabilities shown (the move on the right was played in the actual match). Although blue has the opportunity
to capture the opponent’s 6 with its 9, this move was not considered by DeepNash, likely because the protection
of 9’s identity was assessed to be more important than the material gain.
(cid:4)
(cid:1)
(cid:1)
(cid:3)
; p : otj
ð
(cid:5)
ot; at; r1
ries T ¼
are
the only data the agent will leverage to learn
the parameterized policy.
0≤t<tmax
Þ; Ψt
t r2
t
Þ
Model-free RL with R-NaD
The R-NaD algorithm is scaled by using deep
learning architectures. It performs the same
three algorithmic stages as before in normal
form games: (i) the reward transformation
stage, (ii) the dynamics stage that allows for
convergence to a fixed point, and (iii) the up-
date stage in which the algorithm updates the
policy that defines the regularization function.
The neural architecture consists of the fol-
lowing components: a U-Net torso with resid-
ual blocks and skip connections (25) and four
heads that are smaller replicas of the torso
augmented with final layers to generate an
output of the appropriate shape. The first
DeepNash head outputs the value function
as a scalar, and the three remaining heads
encode the agent’s policy by outputting a
probability distribution over its actions at
deployment and during game play. The
agent architecture is described in detail in
the supplementary materials.
The observation is encoded as a spatial ten-
sor consisting of the following components:
DeepNash’s own pieces, publicly available in-
formation about both the opponent’s and
DeepNash’s pieces, and an encoding of the
40 last moves. This public information repre-
sents the types each piece can still have given
the history of the game. In total, the observa-
tion contains 82 stacked frames encoded in a
single tensor. The observation’s detail is given
in the supplementary materials.
Given the regularization policy pm,reg at
iteration m and a trajectory, the reward
transform used at time step t for player i
if i = Ψt
(cid:1) ηlog
(cid:9)
Þ ¼ ri
t
(cid:8)
a; pð
t;pm;reg
is ri
(cid:8)
(cid:9)
Þ
p ajot
Þ
ð
pm;reg ajot
ð
if i ≠ Ψt.
and ri
t
þ ηlog
p ajot
Þ
ð
ð
pm;reg ajot
Þ
The dynamics stage of the method is com-
posed of two parts. The first part estimates the
value function, which is done through an
adaptation of the n-trace estimator (26) to
the two-player imperfect information case,
resulting in a parameter update direction
Updatevalue. The second part learns the policy
through the Neural Replicator Dynamics
(NeuRD) update (27) using a new estimate of
the state action value based on the n-trace
estimator, resulting in a parameter update di-
rection Updatepolicy. These parts are detailed
in the supplementary materials.
After a fixed number of learning steps, an
approximate fixed point policy, pm,fix, is obtained,
which is then used as the next regularization
Perolat et al., Science 378, 990–996 (2022)
2 December 2022
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. Illustration of
DeepNash bluffing. Shown
is an illustration of Deep
Nash bluffing in three human
(red) versus DeepNash
(blue) matches. (A) Positive
bluffing. (B) Negative bluffing.
(C) DeepNash makes a
scout (2) behave like a spy
and gains material.
policy, pm+1,reg = pm,fix. The three stages are
repeated using a smooth transition from the
reward transformation of iteration m to the
one of step m + 1.
Directly learning with the above-described
method leads to convergence to an empirically
satisfying solution, which, however, is slightly
distorted by low-probability mistakes. Those
mistakes appear because the softmax projec-
tion used to compute the policy from the logits
assigns a nonzero probability to every action.
To alleviate this issue, the policy is fine-tuned
during training by performing additional thresh-
olding and discretization to the action proba-
bilities. The supplementary materials provide
more details on this aspect and also describe a
few additional heuristics applied at test time
that remove obvious mistakes from the policy.
As opposed to the R-NaD model-free training
algorithm, these heuristics are Stratego specific.
Qualitatively, they do remove rare mistakes in
matches against humans, but they do not give
notable quantitative improvements in self-play
(see the supplementary materials).
Results
This section presents an overview of the eval-
uation results of DeepNash against both human
expert players and current state-of-the-art
Stratego bots. For the former, DeepNash has
Perolat et al., Science 378, 990–996 (2022)
2 December 2022
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
been evaluated on the Gravon platform, a well-
known online games server popular among
Stratego players. For the latter, DeepNash has
been tested against eight known AI bots that
play Stratego. A detailed analysis is also pre-
sented with regard to some of the capabilities
of the agent’s game play, including deploy-
ment, bluffing, and trading off of material
versus information.
Evaluation on Gravon
Gravon is an internet platform for human
players that offers several online games, in-
cluding Stratego. It is by far the largest online
platform for Stratego, and is where some of
the strongest players compete. For more de-
tails on the platform and its ranking system,
please refer to the supplementary materials.
DeepNash was evaluated against top human
players over the course of 2 weeks in the be-
ginning of April 2022, resulting in 50 ranked
matches on Gravon. Of these matches, 42 (84%)
were won by DeepNash. In the Classic Stratego
Challenge Ranking 2022 at that time, this cor-
responded to a rating of 1799, which put
DeepNash in third place of all ranked Gravon
Stratego players (the top two ratings were
1868 and 1831). In the all-time Classic Stratego
Ranking, this resulted in a rating of 1778, which
also put DeepNash in the third place of all
ranked Gravon Stratego players (the top two
ratings were 1876 and 1823). The rating for this
leaderboard considers all ranked games going
back to the year 2002.
These results confirm that DeepNash reaches
a human expert level in Stratego only through
self-play learning and without bootstrapping
from existing human data.
Evaluation against state-of-the-art Stratego bots
DeepNash was also evaluated against several
existing Stratego computer programs: Probe
was a three-time winner of the Computer
Stratego World Championship (2007, 2008,
and 2010); Master of the Flag won that cham-
pionship in 2009; Demon of Ignorance is an
opensource implementation of Stratego with
an accompanying AI bot; and Asmodeus, Cel-
sius, Celsius1.1, PeternLewis, and Vixen are
programs that were submitted in an Austra-
lian university programming competition in
2012 (see the supplementary materials for
more details).
As shown in Table 1, DeepNash won the
overwhelming majority of games against all
of these bots despite not having been trained
against any of them and only being trained
using self-play. Therefore, it is not necessarily
expected that the residual losses against some
of these bots would vanish even if the exact
Nash-equilibrium were reached. For example,
in most of the few matches that DeepNash
has lost against Celsius1.1, the latter played a
high-risk strategy of capturing pieces early on
with a high-ranking piece and thus was try-
ing to get a significant material advantage.
Most often, this strategy does not work, but
occasionally it can lead to a win.
Illustration of DeepNash’s abilities
The only goal of the algorithm behind DeepNash
is to learn a Nash equilibrium policy and, by
doing so, to learn qualitative behavior that one
could expect a top player to master.
Indeed, the agent is able to generate a wide
range of deployments, which makes it diffi-
cult for a human player to find patterns to ex-
ploit by adapting their own deployment. We
describe DeepNash’s deployment behavior in
more detail in the supplementary materials
(see the additional results section). Deep-
Nash was able to make nontrivial trade-offs
between information and material, to execute
bluffs, and to take gambles when needed. The
rest of this section illustrates these behaviors
through matches that were played on Gravon.
For convenience, the behavior is described
in a way a human observer might naturally
interpret it, including terms such as “decep-
tion” and “bluffing,” which arguably refer to
mental states that the program does not have.
Trade-off between information
and material
An important tactic in Stratego is to keep as
much information as possible hidden from an
opponent to gain an advantage. During cer-
tain game situations, there will be trade-offs to
be considered in which a player needs to
balance the value of capturing an opponent’s
piece (or even moving a piece), and thus re-
vealing information about their own piece,
versus not capturing a piece (or not moving)
but keeping the identity of a piece hidden.
DeepNash was able to make such trade-offs
in extraordinary ways.
Figure 4A shows a situation in which
DeepNash (in blue) was behind in pieces (it
lost a 7 and an 8) but was ahead in inform-
ation; the opponent in red has its 10, 9, an 8
and two of its 7’s revealed. Valuing inform-
ation and material in Stratego is nontrivial
a priori, but the agent has learned a policy
through self-play that seems to naturally
make this trade-off. In the above example,
DeepNash was behind in material but knew
the identity of many of the opponent’s high-
ranked pieces. On the contrary, almost all
of DeepNash’s remaining pieces had not yet
moved and its opponent was left in the dark.
The value function (ν = 0.403) credited this
information asymmetry as an advantage for
DeepNash (with an expected win rate of ~70%)
despite having lesser material on the board.
This game was won by DeepNash.
The second example in Fig. 4B shows a sit-
uation in which DeepNash had the opportu-
nity of capturing the opponent’s 6 with its 9,
but this move was not considered, probably
because protecting the identity of the 9 was
deemed more important than the material gain.
The situation also illustrates the stochasticity
of DeepNash’s policy during game-play.
Deceptive behavior and bluffing
In addition to being able to value asymmetry
of information, one can also expect the agent to
occasionally bluff to deceive its opponent and
potentially gain an advantage. The situations
shown in Fig. 5, A to C, illustrate this ability.
Figure 5A illustrates positive bluffing, in which
a player pretended that a piece had higher
value than it actually did. DeepNash (blue)
chased the opponent’s 8 with an unknown
piece, a scout (2), pretending it was the 10. The
opponent believed that this piece had a high
chance of being the 10 and guided it next to its
spy (which could capture the 10). In an at-
tempt to capture this piece, however, the op-
ponent lost its spy to DeepNash’s scout.
A second type of bluff, called negative bluf-
fing, is shown in Fig. 5B. In contrast to a
positive bluff, this tactic entails pretending a
piece is of a lower rank. Here, the movement
of the unknown 10 of DeepNash (blue) was
interpreted by the opponent as a positive bluff
because they tried to capture it with a known 8.
DeepNash’s move could have been interpreted
as moving the spy closer to the opponent’s 10,
for example. The opponent instead encountered
DeepNash’s 10 and lost an 8.
A more complex bluff is shown in Fig. 5C,
where DeepNash (blue) brought its unrevealed
scout (2) close to the opponent’s 10, which can
be easily interpreted as a spy. This tactic ac-
tually allowed blue to capture red’s 5 with its 7
a few steps later, thereby gaining material
but also preventing the 5 from capturing the
scout (2), and revealing that it was actually
not the spy.
Conclusion
This work introduces a new game-theoretic
method at scale that allows for AI to play the
imperfect information game Stratego from
scratch in self-play up to a human expert level,
as illustrated by our bot DeepNash. This model-
free learning method combines a deep resid-
ual neural network with the game-theoretical
R-NaD multiagent learning algorithm. No
form of search or explicit opponent modeling
is performed during training, and DeepNash
only relies on the use of some game-specific
heuristics at test time. As such, the method
underlying DeepNash takes a contrasting ap-
proach to state-of-the-art search-based learn-
ing methods that have been successfully applied
to other complex games such as Go and chess
and to imperfect information games such as
poker and Scotland Yard. However, because
of their computational toll and the inherent
complexity of the Stratego game itself, those
Perolat et al., Science 378, 990–996 (2022)
2 December 2022
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
methods are not applicable to such an elab-
orate game.
The core component behind DeepNash is
the at-scale implementation of the R-NaD al-
gorithm. It performs three essential stages in
an iteration of the algorithm: reward trans-
formation starting from a random regularized
policy to define a modified game, subsequent-
ly applying the replicator dynamics on this
modified game to converge to a fixed point
policy, and finally updating the regularization
policy to this new fixed point. Repeatedly ap-
plying this three-stage process yields a strategy
that is empirically difficult to exploit.
Evaluated against other AI bots, DeepNash
achieved a minimum win rate of 97%, and in
the evaluation against human expert players
on the Gravon platform, DeepNash achieved
an overall win rate of 84%, which placed it in
the top-three rank of both the year-to-date
(2022) and all-time leaderboards. This is
an extraordinary result that the Stratego
community did not believe would have been
possible with current techniques, judging
by quotes from Thorsten Jungblut (owner of
the Gravon platform) and Vincent de Boer,
which can be found in the supplementary
materials.
Looking forward, at this stage, there are no
indications of how R-NaD fares beyond zero-
sum two-player settings. However, it is rea-
sonable to assume that it can unlock further
applications of RL methods in real-world mul-
tiagent problems with astronomical state spaces
characterized by imperfect information, which
are currently out of reach for state-of-the-art
AI methods to be applied in an end-to-end
fashion. For example, state-of-the-art methods
on two-player poker (4) have successfully trans-
ferred to six-player poker (5). Many applica-
tions can be found in this larger class of games,
including crowd and traffic modeling, smart
grid, auction design, and market problems.
RE FERENCES AND NOTES
1. M. Campbell, A. J. Hoane Jr., F. Hse, Artif. Intell. 134, 57–83
(2002).
2. D. Silver et al., Science 362, 1140–1144 (2018).
3. M. Moravčík et al., Science 356, 508–513 (2017).
4. N. Brown, T. Sandholm, Science 359, 418–424 (2018).
5. N. Brown, T. Sandholm, Science 365, 885–890 (2019).
6. M. Schmid et al., Player of games. arXiv:2112.03178 [cs.AI]
(2021).
7. Stratego is a trademark of Jumbo Diset Group and is used in
8.
this publication for information purposes only.
J. von Neumann, O. Morgenstern, Theory of Games and
Economic Behavior (Princeton Univ. Press, 1947).
9. O. Vinyals et al., Nature 575, 350–354 (2019).
10. C. Berner et al., Dota 2 with large scale deep reinforcement
learning. arXiv:1912.06680 (2019).
11. M. Jaderberg et al., Science 364, 859–865 (2019).
12. P. R. Wurman et al., Nature 602, 223–228 (2022).
13. V. de Boer, L. J. M. Rothkrantz, P. Wiggers, Int. J. Intell. Games
Simul. 5, 25 (2008).
14. M. Schadd, M. Winands, in Proceedings of the Twenty-First
Benelux Conference on Artificial Intelligence, Eindhoven,
Netherlands, 29 to 30 October 2009 (BNAIC, 2009),
pp. 225–232.
I. Satz, J. Int. Comput. Games Assoc. 34, 27–29 (2011).
15.
16. S. Redeca, A. Groza, in 2018 IEEE 14th International Conference
on Intelligent Computer Communication and Processing
(ICCP), Cluj-Napoca, Romania, 6 to 8 September 2018
(IEEE, 2018), pp. 97–104.
17. S. Mcaleer, J. Lanier, R. Fox, P. Baldi, in Advances in Neural
Information Processing Systems 34 (NeurIPS, 2020), vol. 33,
pp. 20238–20248.
18. J. Nash, Ann. Math. 54, 286–295 (1951).
19. J. Perolat et al., in Proceedings of the 38th International
Conference on Machine Learning, 18 to 24 July 2021
(ICML, 2021), pp. 8525–8535.
20. D. Silver et al., Nature 550, 354–359 (2017).
21. M. Lanctot et al., OpenSpiel: A framework for reinforcement
learning in games. arXiv:1908.09453 [cs.LG] (2019).
22. E. Zeeman, J. Theor. Biol. 89, 249–270 (1981).
23. D. Bloembergen, K. Tuyls, D. Hennes, M. Kaisers, J. Artif. Intell.
Res. 53, 659–697 (2015).
24. H. B. McMahan, in Proceedings of the 14th International
Conference on Artificial Intelligence and Statistics, Fort
Lauderdale, FL, 11 to 13 September 2011 (AISTATS, 2011),
vol. 15, pp. 525–533.
25. K. He, X. Zhang, S. Ren, J. Sun, in Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR), Las Vegas, Nevada, 26 June to 1 July 2016
(CVPR, 2016), pp. 770–778.
26. L. Espeholt et al., in Proceedings of the 35th International
Conference on Machine Learning, Stockholm, Sweden,
10 to 15 July 2018 (ICML, 2018), pp. 1406–1416.
27. D. Hennes et al., in Proceedings of the 19th International
Conference on Autonomous Agents and MultiAgent Systems,
Auckland, New Zealand, 9 to 13 May 2020 (AAMAS, 2020),
pp. 492–501.
28. J. Perolat et al., Figure data for: Mastering the game of
Stratego with model-free multiagent reinforcement learning,
Zenodo (2022); https://doi.org/10.5281/zenodo.7118519.
AC KNOWLED GME NTS
We thank the following DeepMind colleagues for their technical
support and feedback during the development of DeepNash:
S. Mourad, L. Prince, A. Huang, M. Schmidt, M. Bowling,
G. Ostrovski, M. Riedmiller, K. Kavukcuoglu, W. Czarnecki, and
S. Singh. We also thank T. Jungblut, the Gravon platform owner, for
providing access to Gravon, which allowed us to evaluate
DeepNash against human expert players, and for allowing us to use
matches played online for illustration purposes in this manuscript,
as well as the anonymous reviewers for their comments that
improved the quality of this article. Funding: This research was
funded by DeepMind Technologies Ltd. Author contributions:
J.P., K.T., and B.D.V. initiated and oversaw the project, developed
training infrastructure and algorithms, performed agent analysis,
wrote the paper, and provided research direction. D.He., E.T.,
F.S., N.Bu., M.K., and E.L. developed the training infrastructure and
algorithms, performed agent analysis, and wrote the paper. P.M.,
J.T.C., T.A., S.H.C., Z.W., A.G., A.M., S.Oz., M.L., and J.L. developed
the training infrastructure and algorithms and performed agent
analysis. D.Ha. and V.d.B. performed the agent analysis and
provided research direction. S.M., R.E., F.T., T.P., T.E., S.Om., and
R.M. performed the agent analysis. L.S. wrote the paper and
provided research direction. M.R. and B.P. wrote the paper.
N.Be. initiated and oversaw the project. D.S. and S.S. provided
research direction. Competing interests: US provisional patent
application number 63/356,009 related to this work was filed
27 June 2022: “Model-Free Reinforcement Learning with
Regularized Nash Dynamics”; assignee: DeepMind Technologies
Ltd. The authors declare no competing interests. Data and
materials availability: Data for the quantitative figures are
available on Zenodo (28). All other data needed to evaluate the
conclusions in this study are present in the main text or the
supplementary materials. License information: Copyright © 2022
the authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.science.org/about/science-
licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add4679
Related Work
Additional Methods
Additional Results
Tables S1 to S3
Figs. S1 to S9
References (29–50)
Submitted 30 June 2022; accepted 3 November 2022
10.1126/science.add4679
Perolat et al., Science 378, 990–996 (2022)
2 December 2022
7 of 7
|
10.1126_science.add5190
|
Corrected 3 October 2023. See full text.
RES EARCH
BIOGEOGRAPHY
Limited climatic space for alternative ecosystem
states in Africa
Steven I. Higgins1*, Timo Conradi1, Laurence M. Kruger2,3, Robert B. O’Hara4, Jasper A. Slingsby3,5,6
One of the foundational premises of ecology is that climate determines ecosystems. This has been
challenged by alternative ecosystem state models, which illustrate that internal ecosystem dynamics
acting on the initial ecosystem state can overwhelm the influence of climate, and by observations
suggesting that climate cannot reliably discriminate forest and savanna ecosystem types. Using a novel
phytoclimatic transform, which estimates the ability of climate to support different types of plants,
we show that climatic suitability for evergreen trees and C4 grasses are sufficient to discriminate
between forest and savanna in Africa. Our findings reassert the dominant influence of climate on
ecosystems and suggest that the role of feedbacks causing alternative ecosystem states is less prevalent
than has been suggested.
P redicting how global vegetation patterns
will be altered by climate change is a fun-
damental challenge for ecological science
(1). One particular challenge for predic-
tion models is accommodating the possi-
bility that at any single location in environmental
space, alternative ecosystem types may occur. For
example, across the global tropics, forest and
savanna are both observed between 1000
and 2500 mm of mean annual precipitation
(2, 3). The observation that qualitatively dif-
ferent ecosystems (i.e., “ecosystem states”) can
exist under the same environmental conditions
has led to the emergence of the alternative eco-
system state (AES) hypothesis.
The AES hypothesis posits that an AES is a
self-maintaining state that cannot be pre-
dicted from environmental forcing factors
(4–7). Such self-maintaining states are induced
by differences in initial conditions and main-
tained by positive feedback processes. Strictly,
the initial condition or initial ecosystem state
refers to the system state at some time zero. In
practice, whenever the system state is modi-
fied by a historic event (e.g., a stand-replacing
fire, windstorm, or drought), this initial condi-
tion is reset, conceptually resetting the system
to a new time zero. In AES applications such
historical events can have a sufficiently large
effect to move the ecosystem state into the
domain of attraction of an alternative eco-
system state, which is then maintained by
positive feedback processes (8).
Several lines of evidence are consistent with
the AES hypothesis. Theoretical models have
1Plant Ecology, University of Bayreuth, Universitaetsstrasse
30, 95447 Bayreuth, Germany. 2Organization for Tropical
Studies, P.O. Box 33, Skukuza, 1350, South Africa.
3Department of Biological Sciences, University of Cape Town,
South Africa. 4Department of Mathematical Sciences,
Norwegian University of Science and Technology, Trondheim
N-7491 Norway. 5Centre for Statistics in Ecology, the
Environment and Conservation, University of Cape Town,
South Africa. 6Fynbos Node, South African Environmental
Observation Network (SAEON), South Africa.
*Corresponding author. Email: steven.higgins@uni-bayreuth.de
demonstrated the possibility that positive feed-
back processes in the internal system dynamics
acting on the initial ecosystem state can allow
qualitatively different ecosystem states to per-
sist in the same environmental setting (9).
Heuristic models (10–12) and forecasting mod-
els (13) have supported this interpretation, as
have empirical studies using remote sensing
(2, 3, 14, 15), tree basal area (16), and floristic
surveys (17, 18). The plausibility of the AES
hypothesis is thus not in question. However,
the geographical extent of AES regions is
unclear.
Attempts to quantify the geographic extent
of AES (2, 3, 14, 15, 17, 18) have done so by
identifying regions where environmental driv-
ers cannot predict observed differences in eco-
system states. However, such analyses potentially
confound true AES with apparent AES (e.g., fig.
S7). True AES is when observed differences in
ecosystem state are caused by differences in ini-
tial conditions and not by differences in envi-
ronmental drivers. Apparent AES occurs when
differences in observed ecosystem states are
erroneously attributed to differences in initial
conditions, when in fact they are caused by
unmeasured environmental drivers. For ex-
ample, observed differences in ecosystem state
in a landscape exposed to the same climate
might be caused by topographic variation or
soil differences. This means that prediction
uncertainty is not a signature of AES but rather
is the outcome of combinations of model, pa-
rameter, driver, observation, and initial con-
dition uncertainty (19–21) (figs. S6, S7, S8).
For predicting global vegetation patterns it
is of fundamental importance to know whether
the ecosystem state at a particular location in
environmental space can be predicted based
on external forcing by climate system variables
or whether positive feedbacks acting on the
initial ecosystem state might determine eco-
system state (19, 20). Here we revisit the chal-
lenge posed by the alternative ecosystem state
hypothesis for prediction models, by using a
recently proposed method for characterizing
the influence of climate on vegetation (22). As
a case study, we use one of the most widely
studied examples of alternative ecosystem
states: forest and savanna in tropical and
subtropical Africa (2, 3, 17, 23). Specifically,
we test whether a climate-based prediction
model can predict forest and savanna from
a benchmark dataset of floristically defined
savanna and forest sites (17) (hereafter the
AfroTropTree data). The AfroTropTree data
use plant community composition data to
unambiguously assign 753 sites to savanna
and forest categories and thereby avoids obser-
vation uncertainty associated with using
remotely sensed data to classify vegetation
(24, 25).
Phytoclimatic drivers of forest and savanna
We used a plant growth model (fig. S9), spe-
cies distribution data, and climate system data
(temperature, soil moisture, photosynthetically
active radiation, atmospheric CO2; table S4)
to estimate the physiological niches for 4542
African plant species (21). The physiological
niche as conceptualized by this model (fig. S10)
describes how abiotic drivers influence a plant's
growth and its assimilation of carbon and ni-
trogen. We then estimated the climatic suitabil-
ity of geolocations for major plant growth
forms by averaging the niche projections for
all species belonging to each growth form (22).
Conceptually this procedure is a phytoclimatic
transform; that is, it uses the plant growth
model to transform climate variables into plant
suitability variables (21). The phytoclimatic trans-
form produced coherent continental projec-
tions of the suitability of Africa for different
plant growth forms (Fig. 1). The arid parts of
the continent were generally unsuitable for
all growth forms whereas the warm and moist
parts of the continent were suitable for most
growth forms, ensuring that correlations be-
tween the growth form suitability surfaces (e.g.,
evergreen trees and climbers) were conspicu-
ous. Differences were also apparent (e.g., an-
nual versus perennial forbs; evergreen versus
deciduous trees) and some growth forms (an-
nual forbs and C4 grasses) had broader cli-
matic tolerances than others (evergreen trees,
climbers, geophytes, and succulents).
We then used the growth form suitability
surfaces (Fig. 1) as explanatory variables in a
logistic regression model that predicts eco-
system state (forest or savanna) reported in
the AfroTropTree data (17, 21). The logistic regres-
sion revealed that the information contained
in the phytoclimatic transform (Fig. 1 and
table S3) was sufficient to correctly classify
89% of the 678 study sites (table S1). This high
prediction accuracy suggests limited overlap
in the phytoclimatic spaces occupied by forest
and savanna. A similarly high prediction accu-
racy was found when using standard bioclimatic
Higgins et al., Science 380, 1038–1042 (2023)
9 June 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
Corrected 3 October 2023. See full text.
C3 Grass
C4 Grass
Geophyte
Annual forb
Perennial forb
Succulent
Deciduous shrub
Evergreen shrub
Climber
Deciduous tree
Evergreen tree
y
t
i
l
i
b
a
t
i
u
S
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Fig. 1. Climatic suitability surfaces for major plant growth forms. Suitability is the averaged suitability scores of the species belonging to each growth form. See
fig. S1 for the standard errors of these surfaces and fig. S2 for relative growth form suitability surfaces.
variables in the logistic regression model (table
S2), indicating that our findings are not an
artefact of the phytoclimatic variables we used.
The phytoclimatic regression model predicted
that the probability of forest increases with
increasing suitability for evergreen trees and
with decreasing suitability for C4 grasses
(Fig. 2). This prediction is consistent with
work defining forests as systems dominated
by shade-tolerant trees and savannas as sys-
tems co-dominated by C4 grasses and shade-
intolerant trees (26) and with older definitions
that savannas are systems with a continuous
grass layer and a discontinuous tree layer,
whereas forests are systems with a contin-
uous tree layer (27).
When we inspected the 75 prediction errors
(21) (Fig. 3) we found that 45 of them could
potentially be explained by abiotic factors not
included in our analysis (table S3). These
abiotic factors included topographic variation,
Higgins et al., Science 380, 1038–1042 (2023)
9 June 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
Corrected 3 October 2023. See full text.
1.0
0.8
0.6
P
r
o
b
a
b
i
l
i
t
y
o
f
f
o
r
e
s
t
0.4
0.2
0.0
0.0
0.2
S
uita
0.4
0.6
bility for C
4 grass
0.8
1.0
0.8
0.4
0.6
Suitability for evergreen tree
0.2
1.0
0.0
Fig. 2. Probability of forest based on mean
annual rainfall. A logistic regression model esti-
mated using penalized maximum likelihood revealed
that the probability of forest increases with the
suitability for evergreen trees and decreases with
the suitability for C4 grasses. The suitability
information is shown in Fig. 1. Figure 3 shows this
model projected into geographic space. Note that in
the logistic regression model the probability of
savanna is 1 − the probability of forest.
edaphic constraints such as low soil fertility,
the microclimate created by Victoria Falls, flood-
plains, and estuaries (fig. S3). Moreover, a
further 6 prediction errors appeared to have
been observation errors. Observation errors
were diagnosed when Google Earth imagery
of the site clearly contradicted the assignment
in the AfroTropTree database, most likely due
to errors in site coordinates (21). We found no
observation errors of this kind in a random
sample of 75 correctly predicted sites, indicat-
ing that observation errors were more likely in
situations where prediction errors were made.
Overall, this left 24 of the 678 study sites (3.5%)
as potential true AES sites (table S3). The high
predictive accuracy of the model (603 of 678,
89%, Fig. 3) combined with suggestions that
most (51 of 75, 68%, table S3) prediction errors
might be due to factors other than true AES,
together challenge the view that true AES are
widespread in Africa.
It is notable that topographic effects were
associated with 41 of the 75 (55%) model pre-
diction errors (21) (table S3). Topography can
influence the likelihood of forest or savanna
occurrence in many ways. For example, valleys
may have moister microclimates that favor
forest and repress fires (6, 28, 29), and elevation
change can induce orographic rainfall (favoring
forests) and rain shadows (making forests less
likely) (27) (fig. S5). It is also known that subtle
topographic variation associated with catenas
and floodplains can be sufficient to create
wetter sites that can either promote forests or,
when seasonally water logged, exclude trees
True forest prediction
True savanna prediction
False forest prediction
False savanna prediction
t
s
e
r
o
f
f
o
y
t
i
l
b
a
b
o
r
P
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Fig. 3. Probability of forest over Africa. Spatial projection of a logistic regression model for the probability
of forest (Fig. 2). The points show that this model correctly identifies 603 of 678 forest and savanna
sites (17). To enhance visibility of the 75 model errors, the false predictions were plotted on top of true
predictions. Zones with intermediate probability of forest are areas where the prediction uncertainty is
higher and not necessarily true AES zones.
(27, 30) (fig. S4). That is, prediction uncertainty
could be reduced by considering additional
topographic drivers, which would further
reduce the need to invoke true AES when ex-
plaining observed ecosystem states.
Apparent and true AES
Our findings challenge the prevailing view
that forest and savanna are true AES over
large portions of Africa (3, 5–7, 10, 17, 31).
The true AES hypothesis proposes that AES
occurs on homogenous environmental tem-
plates at intermediate resource levels as a
result of positive feedbacks acting on initial
ecosystem states (Fig. 4A). These feedbacks
can involve fire and herbivory being enhanced
by grassland states and retarded by forest
states (10, 11) or due to spatial dynamics of
resource competition acting on a homogenous
environmental template (12). Our analysis sup-
ports an alternative interpretation, which sug-
gests that observations of different ecosystem
states at intermediate positions on resource
gradients may instead be caused by determi-
nistic factors such as topographic gradients
redistributing resources such as moisture in
parts of the landscape, thereby creating micro-
environments that favor either forests or
savannas (Fig. 4B).
Aspects of the true AES hypothesis are
supported by field studies conducted in Africa.
For example, experiments that excluded fire
and grazing in west Africa (32–34) allowed
savanna to transition to forest (for the locations
of these and other sites discussed, see fig. S3).
Conversely, increases in elephant densities in
the Murchison Falls National Park (Uganda)
led to forest transitioning into savanna (35).
Although these studies—by showing that events
can shift ecosystems into alternative states—are
consistent with AES predictions, they cannot be
taken as unequivocal evidence (36). Ideally, we
need to know that the initial ecosystem state
was stable, that an event induced a switch to an
Higgins et al., Science 380, 1038–1042 (2023)
9 June 2023
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
Corrected 3 October 2023. See full text.
A
B
Fig. 4. Alternative ecosystem states along a resource gradient. True
alternative ecosystem states may be observed at intermediate resource levels on a
uniform environmental template due to positive feedbacks in the internal system
dynamics acting on the initial ecosystem state (A). Apparent alternative ecosystem
states occur when observed differences in ecosystem states are erroneously
attributed to differences in initial conditions, when in fact they are caused by
environmental drivers not included in the prediction model. For example, variation in
the environmental template created by topography can redistribute moisture down
slope, thereby creating predictable differences in ecosystem states (B). See fig. S7
for an example of how true and apparent AES can be confounded.
alternative ecosystem state, and that the eco-
system subsequently persisted in this alternative
state (8, 36). For example, although an extreme
fire event consumed a forest and allowed a
savanna state to establish in the Hluhluwe-
iMfolozi Park (South Africa) (37), it is unclear
whether the forest patches destroyed by the
fire were stable vegetation states and only time
will tell whether the savanna will persist as an
alternative ecosystem state or whether the forest
state will reestablish. That is, we need to expand
our observation time windows when evaluating
evidence for true AES. For example, even though
fire exclusion experiments report increases in
fire-sensitive tree species (36), they seldom run
long enough to demonstrate full canopy clo-
sure and exclusion of a flammable grass layer
that can retard fires; indeed, fire exclusion
treatments can be lost to accidental fires even
after decades of fire suppression (38). The value
of an extended time window is further illus-
trated by a 45-year time series of forest-savanna
mosaics in Cameroon, which showed that
forest cover is currently expanding (39), meaning
that the mosaic is unstable, contrary to what
would be expected under the true AES hypo-
thesis (6). In summary, observed transitions
between savanna and forest ecosystem states
are not unequivocal indicators of true AES.
We found that for most locations in Africa
(Fig. 3), feedbacks involving fire, herbivory,
and light competition do not override the
influence of phytoclimatic factors in determin-
ing ecosystem state. This is not to say they are
unimportant; the role of fire and herbivory in
ecosystems is undeniable (40, 41) but they
themselves are influenced by climate and
vegetation, making it difficult to isolate their
role from climatic drivers. Our findings empha-
size that it is only at intermediate positions on
the phytoclimatic gradient where fire and her-
bivory feedbacks may be strong enough that
the ecosystem state is determined by the initial
ecosystem state (Fig. 4A). At such intermediate
positions on the gradient it is however also
possible that topographic gradients redis-
tribute resources such as soil moisture allowing
ecosystem state to be predicted from resource
levels irrespective of initial conditions (Fig. 4B).
Even though feedbacks may not determine
ecosystem states in such situations (Fig. 4B),
they may often have sufficient strength to
produce an abrupt boundary between forest
and savanna states, ensuring that one type
does not gradually blend into the other.
Implications of predictable ecosystem states
Our findings have implications for conserva-
tion practice and climate change mitigation
planning. A concrete example is provided by
the uncertainty associated with whether and
where to plant trees to mitigate climate change.
Previous analyses would suggest that vast
parts of Africa are AES regions and by impli-
cation climatically suitable for forest (2, 3, 17)
or forest restoration (42). The implicit hypoth-
esis is that tree planting and fire suppression
in the savannas of AES regions could manip-
ulate the ecosystem state sufficiently to set up
positive feedback processes that would allow
carbon sequestration on a grand scale in
Africa. However, there is well-justified concern
that such programs overestimate the potential
for carbon sequestration (43) and underplay
threats to native biodiversity (44). Our analy-
sis further suggests that AES regions may be
substantially smaller than previously esti-
mated (2, 3, 17, 31), greatly reducing the carbon
sequestration potential of tree planting in
Africa. For example, the region where the
prediction uncertainty in our analysis is high
is potentially an AES region (Fig. 3). However,
this area of heightened prediction uncertainty
bundles model, parameter, driver, observation,
and initial condition uncertainty. That is, this
ecosystem-uncertain region should not be inter-
preted as a region that could support both
forest or savanna (5); rather it is a region where
we do not know whether it supports forest or
savanna (45). It is likely that Fig. 3 overestimates
the size of the area of high prediction uncer-
tainty because prediction uncertainty could
be reduced by, for example, allowing the model
to consider topographic effects (e.g., fig. S7)
and by reducing observation uncertainty (e.g.,
fig. S8). This means that significant parts of
the environmental template within the area
of high prediction uncertainty shown in
Fig. 3 would not be suitable for tree planting
programs. We therefore caution against using
Fig. 3 to plan tree planting activities without
isolating uncertainty in the initial ecosys-
tem state from the overall prediction uncer-
tainty (46).
Our analyses question the proposition that
for a region as large as Australia (7.5 million km2)
(17) past climatic conditions or past disturbance
events determine whether the vegetation in
Africa is savanna or forest (13). Previous work
appears to have overestimated the geographic
extent of AES regions in Africa. For example,
although previous work (15) presents evidence
for the existence of AES in Africa, their data
indicate that only 6% of the sites included in
their analysis were potentially true AES sites.
In conclusion, using a phytoclimatic transform,
we were able to predict forest and savanna
ecosystem states with substantial accuracy
over most of Africa, challenging the view that
climate cannot predict ecosystem state over
vast regions of Earth (4, 7) and reassuring us
that we can develop climate-forced predictions
of current and future ecosystem states.
REFERENCES AND NOTES
1. G. B. Bonan, S. C. Doney, Science 359, eaam8328 (2018).
2. M. Hirota, M. Holmgren, E. H. Van Nes, M. Scheffer, Science
334, 232–235 (2011).
3. A. C. Staver, S. Archibald, S. A. Levin, Science 334, 230–232 (2011).
4. W. J. Bond, J. Veg. Sci. 16, 261–266 (2005).
Higgins et al., Science 380, 1038–1042 (2023)
9 June 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
5. W. J. Bond, Open Ecosystems: Ecology and Evolution Beyond
the Forest Edge. (Oxford Univ. Press, 2019).
J. G. Pausas, W. J. Bond, Trends Plant Sci. 25, 250–263 (2020).
6.
J. G. Pausas, W. J. Bond, J. Ecol. 109, 3962–3971 (2021).
7.
8. P. Petraitis, Multiple Stable States in Natural Ecosystems
(Oxford Univ. Press, 2013).
I. Noy-Meir, J. Ecol. 63, 459–481 (1975).
9.
10. A. C. Staver, S. A. Levin, Am. Nat. 180, 211–224 (2012).
11.
I. Y. Djeumen, Y. Dumont, A. Doizy, P. Couteron, Ecol. Modell.
440, 109381 (2021).
12. M. Rietkerk et al., Science 374, eabj0359 (2021).
13. G. R. Moncrieff, S. Scheiter, W. J. Bond, S. I. Higgins, New
Phytol. 201, 908–915 (2014).
Corrected 3 October 2023. See full text.
32. R. J. Rose-Innes, “Fire in west African vegetation” in
Proceedings of the Tall Timbers Fire Ecology Conference,
vol. 11, pp. 147 –173, 1972.
33. M. D. Swaine, W. D. Hawthorne, T. K. Orgle, Biotropica 24,
166–172 (1992).
34. D. Louppe, N. Oattara, A. Coulibaly, Commonw. For. Rev. 74,
288–292 (1995).
35. H. K. Buechner, H. C. Dawkins, Ecology 42, 752–766 (1961).
36. E. M. Veenendaal et al., New Phytol. 218, 153–166 (2018).
37. H. Beckett, A. C. Staver, T. Charles-Dominique, W. J. Bond,
J. Ecol. 110, 902–915 (2022).
38. E. Chidumayo, Commonw. For. Rev. 76, 268–272 (1997).
39. L. B. T. Sagang et al., ISPRS J. Photogramm. Remote Sens.
14. J. C. Aleman, A. C. Staver, T. Gillespie, Glob. Ecol. Biogeogr. 27,
188, 142–156 (2022).
792–803 (2018).
15. D.D’Onofrio, J. von Hardenberg, M. Baudena, Glob. Ecol.
Biogeogr. 27, 714–725 (2018).
16. V. de L. Dantas, M. Hirota, R. S. Oliveira, J. G. Pausas, Ecol. Lett.
19, 12–19 (2016).
17. J. C. Aleman et al., Proc. Natl. Acad. Sci. U.S.A. 117,
28183–28190 (2020).
18. P. L. Silva de Mirandaet al ., Glob. Ecol. Biogeogr. 27, 899–912 (2018).
19. M. C. Dietze, Ecol. Appl. 27, 2048–2060 (2017).
20. E. G. Simmonds et al., iScience 25, 105512 (2022).
21. Materials and methods can be found in the supplementary materials.
22. T. Conradi et al., New Phytol. 227, 1294–1306 (2020).
23. S. I. Higgins, S. Scheiter, Nature 488, 209–212 (2012).
24. F. Gerard et al., Glob. Ecol. Biogeogr. 26, 854–859 (2017).
25. S. S. Kumar et al., Remote Sens. 11, 815 (2019).
26. J. Ratnam et al., Glob. Ecol. Biogeogr. 20, 653–660 (2011).
27. H. Walter, Vegetation of the Earth in relation to climate and the
eco-physiological conditions (Springer, 1973).
28. S. W. Wood, B. P. Murphy, D. M. J. S. Bowman, J. Biogeogr. 38,
1807–1820 (2011).
29. W. Bond, B. van Wilgen, Fire and Plants (Population and
Community Biology Series, Springer, 2012).
30. M. Fletcher, S. W. Wood, S. G. Haberle, Ecology 95, 2504–2513
(2014).
31. W. J. Bond, F. I. Woodward, G. F. Midgley, New Phytol. 165,
525–537 (2005).
40. S. Jia et al., Proc. Natl. Acad. Sci. U.S.A. 115, 6237–6242 (2018).
41. A. F. A. Pellegrini et al., Nat. Ecol. Evol. 5, 504–512 (2021).
42. J.-F. Bastin et al., Science 365, 76–79 (2019).
43. Y. Zhou et al., Nature 603, 445–449 (2022).
44. J. W. Veldman et al., Science 366, eaay7976 (2019).
45. R. H. Whittaker, Communities and Ecosystems. Second Edition.
(MacMillan, 1975.
46. J. A. Slingsby, S. Afr. J. Sci. 116, 7684 (2020).
47. S. I. Higgins, pfloek-bt/TTRcodeAESpublic: v1.0.0, Zenodo (2023);
https://doi.org/10.5281/zenodo.7854825.
48. S. I. Higgins, Supplementary data S1 associated with the manuscript
“Limited climatic space for alternative ecosystem states in Africa”,
v1.0.0, Zenodo (2023); https://doi.org/10.5281/zenodo.7854847.
ACKN OWLED GMEN TS
Data sources: We acknowledge the herbaria that contributed data
to this work: A, AAS, AAU, ABH, AD, AJOU, AK, AKPM, ALCB,
AMNH, ARAN, ARM, AS, ASU, B, BA, BAA, BAB, BABY, BACP, BAF,
BAFC, BAI, BC, BCN, BG, BIO, BKF, BM, BOON, BOUM, BR, BRI,
BRIT, BRLU, C, CANB, CAS, CATA, CATIE, CBG, CBM, CDA, CDBI,
CEN, CHR, CHSC, CIB, CICY, CIMI, CNS, COA, COAH, COFC, COI,
COL, CONC, CORD, CP, CS, CSUSB, CTES, CU, CUVC, CVRD, DAV,
DBG, DNA, E, EA, ECON, EMMA, ENCB, ERA, F, FCO, FCQ, FLAS,
FR, FSU, FTG, FULD, FURB, G, GAT, GB, GENT, GEO, GH, GI,
GLM, GMDRC, GMNHJ, GOET, GZU, HAL, HAST, HBG, HBR,
HCIB, HGI, HGM, HIB, HO, HSC, HSS, HU, HUAL, HUJ, HUSA, HYO,
IAC, IBGE, IBK, IBSC, IBUG, ICEL, ICESI, ICN, IEB, INM, INPA,
IPA, IRVC, ISKW, JBAG, JBGP, JCT, JEPS, JOTR, JUA, JYV, K,
KMN, KPM, KSTC, KU, KUN, KYO, L, LA, LAE, LAGU, LBG, LCR, LD,
LEB, LI, LIL, SC, LP, LPAG, LPB, LPS, LSU, LTB, LW, MA, MAF,
MAK, MBK, MBM, MBML, MCNS, MEL, MELU, MERL, MEXU, MGC,
MNHM, MNHN, MO, MPN, MSC, MU, MUB, MVJB, NAC, NAS, NCU,
ND, NE, NH, NHT, NLH, NMNL, NMSU, NSW, NU, NY, NZFRI, O, OBI,
OSA, OSC, P, PAMP, PASA, PE, PERTH, PGM, POM, PY, QMEX, R,
RB, RELC, RM, RSA, S, SACT, SALA, SANT, SAV, SBBG, SD, SDSU,
SEV, SF, SFV, SI, SJSU, SNM, SP, SPF, SRFA, STL, STU, SVG,
SZU, TAI, TAIF, TALL, TAM, TAMU, TAN, TEF, TEX, TI, TKPM,
TNS, TOYA, TRH, TROM, TUB, U, UADY, UAM, UAS, UB, UC,
UCR, UCS, UCSB, UCSC, UESC, UFG, UFMA, UFMT, UFRN,
UFS, UGDA, UJAT, ULM, ULS, UME, UNEX, UNM, UNR, UNSL,
UPNA, UPS, US, USF, USP, UT, UTEP, VAL, VEN, VIT, W, WAG,
WELT, WII, WIS, WOLL, WU, XAL, YAMA, Z. Funding: This work
was supported by the following: BMBF SPACES project
EMSAfrica, grant 01LL1801A (to S.I.H.); National Research
Foundation of South Africa, grant 118593 (to J.A.S.) Author
contributions: Conceptualization: S.I.H., T.C., R.B.O., L.I.K., and
J.S. Methodology: S.I.H., T.C., and R.B.O. Investigation: S.I.H. and
T.C. Writing - original draft: S.I.H. Writing - review and editing: S.I.H.,
T.C., R.B.O., L.M.K., and J.A.S. Competing interests: The authors
declare that they have no competing interests. Data and
materials availability: All data are available in the manuscript,
the supplementary material or deposited in Zenodo (47, 48).
License information: Copyright © 2023 the authors, some
rights reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government
works. https://www.sciencemag.org/about/science-licenses-
journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add5190
Materials and Methods
Figs. S1 to S10
Tables S1 to S4
References (49–78)
Submitted 16 June 2022; accepted 12 May 2023
10.1126/science.add5190
Higgins et al., Science 380, 1038–1042 (2023)
9 June 2023
5 of 5
|
10.1126_science.adc9570
|
RES EARCH
VOCALIZATION
Toothed whales use distinct vocal registers for
echolocation and communication
Peter T. Madsen1, Ursula Siebert2, Coen P. H. Elemans3*
Echolocating toothed whales (odontocetes) capture fast-moving prey in dark marine environments,
which critically depends on their ability to generate powerful, ultrasonic clicks. How their supposedly
air-driven sound source can produce biosonar clicks at depths of >1000 meters, while also producing
rich vocal repertoires to mediate complex social communication, remains unknown. We show that
odontocetes possess a sound production system based on air driven through nasal passages that is
functionally analogous to laryngeal and syringeal sound production. Tissue vibration in different
registers produces distinct echolocation and communication signals across all major odontocete
clades, and thus provides a physiological basis for classifying their vocal repertoires. The vocal fry
register is used by species from porpoises to sperm whales for generating powerful, highly air-efficient
echolocation clicks.
T oothed whales, the odontocetes, have
access to rich marine food resources
down to depths of 2000 m and achieve
a biomass turnover larger than that of
human fisheries combined (1). The key
to this success is their ability to locate, track,
and capture fast-moving prey underwater in
complete darkness at depths of, routinely,
>100 m using echolocation, a feat that crit-
ically depends on the production of short,
powerful, ultrasonic echolocation clicks (here-
after clicks) at rates >300 clicks per second
(2). Paradoxically, odontocetes are thought to
produce clicks with an airflow-driven sound
source in their nose (3), but how they can
produce clicks with less than 10% remain-
ing air volume and pressure-collapsed lungs
at depths beyond 100 m is not understood.
The same source is thought to be sufficiently
plastic to also produce rich, learned acoustic
repertoires essential for mediating complex
social interactions (4). However, if so, how
the same source can generate such a diversity
of acoustic signals to serve both echolocation
and communication remains unknown.
Echolocation clicks are made nasally,
not laryngeally
The location of the odontocete sound source
was initially hypothesized to be the larynx, but
was later designated to the so-called phonic
lips in the nose (5) on the basis of three lines
of evidence: (i) laryngeal muscles are inactive
during clicking (6), (ii) air is pressurized and
recycled in between air sacs surrounding the
nasal passages (7, 8), and (iii) acoustic tri-
angulation localizes the click source 20 to
1Zoophysiology, Department of Biology, Aarhus University,
8000 Aarhus, Denmark. 2Institute for Terrestrial and Aquatic
Wildlife Research, University of Veterinary Medicine
Hannover, 25761 Büsum, Germany. 3Sound Communication
and Behavior Group, Department of Biology, University of
Southern Denmark, 5230 Odense M, Denmark.
*Corresponding author. Email: coen@biology.sdu.dk
70 mm below the blowhole (9). Determining
the odontocete sound source location requires
quantification of phonic lip motion dynamics
with high spatiotemporal resolution to link
tissue motion dynamics to sound generation.
However, because of experimental limitations,
imaging sound-producing events in the odon-
tocete nose at sufficient speed in vivo has
remained very challenging (3).
To test the hypothesis that phonic lips are
the odontocete click source, we imaged their
motion at 7200 frames per second (fps) using
an endoscope while simultaneously measur-
ing air pressure below (psub) and above (psupra)
the phonic lips during click production in
trained Atlantic bottlenose dolphins (Tursiops
truncatus, n = 2) and Harbor porpoises
(Phocoena phocoena, n = 3), in vivo (Fig. 1B)
(10). Dolphins and porpoises produced clicks
exclusively during prolonged bouts of in-
creased psub (Fig. 1D), and 98% were pro-
duced above psub = 3.63 ± 0.14 kPa and
2.12 ± 1.56 kPa, respectively (Fig. 1E), which
were consistent with earlier measurements
(3, 6, 11).
Phonic lips and more deeply situated bi-
lateral nasal plugs (Fig. 1B) were clearly visible
during inspiration, but during pressurization,
dorsal airsac walls moved rostrally and oc-
cluded direct views (Fig. 1F). We observed
click-associated tissue motion with a delay
of 1.0 ± 1.3 ms in dolphins (n = 2) and 1.5 ±
0.36 ms in porpoises (n = 3) after click emis-
sion at the melon (Fig. 1, G to I and movies S1
and S2). Therefore, we likely did not observe
phonic lips directly, but mucosal waves travel-
ing rostrally over the nasal surface. Mucosal
waves are common during tissue vibration–
induced sound production (12, 13). Combining
motion latency with a mucosal wave speed of
~1 m/s (12) predicts the motion source to be
~1 to 2 mm below the observed slit, which is
consistent with the anatomical location of
phonic lips. Thus, our data directly links nasal
tissue motion to sound generation and con-
firms the nasal biosonar sound source.
Echolocation clicks are produced by
colliding phonic lips during flow-induced
voiced sound production
Identifying the physiological mechanism that
drives phonic lip motion is crucial for under-
standing the parameters that determine
acoustic output, from immediate motor con-
trol to morphological changes over evolution-
ary time scales. Two mechanisms have been
suggested for toothed whale sound production:
superfast muscle actuation (14, 15) and air flow–
induced self-sustained vibration (3, 6, 7, 11).
The latter is consistent with the myoelastic-
aerodynamic theory of sound production for
laryngeal and syringeal sound production
(12, 13). Myoelastic-aerodynamic sources re-
quire air pressurization and volume flow,
which are complicated by the highly reduced
air volumes with diving depth and should
thus severely constrain the functional depth
range of toothed whale sound production.
Alternatively, phonic lip motion actuated by
superfast muscles would not be depth depen-
dent and would thus circumvent this con-
straint. Indeed, odontocete genomes imply
the plausibility of superfast muscles to power
rapid phonic lip motion (14) and motor activity
may preceed individual clicks (15). Although
weak superfast muscles may not be capable
of powering heavy phonic lip motion (16),
they may trigger catch-release or stridulation
mechanisms as in fish and aquatic frogs (17).
To test whether a myoelastic-aerodynamic
source or muscle contraction powers click
production, we imaged phonic lips during
sound production in an in vitro preparation of
the nasal complex in dead harbor porpoises,
thereby excluding neural control or muscle
action (Fig. 2) (10). At nasal threshold pres-
sures >5 kPa, we induced emission of click
sequences in six specimens by phonic lip ad-
duction. Imaging of phonic lip kinematics
(4000 fps, 8 bits) showed that anterior and
posterior phonic lips undergo flow-induced,
self-sustained oscillations during click emis-
sion (Fig. 2, B to D). Click production was
tightly associated with phonic lip collision
(delay of 74 ± 204 ms), not opening (delay of
1.70 ± 1.16 ms) (Fig. 2, E and F). Higher-
precision imaging (10,000 fps, 12 bits) showed
that phonic lip collisions occurred only 57 ±
103 ms (n = 2370) prior to click production (Fig.
2, E and F, and movie S3). Thus, phonic lip
collision, not opening, causes tissue excita-
tion that generates pressure waves emanat-
ing as clicks from the melon, which supports
previous speculations (3, 15). A click-triggered,
averaged image upsampled to 100,000 fps
suggests that the phonic lip leading edge
collides before the trailing edge (Fig. 2G). This
is consistent with a caudocranial travelling
Madsen et al., Science 379, 928–933 (2023)
3 March 2023
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Odontocete echolocation clicks
are generated in nasal passages.
(A) Odontocetes evolved the ability to
echolocate (32). (B) In vivo recording
schematic with hydrophone, endoscope,
and pressure catheters placed above
(psupra) and below (psub) the phonic
lips (PLs). NP, nasal plug (23). (C) Harbor
porpoise and bottlenose dolphin
echolocation clicks are radiated during
(D), psub increase. (E) Probability density
functions (PDF) per individual. The
mean (m) psub values per species
during click emission were 5.70 ±
0.75 kPa and 3.91 ± 1.57 kPa for
dolphin and porpoises, respectively.
(F) Full-frame color video (25 fps) during
inspiration (left), showing the posterior
PLs (red arrowheads), psub pressure
catheter, and location of the 7200 fps
videokymography (VKG) linescan.
During psub increase (right), anterior
PLs (yellow arrowheads) quickly move
posterior. (G) Example VKG, summed
VKG intensity, and hydrophone signal in
the nares of two dolphins and (H) three
porpoises with click (blue) and motion
onset (orange vertical lines). (I) Distribu-
tions of click-associated motion delay
(dt) observations per individual.
mucosal wave, critical to sustaining vibra-
tions in myoelastic-aerodynamic sound sources
(12, 13). Because natural porpoise clicks can
be produced in vitro without motor control,
clicks are produced by phonic lip oscillations
conforming to the myoelastic-aerodynamic
theory and not by superfast muscle contraction–
induced motion. Thus, odontocetes do not con-
trol the timing and level of individual clicks,
but modulate click rates and levels by motor
control of phonic lip tension and nasal pressure.
The odontocete acoustic repertoires arise
from distinct vocal registers
communication sounds described qualitatively
as bursts, grunts, and whistles (18). These sounds
are presumably also produced by phonic lips
(6, 9), but it remains puzzling how the same
structures generate such diversity of acoustic
signals. A characteristic of laryngeal myoelastic-
aerodynamic sound production is that vocal
folds exhibit self-sustained oscillation in dis-
tinct patterns, so-called registers, or laryngeal
mechanisms (19) (Fig. 3). However, it is un-
known whether odontocete phonic lips can
vibrate in different registers and, if so, if these
registers could generate the diversity of ob-
served acoustic signals.
Next to clicks, many odontocetes produce rich
repertoires of lower-intensity, lower-frequency
In humans, at least three registers are re-
cognized (Fig. 3A) (19, 20). In vibrational
register M0 (vocal fry), both vocal fold cover
and body are slack, leading to the lowest
frequencies with a short open phase of the
vibratory cycle [open quotient (OQ) of 0 to
0.4] (19). Glottal flow is low and vocal fold
acceleration and tracheal sound are pulsatile
(19, 20). In register M1(chest), vocal folds are
lengthened and the vocal fold body is stiffer
than the cover, leading to higher vibration
frequencies at an OQ between 0.3 and 0.8.
Glottal flow is low and vocal fold accelera-
tion and tracheal sound waveforms are tri-
angular and sinusoidal, respectively (19, 20).
In vibrational register M2 (falsetto), vocal
folds are lengthened further with both body
and cover stiff, leading to higher frequencies
Madsen et al., Science 379, 928–933 (2023)
3 March 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Harbor porpoise echolocation clicks are produced by colliding phonic
lips during flow-induced voiced-sound production. (A) In vitro setup to induce
sound production with removed nasal plug (NP). (B) Still image of open and closed
right PLs filmed at 10,000 fps. Vertical yellow line indicates Digital Kymogram (DKG,
right) location. (C) Signal overview and detail (dashed rectangle) (D) during induced
click. (E) Distributions of PL closing (top) and opening (bottom) events relative to
click emission from the melon show that (F) closing is closely associated with
phonic lip collision, not opening. Colors indicate individuals, with specimen P24152
(10,000 fps, 12 bits) in purple. (G) Averaged DKG (bottom) of 30 upsampled
(100,000 fps) DKG segments aligned to click emission (top).
and an OQ between 0.5 and 0.95 (19). Glot-
tal flow is high and vocal fold acceleration
and tracheal sound waveforms are nearly si-
nusoidal with strong fundamental frequency
(fo) (19).
Thus, at least three independent features
inform a test of our hypothesis that odonto-
cetes vocalize in different registers. (i) Ana-
tomically, layered phonic lips would facilitate
registers by allowing differential tension be-
tween cover and body (19, 21, 22). (ii) Acous-
tically, different register vocalizations should
have distinct sound pressure waveforms and
overlapping, increasing fo ranges (Fig. 3D) (20).
(iii) Physiologically, different register vibra-
tions should have increasing OQ values (19).
To test whether odontocete phonic lip anat-
omy supports register vibrations, we quantified
phonic lip geometry using contrast-enhanced
DiceCT (10). Corroborating earlier work (5, 23),
we observed a superficial cover layer in por-
poise phonic lips with bilaterally paired deeper
fat bodies called bursae (Figs. 3B), which are
present in all odontocetes (5). We propose that
bursae are functionally analogous to the vocal
fold body layer and allow deep tissue rotation
during paired phonic lip oscillation (Fig. 3B).
Furthermore, we propose the cover layer has
two morphological adaptations to optimize
M0 vibration. First, increased height com-
pared with that of laryngeal vocal folds (Fig.
3B), which reduces OQ (22). Second, phonic
lips are covered by ridges (Fig. 3B and fig. S1)
that lead to strong stiffness anisotropy in the
superficial layer, but not in the body. This
anisotropy improves glottal closure and the
ability to maintain adductory position (22)
against driving air pressures as high as 40 to
81 kPa (11).
Madsen et al., Science 379, 928–933 (2023)
3 March 2023
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
To test whether the acoustic diversity of
odontocete phonations supports different re-
gisters, we compiled acoustic repertoires for
representative species from all major clades (10)
(table S1). The repertoires indeed comprise dis-
tinct phonation types of which the waveforms
are pulsatile (“clicks”), harmonic (“bursts”), or
sinusoidal (“whistles”), consistent with the
vocal fold acceleration waveform of registers
M0 to M2 (Fig. 3C). Furthermore, the vocal-
ization types have overlapping, increasing fo
ranges (Fig. 3, D and E, and table S1).
Vocal fry powers echolocation
Next, we tested whether phonic lip vibration
during clicks is consistent with the M0 regis-
ter and measured clicks OQ values in free-
swimming odontocetes (10) (Fig. 4, A to C).
Because we could not film phonic lip kine-
matics in the open ocean, we leveraged the
causal biomechanical interpretation of our
in vitro and in vivo data (Figs. 2 and 3). We mea-
sured the timing of acoustic signatures of phonic
lip opening and closing events on microphone
and hydrophone signals during click emission
in trained porpoises and dolphins (10) (Fig. 4,
A to C), and on hydrophone signals of acoustic
biologging tags during dives in free-swimming
Bottlenosed dolphins, false killer whales, and
sperm whales, down to depths of 1800 m (Fig.
4B and fig. S2). For clicks, the OQ remained
<40% in all species and increased significantly
with click rate (Fig. 4E and table S2). Addition-
ally, opening duration changed with click rates
so that OQ stayed <10% (Fig. 4F). We propose
that odontocetes reduce OQ to increase phonic
lip acceleration, thereby increasing click source
level. Thus, clicks in free swimming odonto-
cetes are produced by phonic lip vibration in
the M0 register, and their OQ is extremely low.
Lastly, we measured an OQ of 0.35 to
0.8 during burst vocalizations in a bottlenose
dolphin (Fig. 4, figs. S3 and S4, and movie
S4), which was consistent with M1 register
vibration and distinct from clicks (M0) in the
same fo range. The similar OQ ranges for M0
and M1 vibration in human laryngeal voice
production (19) suggest either functional con-
vergence between independently evolved
systems or a basal transition property of
myoelastic-aerodynamic sources.
Discussion
We show that odontocetes have evolved an
air-driven nasal sound source that is phys-
ically and functionally analogous to laryngeal
and syringeal sound production in mam-
mals and birds (12, 13). The well-established
myoelastic-aerodynamic theory for voiced sound
production (13) provides a solid physiological
basis to clasify odontocete vocal repertoires.
Odontocetes use vocal registers to generate
low-frequency, low-directionality, low–source
level communication signals (9) and high-
Fig. 3. The odontocete vocal repertoire and anatomy supports different vocal registers. (A) The
human laryngeal vocal registers, M0 to M2. OQ, open quotient. (B) MicroCT scan of porpoise right PL.
(C) Examples of bottlenose dolphin vocal repertoire with spectrogram (top), waveform (bottom left), and
phase-space plot (bottom right) of click, burst, and whistle calls. (D) In humans, vocal registers M0 to
M3 overlap in frequency ranges (table S1). (E) Frequency ranges of different vocalizations across odontocete
clades support three vocal registers (table S1). N.a., no available data.
frequency, directional, high-power echoloca-
tion clicks. Vocal registers have been confirmed
only in humans (19) and crows (24). They ex-
pand the fo range and thus enhance vocal
plasticity—a potential prerequisite for vocal
learning. Bilateral specialization of phonic lip
pairs may also aid in expanding the odontocete
vocal range and allow simultaneous echo-
location and communication (9). Acoustic
source triangulation suggests that dolphins
clicks are produced by right, and M1 and
M2 vocalizations by the left phonic lips (9).
As in songbirds, in which bilateral hemisyrinx
tissues differentiate to produce different fo
ranges (25), odontocetes may use neural or
anatomical specializations to exploit lateral-
ized register vibrations.
We propose five advantages that drove stem
odontocetes to evolve an air-driven nasal
source to replace the ancestral laryngeal source.
Madsen et al., Science 379, 928–933 (2023)
3 March 2023
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. The M0 register produces fast
echolocation clicks in odontocetes.
(A) PL opening and closing dynamics
were reconstructed using acoustic
signatures in restrained and (B) tagged,
freely swimming odontocetes (10).
(C) Delay between (left) PL opening
and microphone pressure decrease
and (right) the latter and hydrophone
preclick onset. (D) Clicks time-
aligned to onset in a freely hunting
35-ton sperm whale at 1800-m depth
(fig S2). (E) OQ remains under 40%
in trained and hunting animals,
which coincides with the maximal
OQ for M0 of 40% in humans.
(F) Open-phase duration decreases
with click rate so that OQ in
free-swimming animals remains
below 10%.
First, a nasal source freed the larynx from sound
production, resulting in an effective valve that
decoupled the lungs and nasal passages. This
decoupling allows odontocetes 80-kPa driving
pressures (11)—one order of magnitude above
laryngeal driving pressures—to power the prod-
uction of the highest biological source levels
(26, 27) without damaging lung tissues. Sec-
ond, the nasal air volume is much smaller than
that of the respiratory system, therefore allow-
ing faster pressure control and air recycling.
Third, the partially bone-lined nasal airspace
provides mechanical resistance to hydrostatic
compression and allows pressure gradient
buildup over phonic lips at depths below ~100 m
where the respiratory air volume matches the
respiratory dead space (28) and lung air can-
not be pressurized. Fourth, a sound source
anterodorsal of the sound-reflecting skull gen-
erates a highly directional forward sound beam,
acoustically isolated from their sensitive ears
(29), which is not possible with laryngeally prod-
uced sounds. Fifth, by decoupling sound emis-
sion from their buccal cavity, odontocetes can
catch and ingest food while echolocating. We
suggest that because of anatomical adapta-
tions to phonic lip closure, nasal-source evo-
lution was primarily driven by selection on
echolocation signals and only secondarily co-
opted for social communication.
Vocal fry (M0) vibration with ultrashort open
times of 0.5 to 2.5 ms uncovers the biomecha-
nical key to making prolonged click trains
suitable for prey pursuit and capture. Because
tissue acceleration directly translates to sound
pressure in water, M0-register tissue accelera-
tion generates short, broadband, powerful pres-
sure transients in tissue and water that provide
ideal long-range, high-frequency echolocation
signals. Odontocetes thus exploited a feature
of air-driven tissue vibration not advantageous
in air. Furthermore, this air-driven system al-
lows click repetition rates up to 500 clicks per
second in the hunting phase prior to prey cap-
ture called the terminal buzz (2) (table S1),
which is critical to active prey pursuit and cap-
ture. Our data show that the maximal buzz rate
is set by the M0-to-M1 transition, which is
determined by complex interplays of phonic
lip size, tissue properties, posturing, and fluid-
structure interaction (21, 22). Just as in echolo-
cating bats (30), the maximal buzz rate is thus
facing a constraint related to sound produc-
tion. Lastly, because of low OQ values, M0 is
the most air-economical register (31). Odonto-
cetes can make clicks with <50 ml of air per
click (8). Such air economy allowed for echo-
location at great depths, which opened the
previously unexplored food niches of the deep
ocean for exploitation.
REFERENCES AND NOTES
J. A. Goldbogen et al., Science 366, 1367–1372 (2019).
1.
2. H. Vance et al., eLife 10, e68825 (2021).
3. T. Cranford et al., J. Exp. Mar. Biol. Ecol. 407, 81–96
(2011).
4. E. Chereskin et al., Curr. Biol. 32, 1657–1663.e4 (2022).
5. T. W. Cranford, M. Amundin, K. S. Norris, J. Morphol. 228,
223–285 (1996).
6. S. Ridgeway et al., in Animal Sonar Systems, R. G. Busnel,
J. F. Fish, Eds. (Plenum, 1980), pp. 239–249.
7. K. Dormer, J. Acoust. Soc. Am. 65, 229–239 (1979).
8.
I. Foskolos, N. Aguilar de Soto, P. T. Madsen, M. Johnson,
Sci. Rep. 9, 15720 (2019).
Madsen et al., Science 379, 928–933 (2023)
3 March 2023
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
9. P. T. Madsen, M. Lammers, D. Wisniewska, K. Beedholm,
25. R. Suthers, S. Zollinger, in Neuroscience of Birdsong,
J. Exp. Biol. 216, 4091–4102 (2013).
10. Materials and methods are available as supplementary
materials.
11. M. Amundin, S. Andersen, J. Exp. Biol. 105, 275–282
(1983).
12. C. P. Elemans et al., Nat. Commun. 6, 8978 (2015).
13. J. G. Švec, H. K. Schutte, C. J. Chen, I. R. Titze, J. Voice 21,
00055-2 (2021).
14. J. H. Lee et al., Sci. Adv. 4, eaat9660 (2018).
15. J. J. Finneran et al., J. Comp. Physiol. A Neuroethol. Sens.
Neural Behav. Physiol. 204, 271–283 (2018).
16. A. F. Mead et al., eLife 6, e29425 (2017).
17. F. Ladich, H. Winkler, J. Exp. Biol. 220, 2306–2317
H. P. Zeigler, P. Marler, Eds. (Cambridge Univ. Press, 2008),
pp. 78–98.
26. L. Jakobsen, J. Christensen-Dalsgaard, P. Juhl,
C. P. H. Elemans, Front. Ecol. Evol. 9, 657254 (2021).
27. B. Møhl, M. Wahlberg, P. T. Madsen, A. Heerfordt, A. Lund,
J. Acoust. Soc. Am. 114, 1143–1154 (2003).
28. A. Fahlman, M. J. Moore, D. Garcia-Parraga, J. Exp. Biol. 220,
1761–1773 (2017).
29. F. H. Jensen, M. Johnson, M. Ladegaard, D. M. Wisniewska,
P. T. Madsen, Curr. Biol. 28, 3878–3885.e3 (2018).
30. C. P. Elemans, A. F. Mead, L. Jakobsen, J. M. Ratcliffe, Science
333, 1885–1888 (2011).
31. R. E. McGlone, T. Shipp, J. Speech Hear. Res. 14, 769–775
(2017).
(1971).
18. B. Jones, M. Zapetis, M. Samuelson, S. Ridgway, Bioacoustics
32. N. D. Pyenson, Curr. Biol. 27, R558–R564 (2017).
29, 399–440 (2020).
19. B. Roubeau, N. Henrich, M. Castellengo, J. Voice 23, 425–438
ACKN OWLED GMEN TS
(2009).
20. H. Hollien, J. Phonetics 2, 125–143 (1974).
21.
I. T. Tokuda, J. Horácek, J. G. Svec, H. Herzel, J. Acoust.
Soc. Am. 122, 519–531 (2007).
22. Z. Zhang, J. Acoust. Soc. Am. 140, 2614–2635 (2016).
23. S. Huggenberger, M. A. Rauschmann, T. J. Vogl,
H. H. Oelschläger, Anat. Rec. (Hoboken) 292, 902–920
(2009).
24. K. K. Jensen, B. G. Cooper, O. N. Larsen, F. Goller, Proc. Biol. Sci.
274, 2703–2710 (2007).
We thank the Harderwijk staff—P. Bunskoek, N. van Elk, H. Goelema,
S. Hearn, J. Mosterd, L. van Ooijen, and J. Wouters—for excellent
support; H. Goelema and L. van Ooijen for training and scoping animals
at Harderwijk; M. Ding for microCT scanning; M. Kieler, K. Beedholm,
P. Tønnesen, and M. Ladegaard for discussions and analytical support;
and W. T. Fitch, M. Johnson, J. Ratcliffe, M. Wahlberg, M. Amundin,
and two anonymous reviewers for discussions and manuscript
feedback. We dedicate this paper to the late Dr. Sam Ridgway. Funding:
Danish research council grant 6108-00355A and Carlsberg Foundation
grant CF16-0405 to P.T.M. and Carlsberg Foundation grant CF14-1096
and Novo Nordisk Foundation grant NNF20OC0063964 to C.P.H.E.
Author contributions: Conceptualization: C.P.H.E. and P.T.M.
Methodology: C.P.H.E. and P.T.M. In vivo and in vitro experimental
design and analysis: C.P.H.E. Tag data acquisition and analysis:
P.T.M. Porpoise head collection: U.S. Visualization: C.P.H.E. Funding
acquisition: C.P.H.E. and P.T.M. Project administration: C.P.H.E.
Writing – original draft: C.P.H.E. and P.T.M. Writing – review and
editing: C.P.H.E., U.S., and P.T.M. Competing interests: Authors
declare that they have no competing interests. Data and materials
availability: All data are available in the main text or the
supplementary materials. Code is available at the Zenodo Repository
(33). License information: Copyright © 2023 the authors, some
rights reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government works.
https://www.science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adc9570
Materials and Methods
Figs. S1 to S5
Tables S1 and S2
References (33–55)
Movies S1 to S4
View/request a protocol for this paper from Bio-protocol.
Submitted 12 May 2022; accepted 9 January 2023
10.1126/science.adc9570
Madsen et al., Science 379, 928–933 (2023)
3 March 2023
6 of 6
|
10.1126_science.add5462
|
RES EARCH
HYDROLOGY
Agricultural expansion raises groundwater and
increases flooding in the South American plains
Javier Houspanossian1,2*†, Raul Giménez1,2, Juan I. Whitworth-Hulse1, Marcelo D. Nosetto1,3,
Wlodek Tych4, Peter M. Atkinson4, Mariana C. Rufino4,5, Esteban G. Jobbágy1*†
Regional effects of farming on hydrology are associated mostly with irrigation. In this work, we show
how rainfed agriculture can also leave large-scale imprints. The extent and speed of farming
expansion across the South American plains over the past four decades provide an unprecedented case
of the effects of rainfed farming on hydrology. Remote sensing analysis shows that as annual crops
replaced native vegetation and pastures, floods gradually doubled their coverage, increasing their
sensitivity to precipitation. Groundwater shifted from deep (12 to 6 meters) to shallow (4 to 0 meters)
states, reducing drawdown levels. Field studies and simulations suggest that declining rooting
depths and evapotranspiration in croplands are the causes of this hydrological transformation. These
findings show the escalating flooding risks associated with rainfed agriculture expansion at
subcontinental and decadal scales.
G lobal demand for grain is leading to
the replacement of large areas of South
America’s native grasslands and forests
as well as cultivated pastures by agri-
cultural crops, mainly soybean (1–3),
providing sustainability and environmental
challenges. Currently, the area covered by an-
nual crops throughout the continent expands
at a rate of 2.1 Mha year−1 (1), particularly in
the sedimentary plains of the Pampas tem-
perate grasslands and the Gran Chaco sub-
tropical forests (4). Flanked by the Andes to
the west and the Brazilian shield to the east,
this region represents one of the flattest and
most hydrologically stagnant plains on the
planet (5, 6). Such flat sedimentary settings
host some of the best farming soils on Earth,
and their hydrology is particularly sensitive
to water balance shifts introduced by land
and water use changes (7, 8). Thus, the ex-
pansion of agriculture on these plains at such
a rate and scale provides both a sustainability
challenge and an outstanding experimental
setting to explore vegetation’s role in mod-
ifying hydrology. This includes the effects of
shifts in the magnitude and timing of water
demand by cultivated canopies as well as the
changing depths of soil explored by rooting
systems (9).
Crop irrigation has introduced some of the
most prominent, large-scale hydrological trans-
formations on Earth, including substantial
water table level rises when surface water is
1Grupo de Estudios Ambientales, CONICET, San Luis,
Argentina. 2Departamento de Geología, National University of
San Luis, San Luis, Argentina. 3Cátedra de Climatología
Agrícola, Universidad Nacional de Entre Ríos, Entre Ríos,
Argentina. 4Lancaster Environment Centre, Lancaster
University, Lancaster, UK. 5Livestock Systems, TUM School
of Life Sciences, Technical University Munich, Freising,
Germany.
*Corresponding author. Email: jhouspa@gmail.com (J.H.);
jobbagy@gmail.com (E.G.J.)
†These authors contributed equally to this work.
the major source and groundwater deple-
tion when its use gains predominance (10, 11),
as detected at continental scales with remote
sensing tools (12, 13). By contrast, subtle changes
in rainfed cultivation can also cause water
excess and water table level rises (7, 8) ob-
servable at local scales and associated with
regional land degradation processes by water-
logging and salinization of both croplands
and natural vegetation relicts in western Aus-
tralia (14), the Sahel (15, 16), and, more re-
cently, the South American Pampas (17) and
the Chaco (18). In this study, we present evi-
dence of such a process taking place at a sub-
continental scale in the central plains of South
America, revealing the critical role that vege-
tation plays in hydrology there and, likely,
over the water storage and fluxes of other
major sedimentary plains across the globe.
Hydrological shifts and their link to land use
change are characterized by a continental-
scale remote sensing analysis of flooding trends
and exploration of their spatial drivers, a re-
gional description of groundwater level regime
shifts based on long-term public records, and
a statistical synthesis of previous local studies
that compares the effects of contrasting vege-
tation types on groundwater dynamics. These
analyses together with simulations using an
ecohydrological model identify the likely mech-
anisms that link rainfed cropland expansion
with rising groundwater levels and increasing
flooding.
Shifting hydrological regimes
We found that the large-scale replacement of
native vegetation and pastures by rainfed crop-
lands that has been taking place over the past
four decades in the major grain-producing
area of South America was accompanied by
increasing flooding in time and space. Fine-
resolution remote sensing imagery shows a
progressive appearance of clusters of newly
flooded land since 1977 (pixels that display
water cover for the first time) at a rate of
~700 km2 year−1 in the central plains that is
unseen in the rest of the continent (Fig. 1A,
black line rectangle). Analysis at a coarser
spatial resolution (0.5° cells for which the pro-
portion of flooded areas incorporated every
year was calculated) (table S1 and fig. S1)
shows that the areas where floods expanded
are zones of extreme flatness [i.e., height
above nearest drainage (19) ≤7 m; darker area
in Fig. 1A] with a relatively high proportion
of rainfed cultivation (i.e., agricultural frac-
tion >25%). Notably, these flooded areas con-
tinued increasing in size after 2000 (time at
which the expanding trend of agriculture
across the plains consolidated), whereas the
rest of the continent exhibited stable flooding
conditions throughout 1985 to 2019. Before
2000, the plains of South America accounted
for 40% (10.44 Mha) of the total flooded area,
with most of it (72%, or 7.52 Mha) located in
wetlands and other less-cultivated land (table
S1). By contrast, more than three-quarters of
the newly flooded areas (i.e., those flooded
for the first time after 2000) are in the plains,
and half of these are highly cultivated land
(1.26 Mha). Boosted regression modeling (BRT)
applied to gridded data (water-covered frac-
tion in 0.5° cells) of the whole continent, con-
sidering the proportion of flooded territory
before 2000 (pre-2000) and that flooded for
the first time after that year (post-2000), shows
that wetland area is the strongest explanatory
covariate with a positive association with flood-
ing in both periods (33.6 and 32.1% of the
variance in pre- and post-2000, respectively).
The fraction of woody vegetation cover shows
the next-strongest association with flooding,
being negative in both periods (14.6 and 25.2%
of the variance in pre- and post-2000, re-
spectively; figs. S2 to S4). Conversely, although
the agricultural fraction is not substantially
influential pre-2000, it becomes the third
variable in importance—explaining 9.9% of
the variance—post-2000, with a positive asso-
ciation with flooded area (figs. S4 and S5). The
importance of agricultural fraction becomes
stronger when the BRT is restricted to the
plains of South America, being the second
variable in importance explaining 19.8% of the
variance of post-2000 flooding (fig. S5). Accord-
ing to the BRT analysis, newly flooded areas
appear only in grid cells with >50% cultivated
area (figs. S4 and S5).
Analysis of the evolution of floods within the
cultivated areas of the central plains reveals
a gradual incorporation of areas newly cov-
ered by water following a north-west direc-
tion. The newly flooded areas expanded away
from the historically flooded environments,
known as the Flooding Pampas (boundary
shown by the black dotted line in Fig. 1C),
into the drier aeolian landscapes of the western
Houspanossian et al., Science 380, 1344–1348 (2023)
30 June 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Changes in the surface
flooded in South America during
the past four decades. (A) Initial
flooded area before 2000 (bubble
size; 1985 to 1999 time period) and
newly flooded area added after
2000 (bubble color; 2000 to 2019
time period) across 0.5° grid cells
in South America based on the
Landsat Surface Water Cover
product with 30-m spatial resolu-
tion. Shades of gray represent the
terrain topography using height
above the nearest drainage
(HAND). (B) Proportion of the area
occupied by agriculture in 2019
based on the Copernicus Land
Cover product, with 100-m spatial
resolution applied to 0.5° grid cells.
Inset area of the central plains
of South America is displayed (black
rectangle). (C) Detail of the intensive
agricultural area of the central
plains of South America in the
Argentine grain belt and the timing
of floods (pentad of first flooding
episode since 1977 based on Land-
sat imagery). The locations
[stations (ST) 1 to 8] of long-term
public water table level monitoring
sites used in this study as well as
aridity index isolines are shown.
Dotted lines indicate the boundary
of the subregion that has been
considered flood prone historically
(the Flooding Pampas). Dashed lines
indicate Mar Chiquita Lake and the
Parana river delta. (D to H) Land-
scape-level detail displaying the pro-
gression of flooding from 1977 to
2019 in five sample areas.
Pampas between the late 1980s and 1990s
and even to the drier and warmer fluvial and
aeolian landscapes of the southern Chaco after
2000 and at a faster rate after 2015 (Fig. 1C).
Within each of these regions, fine-resolution
analysis of flood progression reveals concen-
tric patterns of expanding temporary water
bodies within the cultivated landscape (Fig. 1,
D to H). In the highly cultivated areas of the
plains (grid cells with >25% agriculture cover;
fig. S1) newly flooded areas were annexed in
the four regional flooding cycles observed
between 1977 and 2019 (Fig. 2A), reaching a
maximum area of 3.8 Mha in 2018—almost
doubling the maximum surface flooded in
the first decade of that period. Although all
long-lasting flooding episodes coincided with
periods of relatively high precipitation (Fig.
2B), their magnitude of flooding increased
with time, particularly in the last episode,
although no significant concurrent increase
Houspanossian et al., Science 380, 1344–1348 (2023)
30 June 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Increase in the flooded surface of
the central South American plains in
response to precipitation and agricultural
fraction increases. (A) Annual flooded area in
megahectares of the grain belt of the central
plains of South America (Fig. 1C; Mar Chiquita
Lake and Parana river delta areas were not
considered in the curve) with four steps
of newly flooded area since 1977 colored by the
pentad of first detection (colored bars). Modeled
flooded area subtracting the output of the
annual precipitation-driven dynamic transfer
function (DTF) model is shown as the output
(solid blue line; DTF model fit Rt
with its standard error (dashed blue lines).
(B) Mean annual precipitation (black circles)
and integrated random walk (IRW) smooth
trend estimate for annual precipitation (black
solid line) shown with its approximate 95%
confidence interval (black dashed lines).
(C) DTF time-varying gain parameters (blue
circles) and smoothed trend (blue line)
and their approximate 95% confidence intervals
(blue dotted lines), showing effective time-
varying sensitivity of flooding to precipitation.
Smoothed agricultural land use percentage of
Argentine grain belt (red solid and dotted
lines) is also shown.
2 = 0.9)
in precipitation was observed (fig. S6). The
analysis of last-century precipitation, based
on both aggregated and individual station
data (fig. S7, A and B), shows the previously
documented rise between the 1970s and the
1990s (20) but no sustained trend over the
past 30 years. In this period, there are no
significant trends in the characteristics of
rainfall events, including the frequency and
magnitude of the most extreme events (21).
Although associated with positive precipita-
tion anomalies, floods in the central plains
of South America display an increasing sen-
sitivity to rainfall through time, as shown by
the analysis using a dynamic transfer func-
tion model. The dynamic transfer function
produced a good model fit, explaining just
>90% of the variance of the flooded area (Fig.
2A). The temporal analysis reveals a growing
sensitivity of flooding to positive precipitation
anomalies, indicated by the increase in the
gain parameter of the dynamic transfer func-
tion model linking flooded area to mean an-
nual precipitation (Fig. 2C).
Expanding floods in highly cultivated and
flat landscapes were accompanied by rising
water table levels, as indicated by the long-
term records from eight public monitoring
stations distributed across the study region.
The time series analysis shows a generalized
rise in levels that reaches, with different tim-
ings, a similarly stable shallow groundwater
position (−4 to 0 m) at all sites (Fig. 3 and fig.
S8). Hidden Markov models (22) show that
at all sites, station-level variability is best ex-
plained by two or three states (lower Akaike
information criterion), which suggests shift-
ing regimes (figs. S9 to S16), with deep, transi-
tional, and shallow groundwater level states
(Fig. 3 and fig. S8). Notably, shallow states
did not transition to deeper states even during
years with low precipitation.
The hydrological effects of rainfed cultivation
All available field observations provide inde-
pendent and convergent evidence about the
imprint of vegetation and land use changes
on groundwater levels. The analysis of all pub-
lished studies (seven reports for 19 sites) of
groundwater levels and discharge-recharge
fluxes across paired stands comparing crop-
lands with native or planted forests and pas-
tures of the region shows the consistent effects
of croplands on the hydrology of the plains
(table S2). Pooled together, these observa-
tions indicate higher water table levels (mean
effect of +1.54 m), and, individually, they evi-
dence increased groundwater recharge and/or
decreased discharge in croplands compared
with the rest of the vegetation types at all sites
(Fig. 4). In the subset of studies conducted in
drier sites with water table levels >10 m deep,
salt profiles in the unsaturated zone indicate
a full inhibition of deep drainage in native
forests that lasted millennia until the present.
This no-recharge condition was interrupted
by the onset of deep drainage under neigh-
boring cropland (table S2). The remaining
studies, conducted where water tables are
shallower than 10 m, offer evidence of ground-
water consumption (in most cases through
diurnal water table level fluctuations) down
Houspanossian et al., Science 380, 1344–1348 (2023)
30 June 2023
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Hidden Markov models applied to time series
of water table depth at three sites in the central
plains of South America in the Argentine grain belt,
indicating three alternative hydrological states:
deep, transitional, and shallow. Agricultural fraction
per year and per site are shown, drawn from depart-
mental agricultural statistics corresponding to the
counties where the stations are located. (A) For station
1 (located at Laboulaye, Argentine; ST1 in Fig. 1C), water
table levels rose in the mid-1970s, reaching surface
levels in the 1990s. (B and C) For station 2 (B) (located
in Anguil, Argentine; ST2 in Fig. 1C) and for station 3
(C) (located in Marcos Juarez, Argentine; ST3 in Fig. 1C),
the water table rose in the mid-1970s gradually to
surface levels in the 2000s. Similar situations are
displayed in the rest of the stations (fig. S8). The
vertical bars in (A), (B), and (C) denote the 5th and 95th
percentile estimates of water table depth observed.
A
B
C
)
m
(
h
t
p
e
d
l
e
b
a
t
r
e
t
a
W
0
-3
-6
-9
0
-3
-6
-9
0
-3
-6
-9
Hydrological state
Deep
Shallow
Transitional
Agriculture (%)
25 50 75
s
t
a
t
i
o
n
1
s
t
a
t
i
o
n
2
s
t
a
t
i
o
n
3
1920
1930
1940
1950
1960
1970
Date
1980
1990
2000
2010
2020
Fig. 4. Infield evidence
of water table depth
change under land cover
changes in the South
American plains. (A) Mean
water table depth in paired
plots of croplands and native
dry forests or eucalypt
plantations and of croplands
and natural grasslands or
alfalfa pastures across the
central plains of South
America. Numbers in
parentheses indicate the
studies referenced in table
S2 as follows: (1) Giménez
et al. (2016) (18); (2)
Jobbágy et al. (2021) (33);
(3) Nosetto et al. (2013)
(34); (4) Amdan et al.
(2013) (35); (5)
Nosetto et al. (2015) (36);
and (6) Pal et al. (2021) (37).
(B) Boxplots of water table
depth in forests and/or plantations, grasslands and/or pastures, and croplands. The crossbar within the box indicates the median, the length of the box reflects the interquartile
range, the whiskers show the 95% confidence intervals, and the dots represent the outliers. (C) Comparison between the water table depths in croplands and their adjacent forests
and plantations (n = 4 pairwise comparisons) and grasslands and pastures (n = 3). Small squares and triangles refer to the mean effect size (Hedges’s d) for individual pairwise
comparison, and large squares and triangles represent the overall mean effect size estimate for comparisons between forests and/or plantations and croplands and between
grasslands and/or pastures and croplands, respectively. The horizontal bars around the means refer to the 95% confidence intervals. The absolute values of the means and
confidence intervals (in brackets) are shown on the right. Positive mean effect sizes indicate that the water table depth is shallower in croplands than in forests or plantations and
grasslands or pastures. A mean effect size is significantly different from zero when its 95% confidence interval does not include zero (significance level, ***P < 0.001).
Houspanossian et al., Science 380, 1344–1348 (2023)
30 June 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
to 5- to 9-m depth under native forests, tree
plantations, and perennial pastures but only
down to 3-m depth or less under annual crop-
lands (table S2). These contrasts explain in the
first place how cultivation may trigger water
table level rises during wet periods and limit
water table drawdowns during subsequent
dry periods, reducing the threshold of cumu-
lative water excess needed to cause water-
logging and flooding.
Rising flooding risk is widely assumed to
be a phenomenon associated with unusually
high precipitation events, reduced infiltration
rates, and limited run-off caused by land cover
changes (23, 24), whereby the recovery after
floods is a function of the duration of subse-
quent dry periods (22). Although this is the
case for sloped watersheds, extremely flat sedi-
mentary regions such as the South American
plains with stagnant hydrological systems
favor shallow water table level equilibria under
a broad range of climatic conditions (6). Under
these circumstances, vegetation changes play
a major role in modulating flooding, mostly
through the capacity of plants to draw down
unsaturated and saturated water reserves dur-
ing dry periods (figs. S17 to S19). The two most
critical and interconnected features of vegeta-
tion water use that dictate the speed at and
extent to which water reserves are depleted
are its evaporative capacity and rooting depth
(25). These factors control the depth of the
drawdowns and how much of the water stor-
age capacity can be emptied. Therefore, when
rooting depths and groundwater drawdowns
are reduced, there is a higher likelihood of
saturating the soil profile and bringing water
table levels to the surface during subsequent
wet periods, as suggested by one-dimensional,
long-term hydrological simulations using the
Hydrus model (26) calibrated for the west
Pampas (fig. S19).
Discussion
The replacement of forests and pastures with
rainfed annual crops may have been enough
to create a new dynamic hydrological equi-
librium that, even under low-precipitation
conditions, may not recover the initial deep
groundwater levels, making the whole region
more flood prone. Although ecohydrological
recovery is likely under a scenario of farming
abandonment and native vegetation recovery
and/or cultivated pastures reestablishment,
the effects of the observed hydrological shifts
on these agricultural systems do not seem to
have discouraged annual crop expansion until
the present (1). In the specific case of agricul-
ture and in the short term, the positive effects
of a shallower water table buffering crop pro-
duction during droughts appear to compen-
sate for the negative effects of reduced sowing
area (27). However, as flooding continues to
expand into drier areas with more saline and
erodible soils, this compensatory effect may
disappear (14, 17). Additionally, important
trade-offs with other sectors of the rural econ-
omy have already been reported, including
damages to the infrastructure of dairy farms
and disruptions to the logistics of small agri-
cultural towns. Climate may also be affected
at global-to-local scales by these hydrological
shifts. Increased intermittent waterlogging is
likely to increase methane and water vapor
fluxes to the atmosphere, limiting soil organic
matter sequestration and contributing to glo-
bal warming (28). At a regional level, floods
have been shown to cause more-frequent fog
events, reduce thermal amplitude, and extend
frost-free periods, affecting both cultivated
and natural ecosystems (29). The historical con-
text of low investment in agricultural infra-
structure and taxation (rather than subsidies)
on grain production (30) in Argentina sug-
gests that drainage would not be a viable fu-
ture response to the increase in flooded area,
and neither would nature-based solutions, such
as hydrology-smart rotations and landscape
designs, become feasible (31). Any one of these
approaches needs to consider the critical role
of rooting depths in the stagnant context of
sedimentary plains and the possibilities to
adjust them through crop and cultivar choice,
native vegetation conservation, and hydrolog-
ically informed decision-making. The insights
provided in this work on the connection be-
tween extensive land use change and the grow-
ing flood sensitivity in the South American
plains will help improve our understanding
of hydrological changes in other regions of
the world with similar characteristics, includ-
ing dry and stagnant hydrological settings and
expanding rainfed grain production systems,
such as those of Ukraine or Canada (32). The
findings presented here are critical for future
land use policies that support farming, water
management, and rural towns in smarter and
more-integrated ways.
RE FERENCES AND NOTES
1. X.-P. Song et al., Nat. Sustain. 2021, 784–792 (2021).
2. V. Zalles et al., Sci. Adv. 7, eabg1620 (2021).
3. D. M. Olson et al., Bioscience 51, 933 (2001).
4. P. Potapov et al., Nat. Food 3, 19–28 (2022).
5. K. E. Giller et al., Food Secur. 13, 1073–1099 (2021).
6. Y. Fan, H. Li, G. Miguez-Macho, Science 339, 940–943
(2013).
7. S. Kuppel, J. Houspanossian, M. D. Nosetto, E. G. Jobbágy,
Water Resour. Res. 51, 2937–2950 (2015).
8. N. Ramankutty et al., Annu. Rev. Plant Biol. 69, 789–815
(2018).
9. E. Hauser, P. L. Sullivan, A. N. Flores, D. Hirmas, S. A. Billings,
Earths Futur. 10, e2022EF002897 (2022).
10. B. R. Scanlon, I. Jolly, M. Sophocleous, L. Zhang, Water Resour. Res.
43, W03437 (2007).
11. B. R. Scanlon et al., Proc. Natl. Acad. Sci. U.S.A. 109,
9320–9325 (2012).
12. M. Rodell, I. Velicogna, J. S. Famiglietti, Nature 460, 999–1002
(2009).
13. D. J. MacAllister, G. Krishan, M. Basharat, D. Cuba,
A. M. MacDonald, Nat. Geosci. 15, 390–396 (2022).
14. C. J. Clarke, R. J. George, R. W. Bell, T. J. Hatton, Soil Res. 40,
93 (2002).
15. W. M. Edmunds, C. B. Gaye, J. Hydrol. 156, 47–59
(1994).
16. G. Favreau et al., Water Resour. Res. 45, W00A16
(2009).
17. M. D. Nosetto, E. G. Jobbágy, A. B. Brizuela, R. B. Jackson,
Agric. Ecosyst. Environ. 154, 2–11 (2012).
18. R. Giménez, J. Mercau, M. Nosetto, R. Páez, E. Jobbágy,
Hydrol. Processes 30, 2603–2616 (2016).
19. A. D. Nobre et al., J. Hydrol. 404, 13–29 (2011).
20. R. A. Maenza, E. A. Agosta, M. L. Bettolli, Int. J. Climatol. 37,
445–463 (2017).
21. M. A. Lovino, O. V. Müller, E. H. Berbery, G. V. Müller, Global
Planet. Change 168, 78–97 (2018).
22. T. J. Peterson, M. Saft, M. C. Peel, A. John, Science 372,
745–749 (2021).
23. M. Rogger et al., Water Resour. Res. 53, 5209–5219
(2017).
24. A. Alaoui, M. Rogger, S. Peth, G. Blöschl, J. Hydrol. 557,
631–642 (2018).
25. Y. Fan, G. Miguez-Macho, E. G. Jobbágy, R. B. Jackson,
C. Otero-Casal, Proc. Natl. Acad. Sci. U.S.A. 114, 10572–10577
(2017).
26. J. Šimůnek, M. Šejna, H. Saito, M. Sakai, M. T. van Genuchten,
“The HYDRUS-1D Software Package for Simulating the One-
Dimensional Movement of Water, Heat, and Multiple Solutes in
Variably-Saturated Media” (Department of Environmental
Sciences, University of California–Riverside Research Reports,
2005).
27. J. I. Whitworth-Hulse et al., Agric. Ecosyst. Environ. 349,
108440 (2023).
28. Z. P. Wang, C. W. Lindau, R. D. Delaune, W. H. Patrick Jr., Biol.
Fertil. Soils 16, 163–168 (1993).
29. J. Houspanossian et al., Theor. Appl. Climatol. 131, 111–120
(2016).
30. P. Tittonell et al., Front. Sustain. Food Syst. 4, 584605
(2020).
31. L. A. Garibaldi et al., Conserv. Lett. 14, e12773 (2021).
32. S. Wang, J. Li, Can. J. Rem. Sens. 42, 190–202 (2016).
33. E. G. Jobbágy et al., Hydrol. Processes 35, e14188
(2021).
34. M. D. Nosetto et al., Agric. Water Manage. 129, 120–129
(2013).
35. M. L. Amdan, R. Aragón, E. G. Jobbágy, J. N. Volante,
J. M. Paruelo, Water Resour. Res. 49, 6601–6612 (2013).
36. M. D. Nosetto, R. A. Paez, S. I. Ballesteros, E. G. Jobbágy,
Agric. Ecosyst. Environ. 206, 60–70 (2015).
37. S. Pal et al., Water Resour. Res. 57, e2021WR029704
(2021).
38. J. Houspanossian et al., Supplementary Materials for:
Agricultural expansion raises groundwater and increases
flooding in the South American plains, data set, Zenodo
(2022); https://doi.org/10.5281/zenodo.7962455.
AC KNOWLED GME NTS
Funding: We thank the UUKi Rutherford Fund Strategic Partner
Grants; the UK Department for Business, Energy & Industrial
Strategy (BEIS); CONICET (PIP 363/2020); and the ANPCyT
(PICT 504/2020) for funding this work. We also acknowledge a
grant from the Inter-American Institute for Global Change
Research (IAI) SGP-HW 056 (GovernAgua Project). Author
contributions: Conceptualization: J.H. and E.G.J. Methodology:
J.H., R.G., W.T., J.I.W.-H., M.D.N., P.M.A., and M.C.R. Investigation:
J.H., R.G., and E.G.J. Visualization: J.H., W.T., J.I.W.-H., and E.G.J.
Funding acquisition: M.C.R. and E.G.J. Project administration:
J.H., M.C.R., and E.G.J. Supervision: M.C.R. and E.G.J. Writing –
original draft: J.H., W.T., M.C.R., and E.G.J. Writing – review &
editing: J.H., P.M.A., M.C.R., and E.G.J. Competing interests: The
authors declare that they have no competing interests. Data
and materials availability: All data and code are deposited at
Zenodo (38). License information: Copyright © 2023 the authors,
some rights reserved; exclusive licensee American Association
for the Advancement of Science. No claim to original US
government works. https://www.science.org/about/science-
licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add5462
Materials and Methods
Figs. S1 to S19
Tables S1 and S2
References (39–66)
Submitted 25 July 2022; accepted 24 May 2023
10.1126/science.add5462
Houspanossian et al., Science 380, 1344–1348 (2023)
30 June 2023
5 of 5
|
10.1126_science.add5649
|
manner because widely conserved chitinases
are encoded by both commensal microbes and
mammals, particularly those that consume
chitin (8–13).
Dietary chitin induces gastric distension
and type 2 immune triggering
Chitin activates lung group 2 innate lymphoid
cells (ILC2s) by means of interleukin-25 (IL-25),
IL-33, and thymic stromal lymphopoietin (TSLP)
(14). To test GI responses to dietary chitin, “YRS”
mice that express reporter alleles for ILC2 sig-
nature genes arginase-1 (Yarg; Arg1YFP, where
YFP is yellow fluorescent protein), IL-5 (Red5;
Il5tdTomato), and IL-13 (Smart13; Il13hCD4) on
wild-type (WT) and IL-25, IL-33 receptor (IL-
33R), and thymic stromal lymphopoietin re-
ceptor (TSLPR) triple-knockout (TKO) (15)
backgrounds were fed with standard chow
containing either cellulose (control) or chitin as
fiber (Fig. 1A and table S1). We tested 5 to 20%
chitin, which approximates the dietary com-
position of insectivorous mammals (11, 12).
Food intake was similar, and GI transit time
was unaffected by diet. However, we observed
marked gastric distention and greater stom-
ach contents in chitin-fed versus control-fed
mice (Fig. 1B and fig. S1, A and B), indicating
that dietary fiber influences stomach reten-
tion and stretch. Gastric epithelium rapidly
1Department of Pathology and Immunology, Washington
University School of Medicine, St. Louis, MO, USA. 2Department
of Bioengineering and Therapeutic Sciences, University of
California San Francisco, San Francisco, CA, USA.
*Corresponding author. Email: svandyken@wustl.edu
RES EARCH
IMMUNOMETABOLISM
A type 2 immune circuit in the stomach controls
mammalian adaptation to dietary chitin
Do-Hyun Kim1, Yilin Wang1, Haerin Jung1, Rachael L. Field1, Xinya Zhang1, Ta-Chiang Liu1,
Changqing Ma1, James S. Fraser2, Jonathan R. Brestoff1, Steven J. Van Dyken1*
Dietary fiber improves metabolic health, but host-encoded mechanisms for digesting fibrous
polysaccharides are unclear. In this work, we describe a mammalian adaptation to dietary chitin that is
coordinated by gastric innate immune activation and acidic mammalian chitinase (AMCase). Chitin
consumption causes gastric distension and cytokine production by stomach tuft cells and group 2 innate
lymphoid cells (ILC2s) in mice, which drives the expansion of AMCase-expressing zymogenic chief cells
that facilitate chitin digestion. Although chitin influences gut microbial composition, ILC2-mediated
tissue adaptation and gastrointestinal responses are preserved in germ-free mice. In the absence of
AMCase, sustained chitin intake leads to heightened basal type 2 immunity, reduced adiposity, and
resistance to obesity. These data define an endogenous metabolic circuit that enables nutrient
extraction from an insoluble dietary constituent by enhancing digestive function.
traction, which is evident in bariatric surgical
approaches that counteract overnutrition by
reducing or bypassing gastric digestion (6).
Nutrient availability is also limited by dietary
fiber enrichment because most bulky insolu-
ble polysaccharides are resistant to digestion
by mammalian enzymes and undergo only
limited degradation by distal gut microbes
(7). A notable exception is chitin (b-1,4-poly-N-
acetylglucosamine). One of the most abundant
natural polysaccharides on Earth, chitin is a
component of arthropods and fungi and is an
initiator of type 2 immune responses. We hy-
pothesized that chitin is digested in a distinctive
D ietary fiber intake is associated with a
lower risk of metabolic disorders such as
obesity (1, 2) and type 2 immune activa-
tion has been implicated in metabolic
homeostasis (3–5), but little is known
about how degradation of specific fibers in-
fluences host immunity and metabolism. In
mammals, digestion is initiated in the upper
gastrointestinal (GI) tract and is facilitated by
mechanical forces, neural feedback, and enzy-
matic activities that coordinate chemical and
physical disruption of the food bolus before
passage into the highly absorptive small intes-
tine. Digestion is essential for nutrient ex-
Fig. 1. Innate type 2 immune responses are
triggered by gastric distension and dietary chitin.
(A) Dietary responses in WT or TKO mice on triple-
reporter (YRS) backgrounds. Representative stomach
image (scale bar, 1 cm), stomach size, and luminal
content (B); relative Il25 and Il33 expression in
stomach tuft (CD45−EpCAM+SiglecF+) and epithelial
cells (CD45−EpCAM+) (C); and expression of R5 (Il5)
and S13 (Il13) reporter alleles among stomach ILC2s
(Yarg+, pregated on CD45+Lin−Thy1.2+) (D) in WT
and TKO mice fed the indicated diet for 24 hours. RFP,
red fluorescent protein. (E) Relative stomach gene
expression in WT or TKO mice after Yoda1 adminis-
tration or gastric distension. (F) R5 and S13 expression in
stomach ILC2 after administration of the mechano-
sensitive ion-channel inhibitor GsMTx4 to mice fed
the indicated diet for 12 hours. R5 (G) or S13
(H) expression in stomach and SI ILC2s after vehicle,
Yoda1, and/or NmU administration. Data represent
individual biological replicates except for the data in
(C), which are pooled from two or three mice and
are presented as means ± SD from two or more
independent experiments (n ≥ 3 mice per group).
P values were calculated by unpaired t test [(B), (F),
and (G)], one-way analysis of variance (ANOVA) [(E)
and (H)], or two-way ANOVA with Tukey’s multiple
comparisons test [(C) and (D)]. *P < 0.05; **P < 0.01;
***P < 0.001; ****P < 0.0001; ns, not significant.
Kim et al., Science 381, 1092–1098 (2023)
8 September 2023
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Sustained chitin
intake promotes GI
remodeling and adipose
ILC2 responses. Repre-
sentative stomach histology
images [hematoxylin and
eosin (H&E) stained]
(scale bars, 100 mm)
(A), Ki67-expressing
stomach epithelial cells
(B), stomach tuft cells
(C), and representative
images of SI and small
intestinal length (scale
bar, 1 cm) (D) of the
indicated mice fed the
specified diet for 2 weeks.
(E) Eosinophils, total
ILC2s, and R5-expressing ILC2s per gram of epididymal white adipose tissue (eWAT) in WT and TKO mice fed the specified diet for 2 weeks. Data represent individual biological
replicates and are presented as means ± SD from two or more independent experiments (n ≥ 3 mice per group). P values were calculated by unpaired t test [(A), (C),
and (D)], one-way ANOVA (B), or two-way ANOVA with Tukey’s multiple comparisons test (E). *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001; ns, not significant.
responded to chitin, inducing expression of
the ILC2-activating cytokines Il25 and Il33 in
stomach tuft cells and nontuft epithelial cells,
respectively (Fig. 1C).
Dietary chitin increased stomach IL-5– and
IL-13–producing ILC2s (Fig. 1D) and alterna-
tively activated Arg1+ macrophages, serum IL-5,
and blood eosinophils, whose numbers ex-
panded over time and in proportion with chitin
content (fig. S1, C to E). Few IL-5– or IL-13–
expressing stomach CD4+ T cells were detected,
however, and chitin responses were intact in
Rag1 knockout (Rag1-KO) mice, which lack B
and T cells (fig. S1, F to I), consistent with pre-
dominantly innate immune activation. Chitin
ingestion increased stomach expression of the
gene encoding ILC2-activating neuropeptide
neuromedin U (Nmu) (16, 17) along with the
gene encoding its receptor, Nmur1, among
stomach-resident ILC2s (fig. S1, J and K). The
expression of calcitonin gene–related peptide
(CGRP), another ILC2-activating neuropeptide
(18), was unaltered by dietary chitin (fig. S1L).
Thus, gastric ILC2s may be synergistically ac-
tivated by IL-25 and NMU, similar to intesti-
nal ILC2s (16, 17).
Dietary chitin also increased gastrin and
glucagon-like peptide-1 (GLP-1), GI hormones
that are induced by mechanical stretch after
food ingestion (19). By contrast, serotonin, which
responds to IL-33 (20), was unaffected by chitin
(fig. S1, M and N). In addition, although IL-33
can be released by mechanical perturbation
(21) and drives ILC2 responses to Helicobacter
pylori infection and chemical injury (22, 23),
chitin-induced ILC2 accumulation and eosin-
ophilia were unaffected in Il1rl1-KO mice (fig.
S2, A and B). By contrast, ILC2 activation and
cytokine production was abolished in TKO
mice (Fig. 1, C and D), and eosinophilia was
abrogated in both tuft cell–deficient Pou2f3-
KO and TKO mice (fig. S2, C and D), whereas
chitin-induced stomach distension was main-
tained. Thus, tuft cell–derived IL-25 appears to
be the primary signal in gastric ILC2 responses
to dietary chitin–induced stretch.
We inflated the stomach with air to recapit-
ulate chitin-induced distension (fig. S2E). Within
2 hours, we observed increased expression of
Il25, Nmu, and Edn1, which is induced by trig-
gering the mechanosensitive ion channel Piezo1
(24, 25). Conversely, in vivo administration of
GsMTx-4, a stretch-activated channel inhibitor,
blocked this response (Fig. 1E and fig. S2F) and
abrogated ILC2 cytokine induction after chitin
ingestion (Fig. 1F). Administration of the Piezo1
agonist Yoda1 induced ILC2 IL-5 production in
WT but not TKO mice (Fig. 1G), suggesting that
Piezo1 signaling activates gastric ILC2s through
tissue cytokines, including IL-25. Accordingly,
Yoda1 did not enhance gastric ILC2 responses
before tuft cell development (fig. S2, G and H).
Nmur1 expression was reduced in TKO ILC2s
compared with WT ILC2s (fig. S2I), which is
consistent with prior reports (16) and sug-
gests that gastric ILC2s acquire basal IL-5
expression and receptivity to multiple stretch
signals during development. Combined Yoda1
and NMU administration also increased ILC2
IL-13 and KLRG1 expression compared with
Yoda1 alone (Fig. 1H and fig. S2J). Thus, di-
etary chitin and mechanical stretch initiates
type 2 immune responses that sensitize ILC2s
to additional synergistic neuroimmune acti-
vating signals.
Sustained chitin intake promotes GI
remodeling and adipose ILC2 responses
Dietary chitin remodeled GI tissues, inducing
gastric epithelial proliferation, epithelial and
submucosal thickening, and increased tuft cell
abundance (Fig. 2, A to C). Proliferation was
reduced in TKO mice compared with WT mice
(Fig. 2B), indicating a requirement for IL-25,
IL-33, and TSLP signaling in remodeling. Chitin
lengthened the small intestine (SI), which was
enriched with tuft cells, activated ILC2s, and
eosinophils in WT but not TKO mice (Fig. 2D
and fig. S3, A and B). These SI effects closely
resembled those induced by helminths, protozoa
(Tritrichomonas spp.), and increased luminal
succinate (26–29). However, control and chitin-
fed mice were Tritrichomonas-free, and cecal
succinate levels were unaffected by chitin (fig.
S3C). Thus, dietary chitin appears to initiate a
distinctive type 2 immune circuit within the
GI tract.
Chitin intake also stimulated type 2 immune
responses in metabolically active tissues. Al-
though lung ILC2s and eosinophils were un-
affected by diet (fig. S3, D and E), visceral adipose
from chitin-fed mice contained elevated num-
bers of eosinophils and IL-5–producing ILC2s
(Fig. 2E). Inhibiting tissue lymphocyte egress
by using the immunomodulator FTY720 re-
duced circulating ILCs (fig. S3F) but did not
alter dietary chitin–induced eosinophil and
ILC2 responses (fig. S3, G and H), suggesting
that interorgan ILC2 migration did not medi-
ate adipose effects (30). By contrast, adipose
ILC2 activation and eosinophilia were abro-
gated in TKO mice (Fig. 2E), indicating that
tissue-derived cytokines coordinate local ILC2
responses to chitin. Moreover, mice that lacked
the shared signaling receptor for IL-4 and IL-13,
IL4Ra, failed to induce stomach ILC2s, tuft cells,
SI lengthening, adipose ILC2s, and eosinophils
in response to chitin (fig. S3, I to M). ILC deple-
tion in Rag1-KO mice impaired chitin-induced
gastric ILC2 and tuft cell expansion (fig. S3,
N and O), whereas tuft cells expanded normal-
ly in both Il4-KO and Il5-KO mice (fig. S3P),
which supports a role for IL-13–producing ILC2s
in chitin responses. Finally, GI tissue resilience,
which relies on ILC2s and IL-13 in the absence
of adaptive immunity (31), was enhanced in
Kim et al., Science 381, 1092–1098 (2023)
8 September 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. AMCase is required
for dietary chitin digestion.
(A) Immunostaining of
AMCase-expressing chief cells
in glandular stomach. Magenta,
AMCase; blue, 4′,6-diamidino-2-
phenylindole (DAPI); green,
autofluorescence (scale bars,
50 mm). Stomach ChiaRed+
(CR: AMCase reporter;
CD45−EpCAM+CR+) or total
(CD45−EpCAM+CR−) epithelial
cells were (B) isolated by
fluorescence-activated cell
sorting and (C) analyzed for
relative expression of Gif, Clps,
and Pgc by quantitative
polymerase chain reaction
(qPCR. (D) Analysis of chitin
binding, digestion of soluble
chitooligomer and insoluble
crystalline chitin substrates,
and production of GlcNAc
reaction products by chito-
oligomer oxidase (ChitO) assay
in stomach lavage. (E) Immuno-
blot of chitin-bound AMCase
proteins. The “(+)” indicates
recombinant AMCase control.
(F) Chitinase activity with
soluble substrate in stomach
lavage samples, with and
without predepletion of
AMCase by using insoluble chitin. (G) Gastric fluid pH in mice fasted overnight after 2 weeks on the indicated diet. (H) Digestion of insoluble colloidal chitin after
96-hour incubation with inactivated (heat-treated) or fresh stomach lavage from WT or CC mice (scale bars, 500 mm). (I) ChitO assay for soluble GlcNAc reaction
products in supernatants from insoluble particle digestion in (H). Data points represent individual biological replicates except for the data in (C), which represent
samples pooled from three or four mice. Data represent two or more independent experiments (n ≥ 3 mice per group) and are presented as means ± SD. P values
were calculated by unpaired t test. *P < 0.05; ***P < 0.001; ****P < 0.0001.
chitin-fed Rag1-KO mice infected with the hel-
minth Nippostrongylus brasiliensis compared
with control mice, as marked by increased ILC2s,
serum IL-5, and eosinophils and improved worm
expulsion (fig. S4, A to D).
Because dietary polysaccharides can be de-
graded by intestinal bacteria (7, 8, 32), we
profiled the fecal microbiota from chitin- and
control diet–fed mice. Mice maintained sim-
ilar body weights regardless of diet (fig. S5A),
which indicates equivalent nutrient extraction.
However, chitin intake significantly enriched
Bacteroidetes phyla, whereas Firmicutes were
proportionally decreased in chitin-fed mice
compared with control mice (fig. S5B and table
S2). Thus, chitin alters GI bacterial composition,
a finding that is consistent with prior work on
dietary fiber enrichment (8).
We then tested whether type 2 immune re-
sponses to dietary chitin were dependent on
commensal microbiota using germ-free (GF)
and specific-pathogen–free (SPF) mice. GF and
SPF mice maintained similar body weights re-
gardless of diet, and dietary chitin induced gas-
tric distension, SI lengthening, eosinophilia, ILC2
expansion, and tuft cell hyperplasia in both the
GF and SPF conditions (fig. S5, C to G). These
results indicated that dietary chitin induces in-
nate type 2 immune responses independent of
commensal microbes. To address possible devel-
opmental alterations in GF mice, we also de-
pleted bacteria by administering antibiotics to
adult SPF mice before dietary chitin intake.
Consistent with GF results, stomach, SI, and adi-
pose tissue chitin responses were unaffected by
antibiotics (fig. S5, H to K). Thus, although chitin
alters GI microbial composition and commensal
microorganisms influence tuft cell succinate
responses (27, 28), dietary chitin initiates type
2 immune responses in the absence of com-
mensal microbiota.
Acidic mammalian chitinase is required for
dietary chitin digestion in mammals
Chitinases (EC 3.2.1.14) are widely conserved
enzymes that cleave soluble chitooligomers
and crystalline chitin substrates, liberating N-
acetylglucosamine (GlcNAc). In contrast to in-
soluble dietary chitin fibers, GlcNAc did not
induce gastric distension or GI type 2 immune
responses (fig. S6, A to D), suggesting that un-
digested, insoluble chitin causes mechanical
stretch and type 2 immune triggering. Chiti-
nases can be expressed by microbes, including
gut-resident bacteria (8). However, the lack of
microbial involvement in chitin responses led
us to consider acidic mammalian chitinase
(AMCase), a mammalian chitinase secreted by
respiratory epithelium, salivary glands, and
stomach that has been evolutionarily linked
to chitin consumption (10–12, 33).
AMCase was expressed at the base of gastric
glands in tissues from human sleeve-gastrectomy
patients and 8-week-old mice (Fig. 3A), which
is consistent with RNA sequencing (RNA-seq)
data (fig. S7A) and prior reports (10, 33–35),
which supported chief cells as the main source of
AMCase in the mammalian stomach. We fur-
ther investigated AMCase-expressing cells using
ChiaRed reporter mice, in which tdTomato and
Cre recombinase are knocked into Chia1 (which
encodes AMCase). Homozygous ChiaRed (CC)
mice lack AMCase (36). We lineage-traced
AMCase-expressing cells by crossing ChiaRed
with Rosa26-flox-stop-zsGreen [R26(LSL)-
Kim et al., Science 381, 1092–1098 (2023)
8 September 2023
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Stomach adaptation to dietary chitin is controlled by a type 2 immune
circuit. Relative Chia1 stomach expression in WT (A) and indicated mouse strains
(B) after 2 weeks on the indicated diet. (C) Chia1 expression in stomach tissue
from WT and ILC2-deficient deleter mice after IL-25 administration. (D) Breeding
scheme to obtain CR-Stat6-KO, and CR-Il4rafl/fl mice. (E) Relative Il4ra expression in
ChiaRed+ or total stomach epithelial cells (EpCAM+) from ChiaRed and CR-Il4rafl/fl
mice. (F) Percentage of ChiaRed+ (CR) chief cells out of total stomach epithelial cells
in ChiaRed, CR-Il4rafl/fl, and CR-Stat6-KO mice fed the indicated diet for 2 weeks.
(G) Stomach size of WT and CC mice fed the indicated diet for 2 weeks. S13 (IL-13)–
expressing ILC2s and tuft cells in stomach (H); SI tuft cells, SI and adipose
eosinophils, ILC2s, and S13+ ILC2s (I); eWAT-to–body weight ratio (J), and body
weights (K) of WT and CC mice fed control or chitin diets as indicated. Data
represent individual biological replicates except for the data in (A), (E), (H), and (I),
which are pooled from three to eight mice and are presented as means ± SD from
two or more independent experiments (n ≥ 3 mice per group). P values were
calculated by unpaired t test [(A) to (C) and (E) to (G)] or two-way ANOVA with
Tukey’s multiple comparisons test [(H) to (K)]. *P < 0.05; **P < 0.01; ***P < 0.001;
****P < 0.0001; ns, not significant.
zsGreen] mice. Consistent with antibody stain-
ing, AMCase-expressing ChiaRed+zsGreen+ cells
were localized in gastric glands and concor-
dantly coexpressed ChiaRed and zsGreen with
age (fig. S7B), indicating that AMCase expres-
sion is sustained in terminally differentiated
chief cells. ChiaRed+ cells were also enriched
for mature chief cell markers gastric intrinsic
factor (Gif), colipase (Clps), and pepsinogen C
(Pgc) (Fig. 3, B and C), which supports AMCase
as a marker of zymogenic digestive cells.
We tested the contribution of AMCase to
chitin digestion using GI luminal secretions
from WT and CC (AMCase-deficient) mice.
Chitinase activity was assessed on soluble chito-
oligomers and crystalline chitin substrates (37),
and chitin-binding proteins were isolated with
magnetized chitin (Fig. 3D). WT stomach la-
vage contained AMCase that bound insoluble
chitin (Fig. 3E) and exhibited robust chitinase
activity that was absent in CC lavage and could
be depleted by preincubation with chitin (Fig.
3F), which indicates that AMCase binds chitin
and nonredundantly mediates stomach chiti-
nase activity. Dietary chitin intake also reduced
stomach pH (Fig. 3G), which enhances AMCase
enzymatic activity (10, 33) and suggests a co-
ordinated physiological digestion response that
involves acid-secreting parietal cells. Crystalline
dietary chitin fibers were visibly digested and
converted to soluble GlcNAc by WT stomach
lavage, which reflects highly efficient chitinase
digestive activity. However, this activity was ab-
sent in CC lavage, which retained undigested
crystalline chitin particles and failed to pro-
duce soluble GlcNAc reaction products (Fig. 3,
H and I). Neither control nor chitin chow elic-
ited detectable epithelial ChiaRed or Chia1 ex-
pression in lower GI tissues, including SI and
cecum (fig. S8, A to C). However, WT SI lavage
still contained chitinase activity that was ab-
sent in CC SI lavage, suggesting that AMCase
produced in the stomach transits into the lower
GI tract, where it also makes up the major
source of chitinase activity (fig. S8D). Thus,
although most insoluble dietary polysaccha-
rides consumed by mammals resist digestion
or undergo only limited degradation in the
lower GI tract by commensal microbiota–
derived glycosyl hydrolases (7, 8), dietary chitin
is primarily digested by host-encoded AMCase.
Kim et al., Science 381, 1092–1098 (2023)
8 September 2023
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. Dietary chitin improves
metabolic health in high-fat diet–
induced obesity. (A) Experimental
design. HFD- or CHFD-fed WT or CC
mice were subjected to metabolic cage
analyses (CLAMS), glucose tolerance
tests (GTTs), and insulin tolerance
tests (ITTs). [Part of the illustration was
created with Biorender.com] Body
weights (B), adiposity (C), and food
intake (24-hour average) (D) of
WT and CC mice after 8 weeks on
the indicated diet. Eosinophils and
ILC2s in adipose tissue (E) and tuft
cells and ILC2s in SI and stomach
tissues (F) in WT and CC mice after
13 weeks on the indicated diet.
iWAT, inguinal white adipose tissue.
(G) GTT curves at 10 weeks. (H) ITT
curves at 12 weeks. (I) Heat curve
and averages over 24 hours in
CLAMS cages at 8 weeks. Data repre-
sent individual biological replicates
except for data in (B), (E), and
(F); in these panels, each data
point represents eight pooled mice
and are presented as means ± SD
[(E) and (F)] or means ± SEM [(B) to
(D) and (G) to (I)], from two or more independent experiments (n = 8 mice per group). P values were calculated by one-way ANOVA (C), two-way ANOVA [(E) and (F)],
or two-way ANOVA with repeated measures [(B), (D), and (G) to (I)] with Tukey’s multiple comparisons test. *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001.
Stomach adaptation to dietary chitin
is controlled by a type 2 immune circuit
Because lung AMCase expression is promoted
by type 2 cytokines (36, 38), we tested whether
stomach AMCase is similarly regulated. Indeed,
sustained dietary chitin intake increased Chia1
expression in WT but not Il4ra-KO, Pou2f3-
KO, ILC2-deleter mice that lack ILC2s (15), or
TKO mice (Fig. 4, A and B). Tuft cell–derived
IL-25, ILC2s, and IL-4Ra signaling were there-
fore implicated as major drivers of gastric AMCase
expression. Type 2 triggering was recapitu-
lated by IL-25 administration, which stimu-
lated IL-13 production by ILC2s and stomach
Chia1 expression in WT but not ILC2-deficient
mice (Fig. 4C and fig. S8E). Thus, ILC2-derived
IL-13 is implicated in the expansion of AMCase-
expressing chief cells.
To test this further, we crossed ChiaRed with
Stat6-KO (CR-Stat6-KO) and Il4rafl/fl mice (39)
(CR-Il4rafl/fl), which enabled specific deletion
of IL4Ra from AMCase-expressing cells (Fig.
4D). Il4ra was reduced in ChiaRed+ stomach
epithelial cells from CR-Il4rafl/fl mice compared
with ChiaRed controls, indicating successful
Cre-mediated excision (Fig. 4E). Dietary chitin
increased AMCase-expressing chief cells in
WT ChiaRed mice but failed to occur in both
CR-Stat6-KO and CR-Il4rafl/fl mice (Fig. 4F),
indicating that dietary chitin promotes chief
cell AMCase expression through cell-intrinsic
IL4Ra signaling. We also examined acute gas-
tric injury and transient chief cell depletion by
high-dose tamoxifen treatment (HDT), which
models aspects of human atrophic corpus gas-
tritis and AMCase loss due to H. pylori infec-
tion (40, 41). After HDT, WT mice activated
gastric ILC2s coincident with chief cell recovery,
whereas TKO mice failed to recover Chia1 (fig. S9,
A and B). Thus, restoration of gastric homeosta-
sis and chitin digestive capacity after epithe-
lial injury depends on type 2 circuit activation.
These data suggest that mammals adapt to
dietary chitin by inducing endogenous stomach
type 2 immune responses to boost AMCase pro-
duction. Accordingly, CC mice failed to reduce
gastric distension compared with WT mice after
2 weeks of chitin intake (Fig. 4G). Stomach ILC2
activation and tuft cell hyperplasia were also
sustained in CC versus WT mice over several
weeks (Fig. 4H), which reflects unresolved cir-
cuit activation without AMCase-catalyzed chitin
digestion. Consistent with improved digestion,
WT mice attenuated distal type 2 immune
triggering in the SI and adipose tissues. CC
mice, by contrast, exhibited enhanced and pro-
longed type 2 triggering, characterized by greater
SI tuft cell abundance, increased eosinophils,
and increased IL-13–producing ILC2s in SI and
adipose tissues after 4 weeks of chitin intake
(Fig. 4I). Adipose tissue weight was also reduced
in proportion to body weight in CC mice com-
pared with WT mice, whereas body weights were
similar (Fig. 4, J and K). Thus, both GI and adi-
pose tissue homeostasis are regulated by the
AMCase-mediated adaptation to dietary chitin.
Dietary chitin improves metabolic health
in obesity
We next tested the impact of dietary chitin on
obesity, which is influenced by neuronal-ILC2
interactions and type 2 cytokines (3–5, 42). We
fed WT and CC mice isocaloric high-fat diets
containing either cellulose (control; HFD) or
chitin (CHFD) fiber (Fig. 5A). Food intake was
similar among all groups, and HFD-fed mice
showed comparable body weight gain. How-
ever, CHFD-fed CC mice gained significantly
less weight, with reduced adiposity and fat
mass compared with WT mice (Fig. 5, B to D,
and fig. S10A). Resistance to obesity in CHFD-
fed CC mice was accompanied by adipose ILC2
and eosinophil accumulation (Fig. 5E), which
have been previously linked with metabolic
homeostasis (3–5), suggesting that altered di-
etary chitin digestion could modulate metab-
olism. Indeed, numbers of ILC2s and tuft cells
were elevated in the stomach and SI tissues
of CHFD-fed CC compared with those of WT
mice (Fig. 5F), which reflects sustained type 2
triggering and suggests that AMCase-mediated
dietary chitin digestion contributes to meta-
bolic homeostasis.
Both dietary chitin and AMCase influenced
metabolism in the context of high-fat diets
because CHFD-fed WT and CC mice exhib-
ited significantly improved insulin sensitivity
compared with HFD-fed WT mice (Fig. 5, G
and H). CHFD-fed CC mice exhibited lower
fasting glucose compared with WT mice (fig.
Kim et al., Science 381, 1092–1098 (2023)
8 September 2023
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
S10B), which is consistent with differences in
body weight and suggests that AMCase activ-
ity influences glucose homeostasis after chitin
ingestion. Additionally, light-phase energy ex-
penditure was increased in CC mice, and the
respiratory exchange ratio was increased after
CHFD feeding compared with HFD feeding in
both WT and CC mice, despite no differences
in core body temperature or activity (Fig. 5I
and fig. S10, C to E), which is consistent with
sustained type 2 immune triggering. Indeed,
tuft cell–deficient Pou2f3-KO mice failed to ex-
hibit CHFD-induced effects on insulin sensitivity
(fig. S11, A to G), suggesting that type 2 circuit
initiation is required for some metabolic aspects
of chitin in a high-fat diet. Thus, disruption of
the mammalian stomach’s adaptation to dietary
chitin alters nutrient uptake and metabolic
homeostasis, which manifests in obesity.
Discussion
Chitin consumption has been linked to CHIA
gene selection throughout primate evolution
(11, 12), and chitin-rich fungi and arthropods
are constituents of the diets of both modern
and ancient human populations (13). Mam-
malian glycosyl hydrolase gene selection has
also been linked with starch consumption
(43, 44), but adaptations to specific dietary
fibers after ingestion are mainly ascribed to
shifts in gut microbial composition. As shown
here, mammals encode an endogenous circuit
that enables chitin catabolism and nutrient
extraction through AMCase production. This
pathway is triggered by gastric distension, neuro-
peptide release, and type 2 cytokine produc-
tion caused by insoluble chitin fibers, which
result in GI remodeling and chief cell AMCase
induction over time. This in turn enables en-
hanced chitin digestion. In humans, CHIA var-
iants resulting in lower chitinase activity are
linked with asthma (45, 46), which supports
dual roles for AMCase in mucosal defense and
nutrient extraction, as proposed initially (33).
AMCase-producing chief cells produce addi-
tional digestive enzymes such as pepsinogen
and lipase, which suggests that gastric ILC2s
may coordinate a response that improves over-
all digestion of recalcitrant insect or fungal
foods, perhaps representing a strategy for omni-
vores to adapt to varied diets. Although micro-
bial composition is altered in response to chitin
and other polysaccharides, we show that the
mammalian adaptation does not rely on com-
mensal microbiota and that AMCase is the
primary source of chitinase activity in the GI
tract, thereby distinguishing chitin digestion
from other abundant insoluble polysaccharides
such as cellulose. Our results further elucidate
a mechanism for how chitin initiates type 2
immune initiation by means of mechanical
stretch, thus connecting physical tissue per-
turbation with a branch of immunity increas-
ingly recognized to maintain homeostasis in
response to a wide variety of environmental dis-
ruptions as well as neural and dietary fluctua-
tions. Intriguingly, the mammalian adaptation
to chitin influences innate resistance to helminth
infection and metabolic homeostasis, suggest-
ing that chitin digestive pathways coevolved
with intestinal helminths and may represent a
therapeutic target in metabolic diseases such
as obesity.
RE FERENCES AND NOTES
J. L. Slavin, Nutrition 21, 411–418 (2005).
1.
2. S. K. Gill, M. Rossi, B. Bajka, K. Whelan, Nat. Rev. Gastroenterol.
Hepatol. 18, 101–116 (2021).
3. D. Wu et al., Science 332, 243–247 (2011).
4. A. B. Molofsky et al., J. Exp. Med. 210, 535–549 (2013).
J. R. Brestoff et al., Nature 519, 242–246 (2015).
5.
6. N. T. Nguyen, J. E. Varela, Nat. Rev. Gastroenterol. Hepatol. 14,
160–169 (2017).
7. H. J. Flint, K. P. Scott, S. H. Duncan, P. Louis, E. Forano,
Gut Microbes 3, 289–306 (2012).
J. L. Sonnenburg et al., Science 307, 1955–1959 (2005).
8.
9. M. G. Paoletti, L. Norberto, R. Damini, S. Musumeci, Ann. Nutr.
Metab. 51, 244–251 (2007).
10. M. Ohno et al., Sci. Rep. 6, 37756 (2016).
11. C. A. Emerling, F. Delsuc, M. W. Nachman, Sci. Adv. 4,
eaar6478 (2018).
12. M. C. Janiak, M. E. Chaney, A. J. Tosi, Mol. Biol. Evol. 35,
607–622 (2018).
13. M. C. Wibowo et al., Nature 594, 234–239 (2021).
14. S. J. Van Dyken et al., Immunity 40, 414–424 (2014).
15. S. J. Van Dyken et al., Nat. Immunol. 17, 1381–1387 (2016).
16. V. Cardoso et al., Nature 549, 277–281 (2017).
17. A. Wallrapp et al., Nature 549, 351–356 (2017).
18. P. Sui et al., Science 360, eaan8546 (2018).
19. T. Zhang, M. H. Perkins, H. Chang, W. Han, I. E. de Araujo,
Cell 185, 2478–2494.e28 (2022).
20. Z. Chen et al., Immunity 54, 151–163.e6 (2021).
21. R. Kakkar, H. Hei, S. Dobner, R. T. Lee, J. Biol. Chem. 287,
6941–6948 (2012).
22. N. Satoh-Takayama et al., Immunity 52, 635–649.e4 (2020).
23. A. R. Meyer et al., Gastroenterology 159, 2077–2091.e8 (2020).
24. A. G. Solis et al., Nature 573, 69–74 (2019).
25. Y. Oh et al., Neuron 109, 1979–1995.e6 (2021).
26. J. von Moltke, M. Ji, H.-E. Liang, R. M. Locksley, Nature 529,
221–225 (2016).
30. Y. Huang et al., Science 359, 114–119 (2018).
31. A. E. Price et al., Proc. Natl. Acad. Sci. U.S.A. 107, 11489–11494
(2010).
32. P. J. Turnbaugh et al., Nature 444, 1027–1031 (2006).
33. R. G. Boot et al., J. Biol. Chem. 276, 6770–6778 (2001).
34. M. Suzuki et al., J. Histochem. Cytochem. 50, 1081–1089 (2002).
35. X. Han et al., Cell 172, 1091–1107.e17 (2018).
36. S. J. Van Dyken et al., Cell 169, 497–509.e13 (2017).
37. B. A. Barad et al., Protein Sci. 29, 952–977 (2020).
38. T. A. Reese et al., Nature 447, 92–96 (2007).
39. D. R. Herbert et al., Immunity 20, 623–635 (2004).
40. W. J. Huh et al., Gastroenterology 142, 21–24.e7 (2012).
I. Nookaew et al., BMC Med. Genomics 6, 41 (2013).
41.
42. F. Cardoso et al., Nature 597, 410–414 (2021).
43. G. H. Perry et al., Nat. Genet. 39, 1256–1260 (2007).
44. E. Axelsson et al., Nature 495, 360–364 (2013).
45. M. A. Seibold et al., J. Biol. Chem. 284, 19650–19658
(2009).
46. K. Okawa et al., Mol. Biol. Evol. 33, 3183–3193 (2016).
47. D. H. Kim et al., A type 2 immune circuit in the stomach
controls mammalian adaptation to dietary chitin, version 1.
Dryad (2023); https://doi.org/10.5061/dryad.12jm63z3m.
AC KNOWLED GME NTS
We thank K. Ravichandran for comments on the manuscript;
J. Mills, R. Locksley, H.-E. Liang, J. Kipnis, M. Baldridge, C. Wilen,
and C. Hsieh for advice and reagents; L. Mosby for expert technical
assistance; and J. M. White at the Department of Pathology and
Immunology Gnotobiotic Facility at Washington University in
St. Louis for assistance with GF experiments. Funding: This work
was supported in part by the Bursky Center for Human
Immunology and Immunotherapy Programs, Center for Cellular
Imaging, and Rheumatic Diseases Research Resource-based
Center (NIH P30 AR073752) at Washington University in St. Louis;
NIH DP5 OD028125 (J.R.B.); Burroughs Wellcome Fund CAMS
#1019648 (J.R.B.); NIH R01HL148033, R01AI176660, and
R21AI163640 (S.J.V.D.); and the National Research Foundation of
Korea Award NRF-2020R1A6A3A03037855 (D.-H.K.). Author
contributions: D.-H.K. and S.J.V.D. designed the study and
performed experiments. Y.W. and H.J. performed experiments and
provided advice. R.L.F., X.Z., and J.R.B. performed metabolic
cage experiments, analyzed data, and provided expertise with
metabolic studies. C.M. and T.-C.L. provided human stomach
specimens and expertise, and J.S.F. provided biochemical reagents
and expertise. All authors edited the final manuscript. S.J.V.D.
directed the studies and wrote the manuscript with D.-H.K.
Competing interests: J.R.B. is on the scientific advisory board of
LUCA Science, Inc. The other authors declare that they have no
competing interests. Data and materials availability: All data are
available in the main text or supplementary materials. 16S
sequencing data are available in Dryad (47). License information:
Copyright © 2023 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science. No
claim to original US government works. https://www.science.org/
about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add5649
Materials and Methods
Figs. S1 to S11
Tables S1 to S3
References (48–52)
MDAR Reproducibility Checklist
27. C. Schneider et al., Cell 174, 271–284.e14 (2018).
28. W. Lei et al., Proc. Natl. Acad. Sci. U.S.A. 115, 5552–5557 (2018).
29. M. S. Nadjsombati et al., Immunity 49, 33–41.e7 (2018).
Submitted 20 June 2022; resubmitted 8 May 2023
Accepted 8 August 2023
10.1126/science.add5649
Kim et al., Science 381, 1092–1098 (2023)
8 September 2023
6 of 6
|
10.1126_science.add5787
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
MICROBIOME
Microbial-host-isozyme analyses reveal microbial
DPP4 as a potential antidiabetic target
Kai Wang†, Zhiwei Zhang†, Jing Hang†, Jia Liu†, Fusheng Guo†, Yong Ding, Meng Li, Qixing Nie,
Jun Lin, Yingying Zhuo, Lulu Sun, Xi Luo, Qihang Zhong, Chuan Ye, Chuyu Yun, Yi Zhang, Jue Wang,
Rui Bao, Yanli Pang, Guang Wang*, Frank J. Gonzalez*, Xiaoguang Lei*, Jie Qiao*, Changtao Jiang*
INTRODUCTION: Gut microbiota can regulate
the physiology and pathophysiology of the
host by producing enzymes with functions simi-
lar to those of the host. However, it is difficult to
identify these microbial-host isozymes through
sequencing-based studies because enzymes with
similar functions in different species may lack
sequence conservation. An activity-based func-
tional protein screening framework is more re-
liable for the discovery and characterization of
such microbial-host isozymes, which will help
yield deeper insights into the gut microbiota–
host cross-talk.
RATIONALE: To identify potential microbial-host
isozymes, we set up an enzyme activity screening
platform, including activity assays for 110
enzymes that are functional in various human
diseases. These enzyme activities were measured
in stool-derived ex vivo bacteria communities.
Dipeptidyl peptidase 4 (DPP4) was a prominent
microbial-host isozyme identified in our screen,
but little is known about its pathophysiological
effects on the host. We sought to determine
whether gut microbial–derived DPP4, like host
DPP4 (hDPP4), could decrease active GLP-1 and
thus affect blood glucose homeostasis.
RESULTS: We identified 71 enzymes with posi-
tive activity in the human gut bacteria com-
munities through our enzyme activity screening
platform, most of which were validated in the
protein extracts obtained from feces of germ-
free and specific pathogen–free mice. Among
Microbial host isozyme screening
Microbial host isozymes
Host-derived enzymes
Bacteroides spp.
Microbial DPP4
Active
GLP-1
5.0
mmol/L
Reduced
active
GLP-1
11.1
mmol/L
Sitagliptin
High-throughput screen
Host DPP4
x
Microbial DPP4
Diabetics
High-responder
Low-responder
Dau-d4
Interaction
Efficacy
Microbial DPP4 mediates
sitagliptin efficacy
Microbial DPP4 impairs
glycometabolism
Discovery and inhibition of a gut microbial–host isozyme to regulate host metabolism. Differences in
the gut microbiota may explain why some individuals respond to antidiabetic DPP4 inhibitors but others do
not. An activity-based enzyme activity screening system identified gut microbial DPP4 isozymes that can
decrease active GLP-1 but cannot be inhibited by sitagliptin. High-throughput screening identified Dau-d4 as a
selective inhibitor of microbial DPP4 to increase GLP-1 activity and improve glucose tolerance.
Dau-d4 inhibits microbial
DPP4 activity
these identified enzymes, DPP4 activity had the
highest statistical effect size (Z factor) among the
10 human samples. Through human gut bac-
teria isolation and DPP4 activity screening, we
discovered that microbial DPP4 was mainly
produced by Bacteroides spp. Gut microbial
DPP4 (mDPP4) could degrade active GLP-1(7-37)
in vitro. However, mDPP4 could not affect active
GLP-1 levels in chow-fed mice but could de-
crease active GLP-1 activity and impair glucose
homeostasis in high-fat diet (HFD)–fed mice or
dextran sulfate sodium/indomethacin–treated
mice, suggesting that a damaged gut barrier is
required for mDPP4 to affect the activity of
host GLP-1.
We discovered that the clinical DPP4 inhibitor
sitagliptin failed to efficiently inhibit mDPP4.
And by solving the co-crystal of mDPP4 with
sitagliptin at 1.97-anstrom resolution, we found
differences in the nature of the binding be-
tween the drug and mDPP4 compared with
its binding to hDPP4 that may explain this
difference in inhibitory effects. A sitagliptin
clinical trial (www.clinicaltrials.gov identifer
NCT04495881) among patients with type 2
diabetes (T2D) (n = 57) and a related fecal
microbiota transplant of stool from high res-
ponders and low responders in the present study
to HFD-fed mice demonstrated that mDPP4
could limit the efficacy of sitagliptin in indi-
viduals with T2D and in glucose-intolerant mice.
To identify a selective inhibitor of mDPP4,
we screened ~107,000 compounds, and using
structural modification we identified Dau-d4,
a derivative of daurisoline, that could selectively
inhibit mDPP4 activity compared with hDPP4.
Dau-d4 could increase active GLP-1 levels and
improve glucose metabolism in diabetic mice,
and co-administration of Dau-d4 with sitaglip-
tin further improved blood glucose homeostasis.
CONCLUSION: Here, we developed an activity-
based strategy to identify uncharacterized gut
microbial-host isozymes that provides a deeper
understanding of gut microbiota–host interac-
tions. Gut microbial DPP4 isozyme can impair
host glucose homeostasis, and variations in
microbial DPP4 activities could possibly con-
tribute to the heterogeneous responses to
sitagliptin observed among patients with T2D.
Our findings highlight the promise of de-
veloping therapies that target both host and
gut microbial enzymes to achieve greater
clinical efficacy.▪
The list of author affiliations is available in the full article online.
*Corresponding author. Email: jiangchangtao@bjmu.edu.cn
(C.J.); jie.qiao@263.net (J.Q.); xglei@pku.edu.cn (X.L.);
gonzalef@mail.nih.gov (F.J.G.); wangguangcy@ccmu.edu.cn
(G.W.)
†These authors contributed equally to this work.
Cite this article as K. Wang et al., Science 381, eadd5787
(2023). DOI: 10.1126/science.add5787
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.add5787
Wang et al., Science 381, 501 (2023)
4 August 2023
1 of 1
RES EARCH
R E S E A R C H A R T I C L E
◥
MICROBIOME
Microbial-host-isozyme analyses reveal microbial
DPP4 as a potential antidiabetic target
Kai Wang1,2†, Zhiwei Zhang1,2†, Jing Hang1,3,4†, Jia Liu5†, Fusheng Guo6,7†, Yong Ding1,2, Meng Li1,2,
Qixing Nie1,2, Jun Lin1,2, Yingying Zhuo1,2, Lulu Sun8, Xi Luo1,2, Qihang Zhong1,3,4, Chuan Ye1,2,
Chuyu Yun1,2, Yi Zhang1,2, Jue Wang6, Rui Bao9, Yanli Pang1,3,4, Guang Wang5*, Frank J. Gonzalez8*,
Xiaoguang Lei6,7*, Jie Qiao1,3,4,10*, Changtao Jiang1,2,11*
A mechanistic understanding of how microbial proteins affect the host could yield deeper insights
into gut microbiota–host cross-talk. We developed an enzyme activity–screening platform to
investigate how gut microbiota–derived enzymes might influence host physiology. We discovered
that dipeptidyl peptidase 4 (DPP4) is expressed by specific bacterial taxa of the microbiota.
Microbial DPP4 was able to decrease the active glucagon like peptide-1 (GLP-1) and disrupt glucose
metabolism in mice with a leaky gut. Furthermore, the current drugs targeting human DPP4,
including sitagliptin, had little effect on microbial DPP4. Using high-throughput screening, we
identified daurisoline-d4 (Dau-d4) as a selective microbial DPP4 inhibitor that improves glucose
tolerance in diabetic mice.
T he gut microbiota contributes to host
normo- and pathophysiology by producing
diverse metabolites (1–3) and a variety of
proteins and peptides (4–9). Although the
various functions of most gut microbiota–
derived proteins have not been identified,
several studies have shown that the gut micro-
biota can produce enzymes with functions
similar to those of the host (microbial-host
isozymes) (10–13). However, proteins with
similar functions in the host and in different
bacterial species, so-called isozymes, may lack
sequence conservation, and thus sequencing-
1Department of Physiology and Pathophysiology, School of
Basic Medical Sciences, State Key Laboratory of Female
Fertility Promotion, Center for Reproductive Medicine, Peking
University, Beijing, China. 2Center for Obesity and Metabolic
Disease Research, School of Basic Medical Sciences, State
Key Laboratory of Vascular Homeostasis and Remodeling,
Peking University, Beijing, China. 3Department of Obstetrics
and Gynecology, Peking University Third Hospital, Beijing,
China. 4National Clinical Research Center for Obstetrics and
Gynecology (Peking University Third Hospital), Beijing, China.
5Department of Endocrinology, Beijing Chao-Yang Hospital,
Capital Medical University, Beijing, China. 6Beijing National
Laboratory for Molecular Sciences, Key Laboratory of
Bioorganic Chemistry and Molecular Engineering of Ministry
of Education, Department of Chemical Biology, College of
Chemistry and Molecular Engineering, Synthetic and
Functional Biomolecules Center, Peking University, Beijing,
China. 7Peking-Tsinghua Center for Life Sciences, Peking
University, Beijing, China. 8Laboratory of Metabolism, Center
for Cancer Research, National Cancer Institute, National
Institutes of Health, Bethesda, MD, USA. 9Center of
Infectious Diseases, Division of Infectious Diseases in State
Key Laboratory of Biotherapy, West China Hospital, Sichuan
University, Chengdu, China. 10Beijing Advanced Innovation
Center for Genomics, Beijing, China. 11Center of Basic
Medical Research, Institute of Medical Innovation and
Research, Third Hospital, Peking University, Beijing, China.
*Corresponding author. Email: jiangchangtao@bjmu.edu.cn
(C.J.); jie.qiao@263.net (J.Q.); xglei@pku.edu.cn (X.L.);
gonzalef@mail.nih.gov (F.J.G.); wangguangcy@ccmu.edu.cn (G.W.)
†These authors contributed equally to this work.
based studies to identify such isozymes may
not be a reliable means to identify them or their
biological functions.
To better identify microbial-host isozymes,
we developed an activity-based screening plat-
form as a framework for their discovery and
characterization that may also represent targets
for disease intervention because many of the
enzymes in our screen are known to also be
involved in host pathophysiology. In our screen,
we identified isozymes from the human and
mouse gut microbiota, including dipeptidyl
peptidase 4 (DPP4), an important therapeutic
target for the management of type 2 diabetes
(T2D). DPP4 is responsible for the degradation
of GLP-1(7-36 amide) or GLP-1(7-37), both of
which are called “active” GLP-1 and are secreted
from intestinal enteroendocrine L cells that play
an important role in the coordination of post-
prandial glucose homeostasis, mainly through
the activation of GLP-1 receptor (GLP-1R) (14, 15).
The degradation products of active GLP-1,
GLP-1(9-36 amide) or GLP-1(9-37), were unable
to activate GLP-1R, leading to the generally con-
sidered inactive form of GLP-1 (16). The sum-
mation of the pools of active GLP-1 and inactive
GLP-1 equals the pool of total GLP-1.
Homologous proteins of human DPP4 with
different amino acid sequences have been found
in many other organisms, including bacteria
and fungi. For example, DPP4 has been iden-
tified by others in periodontopathic bacteria,
and when it was injected into mice, their glucose
tolerance was altered, indicating that the micro-
bial DPP4 was functionally similar to host DPP4
(17, 18). However, current clinical DPP4 inhibi-
tors, which target human DPP4, exhibit low
inhibitory activity toward periodontopathic
bacterial DPP4 (19, 20). Although it has been
speculated that the gut microbiota has DPP4-
like activity (21), it is not known which specific
gut commensal bacteria produce DPP4 iso-
zymes or what effects such gut microbiota–
derived DPP4 may have on host health. In our
work, we discovered that in mice with a damaged
gut barrier, gut microbial DPP4 decreased the
pool of active GLP-1 independently of host
DPP4, thus impairing host glucose homeosta-
sis. Clinically relevant DPP4 inhibitors, includ-
ing sitagliptin, that are used to treat T2D have
a weak inhibitory effect on gut microbial DPP4,
which may explain the variable response to
sitagliptin in certain T2D patients. Using high-
throughput screening, we identified Dau-d4, a
derivate of daurisoline (Dau), as a specific micro-
bial DPP4 inhibitor that ameliorated metabolic
dysfunction in mice mainly through its action
on the microbial DPP4–host GLP-1 axis.
Results
Functional screening for microbial-host isozymes
We established a screening platform to iden-
tify microbial-host isozymes that have the poten-
tial to affect host function (Fig. 1A). A major
challenge with a mixed fecal culture approach
is maintaining the diversity and composition
of the microbiota during ex vivo culturing
(13, 22–24). We collected fresh human fecal
samples from three healthy volunteers and
cultured them for 2 days in 10 different types
of media commonly used in gut microbiota
cultivation, followed by 16S ribosomal RNA
(rRNA) gene sequencing. We found that brain
heart infusion (BHI) medium supported the
growth of the bacterial community while main-
taining a composition and diversity similar to
those observed in the fecal samples (fig. S1, A
and B) (22). To reduce overgrowth of facul-
tative anaerobes such as Enterococcus and
Escherichia-Shigella (13, 22), we selected fluo-
romethalone as a supplement for BHI (mBHI
medium) (fig. S1, C and D) (25). By collecting
and culturing 10 fresh stool samples in mBHI
medium, we found that the stool-derived ex vivo
communities were maximally similar to the
original fecal microbiota (fig. S1E).
We selected 110 enzymes that are functional
in various human diseases, including US Food
and Drug Administration–approved drug tar-
gets and other disease target enzymes in the
human key drug-target database SuperTarget
(26) (fig. S1F). Protein extracts from the above
10 stool-derived ex vivo communities were pre-
pared for the activity of microbial-host isozymes
(see the materials and methods for details). A
combination of fecal enzymes may be a better
representation of the in situ enzyme activity
of the microbiota than a separation-based
single enzymatic activity assay. Among 10 tested
human stool–derived ex vivo communities, we
identified 51 enzymes with a statistical effect
size (Z factor) of >0.5 in five or more individual
samples and 20 with a Z factor >0.5 in one to
Wang et al., Science 381, eadd5787 (2023)
4 August 2023
1 of 13
RES EARCH | R E S E A R C H A R T I C L E
A
Healthy individuals
Ex vivo communities
Transferases
Hydrolases
+
Lyases
+
Ligases
Oxidoreductases
Isomerases
+
+
Germ-free mice
Activity-based enzyme screening
Cultivation-based characterization
Microbial-host-isozyme
Conventional mice
B
Dipeptidyl peptidase 4 (EC:3.4.14.5)
Amine oxidase (EC:1.4.3.4)
Biliverdin reductase(EC:1.3.1.24)
Tyrosine aminotransferase (EC:2.6.1.5)
Carboxylesterase (EC:3.1.1.1)
Membrane alanyl aminopeptidase (EC:3.4.11.2)
Carbonyl reductase (EC:1.1.1.184)
Arylesterase (EC:3.1.1.2)
Phosphoglycerate dehydrogenase (EC:1.1.1.95)
Choline O-acetyltransferase (EC:2.3.1.6)
Nicotinamide mononucleotide adenylyltransferase (EC:2.7.7.1)
UDP-glucose 6-dehydrogenase (EC:1.1.1.22)
Cystathionine beta-synthase (EC:4.2.1.22)
Carboxypeptidase U (EC:3.4.17.20)
Purine-nucleoside phosphorylase (EC:2.4.2.1)
Aromatase (EC:1.14.14.14)
Arachidonate 5-lipoxygenase (EC:1.13.11.34)
Amino-acid N-acetyltransferase (EC:2.3.1.1)
Adenosine deaminase (EC:3.5.4.4)
Gamma-glutamyl hydrolase (EC:3.4.19.9)
Kynureninase (EC:3.7.1.3)
Soluble epoxide hydrolase (EC:3.3.2.10)
Memapsin 2 (EC:3.4.23.46)
Cathepsin B (EC:3.4.22.1)
Triacylglycerol lipase (EC:3.1.1.3)
Uridine phosphorylase (EC:2.4.2.3)
Acetylcholinesterase (EC:3.1.1.7)
Carbonic anhydrase (EC:4.2.1.1)
Ornithine aminotransferase (EC:2.6.1.13)
α-Amylase (EC:3.2.1.1)
Thymidine phosphorylase (EC:2.4.2.4)
Alpha-galactosidase (EC:3.2.1.22)
7-dehydrocholesterol reductase (EC:1.3.1.21)
GMP reductase (EC:1.7.1.7)
L-asparaginase (EC:3.5.1.1)
Aldehyde dehydrogenase (EC:1.2.1.3)
Nicotinamide N-methyltransferase (EC:2.1.1.1)
Tryptophan 5-monooxygenas (EC:1.14.16.4)
Methionine aminopeptidase (EC:3.4.11.18)
15-hydroxyprostaglandin dehydrogenase (EC:1.1.1.141)
Phosphoserine transaminase (EC:2.6.1.52)
Trypsin (EC:3.4.21.4)
Ceramide glucosyltransferase (EC:2.4.1.80)
Glucose-6-phosphate dehydrogenase (EC:1.1.1.49)
17beta-estradiol 17-dehydrogenase (EC:1.1.1.62)
Dihydropyrimidinase (EC:3.5.2.2)
Granzyme B (EC:3.4.21.79)
Adenosylhomocysteinase (EC:3.13.2.1)
Gamma-glutamyltransferase (EC:2.3.2.2)
Thioredoxin-dependent peroxiredoxin (EC 1.11.1.24)
GTP cyclohydrolase 1 (EC:3.5.4.16)
Porphobilinogen synthase (EC:4.2.1.24)
Alcohol dehydrogenase (EC:1.1.1.1)
Glutathione-disulfide reductase (EC:1.8.1.7)
Hyaluronoglucosaminidase (EC:3.2.1.35)
L-iditol 2-dehydrogenase (EC:1.1.1.14)
Glutamine synthetase (EC:6.3.1.2)
Caspase-1 (EC:3.4.22.36)
Glycine N-acyltransferase (EC:2.3.1.13)
Glucose-6-phosphate isomerase (EC:5.3.1.9)
L-xylulose reductase (EC:1.1.1.10)
Aldose reductase (EC:1.1.1.21)
5'-nucleotidase (EC:3.1.3.5)
Tyrosine 3-monooxygenase (EC:1.14.16.2)
L-lactate dehydrogenase (EC:1.1.1.27)
Proline dehydrogenase (EC:1.5.5.2)
Protein farnesyltransferase (EC: 2.5.1.58)
Xanthine dehydrogenase (EC:1.17.1.4)
Catalase (EC:1.11.1.6)
IMP dehydrogenase (EC:1.1.1.205)
Adenylyl cyclase (EC:4.6.1.1)
Cerebroside-sulfatase (EC:3.1.6.8)
Creatine kinase (EC:2.7.3.2)
Prolyl oligopeptidase (EC:3.4.21.26)
Dimethylallyltranstransferase (EC:2.5.1.1)
Arginase (EC:3.5.3.1)
Beta-galactosidase (EC:3.2.1.23)
Guanine deaminase (EC:3.5.4.3)
Inositol-phosphate phosphatase (EC:3.1.3.25)
Nitric oxide synthase (EC:1.14.13.39)
Dihydrofolate reductase (EC:1.5.1.3)
Hydroxymethylglutaryl-CoA reductase (EC:1.1.1.34)
Alkaline phosphatase (EC:3.1.3.1)
Sulfite oxidase (EC:1.8.3.1)
Long-chain-fatty-acid—CoA ligase (EC:6.2.1.3)
Phospholipase A2 (EC:3.1.1.4)
Peptidylprolyl isomerase (EC:5.2.1.8)
Beta-glucosidase (EC:3.2.1.21)
Insulysin (EC:3.4.24.56)
CTP synthase (EC:6.3.4.2)
L-glutaminase (EC:3.5.1.2)
Alpha-glucosidase (EC:3.2.1.20)
3-oxo-5-alpha-steroid 4-dehydrogenase (EC:1.3.1.22)
Procollagen-proline 3-dioxygenase (EC:1.14.11.7)
Glutathione transferase (EC:2.5.1.18)
Protein disulfide-isomerase (EC:5.3.4.1)
Acetyl-CoA carboxylase (EC:6.4.1.2)
Peptidyl-dipeptidase (EC:3.4.15.1)
Histidine decarboxylase (EC:4.1.1.22)
Carbamoyl-phosphate synthase (EC:6.3.4.16)
Thymidylate synthase (EC:2.1.1.45)
Malate dehydrogenase (EC:1.1.1.37)
Fructose-bisphosphatase (EC:3.1.3.11)
Chitinase (EC:3.2.1.14)
Lactoylglutathione lyase (EC:4.4.1.5)
3 beta-hydroxy-Delta(5)-steroid dehydrogenase (EC:1.1.1.145)
Dihydropyrimidine dehydrogenase (EC:1.3.1.2)
Tyrosinase (EC:1.14.18.1)
Methylenetetrahydrofolate reductase (EC:1.5.1.20)
Glutamate dehydrogenase (EC:1.4.1.3)
-250
0
0.0
0.5
Z score
1.0
J
HFD 12-week
SPF mice
Dpp4+/+
Control
(Drinking water)
Dpp4-/-
Dpp4-/-
Control
(Drinking water)
Analyze
Antibiotics
Day 0
Day 1
C
D
y
t
i
v
i
t
c
a
4
P
P
D
r
a
l
u
s
s
i
t
l
a
n
i
t
s
e
t
n
i
L
E
]
)
n
i
e
t
o
r
p
g
m
*
n
m
i
(
/
l
o
m
n
[
G
y
t
i
v
i
t
c
a
4
P
P
D
r
a
l
u
s
s
i
t
l
a
n
i
t
s
e
t
n
i
L
E
]
)
n
i
e
t
o
r
p
g
m
*
n
m
i
(
/
l
o
m
n
[
30
25
20
15
10
5
0
Antibiotics+FMT
Antibiotics
Control
40
30
20
10
0
##
***
Antibiotics+FMT
Antibiotics
Control
***
###
+/+
Dpp4
-/-
-/-+Antibiotics
Dpp4
Dpp4
K
y
t
i
v
i
t
c
a
4
P
P
D
r
a
l
u
s
s
i
t
l
a
n
i
t
s
e
t
n
i
L
E
]
)
n
i
e
t
o
r
p
g
m
*
n
m
i
(
/
l
o
m
n
[
30
20
10
0
CD
Control (Drinking water)
Antibiotics
Antibiotics
FMT
Analyze
SPF mice
Control (Drinking water)
HFD
Antibiotics
FMT
Antibiotics
Analyze
Week -12
Week 0
Day 0
Day 1
Day 2
(
D
t
o
F
)
f
o
r
C
D
(
G
t
o
M
)
f
o
r
H
F
D
)
n
i
e
t
o
r
p
/
g
m
g
p
(
2000
1500
1000
500
0
E
1
-
P
L
G
e
v
i
t
c
a
r
a
l
u
s
s
i
t
l
a
n
i
t
s
e
t
n
i
L
E
H
F
12
1
-
P
L
G
e
v
i
t
c
a
a
m
s
a
l
P
)
L
m
g
p
(
/
I
)
L
m
g
p
(
/
1
-
P
L
G
e
v
i
t
c
a
a
m
s
a
l
P
8
4
0
15
10
5
0
Antibiotics+FMT
Antibiotics
Control
***
###
Antibiotics+FMT
Antibiotics
Control
***
###
2500
2000
1500
1000
500
0
Antibiotics+FMT
Antibiotics
Control
Antibiotics+FMT
Antibiotics
Control
M
##
***
###
***
3000
2000
1000
0
1
-
P
L
G
e
v
i
t
c
a
a
m
s
a
l
P
)
L
m
g
p
(
/
25
20
15
10
5
0
+/+
Dpp4
-/-
-/-+Antibiotics
Dpp4
Dpp4
+/+
Dpp4
-/-
-/-+Antibiotics
Dpp4
Dpp4
1
-
P
L
G
e
v
i
t
c
a
r
a
l
u
s
s
i
t
l
a
n
i
t
s
e
t
n
i
L
E
)
n
i
e
t
o
r
p
/
g
m
g
p
(
L
1
-
P
L
G
e
v
i
t
c
a
r
a
l
u
s
s
i
t
l
a
n
i
t
s
e
t
n
i
L
E
)
n
i
e
t
o
r
p
/
g
m
g
p
(
Fig. 1. Microbial-host isozyme activity screening system reveals that
microbial DPP4 can decrease active GLP-1. (A) Schematic representation of
the microbial-host isozyme-screening system. (B) Z factor of the activity of a
given enzyme in each sample, n = 10. A Z factor >0.5 indicates that the activity is
positive. (C) Experimental scheme. SPF mice were fed a chow diet [(D) to (F),
n = 6 mice/group] or a HFD [(G) to (I), n = 6 mice/group] for 12 weeks, divided
into three groups, and administered antibiotics in the drinking water or a
drinking water control for 1 d. The FMT groups were transplanted with SPF fecal
microbiota after antibiotic treatment. (D) Extraluminal (EL) intestinal tissular
DPP4 activity [His-Ala-p-nitroaniline (His-Ala-pNA) as a substrate]. (E) Concen-
trations of EL intestinal tissular active GLP-1. (F) Plasma levels of active
GLP-1. (G) EL intestinal tissular DPP4 activity (His-Ala-pNA as a substrate).
(H) Concentrations of EL intestinal tissular active GLP-1. (I) Plasma levels of
active GLP-1. (J) Experimental scheme for (K) to (M), n = 6 mice/group. Dpp4+/+
and Dpp4−/− mice were fed a HFD for 12 weeks; Dpp4−/− mice were administered
antibiotics in the drinking water or a drinking water control, respectively, for
1 d; and Dpp4+/+ mice were administered with drinking water control. (K) Fecal
DPP4 activity (His-Ala-pNA as a substrate). (L) Concentrations of EL intestinal
tissular active GLP-1. (M) Plasma levels of active GLP-1. Data are presented as
means ± SEM. For (D) to (I), one-way ANOVA with Tukey’s post hoc test
[(D) to (F), (H), and (I)] and Dunnett’s T3 test [(G)]: ***P < 0.001 compared
with the control group, ##P < 0.01, ###P < 0.001 compared with the
antibiotics group. For (K) to (M), one-way ANOVA with Dunnett’s T3 test
(K) and Tukey’s post hoc test [(L) and (M)]: ***P < 0.001 compared
with the Dpp4+/+ group, ##P < 0.01, ###P < 0.001 compared with the
Dpp4−/− group.
Wang et al., Science 381, eadd5787 (2023)
4 August 2023
2 of 13
RES EARCH | R E S E A R C H A R T I C L E
four individual samples (Fig. 1B). It should be
noted that discrepancies in the activity of the
same enzymes between individuals may re-
sult from differences in the composition and
content of their gut microbiota.
We also measured the activities of above 71
potential microbial-host isozymes from fecal
samples obtained from germ-free (GF) and con-
ventionally raised (CONV-R) mice. We found that
the enzymatic activity of 56 (78.9%) microbial-
host isozymes was higher in feces from CONV-R
mice than in that from GF mice (fig. S1G).
Microbial DPP4 decreases active GLP-1 under
leaky gut
In our screen, we found that DPP4 had the
highest average Z factor among the 10 human
samples (Fig. 1B). We also found that the acti-
vity of this enzyme in the feces of GF mice was
less than that of the enzyme extracted from
CONV-R mice (5.5-fold; fig. S1G). Several DPP4
inhibitors have been developed and used in
the clinical management of T2D (14, 27), but
their clinical efficacy varies among patients
for unknown reasons (28). To assess the physio-
logical function of microbial DPP4 in vivo,
chow-fed specific pathogen–free (SPF) mice
were orally treated with a broad-spectrum anti-
biotic cocktail for 1 day (Fig. 1C). The efficiency
of gut bacteria removal was monitored by mea-
suring the relative abundance of the 16S rRNA
gene (fig. S2A). Although antibiotic treatment
significantly decreased fecal DPP4 activity (fig.
S2B), no changes in DPP4 activity or host DPP4
concentrations were observed in the extralu-
minal intestinal tissue (Fig. 1D and fig. S2C).
This is important because it has been reported
that intestinal DPP4 activity is required to
degrade GLP-1 and thus regulate glucose toler-
ance (29). Moreover, active and total GLP-1 levels
in the extraluminal intestinal tissue and plasma
of chow-fed mice did not change after antibiotic
treatment (Fig. 1, E and F, and fig. S2, D and E).
In the intact gut of chow-fed mice, GLP-1 is
secreted from L cells of the intestine and is
released basolaterally directly into the blood-
stream (15). Thus, gut microbial DPP4 may
decrease GLP-1 when it crosses the intestinal
barrier. It is known that high-fat diet (HFD)
feeding can induce lipotoxicity and altera-
tions in the shape and distribution of the gut
microbiota (i.e., gut dysbiosis), which can
damage the intestinal barrier. Such a leaky
gut induced by HFD feeding allows microbial
macromolecules such as lipopolysaccharide
and bacterial proteins into the bloodstream
(30–33). Therefore, we speculate that a HFD-
induced leaky gut allows microbial DPP4 to
decrease the activity of host GLP-1.
To investigate this theory, we tested the effect
of microbial DPP4 on HFD-fed mice (Fig. 1C).
After a 12-week HFD feeding regimen, which
increased intestinal barrier permeability (as
indicated by a fluorescein isothiocyanate–
dextran permeability assay and changes in
the gene expression of tight junction proteins;
fig. S2, F and G), the antibiotic treatment
group showed significantly less DPP4 activity
(including both host and microbial DPP4) in
their feces and extraluminal intestinal tissue
compared with control mice given antibiotic-
free drinking water (Fig. 1G and fig. S2, H and I),
with no change in intestinal host DPP4 levels
(fig. S2J). When lower DPP4 activity was de-
tected, active GLP-1 levels were higher in extra-
luminal intestinal tissue and plasma compared
with the control group (Fig. 1, H and I), whereas
total GLP-1 levels remained unchanged (fig. S2,
K and L). Fecal microbiota transplantation
(FMT) from SPF mice to antibiotic-treated mice
reversed the effect of antibiotics on DPP4
activity and the active GLP-1 levels (Fig. 1, G to I,
and fig. S2, H to L). DPP4 globally deficient
(Dpp4−/−), HFD-fed mice (Fig. 1J) retained
DPP4 activity in the extraluminal intestinal
tissue because of the presence of gut microbial
DPP4 along with a leaky gut despite the loss of
host DPP4 protein. Conversely, DPP4 activity
was absent when the Dpp4−/− mice were treated
with antibiotics, which correlated with greater
pools of active GLP-1 in both the extraluminal
intestinal tissue and the plasma compared with
the control-treated Dpp4−/− mice (Fig. 1, K to M,
and fig. S2, M to Q).
To confirm that microbial DPP4-mediated
decrease of active GLP-1 requires increased
intestinal permeability, we gave chow-fed mice
a low concentration (1%) of dextran sulfate
sodium (DSS) (34). DSS carries a strong neg-
ative charge that induces erosion of the gut
epithelium, so we used it to test whether this
source of damage to the gut, rather than HFD
feeding per se, allows microbial DPP4 to escape
the gut lumen (fig. S3, A and B). We found that
antibiotic treatment in DSS-treated mice showed
similar changes in DPP4 activity and active
GLP-1 levels compared with the HFD-fed mice
(fig. S3, C to K). Moreover, in an indomethacin
drinking model, another mouse model of in-
creased intestinal permeability caused by the
inhibitory effect of indomethacin on protec-
tive prostaglandins (35) (fig. S3, L and M), we
observed similar changes in DPP4 activity and
active GLP-1 levels as with the DSS and HFD
models (fig. S3, N to V). Therefore, in mice with
a damaged gut barrier, the gut microbiota
appears to contribute to DPP4 activity and to
the decrease of active GLP-1 independently of
host DPP4. Gut microbial DPP4 thus appears
to be able to reduce active GLP-1 levels in vivo
under conditions similar to the leaky gut that
often occurs in patients with T2D (36, 37).
Identification of DPP4-producing bacteria
To identify microbial producers of DPP4, we
collected fresh stool samples from 19 healthy
volunteers, cultured the microbial commun-
ities ex vivo, and tested DPP4 activity. We chose
two samples with the highest DPP4 activities
from which to isolate and identify bacterial
strains (fig. S4A). We established a library of
272 culturable isolates representing six phyla
(Fig. 2A and fig. S4B). By using the GLP-1 mimics
as assay substrates, we detected potent DPP4
activity among the Bacteroidetes phylum (Fig.
2A), including Bacteroides thetaiotaomicron,
B. fragilis, B. eggerthii, B. vulgatus, and B. dorei
(Fig. 2A). Among these, B. thetaiotaomicron
showed the greatest activity with His-Ala-p-
nitroaniline and Gly-Pro-p-nitroaniline (Fig. 2A).
Phylogenetic analysis was performed on
24 proteins, including 19 DPP4 homologs from
the above Bacteroides spp., previously reported
DPP4 homologs from periodontal bacteria
(17, 18), and human DPP4 (fig. S4C). We ex-
pressed and purified 14 candidate enzymes
and measured their DPP4 activity (Fig. 2B,
fig. S4D, and table S1). Although sequence
alignment suggested that 19 enzymes in five
Bacteroides spp. were homologs to human
DPP4 (hDPP4), not all the bacterial gene pro-
ducts were functionally similar to hDPP4.
B. thetaiotaomicron peptidase 1, B. fragilis
peptidase 1, B. eggerthii peptidase 1, B. vulgatus
peptidase 1, and B. dorei peptidase 1 (btDPP4,
bfDPP4, beDPP4, bvDPP4, and bdDPP4, respec-
tively) showed potent DPP4 activity, whereas
others had no significant DPP4 activity (Fig.
2B). All of the Bacteroides spp. DPP4 fell into
a clade distinct from that of oral bacterial
DPP4 (fig. S4C). Sequence analysis also showed
that the Bacteroides DPP4 protein homologs
contain an N-terminal signal peptide, a dipep-
tidyl IV domain, and a peptidase_S9 domain
(fig. S4E). By searching with our proven micro-
bial DPP4 proteins in both the oral cavity and
stool in the Human Microbiome Project (HMP-
1-1), we found that Bacteroides-encoded DPP4
is mainly distributed to the stool, with little
occurrence in the oral environment (fig. S4,
F to I). It has been reported that bacterial
species have extensive functional variation be-
tween strains (38). Using profiling DPP4
enzyme–encoding genes in the genome of 1245
strains from Bacteroides spp. obtained from the
National Center for Biotechnology Information
(NCBI), we found that such genes occur widely
among the sequenced Bacteroides spp. strains
(fig. S4J and table S2).
Using enzyme kinetic assays, we found that
the ratio of the apparent unimolecular rate
constant to the Michaelis constant (kcat/Km) of
btDPP4 was greater than that of other micro-
bial DPP4 isozymes and hDPP4 (Fig. 2, C and D).
Further, btDPP4 showed the highest activity
toward GLP-1 among all of the DPP4 pro-
teins tested, including hDPP4 (Fig. 2E).
Microbial DPP4 disrupts host glucose homeostasis
To explore the role of microbial DPP4 in host
glucose metabolism, Bacteroides DPP4 proteins
from five species were expressed in E. coli
Wang et al., Science 381, eadd5787 (2023)
4 August 2023
3 of 13
RES EARCH | R E S E A R C H A R T I C L E
A
Species
Adlercreutzia caecimuris
Bifidobacterium adolescentis
Bifidobacterium bifidum
Bifidobacterium faecale
Bifidobacterium longum
Strain
Phylum
Bifidobacterium pseudocatenulatum
Actinobacteria
Collinsella aerofaciens
Eggerthella lenta
Alistipes shahii
Alistipes finegoldii
Bacteroides acidifaciens
Bacteroides caccae
Bacteroides dorei
Bacteroides eggerthii
Bacteroides fragilis
Bacteroides koreensis
Bacteroides kribbi
Bacteroides stercoris
Bacteroides thetaiotaomicron
Bacteroides uniformis
Bacteroides vulgatus
Bacteroides ovatus
Bacteroides xylanisolvens
Parabacteroides distasonis
Parabacteroides faecis
Parabacteroides goldsteinii
Parabacteroides merdae
Porphyromonas capnocytophagoides
Porphyromonas endodontalis
Porphyromonas gingivalis
Amedibacillus dolichus
Anaerostipes caccae
Bacillus aerius
Butyricicoccus faecihominis
Clostridium butyricum
Clostridium disporicum
Clostridium fallax
[Clostridium] innocuum
Clostridium perfringens
Clostridium saudiense
[Clostridium] symbiosum
Clostridium tertium
Coprobacillus cateniformis
Enterococcus avium
Enterococcus durans
Enterococcus faecalis
Enterococcus faecium
Enterococcus gallinarum
Enterococcus hirae
Eubacterium budayi
[Eubacterium] eligens
Eubacterium limosum
Faecalibacterium prausnitzii
Faecalicoccus pleomorphus
Finegoldia magna
Flavonifractor plautii
Hungatella effluvii
Hungatella hathewayi
Lactobacillus johnsonii
Lactobacillus salivarius
Lactococcus garvieae
Lactococcus lactis
Lactococcus plantarum
Murimonas intestini
Paraclostridium benzoelyticum
Paeniclostridium sordellii
Ruminococcus bicirculans
Ruminococcus faecis
[Ruminococcus] gnavus
[Ruminococcus] torques
Staphylococcus epidermidis
Fusobacterium nucleatum
Citrobacter freundii
Citrobacter sedlakii
Cupriavidus necator
Enerobacter asburiae
Enterobacter ludwigii
Escherichia coli
Escherichia fergusonii
Hafnia alvei
Kerstersia gyiorum
Klebsiella oxytoca
Pelomonas aquatica
Shigella flexneri
Shigella sonnei
Yokenella regensburgei
Bacteroidetes
Firmicutes
Fusobacteria
Proteobacteria
Akkermansia muciniphila
Verrucomicrobia
Blank
0
5
10
15
20
0
5
10
15
20
Activity for Gly-Pro-pNA
[nmol/(min*OD600)]
Activity for His-Ala-pNA
[nmol/(min*OD600)]
B
y
t
i
v
i
t
c
a
4
P
P
D
]
)
n
i
e
t
o
r
p
g
µ
*
n
m
i
(
/
l
o
m
n
[
60
40
20
0
2
3
2
2
3
1(bfDPP4)
1(bdDPP4)
1(bvDPP4)
peDPP4
)4PPDtb(
tfDPP4
piDPP4
hDPP4
pgDPP4
peptidase
peptidase
peptidase
peptidase
peptidase
1
peptidase
peptidase
peptidase
BT
BT
BE peptidase 1 (beDPP4)
BD
BV
BE
BF
BF
BF
peptidase
BT
60
)
n
m
i
/
l
o
m
µ
(
v
40
20
0
0
C
D
hDPP4
btDPP4
bfDPP4
beDPP4
bvDPP4
bdDPP4
pgDPP4
peDPP4
tfDPP4
piDPP4
500
1000
1500
Substrate concentration (µM )
)
1
-
n
m
i
1
-
M
µ
(
m
K
t/
a
c
k
50
40
30
20
10
0
hDPP4
btDPP4
bfDPP4
beDPP4
bvDPP4
bdDPP4
pgDPP4
peDPP4
tfDPP4
piDPP4
E
)
7
3
-
1
(
1
-
P
L
G
t
s
n
i
a
g
a
y
t
i
v
i
t
c
A
]
)
n
i
e
t
o
r
p
g
µ
*
n
m
i
(
/
l
o
m
p
[
1000
800
600
400
200
0
control
hDPP4
btDPP4
bfDPP4
beDPP4
bvDPP4
bdDPP4
pgDPP4
peDPP4
tfDPP4
piDPP4
enzyme
No
Fig. 2. Identification of DPP4-producing bacteria and microbial DPP4.
(A) Testing of the in vitro DPP4 activity of bacteria strains isolated from human
stools [His-Ala-pNA and Gly-Pro-p-nitroaniline (Gly-Pro-pNA) as substrates],
n = 3 biological replicates per group. (B) Heterologous expression and
purification of candidate DPP4s from five Bacteroides spp. (B. thetaiotaomicron,
B. fragili, B. eggerthii, B. dorei, and B. vulgatus); previously reported DPP4
homologous from periodontal bacteria (Porphyromonas gingivalis, P. endodon-
talis, Tannerella forsythia, and Prevotella intermedia) and hDPP4. Protein DPP4
activity was tested in vitro (His-Ala-pNA as a substrate), n = 3 biological
replicates per group. (C) Michaelis–Menten saturation curves for hDPP4 and
nine active microbial DPP4 homologs (His-Ala-pNA as a substrate), n = 3
biological replicates per group. (D) kcat/Km of hDPP4 and nine active microbial
DPP4 homologs calculated from the Michaelis–Menten saturation curve in
(C), n = 3 biological replicates per group. (E) Activity of hDPP4 and nine active
microbial DPP4 homologs against GLP-1(7-37), n = 3 biological replicates
per group. Data are presented as means ± SEM.
Wang et al., Science 381, eadd5787 (2023)
4 August 2023
4 of 13
RES EARCH | R E S E A R C H A R T I C L E
Nissle 1917 (EcN) by genomic integration
through CRISPR-Cas9. First, compared with
the phosphate-buffered saline (PBS)–gavaged
group, EcN colonization had no significant
effect on DPP4 activity or on active GLP-1 levels
(fig. S5, A to D). Next, a single dose of EcN
expressing different microbial DPP4 constructs
was orally administered to HFD-fed mice (Fig.
3A). All groups were successfully colonized 7
days after administration (fig. S5E), and all of
the experimental groups showed significantly
higher fecal DPP4 activity than control mice
colonized with EcN treatment (fig. S5F). Higher
DPP4 activity in extraluminal intestinal tissue
was observed only in the EcN::btDpp4, EcN::
bfDpp4, and EcN::bvDpp4 colonized groups
(Fig. 3B), and there was similar intestinal host
DPP4 levels in all of the colonized groups com-
pared with the control EcN group (fig. S5G).
As DPP4 activity changed, the levels of active
GLP-1 in the extraluminal intestinal tissue and
plasma were significantly lower in the EcN::
btDpp4, EcN::bfDpp4, and EcN::bvDpp4 groups
(Fig. 3, C and D), with unchanged levels of total
A
E
H
L
1. EcN
2. EcN::btDpp4
3. EcN::bfDpp4
4. EcN::beDpp4
5. EcN::bvDpp4
6. EcN::bdDpp4
Analyze
Week 0
Week 1
HFD
12-week
SPF mice
***
*
*
50
40
30
20
10
0
EcN
EcN::bvDpp4
EcN::beDpp4
EcN::bdDpp4
EcN::bfDpp4
EcN::btDpp4
C
1
-
P
L
G
e
v
i
t
c
a
r
a
l
u
s
s
i
t
l
a
n
i
t
s
e
t
n
i
L
E
)
n
i
e
t
o
r
p
/
g
m
g
p
(
*
*
***
1500
1000
500
0
EcN
EcN::bdDpp4
EcN::bvDpp4
EcN::beDpp4
EcN::bfDpp4
EcN::btDpp4
B
y
t
i
v
i
t
c
a
4
P
P
D
r
a
l
u
s
s
i
t
l
a
n
i
t
s
e
t
n
i
L
E
]
)
n
i
e
t
o
r
p
g
m
*
n
m
i
(
/
l
o
m
n
[
F
n
i
l
u
s
n
i
a
m
s
a
P
l
)
L
m
g
n
(
/
6
4
2
0
*
***
***
EcN
EcN::btDpp4
EcN::bfDpp4
EcN::beDpp4
EcN::bvDpp4
EcN::bdDpp4
0 15
0 15
0 15
0 15
0 15
0 15
Time (min)
600
400
200
)
L
d
/
g
m
(
e
s
o
c
u
l
g
d
o
o
B
l
0
0
**
*
*
*
EcN
EcN::btDpp4
EcN::bfDpp4
EcN::beDpp4
EcN::bvDpp4
EcN::bdDpp4
20
40
60
80
100
Time (min)
1. EcN
2. EcN::btDpp4
HFD
12-week
Analyze
SPF Dpp4-/- mice
Week 0
Week 1
I
y
t
i
v
i
t
c
a
4
P
P
D
r
a
l
u
s
s
i
t
l
a
n
i
t
s
e
t
n
i
L
E
]
)
n
i
e
t
o
r
p
g
m
*
n
m
i
(
/
l
o
m
n
[
30
20
10
0
***
M
y
t
i
v
i
t
c
a
4
P
P
D
r
a
l
u
s
s
i
t
l
a
n
i
t
s
e
t
n
i
L
E
]
)
n
i
e
t
o
r
p
g
m
*
n
m
i
(
/
l
o
m
n
[
Short-term
Control
(Anaerobic PBS)
Analyze
Week 0 Week 1
HFD 12-week
SPF mice
Bt
Long-term
P
Week 0
Bt- Dpp4
Analyze
Week 6
-/-+EcN
-/-+EcN::btDpp4
Dpp4
Dpp4
**
##
60
40
20
0
Control Bt
Bt-
Dpp4
###
***
1
-
P
L
G
e
v
i
t
c
a
a
m
s
a
P
l
)
L
m
g
p
(
/
10
8
6
4
2
0
1
-
P
L
G
e
v
i
t
c
a
a
m
s
a
P
l
)
L
m
g
p
(
/
25
20
15
10
5
0
***
-/-+EcN
-/-+EcN::btDpp4
Dpp4
Dpp4
##
*
1
-
P
L
G
e
v
i
t
c
a
a
m
s
a
l
P
)
L
m
g
p
(
/
10
8
6
4
2
0
Control Bt
Bt-
Dpp4
Control
*
Bt
Bt Dpp4
J
N
Q
)
L
d
/
g
m
(
e
s
o
c
u
l
g
d
o
o
B
l
500
400
300
200
100
0
0
D
G
K
*
*
***
1
-
P
L
G
e
v
i
t
c
a
a
m
s
a
l
P
)
L
m
g
p
(
/
10
8
6
4
2
0
EcN
EcN::bdDpp4
EcN::bvDpp4
EcN::beDpp4
EcN::bfDpp4
EcN::btDpp4
***
**
**
)
n
m
i
*
L
d
g
m
/
(
T
T
G
O
f
o
C
U
A
40000
30000
20000
10000
0
EcN
EcN::bdDpp4
EcN::bvDpp4
EcN::beDpp4
EcN::bfDpp4
EcN::btDpp4
n
i
l
u
s
n
i
a
m
s
a
l
P
)
L
m
g
n
(
/
12
8
4
0
Dpp4-/-+EcN
Dpp4-/-+EcN::btDpp4
*
0
15
0
15
Time (min)
O
n
i
l
u
s
n
i
a
m
s
a
l
P
)
L
m
g
n
(
/
R
6
4
2
0
T
T
G
O
f
o
C
U
A
i
)
n
m
*
L
d
g
m
/
(
Control
Bt
Bt Dpp4
###
***
0
15
0
15
0
15
Time (min)
***
##
35000
30000
25000
20000
15000
10000
5000
0
(A to R) for HFD
Control Bt
Bt-
Dpp4
20
40
60
80
100
Time (min)
Control Bt
Bt-
Dpp4
Wang et al., Science 381, eadd5787 (2023)
4 August 2023
5 of 13
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Microbial DPP4 decreases the pools of active GLP-1 and disrupts
glucose metabolism in HFD-fed mice. (A) Experimental scheme for (B) to (G),
n = 6 mice/group. Mice were fed a HFD for 12 weeks, divided into six groups, and
given a single gavage of GFP-containing EcN, EcN::btDpp4, EcN::bfDpp4, EcN::
beDpp4, EcN::bvDpp4, or EcN::bdDpp4 for 1 week. (B) EL intestinal tissular DPP4
activity (His-Ala-pNA as a substrate). (C) Concentrations of EL intestinal tissular
active GLP-1 levels. (D) Plasma levels of active GLP-1. (E) Glucose-stimulated
insulin levels. (F and G) Oral glucose tolerance test (OGTT) curve (F) and its area
under the curve (AUC) (G). (H) Experimental scheme for (I) to (K), n = 6 mice/
group. Dpp4−/− mice were fed a HFD for 12 weeks and administered a single
gavage of EcN or EcN::btDpp4 for 1 week. (I) EL intestinal tissular DPP4 activity
(His-Ala-pNA as a substrate). (J) Plasma levels of active GLP-1. (K) Glucose-
stimulated insulin levels. (L) Experimental scheme. Mice were fed a HFD for
12 weeks, divided into three groups, and administered the same dose of
anaerobic PBS control, Bt, or Bt-DDpp4 by gavage for short-term or long-term
studies. For the short-term experiments in (M) to (O), n = 5 mice/group were
given a single gavage and lasted for 1 week; for the long-term experiments in
(P) to (R), n = 6 mice/group were gavaged once per week and lasted for 6
weeks. (M) EL intestinal tissular DPP4 activity (His-Ala-pNA as a substrate).
(N) Plasma levels of active GLP-1. (O) Glucose-stimulated insulin levels.
(P) Plasma levels of active GLP-1. (Q and R) OGTT curve (Q) and its AUC (R).
Data are presented as means ± SEM. For (B) to (G), one-way ANOVA with
Tukey’s post hoc test: *P < 0.05, **P < 0.01, ***P < 0.001 compared with the
EcN. For (I) to (K), two-tailed Student’s t test: *P < 0.05, ***P < 0.001 compared
with the Dpp4−/− + EcN group. For (M), (N) to (P), (R) and (Q), one-way ANOVA
with Dunnett’s T3 test (M) or Tukey’s post hoc test [(N) to (P) and (R)] and
Kruskal-Wallis test (Q): *P < 0.05, **P < 0.01, ***P < 0.001 compared with the
control group, ##P < 0.01, ###P < 0.001 compared with the Bt group.
GLP-1 (fig. S5, H and I). Moreover, the EcN::
btDpp4, EcN::bfDpp4, and EcN::bvDpp4 groups
had lower glucose-induced insulin levels and
poorer glucose tolerance than control EcN-
gavaged mice (Fig. 3, E to G). Colonization ex-
periments gavaging EcN-btDPP4 into Dpp4−/−
mice were used to exclude the possibility that
host DPP4 was decreasing active GLP-1 (Fig.
3H and fig. S5J). As expected, similar effects of
EcN::btDpp4 were observed in Dpp4−/− mice
compared with wild-type mice (Fig. 3, I to K,
and fig. S5, K to P), but host DPP4 in the intes-
tine was not detectable in either group.
To explore the role of btDPP4 in the effect of
B. thetaiotaomicron on glucose homeostasis,
we constructed a Dpp4-deficient strain of
B. thetaiotaomicron JS071 (Bt-DDpp4) (fig. S6A)
and used it to colonize wild-type SPF-housed
mice on a HFD and compare with wild-type
B. thetaiotaomicron (Fig. 3L). Bt-DDpp4 showed
no DPP4 activity toward GLP-1, but also showed
no growth deficit compared with wild-type
B. thetaiotaomicron JS071 (Bt) (fig. S6, B to D).
One week after the colonization of HFD-fed
mice with Bt (fig. S6, E and F), we found that
DPP4 activities in the feces and extraluminal
intestinal tissue were both higher than in the
control group given anaerobic PBS (Fig. 3M
and fig. S6G), whereas the host intestinal DPP4
concentrations were identical among the dif-
ferent treatment groups (fig. S6H). Active
GLP-1 concentrations in extraluminal intestinal
tissue and plasma in the Bt group were lower
than those in the control (Fig. 3N and fig. S6I),
although total GLP-1 levels were unchanged
(fig. S6, J and K). After wild-type Bt treatment,
mice exhibited lower glucose-stimulated insu-
lin levels and poorer glucose tolerance than
the anaerobic PBS–treated control mice (Fig. 3O
and fig. S6, L and M). Colonization with Bt-
DDpp4 blunted these various effects of Bt on
DPP4 activity, active GLP-1 levels, and glucose
homeostasis (Fig. 3, M to O, and fig. S6, F to
M). Long-term colonization of Bt for 6 weeks
(Fig. 3L) resulted in similar effects as 1 week
of colonization, and body weight and food
intake remained unchanged between the two
groups (Fig. 3, P to R, and fig. S6, N to W).
To eliminate the effect of endogenous bacteria
on Bt-induced glucose disruption, we colonized
GF mice treated with 1% DSS with wild-type
Bt or Bt-DDpp4 (fig. S7, A to C). Like the effects
in SPF mice, gavage with either wild-type Bt or
purified btDPP4 delivered by L100-55 resulted
in higher DPP4 activity, lower active GLP-1 levels,
and impaired glucose tolerance compared with
controls (fig. S7, D to M); these effects were not
apparent in the Bt-DDpp4 colonization group
(fig. S7, D to M).
Because the above sequence analysis showed
that btDPP4 contained an N-terminal signal
peptide (fig. S4E), we next investigated whether
Bt impaired host glucose homeostasis through
the secretion of btDPP4. We observed DPP4
activity in the B. thetaiotaomicron JS071 cul-
ture supernatant (fig. S8, A and B). Further-
more, a btDPP4 signal peptide knockout strain,
Bt-DSP-Dpp4, was established (fig. S8, C and D).
As expected, no DPP4 activity was detected
in the culture supernatant of Bt-DSP-Dpp4,
whereas the bacterial pellets showed higher
DPP4 activity in the Bt-DSP-Dpp4 strain com-
pared with wild-type Bt (fig. S8E). We next
gavaged HFD-fed mice with either Bt or Bt-
DSP-Dpp4, leading to similar levels of abun-
dance after gavage (fig. S8, F and G). Only
the Bt group showed a significant increase in
DPP4 activity in extraluminal intestinal tissue
and a decrease of active GLP-1 (fig. S8, H to J)
compared with control.
Clinical DPP4 inhibitors do not inhibit
microbial DPP4
We next investigated whether this interking-
dom pathway is susceptible to inhibitors that
target hDPP4. We found that the clinical DPP4
inhibitor sitagliptin (39) was ~78-fold less active
against btDPP4 than hDPP4 (Fig. 4A and fig.
S9A). All of the clinically relevant DPP4 inhi-
bitors that we tested showed weak inhibition of
DPP4 activity and did not affect the growth of
B. thetaiotaomicron (fig. S9).
We then solved the apo structure of btDPP4
to a resolution of 1.92 Å (Fig. 4B and table S3).
There are two dimers in one crystallographic
unit cell. Each monomer comprises two struc-
tural domains: an N-terminal, eight-bladed
b-propeller with open “Velcro” topology and a
C-terminal a/b-hydrolase catalytic domain
(Fig. 4B). The dimeric interface of the btDPP4
structure is predominantly mediated by hydro-
phobic interactions of the catalytic domain and
two protruding antiparallel b-sheets from pro-
peller blade IV (Fig. 4B). With an amino acid
sequence identity of 32% to hDPP4, the overall
architecture of btDPP4 could be superimposed
on hDPP4 [Protein Data Bank (PDB) ID 1R9N]
with a root-mean-square deviation (RMSD) of
1.4 Å over 611 Ca atoms (fig. S10A, left). The
structure of btDPP4 resembles previously solved
structures of pgDPP4 (PDB ID 5OLJ) from
Porphyromonas gingivalis (19) (RMSD = 0.9 Å),
smDPP4 (PDB ID 2ECF) from Stenotrophomonas
maltophilia (40) (RMSD = 1.8 Å), and pmDPP4
(PDB ID 5YP1) from Pseudoxanthomonas
mexicana WO24 (20) (RMSD = 1.6 Å) (fig.
S10B). Structural comparison of the catalytic
domains between btDPP4 and hDPP4 indi-
cated that the conserved S606, D681, and H713
residues in btDPP4 act as the catalytic triad
(fig. S10A, right). A hydrophobic pocket near
Y638 may serve as the S1 pocket, and the S2
subsite is formed by a hairpin R-loop contain-
ing R113, an E201-E202 motif, Y524, and Y642
(fig. S10A, right). Subsequent site-directed
mutagenesis of btDPP4 followed by enzymatic
assays showed that most of the above residues
(except for R113) resulted in a loss of activity
(fig. S10C), thus verifying their roles in btDPP4
function.
To elucidate the molecular basis for the weaker
inhibitory effect of sitagliptin on btDPP4 rela-
tive to hDPP4, we co-crystalized btDPP4 with
sitagliptin and solved the complex structure to
a resolution of 1.97 Å (Fig. 4C, fig. S10D, and
table S3). Both a 2mFo-DFc map and an omit
map indicated the binding of sitagliptin at this
site (Fig. 4C and fig. S10D). Structural compa-
rison with the sitagliptin-bound hDPP4 (PDB
ID 1X70) (41) revealed common and distinct
recognition features between btDPP4 and hDPP4
(Fig. 4D). The 2,4,5-trifluorophenyl moiety and
the oxobutan-amine of sitagliptin (fig. S10E)
adopted nearly identical conformations and
Wang et al., Science 381, eadd5787 (2023)
4 August 2023
6 of 13
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. The clinically relevant DPP4 inhibitor
sitagliptin exhibits a weaker effect on btDPP4
than on hDPP4. (A) In vitro inhibition curve of
sitagliptin against either hDPP4 or btDPP4. (B) The
overall dimer structure of btDPP4 shown in a
cartoon representation. One protomer is shown in
gold and one in pale green. The dimeric interface is
shown in orange and deep green. (C) 2mFo-DFc
map (blue mesh, s = 1.5) of btDPP4-bound
sitagliptin. The binding pocket is shown as a gray
surface, and the side chains of some key residues
are shown as sticks. (D) Active-site super-
imposition between sitagliptin-bound btDPP4 (pale
green) and sitagliptin-bound hDPP4 (gray, PDB:
1X70), with key residues labeled as one-letter
codes. Both inhibitors are shown as ball-and-stick
figures. The trifluoromethyl group of the bound
sitagliptin in btDPP4 flipped by ~40° to avoid steric
clash with btDPP4-E342 compared with that in
hDPP4. The same color schemes were applied to
all figures. F, phenylalanine; S, serine; R, arginine;
E, glutamic acid; Y, tyrosine. (E) Schematic
diagram of the clinical trial design for evaluating
the effects of microbial DPP4 on the clinical
efficacy of sitagliptin. (F) The change of HbA1C
over 3 months of sitagliptin intervention between
the SHR group (n = 34) and the SLR group (n =
23). (G) Fecal DPP4 activity of the SHR and SLR
groups (His-Ala-pNA as a substrate).
(H) Volcano plot of metagenomic sequencing data
of the SHR and SLR groups. The graph depicts the
average log2 ratio of abundances between both
groups for each individual species and the
corresponding adjusted P value. The adjusted
P value threshold was calculated using moderated
Student’s t test followed by Benjamini–Hochberg
multiple test correction false discovery rate (FDR).
(I) Taxonomic cladogram generated from linear
discriminant analysis (LDA) effect size of meta-
genomic sequencing data. Blue indicates enriched
taxa in the SHR group; pink represents enriched
taxa in the SLR group. (J) Relative levels of
microbial DPP4 mRNA in human feces as detected
by RT-qPCR. (K) Correlation analysis between
microbial DPP4 mRNA levels and fecal DPP4
activity, the change (before treatment minus after
treatment) of fasting blood glucose, HbA1c, fasting
insulin, and fasting C-peptide. Spearman’s rank
test: *Q < 0.05, **Q < 0.01, Q value (FDR-adjusted
P value). Data are presented as means ± SEM. For
(G), two-tailed Students’ t test and for (J), Mann-
Whitney U test: **P < 0.01, ***P < 0.001.
A
150
)
%
(
e
t
a
r
n
o
i
t
i
b
i
h
n
I
C
E
H
)
e
u
l
a
v
-
P
.
j
d
a
(
0
1
g
o
L
-
J
A
N
R
m
4
p
p
D
l
a
i
i
b
o
r
c
m
e
v
i
t
a
l
e
R
5
4
3
2
1
0
hDPP4
btDPP4
100
IC50=73.4 ± 14.3 nM
50
0
-4
IC50=5.73 ± 2.37 µM
-2
0
2
4
Log10[Sitagliptin] µM
R113
E201
E202
S606
E342
Y524
F
)
%
(
C
1
A
b
H
15
10
6.5
5
0
3-month sitagliptin
treatment
Sitagliptin High Responder
(SHR)
Patients with
newly diagnosed T2D
Sitagliptin Low Responder
(SLR)
Metagenomic
analysis
Glucose homeostasis
assessment
Gut microbial
community
Sitagliptin
effectiveness
I
6
4
2
B. thetaiotaom icron
P. distasonis
B. uniformis
B. ovatus
B. vulgatus
Mogibacterium diversum
Actinomyes sp S6 Spd3
adj. P-value=0.05
B
Catalytic
domain
β-propeller
domain
D
V/VI
linker
motif
R358
Blade IV
Y524
Y547
F357
Y642
Y666
E342
~40°
S209
S606
S630
E202
E206
E201
E205
Pre-treatment
Post-treatment
G
y
t
i
v
i
t
c
a
4
P
P
D
l
a
c
e
F
]
)
s
e
c
e
f
g
m
*
n
m
i
(
/
l
o
m
n
[
SHR
SLR
150
100
50
0
***
SHR
SLR
Bacteroidia
Bacteroidales
Bacteroidetes
Bacteroidaceae
Bacteroides
Bacteroides thetaiotaomicron
Porphyromonadaceae
Parabacteroides
Clostridium clostridioforme
Clostridium symbiosum
Sutterellaceae unclassified
Flavonifractor plautii
Flavonifractor
SHR
SLR
Bacillales
Bacillales noname
Gemella
Bacteroidetes
Bacteroidia
Bacteroidales
Porphyromonadaceae
Parabacteroides
Bacteroiaceae
B. thetaiotaomicron
Clostidiales noname
s
ute
Firmic
ctor
nifra
o
v
Fla
m
u
i
r
e
t
c
a
b
i
r
O
A
c
t
i
n
o
m
y
c
e
t
a
c
e
a
e
A
c
t
i
n
o
m
y
c
e
t
a
l
e
s
Haemophilus
-20
-10
0
10
20
Sutterellaceae unclassified
Log 2 fold
***
K
-6.0
-4.8 -3.6
-2.4
SHR
SLR
***
**
bt Dpp4
bfDpp4
beDpp4
bvDpp4
bdDpp4
**
*
**
*
**
*
btDpp4
bfDpp4
beDpp4
bvDpp4
Dpp4
bd
Fasting blood glucose change
Fecal DPP4 activity
HbA1C
change
Fasting C-peptide change
Fasting insulin change
Streptococcus sanguinis
Actinomyces graevenitzii
Granulicatella adiacens
Actinomyces odontolyticus
Actinomycetales
Oribacterium
Bacillales
Oribacterium sinus
Haemophilus parainfluenzae
Haemophilus
Bacillales noname
Clostridiales noname
Gemella
Bifidobacterium longum
Actinobacteria
Clostridiales
Actinomycetaceae
Firmicutes
-1.2
0
LDA SCORE (log 10)
1.2
2.4
3.6
4.8
6.0
l
n
o
i
t
a
e
r
r
o
c
k
n
a
r
'
s
n
a
m
r
a
e
p
S
0.5
0
-0.5
conservatively interacted with E201, E202, Y524,
S606, and Y642 in btDPP4, which correspond to
E205, E206, Y547, S630, and Y666 in hDPP4
(Fig. 4D). In the hDPP4-sitagliptin structure,
the trifluoromethyl triazolopyrazinyl (TMTP) of
sitagliptin forms additional interactions, includ-
ing p-p stacking with F357 and hydrogen bonds
with S209 and R358. However, in the btDPP4-
sitagliptin structure, this TMTP moiety under-
goes ~40° of flipping to avoid steric hindrance
with E342, and thus disallows the formation of
stable interactions to the small helix linking
blade V and blade VI (the V/VI linker motif)
(Fig. 4D). Sequence alignment revealed that
the TMTP-stabilizing residues (F357, S209, and
R358) in hDPP4 are not conserved in btDPP4,
whereas the protruding btDPP4-E342 residue
that occludes the trifluoromethyl group is miss-
ing in hDPP4 (fig. S10F). Therefore, the weak
TMTP contact observed in the btDPP4-sitagliptin
complex structure likely accounts for the
decreased inhibitory effect of sitagliptin on
btDPP4 (Fig. 4A), and the variable V/VI linker
motif provides a structural basis for explaining
the distinct inhibitory efficiencies of gliptins
against different DPP4 homologs. Thus, inhibi-
tion of hDPP4 alone may not be able to fully
Wang et al., Science 381, eadd5787 (2023)
4 August 2023
7 of 13
RES EARCH | R E S E A R C H A R T I C L E
restore the concentration of active GLP-1, and
it seems necessary to inhibit both human and
microbial DPP4 to improve glucose homeostasis
more effectively.
Microbial DPP4 limits the efficacy of sitagliptin
Because of the weaker inhibitory effects of cli-
nical DPP4 inhibitors toward microbial DPP4,
we next tested the effect of gut microbial DPP4
on the therapeutic effectiveness of clinical DPP4
inhibitors. Serum and stool samples were col-
lected from 57 patients with newly diagnosed
T2D before and after sitagliptin treatment for
3 months (fig. S11A) (42), and changes in gly-
cated hemoglobin (HbA1c) were monitored to
assess the therapeutic effect of sitagliptin (Fig.
4E, and table S4). After sitagliptin intervention,
high interpersonal variability in glucose homeo-
stasis assessment of HbA1c was observed, indi-
cating heterogeneous responses to sitagliptin
among the cohort. According to the diagnostic
criteria for T2D (43) and previous sitagliptin
clinical trials (44–47), we defined the patients
whose HbA1c decreased to ≤6.5% or decreased
by >1% during sitagliptin treatment as sitagliptin
high-responders (SHRs, n = 34); the other pa-
tients were defined as sitagliptin low-responders
(SLRs, n = 23) (Fig. 4F). There was no significant
difference in the baseline of physiological indi-
cators, including glucose metabolism, between
SHRs and SLRs before sitagliptin treatment (fig.
S11B). SHRs exhibited greater improvement in
fasting blood glucose than did SLRs, with equal
levels in fasting insulin and fasting C-peptide
(fig. S11, C to E). The baseline DPP4 activity in
stool samples of SLRs was significantly higher
than that in SHRs (Fig. 4G).
Further exploring the gut microbiota of the
two categories of patients (table S5), we did not
observe any overall difference in the diversity
indices between the two groups (fig. S11, F and
G); however, we did observe an enrichment of
B. thetaiotaomicron among the SLRs (Fig. 4H).
Linear discriminant analysis effect size analy-
sis also showed the Bacteroides genus and
B. thetaiotaomicron as distinguishable markers
between the SHRs and SLRs groups (Fig. 4I),
mirroring the DPP4 activity analysis performed
in the bacterial isolates (Fig. 2A). Both gene reads
and mRNA levels of btDPP4 and bvDPP4 were
enriched in the SLR group (Fig. 4J and fig. S11H).
btDPP4 (both gene reads and mRNA levels)
negatively correlated with the improvements
of fasting blood glucose and HbA1c and showed
a positive correlation with fecal DPP4 activity
(Fig. 4K and fig. S11I), indicating that the
abundance of btDPP4 and bvDPP4 may reflect
the fecal DPP4 activity in individuals with T2D.
To investigate the potential effect of the gut
microbiota (especially B. thetaiotaomicron and
btDPP4) on the clinical efficacy of sitagliptin,
we performed FMT experiments transferring
human stool into HFD-fed mice pretreated with
antibiotic for 1 day and then with sitagliptin for
6 weeks (fig. S12A). We obtained a bacteriophage
strain (BtP) that selectively kills B. thetaiotaomicron
(48, 49) (fig. S12, B to E). As a control, we used T7
phage (50), which targets E. coli, an abundant,
non–DPP4-producing microbe in our cohort
(fig. S12F). As expected, mice transplanted with
the gut microbiota from SLRs had a higher
abundance of B. thetaiotaomicron (fig. S12E),
as well as elevated DPP4 activity (fig. S12,
G and H) and no change in the intestinal host
DPP4 content (fig. S12I), and the mice showed
lower active GLP-1 concentrations and impaired
glucose tolerance (fig. S12, J to M). T7 phage
treatment substantially reduced the abundance
of E. coli in the feces (fig. S12F); however, the
glucose homeostasis was not altered (fig. S12,
G to M). By contrast, in the SLR + BtP group,
clearance of B. thetaiotaomicron (fig. S12E)
reversed the negative effects induced by SLR
FMT on sitagliptin treatment (fig. S12, G to M).
Dau-d4 is a specific microbial DPP4 inhibitor
Our results indicated the potential of thera-
peutically targeting microbial DPP4 to treat T2D.
Therefore, to identify inhibitors of microbial
DPP4, we developed a high-throughput drug-
screening system using the btDPP4 protein and
tested the inhibitory effects of a small-molecule
library containing ~107,000 compounds (Fig. 5A).
At a concentration of 10 mM, 10 compounds
inhibited btDPP4 activity by >90% (hit rate
0.01%; Fig. 5B); the low hit rate indicated that
the screen was selective. We measured the
median inhibitory concentration (IC50) values
(fig. S13) and among the 10 hits, Dau, an alka-
loid natural product found in a variety of medi-
cinal herbs, including Menispermum dauricum
(51), showed the strongest inhibitory activity
against btDPP4 (IC50 = 0.37 mM). It also had
high selectivity because almost no inhibition of
hDPP4 was observed (fig. S13B).
A series of Dau derivatives were obtained
for preliminary structure-activity relationship
studies (fig. S14, A and B). We found that Dau-d4
(Fig. 5C and fig. S14C), synthesized by methoxyl-
ation of the hydroxyl groups, showed improved
btDPP4 inhibitory activity (IC50 = 88 nM) over
the native molecule and no hDPP4 inhibitory
activity (Fig. 5D). Using isothermal titration
calorimetry, we found that Dau-d4 bound to
btDPP4 at a KD of 0.27 mM (Fig. 5E), whereas
sitagliptin bound btDPP4 at a KD of 12.6 mM
(fig. S14D).
Dau-d4 showed strong inhibitory effects
against bfDPP4, beDPP4, bvDPP4, and bdDPP4
(Fig. 5F). It also inhibited DPP4 activity in
Bacteroides spp. at the strain level (fig. S14E)
and in ex vivo communities obtained from
healthy volunteers (fig. S14F). By contrast, sitag-
liptin had little effect against Bacteroides spp.
at the protein, strain, and ex vivo levels (fig.
S14, E to G).
Next, we determined the co-crystallized struc-
ture of the btDPP4-Dau-d4 complex to a reso-
lution of 2.74 Å (Fig. 5G, fig. S15A, and table S3).
Because the tetrahydroisoquinoline moieties on
both sides of Dau-d4 (fig. S15B) are too large to
be coordinated into the S1 pocket (Fig. 5G), the
binding sites for Dau-d4 to btDPP4 must be dis-
tinct from that of sitagliptin (fig. S15C). Specifi-
cally, the tetrahydroisoquinoline A moiety and
the phenyl A ring form two p-cation interactions
with the side chains of S2 subsite residues R113
and Y524 (Fig. 5H). The other half of the tetra-
hydroisoquinoline and phenyl moiety forms
hydrophobic interactions with W605 and Y716
and hydrophilic interactions with R721, com-
pletely occluding the binding of substrate p1′ and
p2′ residues at the S1′ and S2′ subsites (Fig. 5H).
The importance of these residues was confirmed
by mutations to Ala that showed marked reduced
binding affinities for Dau-d4 (Fig. 5I).
Because Dau-d4 is more selective for btDPP4
than for hDPP4, we superimposed the btDPP4-
Dau-d4 complex structure with hDPP4 and
found that several residues surrounding Dau-d4
in btDPP4 were not conserved in hDPP4 (fig.
S15D). Among them, P110/Y716/R721 from
bunch 1, Y522/Q531 from bunch 2, and the
V/VI linker motif were the major contributors
to the specificity of Dau-d4 for btDPP4 (fig.
S15D). Individual mutations of these sites to
the corresponding residues in hDPP4 attenu-
ated Dau-d4 binding affinity to varied extents
(fig. S15E). By contrast, we were also able to
engineer hDPP4 to acquire the capacity for
Dau-d4 binding by introducing the single
mutation of either hDPP4-K554 to glutamine
(K554Q) or hDPP4-F357 to glutamic acid (F357E)
(fig. S15F). This conversion of hDPP4 giving it
the ability to effectively bind Dau-d4 was even
more pronounced when a chimeric hDPP4
protein containing the btDPP4-V/VI linker mo-
tif replacement (hDPP4-V354~P362→btDPP4-
I340~I348) was tested (fig. S15F). These three
hDPP4 mutants that could effectively bind
Dau-d4 retained their activity toward GLP-1(7-37)
(fig. S15G). Similar to the binding assay results
(fig. S15F), Dau-d4 showed strong inhibition
of these three mutants of hDPP4 (fig. S15H).
These results show that different residue com-
positions around the active sites, especially the
V/VI linker motif, are the primary elements
that determine the specificity of inhibition of
btDPP4 by Dau-d4 versus its targeting of hDPP4.
Dau-d4 improves glucose homeostasis
We next performed Dau-d4 treatment of SPF-
raised mice on a HFD colonized with an an-
aerobic PBS control, Bt, or Bt-DDpp4 to further
test whether Dau-d4 promotes glucose ho-
meostasis through a btDPP4-dependent path-
way (Fig. 6A). After successful colonization by
Bt and Bt-DDpp4 (fig S16, A and B), we found
that Dau-d4 decreased DPP4 activity in the
feces and extraluminal intestinal tissue (Fig.
6B and fig. S16C) without changing the intes-
tinal host DPP4 content (fig. S16D). Dau-d4
Wang et al., Science 381, eadd5787 (2023)
4 August 2023
8 of 13
RES EARCH | R E S E A R C H A R T I C L E
A
Library of 107,000
small molecular compounds
384-well plates (10 µM)
against btDPP4
B
150
100
50
0
-50
)
W
µ
(
P
D
)
l
o
m
/
J
k
(
)
%
(
e
t
a
r
n
o
i
t
i
b
h
n
i
I
E
G
IC50 measurement
Structure modification
Multilevel activity
verification
In vivo
evaluation
C
O
O
N
N
O
O
O
O
40000
80000
120000
Dau-d4
Individual compounds
Time (min)
)
W
µ
(
P
D
)
l
o
m
/
J
k
(
Time (min)
Dau-d4—btDPP4
K D = 0.27 ± 0.03 (µM)
∆H = -210 ±
4.67 (kJ/mol
Dau-d4—hDPP4
No binding
Molar ratio
Molar ratio
D
150
)
%
(
e
t
a
r
n
o
i
t
i
b
i
h
n
I
100
50
0
-4
F
150
)
%
(
e
t
a
r
n
o
i
t
i
b
i
h
n
I
100
50
0
-4
btDPP4
hDPP4
IC50 = 88.0 ± 6.35 nM
IC50 > 100 µ M
(Maximum dissolution
concentration)
-2
0
Log10[Dau-d4] µM
2
4
bfDPP4
beDPP4
bvDPP4
bdDPP4
-2
0
2
4
Log10[Dau-d4] µM
Protein
bfDPP4
IC50 (µM)
0.22 ± 0.02
beDPP4
0.20 ± 0.02
bvDPP4
0.42 ± 0.04
bdDPP4
0.11 ± 0.02
H
I
R113
E201
E342
E202
W605
S606
Y524
Y716
R113
p3’/P
p1/P
p2/Y
R721
p1’/S
p2’/K
W605
Y725
Y524
Site mutation
WT
W605A
Y524A
R113A
Y716A
R721A
K
D (µM)
0.27
>100
>100
>100
1.49
7.43
Fig. 5. High-throughput screening identifies Dau-d4 as a selective inhibitor
of microbial DPP4. (A) Schematic diagram illustrating the study design for
the high-throughput screening and evaluation of microbial DPP4 inhibitors.
(B) Distribution of the inhibition rates against btDPP4 of the individual
compounds from a 107,000 small-molecule chemical library. (C) Chemical
structure of Dau-d4. (D) Inhibition curve of Dau-d4 against btDPP4 or hDPP4,
n = 3 biological replicates. (E) Isothermal titration calorimetry (ITC)–mediated
determination of btDPP4 and hDPP4 binding to Dau-d4. (F) Inhibition curve of
Dau-d4 against bfDPP4, beDPP4, bvDPP4, or bdDPP4 (top) and the quantitation
of the IC50 (bottom), n = 3 biological replicates. (G) 2mFo-DFc map (blue mesh,
s =1.5) of btDPP4-bound Dau-d4 structure with btDPP4 shown as a gray surface.
(H) Dau-d4 binding competes with the substrate binding at the active site.
Substrate p2, p1, p1′, p2′, and p3′ residues (Y, tyrosine; P, proline; S, serine; K,
lysine) are indicated and shown in salmon (PDB code: 1R9N). (I) KD values of
Dau-d4 binding against btDPP4 among different loss-of-function mutations, n = 3
biological replicates. WT, wild-type btDPP4; W, tryptophan; R, arginine; A, alanine.
Wang et al., Science 381, eadd5787 (2023)
4 August 2023
9 of 13
RES EARCH | R E S E A R C H A R T I C L E
treated mice increased the pools of active
GLP-1 (Fig. 6C and fig. S16, E to G), and this
was associated with higher glucose-stimulated
insulin levels and improved responses on an
oral glucose tolerance test (Fig. 6D and fig.
S16, H and I). Bt colonization blunted the
effects of Dau-d4, whereas Bt-DDpp4 did not
(Fig. 6, B to D, and fig. S16, B to I). Dau-d4
remained effective in Dpp4−/− mice in terms of
the beneficial effects on glucose homeostasis
(Fig. 6, E to H, and fig. S16, J to M) because it
targets microbial-derived DPP4. Similar ef-
fects of Dau-d4 were observed in both DSS
and indomethacin mouse models of leaky gut
(fig. S17, A to R).
We next assessed the chronic effects of Dau-d4.
Mice fed a HFD for 12 weeks were subsequently
treated with Dau-d4 for an additional 6 weeks
(Fig. 6I). Dau-d4 did not significantly influence
body weight or food intake during the 6-week
treatment period compared with PBS treatment
(fig. S18, A and B). However, Dau-d4 treatment
resulted in significantly lower DPP4 activity in
the feces and extraluminal intestinal tissue and
greater intestinal and plasma active GLP-1 levels,
and Dau-d4 treatment improved glucose toler-
ance compared with PBS control (Fig. 6, J to L
B
y
t
i
v
i
t
c
a
4
P
P
D
r
a
l
u
s
s
i
t
l
a
n
i
t
s
e
t
n
i
L
E
]
)
n
i
e
t
o
r
p
g
m
*
n
m
i
(
/
l
o
m
n
[
F
J
A
E
I
1. Control (PBS with 1% DMSO)
2. Bt
3. Dau-d4
4. Dau-d4+Bt
5. Dau-d4+Bt- Dpp4
HFD 12-week
Analyze
SPF mice
Week 0
Week 1
1. Control (PBS with 1% DMSO)
2. Dau-d4
HFD 12-week
Analyze
SPF Dpp4-/- mice
Week 0
Week 1
1. Control (PBS with 1% DMSO)
2. Dau-d4
HFD 12-week
Analyze
SPF mice
Week 0
Week 6
***
###
***
$$$
50
40
30
20
10
0
Control Bt
Dpp4
Dau-d4+Bt
Dau-d4
Dau-d4+Bt
y
t
i
v
i
t
c
a
4
P
P
D
r
a
l
u
s
s
i
t
l
a
n
i
t
s
e
t
n
i
L
E
]
)
n
i
e
t
o
r
p
g
m
*
n
m
i
(
/
l
o
m
n
[
1
-
P
L
G
e
v
i
t
c
a
a
m
s
a
l
P
)
L
m
g
p
(
/
15
10
5
0
15
10
5
0
***
-/-
-/-+Dau-d4
Dpp4
Dpp4
**
Control
Dau-d4
C
G
1
-
P
L
G
e
v
i
t
c
a
a
m
s
a
l
P
)
L
m
g
p
(
/
15
10
5
0
$$
***
##
*
Control Bt
Dpp4
Dau-d4+Bt
Dau-d4
Dau-d4+Bt
1
-
P
L
G
e
v
i
t
c
a
a
m
s
a
l
P
)
L
m
g
p
(
/
25
20
15
10
5
0
***
-/-
-/-+Dau-d4
Dpp4
Dpp4
***
**
Control
Dau-d4
***
***
20
40
60
80
100
Time (min)
K
)
L
d
/
g
m
(
e
s
o
c
u
l
g
d
o
o
B
l
500
400
300
200
100
***
0
0
(A to L) for HFD
N
O
D
)
L
m
g
n
(
/
n
i
l
u
s
n
i
a
m
s
a
l
P
8
6
4
2
0
H
Control
Bt
Dau-d4
Dau-d4+Bt
Dau-d4+Bt Dpp4
$
*
#
**
0 15
0 15
0 15
0 15
0 15
Time (min)
Dpp4-/-
Dpp4-/-+Dau-d4
**
n
i
l
u
s
n
i
a
m
s
a
l
P
)
L
m
g
n
(
/
15
10
5
0
0
15
0
15
Time (min)
32000
24000
16000
***
i
)
n
m
*
L
d
g
m
/
Control
Dau-d4
$$$
###
***
***
(
8000
0
2000
1500
1000
500
0
Control
Dau-d4 Sit
Dau-d4+Sit
L
P
S
T
T
G
O
f
o
C
U
A
1
-
P
L
G
e
v
i
t
c
a
r
a
l
u
s
s
i
t
l
a
n
i
t
s
e
t
n
i
L
E
)
n
i
e
t
o
r
p
/
g
m
g
p
(
M
1. Control (PBS with 1% DMSO)
2. Dau-d4
3. Sit
4. Dau-d4+Sit
Analyze
with CD
Week 0
Week 6
Q
(M to S) for CD
***
$$$
y
t
i
v
i
t
c
a
4
P
P
D
l
a
c
e
F
]
)
s
e
c
e
f
g
m
*
n
m
i
(
/
l
o
m
n
[
200
150
100
50
0
Control
Dau-d4 Sit
Dau-d4+Sit
$$$
###
***
***
1
-
P
L
G
e
v
i
t
c
a
a
m
s
a
l
P
)
L
m
g
p
(
/
20
15
10
5
0
Control
Dau-d4 Sit
Dau-d4+Sit
y
t
i
v
i
t
c
a
4
P
P
D
r
a
l
u
s
s
i
t
l
a
n
i
t
s
e
t
n
i
L
E
]
)
n
i
e
t
o
r
p
g
m
*
n
m
i
(
/
l
o
m
n
[
40
30
20
10
0
***
**
$$$
##
Control
Dau-d4 Sit
Dau-d4+Sit
Control
Dau-d4
Sit
Dau-d4+Sit
R
500
400
e
s
o
c
u
l
g
d
o
o
B
l
)
L
d
g
m
/
(
300
200
100
0
0
**
**
#
$
#
$
*
*
$$
*
T
T
G
O
f
o
C
U
A
20
40
60
80
100
Time (min)
40000
30000
20000
*
*
$$
##
i
)
n
m
*
L
d
g
m
/
(
10000
0
Control
Dau-d4 Sit
Dau-d4+Sit
Wang et al., Science 381, eadd5787 (2023)
4 August 2023
10 of 13
RES EARCH | R E S E A R C H A R T I C L E
Fig. 6. Dau-d4 improves host glucose metabolism. (A) Experimental scheme
for (B) to (D), n = 6 mice/group. Mice were fed a HFD for 12 weeks, divided into
five groups, and treated with anaerobic PBS control, Bt single colonization
(Bt group), Dau-d4 (10 mg/kg) gavage (Dau-d4 group), Dau-d4 (10 mg/kg)
combined with Bt single colonization (Dau-d4 + Bt group), or Dau-d4 (10 mg/kg)
combined with Bt-DDpp4 colonization (Dau-d4 + Bt-DDpp4 group), respectively,
for 1 week. (B) EL intestinal tissular DPP4 activity (His-Ala-pNA as a substrate).
(C) Plasma levels of active GLP-1. (D) Glucose-stimulated insulin levels.
(E) Experimental scheme for (F) to (H), n = 6 mice/group. Dpp4−/− mice were
fed a HFD for 12 weeks, divided into two groups, and administered a PBS control
containing 1% DMSO or Dau-d4 (10 mg/kg) for 1 week, n = 6 mice/group.
(F) EL intestinal tissular DPP4 activity (His-Ala-pNA as a substrate). (G) Plasma
levels of active GLP-1. (H) Glucose-stimulated insulin levels. (I) Experimental
scheme for (J) to (L), n = 6 mice/group. Mice were fed a HFD for 12 weeks,
divided into two groups, and then administered a PBS control containing 1%
dimethylsulfoxide (DMSO) or Dau-d4 (10 mg/kg) for 6 weeks. (J) Plasma levels
of active GLP-1. (K and L) OGTT curve (K) and its AUC (L). (M) Experimental
scheme for (N) to (S), n = 6 mice/group. The ob/ob mice were divided into
four groups and treated with 10 mg/kg Dau-d4, 10 mg/kg sitagliptin, or 10 mg/kg
Dau-d4 and 10 mg/kg sitagliptin or the same volume of PBS control containing
1% DMSO for 6 weeks. (N) Fecal DPP4 activity. (O) EL intestinal tissular DPP4
activity (His-Ala-pNA as a substrate). (P) Concentrations of EL intestinal tissular
active GLP-1 levels. (Q) Plasma levels of active GLP-1. (R and S) OGTT curve (R)
and its AUC (S). Data are presented as means ± SEM. For (B) to (D), one-way
ANOVA with Tukey’s post hoc test: *P < 0.05, **P < 0.01, ***P < 0.001 compared
with the control; #P < 0.05, ##P < 0.01, ###P < 0.001 compared with the
Dau-d4 group; and $P < 0.05, $$P < 0.01, $$$P < 0.001 compared with the
Dau-d4 + Bt group. For (F) to (L), two-tailed Student’s t test: *P < 0.05, **P < 0.01,
***P < 0.001. For (N) to (S), one-way ANOVA with Tukey’s post hoc test [(N),
(P), and (Q)] or Dunnett’s T3 test [(O), (R), and (S)]: *P < 0.05, **P < 0.01,
***P < 0.001 compared with the control, #P < 0.05, ##P < 0.01, ###P < 0.001
compared with the Dau-d4 group.
and fig. S18, C to F). Dau-d4 could significantly
improve glucose homeostasis of ob/ob mice, a
genetic model of obesity, without altering body
weight and food intake (fig. S18, G and H). The
effect in ob/ob mice was associated with an
inhibition of the decrease of active GLP-1 (Fig. 6,
M to S, and fig. S18, G to I), and the combination
of Dau-d4 and sitagliptin in ob/ob mice further
improved glucose tolerance without altering
body weight, food intake, or intestinal levels of
host DPP4 (Fig. 6, M to S, and fig. S18, G to I).
Furthermore, Dau-d4 improved the respon-
siveness of mice treated with sitagliptin that
had received FMT with gut microbiota from
SLRs (fig. S19).
Dau-d4 had no significant inhibitory effects
on the growth of several gut bacteria (fig.
S20A) and had no toxic effects on several host
cell lines (fig. S20B). To explore the in vivo
safety of Dau-d4, we administered Dau-d4 to
SPF mice by gavage with the drug at an effec-
tive dose (10 mg/kg) and a high dose (100 mg/kg)
for 6 weeks, and found that neither dose resulted
in any notable adverse effects, including loss of
body weight, changes in food intake, or changes
in liver and kidney function (alanine amino-
transferase, aspartate transaminase, urea nitro-
gen, and creatinine; fig. S20, C to H). To further
explore the stability of Dau-d4 in the native gut
microbiota setting, we found Dau-d4 levels
remain unchanged for 4 days in ex vivo fecal
communities anaerobically (fig. S20I). We then
gave Dau-d4 orally to both SPF and GF mice and
measured fecal Dau-d4 concentrations at differ-
ent times. Although the fecal Dau-d4 concentra-
tion curve of GF mice had an overall backward
shift, perhaps because of the known slower
gastrointestinal motility and gastric emptying
of GF mice (52), the area under the curve for
Dau-d4 concentration was not significantly
different between the two groups (fig. S20, J
and K). Pharmacokinetics analysis showed that
Dau-d4 had a bioavailability of 6.43% (fig. S20L),
indicating that Dau-d4 has poor oral absorption
and may be a gut-restricted inhibitor. Together,
these results suggest that Dau-d4 may be a safe
and effective treatment for T2D, depending on
the gut microbiota profile of the patients.
To engineer a more ideal gut-restricted inhib-
itor with better potential safety and tissue
targeting, we synthesized a Dau-d4 derivative,
Dau-d23 (fig. S21A), which also has potential
inhibitory activity against btDPP4 (fig. S21B).
Using pharmacokinetics analysis, we found that
Dau-d23 is a more gut-restricted compound
than Dau-d4 (fig. S21C), with a bioavailability
of 3.61%. Moreover, Dau-d23 remained effec-
tive in HFD-fed mice in terms of the beneficial
effects on glucose homeostasis (fig. S21, D to K).
Discussion
The enzymes produced by the gut microbiota
that perform similar functions to host enzymes
may play a role in modulating host health and
diseases (13, 53, 54), but they remain largely
unexplored. By developing an experimental
workflow, we explored the role of microbial
enzymes in the disruption of host metabolic
homeostasis and clinical treatment options.
Although our ex vivo cultures produced several
microbial-host isozymes, because of the limita-
tions of enzyme assays, some of the substrates
that we used were fluorogenic or other mimics,
which cannot fully represent the activity of
human enzymes. Also, we should note that our
enzyme activity workflow can only measure
the overall activity of each isozyme at a single
substrate concentration, and the source of the
isozyme, its gene, and its kinetic analysis deserve
further attention. Moreover, our workflow was
mainly based on stool samples; however, it has
been reported that fecal microbiota compo-
sition does not accurately reflect the intestinal
environments such as the colon (55, 56). There-
fore, our microbial-host isozyme system only
simulates the activity of fecal microbiota, and
the activity of different enzymes in diverse
intestinal segments still needs to be further
evaluated.
Our microbial-host isozyme screening plat-
form identified DPP4 as a major microbial-host
isozyme that targets GLP-1, a potent incretin
involved in glucose homeostasis. Previous stud-
ies revealed the role of gut microbiota in the
regulation of host GLP-1 by other mechanisms
(57, 58), which laid the foundation for studying
the microbial factors regulating GLP-1. It has
been speculated that the gut microbiota may
also include community members with DPP4-
like activity (21). Our question was, if GLP-1 is
mainly produced in the extraluminal intestinal
tissue whereas microbial DPP4 is produced by
gut microbiota in the gut lumen, then how does
the microbial DPP4 regulate the bioavailability
of GLP-1 in the host?
Gut microbial proteins can cross the intes-
tinal barrier if its integrity has been compro-
mised and enter the bloodstream (59). However,
apart from promoting gut leakiness, a HFD can
have highly pleiotropic effects on the host,
including elevated bile acid concentrations,
overgrowth of proteobacteria, stimulation of
subclinical inflammatory responses, and altered
circadian behavior. We used low-dose DSS and
indomethacin as alternative models for intesti-
nal barrier damage and observed similar effects
as for HFD on elevation of microbial DPP4,
decrease of extraluminal intestinal and plasma
levels of active GLP-1, and metabolic readouts.
However, there are also other possible explana-
tions apart from our speculation. It cannot be
completely excluded that microbial DPP4 regu-
lates a microbiota-derived signal that is de-
tected on the apical border of L cells and alters
GLP-1 secretion. Unchanged total GLP-1 levels
in plasma are also determined by renal clear-
ance of the inactivated peptide rather than cleav-
age rate. Although all of the treatments tested
(HFD, DSS, and indomethacin) increase intes-
tinal permeability, they also cause inflammatory
responses in the gut to varying degrees, which may
have indirect effects on the GLP-1 secretion rate.
By showing the benefit of pharmacological
inhibition of microbial DPP4, our results high-
light the potential value of targeting this micro-
bial enzyme in the clinical treatment of patients
with T2D who show poor responses to currently
used DPP4 inhibitors. Such a “drugs for bugs”
Wang et al., Science 381, eadd5787 (2023)
4 August 2023
11 of 13
RES EARCH | R E S E A R C H A R T I C L E
approach in the management of this common
chronic disease could be expanded to other
similar human diseases.
Materials and methods summary
Detailed materials and methods can be found
in the supplementary materials, including me-
thods for the collection of human samples,
mouse treatments, gut bacteria isolation and
culture, microbial-host isozyme assays, DPP4
activity measurement, growth curves of related
bacterial stains, phylogenetic analyses, reverse
transcription quantitative polymerase chain
reaction (RT-qPCR), 16S rRNA gene amplicon
sequencing and analysis, metagenomic sequenc-
ing and analysis, the construction of a microbial
DPP4–expressing E. coli strain, the deficiency of
DPP4 and SP-DPP4 in B. thetaiotaomicron, mea-
surement of mouse metabolic phenotype and
intestinal permeability, purification of micro-
bial DPP4, crystallization and structure deter-
mination of btDPP4, prevalence analysis
microbial DPP4, bacteriophage isolation and
host range assessment, high-through screening
for btDPP4 inhibitors, synthesis of Dau-d4 and
Dau-d23, and isothermal titration calorime-
try assay.
RE FE RENCES AND N OT ES
1. K. A. Krautkramer, J. Fan, F. Bäckhed, Gut microbial metabolites as
multi-kingdom intermediates. Nat. Rev. Microbiol. 19, 77–94
(2021). doi: 10.1038/s41579-020-0438-4; pmid: 32968241
2. A. Agus, K. Clément, H. Sokol, Gut microbiota-derived metabolites
as central regulators in metabolic disorders. Gut 70, 1174–1182
(2021). doi: 10.1136/gutjnl-2020-323071; pmid: 33272977
3. B. B. Williams et al., Discovery and characterization of gut
microbiota decarboxylases that can produce the
neurotransmitter tryptamine. Cell Host Microbe 16, 495–503
(2014). doi: 10.1016/j.chom.2014.09.001; pmid: 25263219
4. C. Gil-Cruz et al., Microbiota-derived peptide mimics drive
lethal inflammatory cardiomyopathy. Science 366, 881–886
(2019). doi: 10.1126/science.aav3487; pmid: 31727837
5. Y. Ma et al., Identification of antimicrobial peptides from the
human gut microbiome using deep learning. Nat. Biotechnol.
40, 921–931 (2022). doi: 10.1038/s41587-022-01226-0;
pmid: 35241840
6. M. Zimmermann, M. Zimmermann-Kogadeeva, R. Wegmann,
A. L. Goodman, Separating host and microbiome contributions
to drug pharmacokinetics and toxicity. Science 363, eaat9931
(2019). doi: 10.1126/science.aat9931; pmid: 30733391
J. Balaich et al., The human microbiome encodes resistance to
the antidiabetic drug acarbose. Nature 600, 110–115 (2021).
doi: 10.1038/s41586-021-04091-0; pmid: 34819672
7.
8. R. Legrand et al., Commensal Hafnia alvei strain reduces food
intake and fat mass in obese mice-a new potential probiotic for
appetite and body weight management. Int. J. Obes. 44,
1041–1051 (2020). doi: 10.1038/s41366-019-0515-9;
pmid: 31911661
9. P. Déchelotte et al., The probiotic strain H. alvei HA4597®
improves weight loss in overweight subjects under moderate
hypocaloric diet: A proof-of-concept, multicenter randomized,
double-blind placebo-controlled study. Nutrients 13, 1902 (2021).
doi: 10.3390/nu13061902; pmid: 34205871
10. V. Maini Rekdal, E. N. Bess, J. E. Bisanz, P. J. Turnbaugh,
E. P. Balskus, Discovery and inhibition of an interspecies gut
bacterial pathway for Levodopa metabolism. Science 364,
eaau6323 (2019). doi: 10.1126/science.aau6323; pmid: 31196984
11. M. P. O’Donnell, B. W. Fox, P. H. Chao, F. C. Schroeder, P. Sengupta,
A neurotransmitter produced by gut bacteria modulates host
sensory behaviour. Nature 583, 415–420 (2020). doi: 10.1038/
s41586-020-2395-5; pmid: 32555456
12. N. Pernigoni et al., Commensal bacteria promote endocrine
resistance in prostate cancer through androgen biosynthesis.
Science 374, 216–224 (2021). doi: 10.1126/science.abf8403;
pmid: 34618582
13. B. Javdan et al., Personalized mapping of drug metabolism by
the human gut microbiome. Cell 181, 1661–1679.e22 (2020).
doi: 10.1016/j.cell.2020.05.001; pmid: 32526207
14. E. E. Mulvihill, D. J. Drucker, Pharmacology, physiology, and
mechanisms of action of dipeptidyl peptidase-4 inhibitors.
Endocr. Rev. 35, 992–1019 (2014). doi: 10.1210/er.2014-1035;
pmid: 25216328
15. F. M. Gribble, F. Reimann, Metabolic messengers: Glucagon-
like peptide 1. Nat. Metab. 3, 142–148 (2021). doi: 10.1038/
s42255-020-00327-x; pmid: 33432200
16. G. Cantini, A. Di Franco, E. Mannucci, M. Luconi, Is cleaved
glucagon-like peptide 1 really inactive? Effects of GLP-1(9-36)
on human adipose stem cells. Mol. Cell. Endocrinol. 439, 10–15
(2017). doi: 10.1016/j.mce.2016.10.013; pmid: 27746194
17. Y. Ohara-Nemoto et al., Degradation of incretins and
modulation of blood glucose levels by periodontopathic
bacterial dipeptidyl peptidase 4. Infect. Immun. 85, e00277-17
(2017). doi: 10.1128/IAI.00277-17; pmid: 28630069
18. Y. Ohara-Nemoto et al., Distribution of dipeptidyl peptidase
(DPP) 4, DPP5, DPP7 and DPP11 in human oral microbiota-
potent biomarkers indicating presence of periodontopathic
bacteria. FEMS Microbiol. Lett. 365, fny221 (2018).
doi: 10.1093/femsle/fny221; pmid: 30203018
19. D. Rea et al., Crystal structure of Porphyromonas gingivalis
dipeptidyl peptidase 4 and structure-activity relationships
based on inhibitor profiling. Eur. J. Med. Chem. 139, 482–491
(2017). doi: 10.1016/j.ejmech.2017.08.024; pmid: 28826083
20. S. Roppongi et al., Crystal structures of a bacterial dipeptidyl
peptidase IV reveal a novel substrate recognition mechanism
distinct from that of mammalian orthologues. Sci. Rep. 8, 2714
(2018). doi: 10.1038/s41598-018-21056-y; pmid: 29426867
21. M. Olivares et al., The potential role of the dipeptidyl
peptidase-4-like activity from the gut microbiota on the host
health. Front. Microbiol. 9, 1900 (2018). doi: 10.3389/
fmicb.2018.01900; pmid: 30186247
22. A. Aranda-Díaz et al., Establishment and characterization of
stable, diverse, fecal-derived in vitro microbial communities
that model the intestinal microbiota. Cell Host Microbe 30,
260–272.e5 (2022). doi: 10.1016/j.chom.2021.12.008;
pmid: 35051349
23. C. D. Robinson, J. M. Auchtung, J. Collins, R. A. Britton,
Epidemic Clostridium difficile strains demonstrate increased
competitive fitness compared to nonepidemic isolates.
Infect. Immun. 82, 2815–2825 (2014). doi: 10.1128/
IAI.01524-14; pmid: 24733099
aluminium across Caco-2 cell monolayers. Chem. Biol. Interact. 181,
272–278 (2009). doi: 10.1016/j.cbi.2009.06.016; pmid: 19576870
34. A. Jackson et al., Intestinal barrier dysfunction in the absence of
systemic inflammation fails to exacerbate motor dysfunction and
brain pathology in a mouse model of Parkinson’s disease.
Front. Neurol. 13, 882628 (2022). doi: 10.3389/
fneur.2022.882628; pmid: 35665034
35. C. O. Elson, R. B. Sartor, G. S. Tennyson, R. H. Riddell,
Experimental models of inflammatory bowel disease.
Gastroenterology 109, 1344–1367 (1995). doi: 10.1016/0016-
5085(95)90599-5; pmid: 7557106
36. C. A. Thaiss et al., Hyperglycemia drives intestinal barrier
dysfunction and risk for enteric infection. Science 359,
1376–1383 (2018). doi: 10.1126/science.aar3318;
pmid: 29519916
37. L. Zhao et al., Elevated levels of circulating short-chain fatty
acids and bile acids in type 2 diabetes are linked to gut barrier
disruption and disordered gut microbiota. Diabetes Res. Clin.
Pract. 169, 108418 (2020). doi: 10.1016/j.diabres.2020.108418;
pmid: 32891692
38. K. T. Konstantinidis, A. Ramette, J. M. Tiedje, The bacterial
species definition in the genomic era. Philos. Trans. R. Soc.
Lond. B Biol. Sci. 361, 1929–1940 (2006). doi: 10.1098/
rstb.2006.1920; pmid: 17062412
39. N. Li et al., Recent progress of the development of dipeptidyl
peptidase-4 inhibitors for the treatment of type 2 diabetes
mellitus. Eur. J. Med. Chem. 151, 145–157 (2018). doi: 10.1016/
j.ejmech.2018.03.041; pmid: 29609120
40. Y. Nakajima et al., Dipeptidyl aminopeptidase IV from
Stenotrophomonas maltophilia exhibits activity against a substrate
containing a 4-hydroxyproline residue. J. Bacteriol. 190, 7819–7829
(2008). doi: 10.1128/JB.02010-07; pmid: 18820015
41. D. Kim et al., (2R)-4-oxo-4-[3-(trifluoromethyl)-5,6-dihydro
[1,2,4]triazolo[4,3-a]pyrazin-7(8H)-yl]-1-(2,4,5-trifluorophenyl)
butan-2-amine: A potent, orally active dipeptidyl peptidase IV
inhibitor for the treatment of type 2 diabetes. J. Med. Chem.
48, 141–151 (2005). doi: 10.1021/jm0493156; pmid: 15634008
42. G. Wang, “Sitagliptin for the treatment of type 2 diabetes”
(www.ClinicalTrials.gov identifier NCT04495881, 2020);
https://clinicaltrials.gov/study/NCT04495881.
43. Chinese Medical Association Diabetes Branch, Guideline for the
prevention and treatment of type 2 diabetes mellitus in China
(2020 edition). Chinese Journal of Endocrinology and
Metabolism. 37, 311–398 (2021). doi: 10.3760/cma.j.cn311282-
20210304-00142
24. J. M. Auchtung, C. D. Robinson, R. A. Britton, Cultivation of
44. T. R. Pieber et al., Efficacy and safety of oral semaglutide with
stable, reproducible microbial communities from different fecal
donors using minibioreactor arrays (MBRAs). Microbiome 3, 42
(2015). doi: 10.1186/s40168-015-0106-5; pmid: 26419531
25. L. Maier et al., Extensive impact of non-antibiotic drugs on
human gut bacteria. Nature 555, 623–628 (2018).
doi: 10.1038/nature25979; pmid: 29555994
26. S. Günther et al., SuperTarget and Matador: Resources for
exploring drug-target relationships. Nucleic Acids Res. 36,
D919–D922 (2008). doi: 10.1093/nar/gkm862; pmid: 17942422
27. R. E. Amori, J. Lau, A. G. Pittas, Efficacy and safety of incretin
therapy in type 2 diabetes: Systematic review and meta-
analysis. JAMA 298, 194–206 (2007). doi: 10.1001/
jama.298.2.194; pmid: 17622601
28. American Diabetes Association, Pharmacologic approaches to
glycemic treatment: Standards of medical care in diabetes-
2018. Diabetes Care 41, S73–S85 (2018). doi: 10.2337/dc18-
S008; pmid: 29222379
29. E. E. Mulvihill et al., Cellular sites and mechanisms linking
reduction of dipeptidyl peptidase-4 activity to control of
incretin hormone action and glucose homeostasis. Cell Metab.
25, 152–165 (2017). doi: 10.1016/j.cmet.2016.10.007;
pmid: 27839908
30. P. D. Cani et al., Changes in gut microbiota control metabolic
endotoxemia-induced inflammation in high-fat diet-induced
obesity and diabetes in mice. Diabetes 57, 1470–1481 (2008).
doi: 10.2337/db07-1403; pmid: 18305141
31. H. Gao et al., Accumulation of microbial DNAs promotes to
islet inflammation and b cell abnormalities in obesity in mice.
Nat. Commun. 13, 565 (2022). doi: 10.1038/s41467-022-
28239-2; pmid: 35091566
32. T. Suzuki, H. Hara, Dietary fat and bile juice, but not obesity,
are responsible for the increase in small intestinal permeability
induced through the suppression of tight junction protein
expression in LETO and OLETF rats. Nutr. Metab. (Lond.) 7, 19
(2010). doi: 10.1186/1743-7075-7-19; pmid: 20222989
33. B. Aspenström-Fagerlund, B. Sundström, J. Tallkvist, N. G. Ilbäck,
A. W. Glynn, Fatty acids increase paracellular absorption of
flexible dose adjustment versus sitagliptin in type 2 diabetes
(PIONEER 7): A multicentre, open-label, randomised, phase 3a
trial. Lancet Diabetes Endocrinol. 7, 528–539 (2019).
doi: 10.1016/S2213-8587(19)30194-9; pmid: 31189520
45. R. E. Pratley et al., Liraglutide versus sitagliptin for patients
with type 2 diabetes who did not have adequate glycaemic
control with metformin: A 26-week, randomised, parallel-group,
open-label trial. Lancet 375, 1447–1456 (2010). doi: 10.1016/
S0140-6736(10)60307-8; pmid: 20417856
46. R. E. Pratley et al., Efficacy and safety of switching from the DPP-4
inhibitor sitagliptin to the human GLP-1 analog liraglutide after
52 weeks in metformin-treated patients with type 2 diabetes: A
randomized, open-label trial. Diabetes Care 35, 1986–1993 (2012).
doi: 10.2337/dc11-2113; pmid: 22851600
47. P. Aschner et al., Effect of the dipeptidyl peptidase-4 inhibitor
sitagliptin as monotherapy on glycemic control in patients
with type 2 diabetes. Diabetes Care 29, 2632–2637 (2006).
doi: 10.2337/dc06-0703; pmid: 17130196
48. A. J. Hryckowian et al., Bacteroides thetaiotaomicron-infecting
bacteriophage isolates inform sequence-based host range
predictions. Cell Host Microbe 28, 371–379.e5 (2020).
doi: 10.1016/j.chom.2020.06.011; pmid: 32652063
49. N. T. Porter et al., Phase-variable capsular polysaccharides and
lipoproteins modify bacteriophage susceptibility in Bacteroides
thetaiotaomicron. Nat. Microbiol. 5, 1170–1181 (2020).
doi: 10.1038/s41564-020-0746-5; pmid: 32601452
50. H. Yue, Y. Li, M. Yang, C. Mao, T7 phage as an emerging
nanobiomaterial with genetically tunable target specificity.
Adv. Sci. (Weinh.) 9, e2103645 (2022). doi: 10.1002/
advs.202103645; pmid: 34914854
51. X. Pan, C. Hu, F. Zeng, S. Zhang, J. Xu, Isolation and identification
of alkaloids form Menispermum dauricum growing in Xianning.
Zhong Yao Cai 21, 456–458 (1998). pmid: 12569837
52. A. Ignacio et al., Small intestinal resident eosinophils maintain
gut homeostasis following microbial colonization. Immunity 55,
1250–1267.e12 (2022). doi: 10.1016/j.immuni.2022.05.014;
pmid: 35709757
Wang et al., Science 381, eadd5787 (2023)
4 August 2023
12 of 13
RES EARCH | R E S E A R C H A R T I C L E
53. R. Zhao et al., Aspirin reduces colorectal tumor development in
mice and gut microbes reduce its bioavailability and
chemopreventive effects. Gastroenterology 159, 969–983.e4
(2020). doi: 10.1053/j.gastro.2020.05.004; pmid: 32387495
54. C. F. Deacon, Dipeptidyl peptidase 4 inhibitors in the treatment
of type 2 diabetes mellitus. Nat. Rev. Endocrinol. 16, 642–653
(2020). doi: 10.1038/s41574-020-0399-8; pmid: 32929230
55. K. Martinez-Guryn, V. Leone, E. B. Chang, Regional diversity of the
gastrointestinal microbiome. Cell Host Microbe 26, 314–324
(2019). doi: 10.1016/j.chom.2019.08.011; pmid: 31513770
56. D. Shalon et al., Profiling the human intestinal environment
under physiological conditions. Nature 617, 581–591 (2023).
doi: 10.1038/s41586-023-05989-7; pmid: 37165188
57. S. L. LeValley, C. Tomaro-Duchesneau, R. A. Britton,
Degradation of the incretin hormone glucagon-like peptide-
1 (GLP-1) by Enterococcus faecalis metalloprotease GelE.
MSphere 5, e00585-19 (2020). doi: 10.1128/mSphere.00585-
19; pmid: 32051237
58. C. Tomaro-Duchesneau et al., Discovery of a bacterial peptide as
a modulator of GLP-1 and metabolic disease. Sci. Rep. 10, 4922
(2020). doi: 10.1038/s41598-020-61112-0; pmid: 32188864
59. J. Breton et al., Gut commensal E. coli proteins activate host
satiety pathways following nutrient-induced bacterial growth.
Cell Metab. 23, 324–334 (2016). doi: 10.1016/
j.cmet.2015.10.017; pmid: 26621107
ACKN OWLED GMEN TS
We thank the x-ray crystallography platform of Tsinghua University
Technology Center for Protein Sciences and S. Fan for providing
support and J. Ren (State Key Laboratory of Mycology, Institute
of Microbiology, Chinese Academy of Sciences) for help in obtaining the
NMR data. Schematic diagrams were generated by Biorender.com.
Author contributions: C.J. conceived of the original concept. K.W.,
Z.Z., J.H., J.L., F.G., Y.D., M.L., Q.N., J.L., Y.Z., L.S., X.L., Q.Z., C.Ye.,
C.Yun, Y.Z., J.W., R.B., and Y.P. performed the experiments and
analyzed the data. C.J., J.Q., F.J.G., X.L., and G.W. supervised the
study. K.W., Z.Z., J.H., and C.J. wrote the manuscript with input from
all authors. All authors edited the manuscript and approved the final
manuscript. Funding: This work was supported by the National
Natural Science Foundation of the Peoples’ Republic of China (grant
nos. 82288102, 82130022, and 31925021), the National Key Research
and Development Program of China (grant nos. 2018YFA0800700,
2022YFA0806400, and 2022YFC3401500), the National Natural
Science Foundation of the Peoples’ Republic of China (grant nos.
91857115, 92149306, 82071658, 81921001, 22193073, 92253305, and
32000813), the National Cancer Institute Intramural Research Program
(grant no. ZIABC005708), the Beijing Nova Program (grant no.
Z201100006820010), the Beijing Outstanding Young Scientist Program
(grant no. BJJWZYJH01201910001001), and the National Postdoctoral
Program for Innovative Talents (grant no. BX20190019). Competing
interests: The authors declare no competing interests. Data and
materials availability: All data are available in the main text or
supplementary materials. Metagenome sequencing data, 16S rRNA
gene amplicon sequencing data, phylogenesis data, high-throughput
drug screening data, clinical data, supplementary methods for enzyme
activity detection, and source data have been deposited into the Dryad
repository. The National Center for Biotechnology Information public
reference database was used in this study. The structure data have been
deposited to the Protein Data Bank with accession codes 7Y4F (apo
btDPP4), 7Y4G (sitagliptin-bound btDPP4), and 8HAY (Dau-d4-bound
btDPP4). License information: Copyright © 2023 the authors, some
rights reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government works.
https://www.science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add5787
Materials and Methods
Figs. S1 to S21
References (60–86)
Tables S1 to S5
MDAR Reproducibility Checklist
Submitted 21 June 2022; resubmitted 1 February 2023
Accepted 14 June 2023
10.1126/science.add5787
Wang et al., Science 381, eadd5787 (2023)
4 August 2023
13 of 13
|
10.1126_science.add7347
|
RES EARCH
CRISPR
RNA-triggered protein cleavage and cell growth
arrest by the type III-E CRISPR nuclease-protease
Kazuki Kato1†, Sae Okazaki1†, Cian Schmitt-Ulms2†, Kaiyi Jiang2,3†, Wenyuan Zhou2,
Junichiro Ishikawa1, Yukari Isayama1, Shungo Adachi4, Tomohiro Nishizawa5, Kira S. Makarova6,
Eugene V. Koonin6, Omar O. Abudayyeh2*, Jonathan S. Gootenberg2*, Hiroshi Nishimasu1,7,8,9*
The type III-E CRISPR–Cas7-11 effector binds a CRISPR RNA (crRNA) and the putative protease Csx29
and catalyzes crRNA-guided RNA cleavage. We report cryo–electron microscopy structures of the
Cas7-11–crRNA–Csx29 complex with and without target RNA (tgRNA), and demonstrate that tgRNA
binding induces conformational changes in Csx29. Biochemical experiments revealed tgRNA-dependent
cleavage of the accessory protein Csx30 by Csx29. Reconstitution of the system in bacteria showed
that Csx30 cleavage yields toxic protein fragments that cause growth arrest, which is regulated by
Csx31. Csx30 binds Csx31 and the associated sigma factor RpoE (RNA polymerase, extracytoplasmic E),
suggesting that Csx30-mediated RpoE inhibition modulates the cellular response to infection. We
engineered the Cas7-11–Csx29–Csx30 system for programmable RNA sensing in mammalian cells.
Overall, the Cas7-11–Csx29 effector is an RNA-dependent nuclease-protease.
P rokaryotic CRISPR-Cas systems provide
adaptive immunity against foreign nu-
cleic acids, including phages and mobile
genetic elements, through diverse mech-
anisms of programmed nucleic acid
cleavage. CRISPR-Cas systems are divided
into two classes based on the number of com-
ponents in the effector complexes responsible
for defense through the cleavage of invading
nucleic acids programmed by a CRISPR RNA
(crRNA) guide (1, 2). In class 1 systems, which
encompass types I, III, and IV, target nucleic
acids are degraded by multiprotein effector
complexes, whereas the effectors of class 2
systems (types II, V, and VI) are single-protein,
multidomain Cas proteins (Cas9, Cas12, and
Cas13, respectively). CRISPR-Cas systems also
deploy a wide array of accessory proteins that
enhance and modulate the antiviral activity of
the primary effector nuclease (3–7).
Unlike typical class 1 effectors, the type III-E
effector Cas7-11 (also known as gRAMP) is a
1Structural Biology Division, Research Center for Advanced
Science and Technology, The University of Tokyo, 4-6-1
Komaba, Meguro-ku, Tokyo 153-8904, Japan. 2McGovern
Institute for Brain Research at MIT, Massachusetts Institute of
Technology, Cambridge, MA 02139, USA. 3Department of
Biological Engineering, Massachusetts Institute of Technology,
Cambridge, MA 02139, USA. 4Cellular and Molecular
Biotechnology Research Institute, National Institute of Advanced
Industrial Science and Technology, 2-4-7 Aomi, Koto-ku, Tokyo
135-0064, Japan. 5Graduate School of Medical Life Science,
Yokohama City University, 22-2 Seto, Kanazawa-ku, Yokohama,
Kanagawa 236-0027, Japan. 6National Center for Biotechnology
Information, National Library of Medicine, National Institutes
of Health, Bethesda, MD 20894, USA. 7Department of Chemistry
and Biotechnology, Graduate School of Engineering, The
University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656,
Japan. 8Department of Biological Sciences, Graduate School of
Science, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo
113-0033, Japan. 9Inamori Research Institute for Science, 620
Suiginya-cho, Shimogyo-ku, Kyoto 600-8411, Japan.
*Corresponding author. Email: omar@abudayyeh.science
(O.O.A.); jgoot@mit.edu (J.S.G.); nisimasu@g.ecc.u-tokyo.ac.jp (H.N.)
†These authors contributed equally to this work.
single-protein, multidomain effector that con-
sists of four Cas7 domains (Cas7.1 to Cas7.4)
and a Cas11 domain (8, 9). Cas7-11 associates
with a crRNA and cleaves complementary
single-stranded RNA (ssRNA) targets at two
defined positions using the Cas7.2 and Cas7.3
domains. Whereas the type VI effector Cas13
displays promiscuous RNase activity (10, 11),
Cas7-11 exhibits precise, guide RNA–dependent
tgRNA cleavage activity in human cells and has
been used as a new RNA-targeting tool with
high specificity and low cell toxicity (8, 9). The
type III-E locus encodes multiple conserved
accessory proteins, including Csx29 (a caspase-
like putative protease with fused TPR and
CHAT domains), Csx30 and Csx31 (proteins
with unknown functions), and RpoE (an alter-
native sigma factor). Cas7-11 forms a complex
with Csx29 (8, 9), suggesting a potential mech-
anism of RNA-guided protease activity for
antiviral immunity. Recently, we reported the
cryo–electron microscopy (cryo-EM) structure
of Desulfonema ishimotonii Cas7-11 complexed
with its cognate crRNA and tgRNA (12), pro-
viding mechanistic insights into pre-crRNA
processing and tgRNA cleavage. Nonetheless,
the means by which Cas7-11 cooperates with
the other proteins encoded in the type III-E
locus (Csx29, Csx30, Csx31, and RpoE) and
binds Csx29 to potentially activate its prote-
ase activity remain unknown.
Results
Structures of Cas7-11 complexed with Csx29
We sought to solve the structure of the Cas7-
11–Csx29 effector complex to gain insights into
its action mechanism. We coexpressed the cat-
alytically inactive D. ishimotonii Cas7-11 mutant
(referred to as Cas7-11 for simplicity), with the
D429A (Cas7.2) and D654A (Cas7.3) mutations
introduced to prevent tgRNA cleavage by Cas7-
11, together with Csx29 and a crRNA tran-
scribed from a CRISPR array containing two
repeat-spacer units and then purified the Cas7-
11–crRNA–Csx29 complex. We determined cryo-
EM structures of this Cas7-11–crRNA–Csx29
complex with and without a tgRNA at 2.5-
and 2.8-Å resolutions, respectively (Fig. 1, A
to D, figs. S1 to S3, table S1, and movie S1). In
both structures, Cas7-11 adopts a modular
architecture consisting of four Cas7 domains
(Cas7.1 to Cas7.4) with a zinc finger (ZF) motif,
a Cas11 domain, an insertion (INS) domain
within the Cas7.4 domain, a C-terminal ex-
tension domain, and four interdomain linkers
(L1 to L4) (Fig. 1, C and D), as in the Csx29-
unbound Cas7-11–crRNA–tgRNA structure (12)
(fig. S4).
The 15-nucleotide (nt) 5′ tag region [U(−15)–
C(−1)] in the 38-nt crRNA [U(−15)–A23] is
anchored by the Cas7.1 and Cas7.2 domains
in both structures (Fig. 1, C and D, and figs.
S4, B and C, and S5A), as in the Csx29-unbound
Cas7-11–crRNA–tgRNA structure (12) (fig. S4A).
The 23-nt crRNA spacer region is recognized
by the Cas7.2 to Cas7.4 domains in the tgRNA-
free structure (Fig. 1C and fig. S4B), whereas
the crRNA spacer region (C1 to A23, except for
U4 and C10) hybridizes with the tgRNA (G1 to
U23, except for A4 and G10) to form a guide-
target duplex in the tgRNA-bound structure
(Fig. 1D and fig. S4C). A(−3) in the 5′ tag (6 nt
downstream of the first flipped-out spacer nu-
cleotide) is flipped out because of its interac-
tion with the thumb-like b-hairpin in the Cas7.1
domain (fig. S5B), similar to the equivalent
nucleotide C(−1) in the type III-A Csm effector
complex (13, 14) (fig. S5C). Nonetheless, unlike
the Csm complex, A(−2) and C(−1), which are
located upstream of A(−3), cannot base pair
with the tgRNA in Cas7-11 because of the pres-
ence of the L2 linker (fig. S5B), explaining the
distinct RNA cleavage patterns between the
Cas7-11 and Csm effector complexes. In the pres-
ent structures, the peripheral region (residues
1043 to 1124) of the INS domain was less re-
solved in the density map, probably because of
its flexibility (fig. S3), and was excluded from
the final models (fig. S4).
Csx29 structure
Csx29 consists of a tetratricopeptide repeat
(TRP) domain (residues 1 to 422) and a Cas-
pase HetF Associated with TPRs (CHAT) pro-
tease domain (residues 423 to 751) (Fig. 2A). The
TRP domain can be divided into an N-terminal
domain (NTD) (residues 1 to 64), seven TPR
units (TPR1 to TPR7), and a central region re-
ferred to as the activation region (AR). The
NTD adopts a three-helix bundle structure and
interacts with the Cas7.4 domain of Cas7-11
(Fig. 2B). In Csx29, each TPR unit contains two
a helices, similar to canonical TPR-containing
proteins in which the TPRs interact with their
protein targets (15). Correspondingly, TPR1 and
TPR2 of Csx29 interact with the L2 linker of
Kato et al., Science 378, 882–889 (2022)
25 November 2022
1 of 8
RES EARCH | R E S E A R C H A R T I C L E
A
1
238
260
367
402
582
625
824
891
979
1294
1507
1601
Cas7-11
Cas7.1
L1
Cas11
L2
Cas7.2
L3
Cas7.3
L4
SNI
Cas7.4
ETC
1
65
AR1
AR2
423
552
751
Csx29
DTN
TPR1–6
7
DPP
APD
TPR domain
CHAT protease domain
B
tgRNA
crRNA
PFS
−6
GC
G
−1
A A C
1
G
3
5
U
−15
U G
A U G U C A C G G
−10
−5
A C C
1
−1
A
−3
C
Cas7-11–crRNA–Csx29
5 tag
Cas7.1
crRNA
5
TPR
Cas11
NTD
INS
A
U
3
A
U
3
4
A
U
4
5
C
G
5
A
A
C
U
U
G
9
A
U
9
10
G
C
10
11
A
U
11
A
G
C
U
C
G
15
U
A
15
Spacer
G
U
A
C
C
A
U
G
20
C
G
20
27
23
U U A G G
5
C
A
G
U
3
A
23
Csx29
PPD
TPR
NTD
PPD
APD
L3
Cas7.2
Cas7.3
1042
1125
3
90°
APD
CTE
L2
5
L4
3
Cas7.1
L3
Cas11
Cas7.4
INS
D
Cas7-11–crRNA–Csx29–tgRNA
L4
Cas7.4
crRNA
Cas7.1
5
TPR
Cas11
INS
NTD
PPD
Cas7-11
Csx29
TPR
NTD
PPD
L3
Cas7.2
Cas7.3
1042
1125
3
90°
tgRNA
CTE
L2
5
L4
5
3
L4
Cas7.4
Cas7-11
Cas7.1
L3
Cas11
Cas7.4
INS
Fig. 1. Cryo-EM structures of the Cas7-11–crRNA–Csx29 complexes with
and without the tgRNA. (A) Domain structures of Cas7-11 and Csx29.
(B) Nucleotide sequences of the crRNA and its tgRNA. The pre-crRNA processing site
and the tgRNA cleavage sites are indicated by cyan and yellow/green triangles,
respectively. Disordered nucleotides are indicated by dotted lines. (C and D) Overall
structures of Cas7-11–crRNA–Csx29 (C) and Cas7-11–crRNA–Csx29–tgRNA (D).
The bound zinc ions are shown as orange spheres. The disordered L1 and L2 linkers
are not shown for clarity. PPD, pseudo-protease domain.
Cas7-11 (Fig. 2B). The CHAT domain of Csx29
consists of a central 11-stranded mixed b-sheet
and flanking a-helices, and can be divided into
a pseudo-protease domain (residues 423 to 551)
and an active-protease domain (APD) (residues
552 to 751) with the conserved putative catalytic
residues H615 and C658 (Fig. 2A). A Dali search
(16) confirmed that the CHAT domain of Csx29
structurally resembles caspase-like cysteine pro-
teases such as human separase (17) (fig. S6). In
the Cas7-11–crRNA–Csx29 structure, the AR con-
sists of two regions, AR1 (a b-hairpin between
TPR6 and TPR7) and AR2 (a b-strand and a
helix-loop-helix after TPR7), and interacts with
TPR1–TPR6 and the APD (Fig. 2A).
Interactions between Cas7-11 and Csx29
In the Cas7-11–crRNA–Csx29 structure, Csx29
interacts with multiple regions of Cas7-11 (Fig.
2B and fig. S7A). The L2 linker (residues 367 to
401) and a-helical insertion (residues 1313 to
1341) in the Cas7.4 ZF motif, which are dis-
ordered in the Csx29-unbound Cas7-11 structure
(12) (fig. S7B), are ordered and form interactions
with Csx29 in the Cas7-11–crRNA–Csx29 struc-
ture (Fig. 2B and fig. S7, C and D). This a-helical
insertion in the ZF motif is unique to Cas7.4
(fig. S7C). The NTD of Csx29 mainly interacts
with the Cas7.4 domain of Cas7-11 (Fig. 2B and
figs. S7A and S8A). In addition, the Cas7.2
thumb-like b-hairpin and the L2/L4 linkers
reinforce the binding to the Csx29 NTD (Fig.
2B and fig. S8, B and C).
TPR1 and TPR2 interact with Cas7.3 (ZF) and
L2 of Cas7-11 (Fig. 2B and figs. S7A and S8D).
TPR2 also interacts with Cas7.1 (thumb-like
Kato et al., Science 378, 882–889 (2022)
25 November 2022
2 of 8
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Interaction between Cas7-11 and Csx29. (A) Structure of Csx29 in the
Cas7-11–crRNA–Csx29 complex. (B) Interface between Cas7-11 and Csx29 in
the Cas7-11–crRNA–Csx29 complex. The Cas11 and INS domains are omitted
for clarity. (C) Location of the Csx29 active site. The catalytic residue H615 of the
Csx29 protease is colored red. (D and E) Interfaces between Cas7-11 and Csx29 in
Cas7-11–crRNA–Csx29 (D) and Cas7-11–crRNA–Csx29–tgRNA (E). Csx29 is shown as
a surface representation, except for the AR, which is shown as a ribbon representation.
The AR and APD are disordered in the Cas7-11–crRNA–Csx29–tgRNA structure in (E).
b-hairpin) and Cas7.2 (ZF) (Fig. 2B). TPR3 to
TPR7 do not contact Cas7-11. The CHAT pro-
tease domain of Csx29 interacts with Cas7.1
(thumb-like b-hairpin), Cas7.2 (ZF), Cas7.3 (ZF),
and L2 of Cas7-11 (Fig. 2B and fig. S7A). The
protease active site of Csx29 is near the Cas7.2
domain of Cas7-11 (Fig. 2C), suggesting lim-
ited accessibility for the peptide substrate in
this conformation. Furthermore, in contrast
to the separase–securin structure (17), the side
chain of the catalytic residue C658 is buried
inside the CHAT domain in the present struc-
ture (fig. S6), indicating that a structural re-
arrangement of C658 would be required for
the substrate cleavage. These observations sug-
gest that the Cas7-11–crRNA–Csx29 structure
represents the inactive state of the Csx29 puta-
tive protease.
Kato et al., Science 378, 882–889 (2022)
25 November 2022
3 of 8
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. tgRNA-triggered Csx30 cleavage by Csx29. (A) Schematic of the RNA-
triggered Csx30 cleavage by the Cas7-11–crRNA–Csx29 complex. TR, tgRNA
without a PFS; CTR, cognate tgRNA with a nonmatching PFS; NTR, noncognate
tgRNA with a matching PFS. (B) RNA-triggered Csx30 cleavage by the Cas7-11–
crRNA–Csx29 complex. The Cas7-11–crRNA–Csx29 complex was incubated with
Csx30 at 37°C for 10 min in the presence or absence of the tgRNA (CTR). The
wild-type (W) and catalytically deactivated (D) versions of Cas7-11 and Csx29
were used. (C) Effects of the complementarity between the crRNA 5′ tag and tgRNA
PFS on the Csx30 cleavage. The dCas7-11–crRNA–Csx29 complex was incubated
with Csx30 at 37°C for 5, 10, or 15 min in the presence of the tgRNA (TR, CTR, or
NTR). (D) Proteolytic cleavage site in Csx30. The Csx30 site cleaved by Csx29
is indicated by a triangle. The Csx30 structure was predicted using AlphaFold2
(18), and the Ca atoms of M427 and K428 at the cleavage site are indicated
by spheres. (E) Csx29-mediated cleavage of the Csx30 mutants. The dCas7-11–
crRNA–Csx29 complex was incubated with the Csx30 mutants at 37°C for
10 min, in the presence or absence of the tgRNA (CTR). In (B), (C), and (E), the
proteins were analyzed by SDS-polyacrylamide gel electrophoresis, and the gel
was stained with Coomassie brilliant blue.
tgRNA binding–induced structural change in the
Cas7-11–Csx29 complex
Comparison of the Cas7-11–crRNA–Csx29 struc-
tures with and without the tgRNA revealed dif-
ferent Csx29 conformations (Fig. 2, D and E,
and movie S2). In the tgRNA-free structure,
TPR1 and TPR2 of Csx29 interact with Cas7.3
and Cas7.1/Cas7.2 of Cas7-11, respectively (Fig. 2,
B and D, and fig. S9A). By contrast, in the
tgRNA-bound structure, TPR1 and TPR2 of
Csx29 are farther away from Cas7-11 and do
not interact with Cas7.1–Cas7.3 of Cas7-11
because of the binding of the tgRNA 3′ region
between Cas7-11 and Csx29 (Fig. 2E and fig.
S9B). In the 6-nt protospacer flanking sequence
(PFS) of the tgRNA, only C(−1) and A(−2) are
well resolved in the density map and interact
with Cas7-11 (L2/Cas7.3) and Csx29 (TPR1/TPR2)
Kato et al., Science 378, 882–889 (2022)
25 November 2022
4 of 8
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Effects of Csx30 and Csx31 on bacterial cell growth. (A) Schematic
of bacterial growth assays for studying the Csx30 and Csx31 functions. (B and
C) Growth curves (B) and end-point analyses (C) of E. coli expressing full-
length Csx30, the N-terminal fragment (residues 1 to 427) of Csx30 (Csx30-1), or
the C-terminal fragment (residues 428 to 565) of Csx30 (Csx30-2). (D and
E) Growth curves (D) and end-point analyses (E) of E. coli expressing Csx30-1,
full-length Csx30 and Csx31, or Csx30-1 and Csx31. In (B) to (E), growth was
compared between induced and uninduced expression conditions. In (C) and
(E), significance was calculated by two-tailed Student’s t test (****P < 0.0001; n.s.,
not significant). Data are shown as mean ± SEM (n = 3). (F) Heatmap comparing
the survival percentages of bacteria expressing Csx30-1, Csx30-2, full-length Csx30
and Csx31, full-length Csx30 alone, Csx30-1 and Csx31, or Csx30-2 and Csx31,
cultured at three different temperatures. Percent survival was calculated by the ratio
of optical density at 600 nm (OD600) of the bacterial culture under the induced
conditions over the OD600 for the noninduced conditions. Color scale shows percent
survival from 0 to 100%. (G) Confocal images of E. coli expressing EGFP alone,
EGFP-Csx30, or EGFP-Csx31 and unlabeled Csx30. White outlines indicate the
shapes of individual E. coli cells. (H) Schematic of the mammalian application of
the Cas7-11–Csx29–Csx30 degron reporter system for RNA sensing in live cells.
(I) Citrine fluorescence of HEK293FT cells transfected with either the Gluc target or
pUC19 control target in the presence of the Cas7-11–Csx29–Csx30 degron reporter.
Significance was calculated with two-tailed Student’s t test (****P < 0.0001).
Data are shown as mean ± SEM (n = 3). (J) RNA-triggered Csx30 reporter cleavage
in HEK293FT cells. The N-terminally FLAG-tagged citrine–Csx30–degron reporter
was transfected either with or without the Gluc target and with a targeting or
nontargeting (NT) guide. Forty-eight hours after transfection, total protein was
extracted from the transfected HEK293FT cells and analyzed by Western blot with an
anti-FLAG antibody.
Kato et al., Science 378, 882–889 (2022)
25 November 2022
5 of 8
RES EARCH | R E S E A R C H A R T I C L E
(fig. S9, B and C). The nucleobases C(−1) and
A(−2) stack with R375 (L2) and Y718 (Cas7.3),
respectively (fig. S9, B and C). In addition, the
phosphate groups between A(−3) and A(−2)
and between A(−2) and C(−1) interact with
R131 (TPR2) and R145 (TPR2), respectively
(fig. S9B). These interactions induce a kink
turn between A(−2) and C(−1) in the PFS, there-
by projecting the tgRNA nucleotides down-
stream of position −2 toward the AR of Csx29.
There are also structural differences in the
AR–APD of Csx29 between the RNA-free and
RNA-bound structures. Whereas the AR exten-
sively interacts with TPR1 to TPR5 and APD
in the tgRNA-free structure (Fig. 2D and fig.
S9A), the AR-APD of Csx29 is not resolved in
the density map in the tgRNA-bound structure
(Fig. 2E and fig. S3B). In addition, the pseudo-
protease domain of Csx29 in the tgRNA-bound
structure exhibits weaker density compared
with that in the tgRNA-free structure (fig. S3,
A and B). These observations suggest that
tgRNA binding increases the conformational
flexibility of the CHAT protease domain of
Csx29, and this conformational change releases
the steric block of the Csx29 active site, allow-
ing access to the substrate protein. Structural
comparisons between the two Cas7-11–Csx29
complexes suggest steric clash between the
tgRNA PFS and the Csx29 AR (fig. S9D), high-
lighting the importance of the PFS for the
tgRNA-induced conformational change in
Csx29. These structural observations indi-
cate that Csx29 is a tgRNA-triggered protease.
tgRNA-triggered Csx30 cleavage by Csx29
Given that Csx30 and Csx31 are encoded togeth-
er with Cas7-11 and Csx29 in the D. ishimotonii
CRISPR locus and are highly conserved among
the type III-E systems (8), we hypothesized that
Csx29 could target either Csx30 or Csx31. To
test this hypothesis, we attempted to prepare
the recombinant Csx30 and Csx31 proteins
and determine whether they are cleaved by
Csx29 in a tgRNA-dependent manner. Csx30
was purified as a soluble protein, whereas
Csx31 was expressed in an insoluble fraction.
We examined the in vitro cleavage of Csx30 by
Cas7-11–crRNA–Csx29 in the absence and pres-
ence of the tgRNA, and found that Cas7-11–
crRNA–Csx29 cleaved Csx30 into two fragments,
Csx30-1 (~50 kDa) and Csx30-2 (~15 kDa), only
in the presence of the tgRNA (Fig. 3, A and B).
The H615A/C658A mutations in Csx29 abol-
ished the Csx30 cleavage (Fig. 3B) but did not
affect the tgRNA cleavage by Cas7-11 (fig. S10A),
indicating the separable nuclease and protease
activities. Furthermore, the D429A/D654A cat-
alytic mutations in Cas7-11 abolished tgRNA
cleavage (fig. S10A), as previously observed
(12), and, unexpectedly, improved the Csx30
cleavage by Csx29 (Fig. 3B and fig. S10B). This
improvement in the proteolytic activity sug-
gests that the tgRNA dissociates from the ef-
fector complex after the Cas7-11–mediated
cleavage, and that the Csx29 protease is only
active when a tgRNA is bound to the Cas7-11–
Csx29 complex. Thus, Csx30 is cleaved by the
CHAT protease domain of Csx29 in a tgRNA-
dependent manner.
The base complementarity between the
crRNA 5′ tag and a tgRNA PFS regulates the
activities of the type III-A Csm effector com-
plex to avoid an autoimmune response in the
type III-A system (13, 14). Thus, we examined
the effects of the PFS in the tgRNA on Csx30
cleavage using a tgRNA without a PFS, a cog-
nate tgRNA with a nonmatching PFS, or a non-
cognate tgRNA with a matching PFS (Fig. 3A).
Csx30 was cleaved by the Cas7-11–Csx29 com-
plex efficiently in the presence of cognate
tgRNA with a nonmatching PFS, but not the
tgRNA without a PFS or the noncognate tgRNA
with a matching PFS (Fig. 3C), consistent with
our observation that a nonmatching PFS plays
a role in the structural changes and protease
activation of Csx29. These findings suggest a
potential mechanism for self-targeting avoid-
ance in the type III-E system, as in the type III-
A system.
The N-terminal analysis of the Csx30-2 frag-
ment showed that it begins with K428 (fig.
S11), demonstrating that Csx30 is cleaved by
Csx29 between M427 and K428 (Fig. 3D). Struc-
ture prediction using AlphaFold2 (18) indi-
cated that Csx30 consists of an NTD and a
C-terminal domain (CTD), which are connected
by a linker region. The NTD (residues 1 to 377)
contains two a-helical subdomains, whereas
the CTD (residues 418 to 565) comprises a core
b-barrel with flanking a helices (Fig. 3D). The
cleavage site between M427 and K428 is lo-
cated in a b-hairpin within the Csx30 CTD
(Fig. 3D). We examined the in vitro Csx29-
mediated cleavage of eight Csx30 mutants,
in which residues V425 to K431 were indi-
vidually replaced with alanine. The M427A and
G416A substitutions substantially and slightly
reduced the Csx30 cleavage, respectively,
whereas the other substitutions had almost no
effect (Fig. 3E). Accordingly, Csx29 seems to
primarily recognize M427 at the P1 site with-
in the AVGM|KKDK sequence in Csx30 and
cleaves Csx30 between M427 (P1) and K428
(P1′). Thus, the Cas7-11–Csx29 complex catalyzes
the tgRNA-triggered proteolysis of Csx30.
Effects of Csx30 and Csx31 on bacterial
cell growth
To explore the physiological relevance of the
Csx29-mediated Csx30 cleavage, we overex-
pressed the full-length Csx30 (referred to as
Csx30 for simplicity), the N-terminal fragment
of Csx30 (residues 1 to 427, Csx30-1), or the
C-terminal fragment of Csx30 (residues 428
to 565, Csx30-2) in Escherichia coli, and mon-
itored the cell growth (Fig. 4A). Overexpres-
sion of Csx30 substantially inhibited cell
growth compared with uninduced controls
(Fig. 4, B and C, and fig. S12, A and B). Over-
expression of Csx30-1 similarly caused pro-
nounced growth suppression, whereas Csx30-2
displayed only mild inhibition (Fig. 4, B and C,
and fig. S12, A and B), indicating that Csx30-1
is necessary and sufficient for the observed
growth effects of the full-length Csx30. Be-
cause the AlphaFold2 structural prediction
(18) suggested that Csx30 and Csx31 have op-
positely charged surfaces and could electro-
statically interact with each other (fig. S12C),
we also explored the effect of Csx31 on bac-
terial growth. Overexpression of Csx31 res-
cued the Csx30-mediated growth defect, but
could not completely eliminate the Csx30-1–
induced growth suppression (Fig. 4, D and E,
and fig. S12, B and D). These data indicate
that Csx31 interacts with Csx30 and regulates
Csx30-induced growth suppression, whereas
the generation of the Csx30-1 and Csx30-2
fragments by the Cas7-11–Csx29 protease in-
terferes with this regulation.
Interactions among Csx30, Csx31, and RpoE
The common co-occurrence of Cas7-11, Csx30,
Csx31, and the stress-associated sigma factor
RpoE (19) in type III-E CRISPR loci (2, 8) sug-
gests interplay among these four proteins,
such that the observed Csx30-induced growth
suppression might be caused by interactions
with endogenous E. coli RpoE (EcRpoE). Given
the involvement of EcRpoE in cellular heat
shock responses (20, 21), we hypothesized that
the growth defects might be more pronounced
at higher temperatures because of the inhibi-
tion of EcRpoE by Csx30 and Csx31, and tested
the effects of Csx30 and Csx31 in E. coli at
temperatures ranging from 30° to 42°C. In
agreement with our hypothesis, the growth
suppression by Csx30 was stronger at higher
temperatures across all tested combinations
(Fig. 4F), implicating EcRpoE in the growth
defects caused by the overexpression of Csx30
and Csx31.
To examine the direct interactions among
Csx30, Csx31, and D. ishimotonii RpoE (DiRpoE),
we coexpressed the three proteins in E. coli
and analyzed complex formation by gel-filtration
chromatography. Csx30, Csx31, and DiRpoE
eluted as a single peak (fig. S13A), indicating
that they form a stable complex. Like the iso-
lated Csx30, Csx30 in the Csx30–Csx31–DiRpoE
complex was cleaved by Cas7-11–Csx29, and
Csx30-1, Csx31, and DiRpoE co-eluted from the
column (fig. S13B), indicating that Csx30-1,
Csx31, and RpoE remain complexed after Csx30
cleavage, whereas Csx30-2 dissociates from
the complex. Consistently, structural predic-
tion implied that Csx30, Csx31, and DiRpoE
form a ternary complex in which the Csx30
NTD extensively interacts with DiRpoE (fig.
S13C). DiRpoE is structurally similar to EcRpoE
(fig. S13D), suggesting that the observed cell
Kato et al., Science 378, 882–889 (2022)
25 November 2022
6 of 8
RES EARCH | R E S E A R C H A R T I C L E
growth inhibition could be mediated through
Csx30–EcRpoE interactions, similar to the
mechanism of the anti-sigma factor RseA (22).
However, D. ishimotonii lacks RseA homologs
(23), so DiRpoE probably mediates the tran-
scriptional response through distinct mecha-
nisms that remain to be characterized. A Dali
search (16) revealed structural similarity be-
tween the Csx30 CTD and pore-forming pro-
teins in type IV secretion systems, such as
CagX (24) (fig. S13E). Given the reduced growth
effects of full-length Csx30 in E. coli compared
with Csx30-1, the Csx30 CTD might function
as a membrane anchor rather than as a pore-
forming protein, consistent with the role of
the membrane-localized RseA (19, 23). The
CTD and NTD of Csx30 are connected through
a flexible linker, suggesting that the Csx29-
mediated cleavage releases the N-terminal
fragment of Csx30 (Csx30-1) into the cyto-
plasm, thereby modulating gene expression
through RpoE suppression. Csx30 NTDs are
highly conserved (fig. S14), whereas Csx30
CTDs are divergent and can be divided into
seven distinct groups (fig. S15), of which two
are homologous protein domains found in
other contexts. One is an uncharacterized
DUF4384 family member that is often fused
to different proteases (25). The other CTD is
similar to the pilus assembly protein PilP,
which forms a periplasmic ring of bacterial
type IV pili (26). These observations high-
light the mechanistic diversity of the Csx30-
mediated RpoE interactions and programmed
gene expression modulation.
Localization of Csx30 and Csx31 in bacterial cells
To explore the growth suppression associated
with the expression of Csx30, the putative
membrane localization of the Csx30 CTD, and
the corresponding regulatory function of Csx31,
we imaged Csx30 and Csx31 by fusing bacterial
codon-optimized enhanced GFP (EGFP) to the
N termini of both proteins. Protein localization
in E. coli transformed with a plasmid express-
ing EGFP-Csx30, plasmids expressing EGFP-
Csx31 and unlabeled Csx30, or a plasmid
expressing EGFP alone was imaged. Both
labeled Csx30 alone and labeled Csx31 co-
expressed with Csx30 localized to individual
foci, whereas EGFP diffused throughout the
cells (Fig. 4G). These results suggest the direct
interaction between Csx30 and Csx31 through
colocalization at foci in bacterial cells before
Csx29-mediated Csx30 cleavage.
Engineering Csx29 and Csx30 for programmable
RNA sensing in mammalian cells
The programmable transcript-activated prote-
ase activity of the Cas7-11–Csx29–Csx30 system
could enable multiple applications in mam-
malian cells, including transcript sensing. We
codon-optimized Cas7-11, Csx29, and Csx30 for
mammalian cells and placed the Csx30 pro-
tein sequence between a citrine protein and a
dihydrofolate reductase (DHFR) degron (27),
which would eliminate citrine fluorescence
unless Csx30 was cleaved by Cas7-11–Csx29
because of the sequence-specific recognition
of a target sequence (Fig. 4H). We transfected
human embryonic kidney (HEK) 293FT cells
with either a targeting or nontargeting guide
RNA toward a Gaussia luciferase (Gluc) tar-
get to test the activation of the Cas7-11–Csx29–
Csx30 system. In the presence of the Gluc
mRNA target, we observed threefold higher
citrine fluorescence with the targeting, but
not nontargeting, guide RNA (Fig. 4I), indicat-
ing that Csx29 is activated and cleaves the
DHFR degron from the citrine reporter. To
validate that the increase in citrine fluores-
cence was caused by the cleavage of Csx30 in
the reporter, we analyzed the total protein from
the HEK293FT cells by Western blot using
an anti-FLAG antibody and visualized the
N-terminally FLAG-tagged reporter. The molec-
ular mass of the reporter protein decreased
from ~110 to 78 kDa only in the presence of
the tgRNA and targeting guide, indicating
the Csx29-mediated cleavage of Csx30 in the
reporter (Fig. 4J and fig. S16). These results
demonstrate that the Cas7-11–Csx29–Csx30
system is reprogrammable in mammalian cells
and can be used as a protease-based RNA-
guided posttranslational sensor.
Discussion
Our structural and functional analyses of the
type III-E CRISPR-Cas systems revealed nota-
ble complexity and fine-tuned regulation. The
effects of Csx30 and Csx31 on bacterial growth
suggested that Csx29-mediated Csx30 cleavage
releases the N-terminal fragment of Csx30
bound to Csx31, inhibiting host cell growth
(fig. S17). Furthermore, our biochemical and
structural analyses indicated that Csx30, Csx31,
and RpoE can form a ternary complex in which
Csx30 extensively interacts with RpoE, sug-
gesting that RpoE inhibition by Csx30 contrib-
utes to the observed cell growth arrest, akin to
the Cas13 collateral activity-based cell growth
arrest (28). Csx30 cleavage by Csx29 could also
facilitate the dissociation of RpoE from Csx30,
allowing RpoE to engage in a transcriptional
response to viral infection. By leveraging the
programmable nature of this system, we de-
veloped a molecular RNA sensor for transcripts
in mammalian cells, demonstrating the po-
tential of this Cas7-11–Csx29–Csx30 system for
sensing and therapeutic applications, analo-
gous to recently developed mammalian RNA
sensors (29, 30).
Thus, in type III-E CRISPR-Cas systems, the
Cas7-11–Csx29 effector complex likely degrades
the ssRNA transcripts of phage genes and
stimulates potentially toxic host cell stress
responses through the Csx29-mediated Csx30
cleavage (fig. S17). This type of programmed
growth suppression, through cell death or
growth arrest, appears analogous to that caused
by gasdermins, the bacterial membrane pore-
forming toxins that are switched on by the
release of auto-inhibitory peptides by asso-
ciated proteases that become activated during
phage infection (31, 32). Moreover, given the
high diversity of Csx30 CTDs (fig. S15), further
explorations of other subtype III-E systems
might reveal additional functions associated
with Cas7-11–mediated tgRNA recognition.
Among the CRISPR-Cas systems, a biolog-
ical, if not mechanistic, analogy can be found
in the type VI systems, where the Cas13–crRNA
effector complex recognizes complementary
phage mRNAs and cleaves both phage (specif-
ically and in cis) and host (indiscriminately and
in trans) transcripts, thus stalling cell growth
and with it the infectious cycle (28). Similarly,
in some type III systems, the CRISPR-Lon pro-
tease is activated by cyclic oligoadenylates
upon RNA recognition by the effector com-
plexes and cleaves the associated CRISPR-T
protein specifically, releasing a toxic frag-
ment (33). Our characterization of the sub-
type III-E system highlights the remarkable
diversity of CRISPR-associated functions acti-
vated by programmable nucleic acid recog-
nition, motivating continued exploration of
CRISPR-associated proteins and their pro-
grammable functions that could be useful for
biological applications. Our findings that the
type III-E Cas7-11–Csx29 effector complex is an
RNA-triggered nuclease-protease establish a
distinct paradigm of prokaryotic signal trans-
duction in viral immunity and could pave the
way for the development of new RNA/protein-
targeting technologies, including in vitro diag-
nostics and cellular RNA sensing.
REFERENCES AND NOTES
F. Hille et al., Cell 172, 1239–1259 (2018).
1.
2. K. S. Makarova et al., Nat. Rev. Microbiol. 18, 67–83
(2020).
3. M. Kazlauskiene, G. Kostiuk, Č. Venclovas, G. Tamulaitis,
V. Siksnys, Science 357, 605–609 (2017).
4. O. Niewoehner et al., Nature 548, 543–548 (2017).
5. C. Rouillon, J. S. Athukoralage, S. Graham, S. Grüschow,
M. F. White, eLife 7, e36734 (2018).
6. W. Han, S. Pan, B. López-Méndez, G. Montoya, Q. She,
Nucleic Acids Res. 45, 10740–10750 (2017).
7. A. R. VanderWal, J.-U. Park, B. Polevoda, E. H. Kellogg,
M. R. O’Connell, CRISPR-Csx28 forms a Cas13b-activated
membrane pore required for robust CRISPR-Cas adaptive
immunity. bioRxiv 466367 [Preprint] (2021); https://doi.org/
10.1101/2021.11.02.466367.
8. A. Özcan et al., Nature 597, 720–725 (2021).
9. S. P. B. van Beljouw et al., Science 373, 1349–1353
(2021).
10. O. O. Abudayyeh et al., Nature 550, 280–284 (2017).
11. J. S. Gootenberg et al., Science 356, 438–442 (2017).
12. K. Kato et al., Cell 185, 2324–2337.e16 (2022).
13. L. You et al., Cell 176, 239–253.e16 (2019).
14. N. Jia et al., Mol. Cell 73, 264–277.e5 (2019).
15. N. Zeytuni, R. Zarivach, Structure 20, 397–405 (2012).
16. L. Holm, Methods Mol. Biol. 2112, 29–42 (2020).
17. J. Yu et al., Nature 596, 138–142 (2021).
18. J. Jumper et al., Nature 596, 583–589 (2021).
19. L. Connolly, A. De Las Penas, B. M. Alba, C. A. Gross,
Genes Dev. 11, 2012–2021 (1997).
Kato et al., Science 378, 882–889 (2022)
25 November 2022
7 of 8
RES EARCH | R E S E A R C H A R T I C L E
20. J. Mecsas, P. E. Rouviere, J. W. Erickson, T. J. Donohue,
C. A. Gross, Genes Dev. 7 (12b), 2618–2628 (1993).
21. K. Hiratsu, M. Amemura, H. Nashimoto, H. Shinagawa,
K. Makino, J. Bacteriol. 177, 2918–2922 (1995).
22. E. A. Campbell et al., Mol. Cell 11, 1067–1078 (2003).
23. V. Schnaars et al., Microb. Physiol. 31, 1–20 (2021).
24. J. M. Chung et al., eLife 8, e47644 (2019).
25. S. Lu et al., Nucleic Acids Res. 48, D265–D268 (2020).
26. Y.-W. Chang et al., Science 351, aad2001 (2016).
27. X. J. Gao, L. S. Chong, M. S. Kim, M. B. Elowitz, Science 361,
1252–1258 (2018).
28. A. J. Meeske, S. Nakandakari-Higa, L. A. Marraffini, Nature 570,
241–245 (2019).
29. K. Jiang et al., Nat. Biotechnol. 10.1038/s41587-022-01534-5
(2022).
30. E. M. Zhao et al., Nat. Biotechnol. 40, 539–545
(2022).
31. A. G. Johnson et al., Science 375, 221–225 (2022).
32. C. Clavé et al., Proc. Natl. Acad. Sci. U.S.A. 119, e2109418119
(2022).
33. C. Rouillon et al., SAVED by a toxin: Structure and function of
the CRISPR Lon protease. bioRxiv 471393 [Preprint] (2021);
https://doi.org/10.1101/2021.12.06.471393.
ACKN OW LEDG MEN TS
We thank M. Hiraizumi and the staff scientists at The University
of Tokyo’s cryo-EM facility, especially Y. Sakamaki, for help with
cryo-EM data collection, and N. Okumura for help with the
N-terminal analysis. Funding: S.A. is supported by the Japan
Society for the Promotion of Science (JSPS) (KAKENHI grant
21H05736). K.S.M. and E.V.K are supported through the National
Institutes of Health’s (NIH’s) Intramural Research Program
(National Library of Medicine). H.N. is supported by the Platform
Project for Supporting Drug Discovery and Life Science Research
Basis for Supporting Innovative Drug Discovery and Life Science
Research (BINDS) from the Japanese Agency for for Medical
Research and Development (AMED) (grant JP21am0101115 support
number 2792 and grant JP21wm0325048h0001), JSPS (KAKENHI
grants 20K20579 and 21H05281), the Takeda Medical Research
Foundation, and the Inamori Research Institute for Science.
J.S.G. and O.O.A. are supported by the NIH (grants 1R21-AI149694,
R01-EB031957, and R56-HG011857), the McGovern Institute
Neurotechnology (MINT) program; the K. Lisa Yang and
Hock E. Tan Center for Molecular Therapeutics in Neuroscience,
the G. Harold & Leila Y. Mathers Charitable Foundation, the
MIT John W. Jarve (1978) Seed Fund for Science Innovation,
FastGrants, the Cystic Fibrosis Foundation, Google Ventures, by a
Longevity Impetus Grant from the Norn Group, the NHGRI/TDCC
Opportunity Fund, and the McGovern Institute. Author
contributions: Conceptualization: K.K., O.O.A., J.S.G., H.N.;
Funding acquisition: O.O.A., J.S.G., H.N.; Investigation: K.K., S.O.,
C.S., K.J., W.Z., J.I., Y.I., S.A., T.N., K.S.M., E.V.K., O.O.A., J.S.G.,
H.N.; Supervision: O.O.A., J.S.G., H.N.; Writing – original draft:
K.K., C.S., K.J., O.O.A., J.S.G., H.N.; Writing – review and editing:
E.V.K., O.O.A., J.S.G., H.N. Competing interests: The authors have
filed a patent application related to this work. J.S.G. and O.O.A.
are co-founders of Sherlock Biosciences, Proof Diagnostics,
Moment Biosciences, and Tome Biosciences. H.N. is an adviser for
Moment Biosciences. Data and materials availability: Relevant
plasmids are available from the corresponding authors under a
material transfer agreement with MIT or Addgene. The EM
densities have been deposited in the Electron Microscopy Data
Bank under the accession codes 33695 and 34218. The model
coordinates have been deposited in the Protein Data Bank under
the accession codes 7Y7X and 8GS2. License information:
Copyright © 2022 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science. No
claim to original US government works. https://www.science.org/
about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add7347
Materials and Methods
Figs. S1 to S17
Tables S1 to S2
References (34–49)
MDAR Reproducibility Checklist
Movies S1 and S2
View/request a protocol for this paper from Bio-protocol.
Submitted 30 June 2022; accepted 24 October 2022
10.1126/science.add7347
Kato et al., Science 378, 882–889 (2022)
25 November 2022
8 of 8
|
10.1126_science.add6575
|
RES EARCH
MAGNETISM
Topological kinetic crossover in a nanomagnet array
Xiaoyu Zhang1, Grant Fitez1, Shayaan Subzwari1, Nicholas S. Bingham1†, Ioan-Augustin Chioar1,
Hilal Saglam1, Justin Ramberger3, Chris Leighton3, Cristiano Nisoli2*, Peter Schiffer1,4*
Ergodic kinetics, which are critical to equilibrium thermodynamics, can be constrained by a system’s
topology. We studied a model nanomagnetic array in which such constraints visibly affect the behavior of the
magnetic moments. In this system, magnetic excitations connect into thermally active one-dimensional
strings whose motion can be imaged in real time. At high temperatures, our data showed the merging,
breaking, and reconnecting of strings, resulting in the system transitioning between topologically distinct
configurations. Below a crossover temperature, the string motion is dominated by simple changes in
length and shape. In this low-temperature regime, the system is energetically stable because of its
inability to explore all possible topological configurations. This kinetic crossover suggests a generalizable
conception of topologically broken ergodicity and limited equilibration.
O ur understanding of equilibrium ther-
modynamics relies on simplifying as-
sumptions about kinetics—for example,
the ergodic hypothesis, which postu-
lates that a system can explore all en-
ergetically equivalent configurations (1). When
kinetics are constrained (2–7), however, ergo-
dicity can break down on relevant timescales,
causing memory effects, glassiness, nonequili-
bration, or slow equilibration (8).
Topological constraints on kinetics are of
particular interest in the context of the critical
role of topology in much of modern physics
and have been examined within the context
of various mathematical models (2, 4). The
constraints can arise from a partition of the
accessible states (or phase space) into topo-
logically distinct subsets, often called topo-
logical sectors (9, 10). Kinetics within a sector
are topologically trivial; they do not alter the
system’s topology, contrasting with nontrivial
kinetics that allow the system to cross through
sectors and change the system’s topology. If
the latter are constrained, then the system
cannot explore all of its configurations and
should show deviations from reversible equi-
librium thermodynamics. In experimental sys-
tems, topologically constrained kinetics were
invoked early in the context of soft-matter
systems—for example, in macromolecule elas-
ticity, foams, cellular patterns, and the kinetics
of continuous networks (2)—and this broad
topic is closely connected to fundamental
questions around equilibration, ergodicity,
and memory.
1Department of Applied Physics, Yale University, New Haven,
CT 06511, USA. 2Theoretical Division and Center for
Nonlinear Studies, MS B258, Los Alamos National
Laboratory, Los Alamos, NM 87545, USA. 3Department of
Chemical Engineering and Materials Science, University of
Minnesota, Minneapolis, MN 55455, USA. 4Department of
Physics, Yale University, New Haven, CT 06511, USA.
*Corresponding author. Email: peter.schiffer@yale.edu (P.S.);
cristiano@lanl.gov (C.N.)
†Present address: Department of Physics, University of Maine,
Orono, ME 04469, USA.
Strings in Santa Fe ice
We report the direct visualization of topolog-
ically constrained kinetics in a designed nano-
magnet array. Such arrays, known broadly as
artificial spin ice systems, have displayed a
range of exotic collective phenomena and
technological potential (11, 12). The particu-
lar array geometry that we studied is Santa
Fe ice (SFI) (13, 14). The structure of SFI is
shown in Fig. 1, A and B, where each element
is a single-domain ferromagnetic island with
the moment oriented along the long axis as a
result of shape anisotropy. The islands are
arranged in rectangular plaquettes, with pairs
of interior plaquettes (Fig. 1, A, C, and D, light
blue shaded areas) surrounded by six periphe-
ral plaquettes. Much of the physics of this sys-
tem can be understood through the moment
configurations at the vertices where islands
converge in the lattice (13, 14). The SFI lattice
geometry requires that some fraction of the
vertices are not in the lowest local energy
state, even in the collective ground state. These
local excitations have previously been dubbed
“unhappy” vertices (13, 15) and are indicated
with red dots in Fig. 1, C and D. SFI lends itself
to the study of topological kinetics because,
as we previously showed (14), the lattice ge-
ometry forces unhappy vertices to form one-
dimensional (1D) strings that can either end
within interior plaquettes or form closed loops
[(16), section S3]. In Fig. 1, C and D, we show
these strings as olive lines connecting the un-
happy vertices in two realizations of the highly
degenerate, disordered ground state of the sys-
tem (13, 14). Such strings have naturally com-
plex topological properties (17), as is readily
apparent in a bowl of spaghetti or udon (18, 19).
To probe the physics of SFI, we studied sam-
ples composed of stadium-shaped permalloy
islands with designed lateral dimensions of
470 by 170 nm, thickness of ~2.5 nm, and lat-
tice spacing of a = 600, 700, and 800 nm (Fig.
1A). The specifics of sample fabrication and
characterization are discussed in (16), section
S1, and have been described previously in (14),
in which we also showed that the string pop-
ulation and length distributions are thermally
activated at the highest experimentally acces-
sible temperatures. We measured the moment
configuration in our samples using x-ray mag-
netic circular dichroism photoemission elec-
tron microscopy (XMCD-PEEM) (16), which
yielded real-space images of the island mag-
netic moments (Fig. 1B). When we increased
temperature, thermal excitations caused a frac-
tion of the moments to visibly flip in orienta-
tion between images, with the rate of flipping
increasing with increasing temperature, and
we could therefore observe the thermal kine-
tics of this system. We show results for the a =
600 nm sample, in which the interisland in-
teractions are strongest, in the section on
quantifying string motion; the data are con-
sistent for the other lattice spacings (16). All
data are derived from the same XMCD-PEEM
images taken on the same samples that were
analyzed in (14).
Because the energy cost of a string increases
with its length, the disordered ground state
corresponds to strings of three unhappy ver-
tices, each on three-island vertices. Therefore,
in the ground state, strings are constrained to
connecting near-neighbor interior plaquettes
(Fig. 1, C and D). In excited states of the sys-
tem, strings can be longer and can connect in-
terior plaquettes that are not neighboring,
or they can form closed loops. In our ex-
perimental data, the system approached the
disordered ground state at the lowest temper-
atures, with a residual population of excita-
tions that increased in number for the larger
lattice spacings (14).
Types of string motion
In considering string kinetics, one must con-
sider the possible motions of the strings with-
in the geometrical constraints imposed by the
SFI structure. For any string configuration, the
strings can be continuously deformed by bend-
ing, elongating, or shrinking them while keep-
ing the endpoints the same. Such deformations
do not change the topology of the string con-
figuration and thus keep the system within a
single topological sector of available configura-
tions, also known as a homotopy class. In the
case of SFI, a homotopy class is defined as
the set of all string configurations connecting
the same interior plaquettes, in which any
member of the set can be transformed to any
other member of the set through continuous
deformations only. Because the SFI ground
state can be realized with many different sets
of string connections among the plaquettes, the
low-energy configurations of the system can be
partitioned into different homotopy classes
(Fig. 1, C and D).
By contrast, string configurations that can-
not be obtained by continuous bending or
deformation of the strings are in different
Zhang et al., Science 380, 526–531 (2023)
5 May 2023
1 of 6
A
C
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. The Santa Fe ice (SFI) geometry.
(A) Schematic of the lattice, where
each element represents a single-
domain nanomagnet and the lattice
spacing is a. The unit cell
(indicated in yellow) has four
interior plaquettes (indicated
by light blue shading) that
are separated by two-island
vertices and surrounded by
12 peripheral plaquettes. (B) XMCD-
PEEM image of SFI, in which all of
the islands are either black or white,
indicating the magnetic moment
direction through its component
projected onto the incident
x-ray beam direction (red arrow).
The lattice is slightly offset
from the structure in (A). (C and
D) Illustrations of two SFI disordered
ground states, where the excited
vertices are indicated by circular
red dots. They are in different
homotopy classes because the
interior plaquettes that are
connected by strings are distinct.
B
2a
X-ray
D
homotopy classes. For example, if two string
configurations have different interior plaquettes
attached to each other, there is no continuous
way to transform the strings from one con-
figuration to another. Strings must be created,
eliminated, combined, or cut to transform be-
tween such configurations, thus breaking con-
tinuity and changing the topology.
In terms of energy, each homotopy class has
an energy minimum obtained by continuously
deforming and shrinking all the strings in any
configuration belonging to the class, so that
strings have the minimum number of lowest-
energy vertices connecting with the string
endpoints. If strings in a homotopy class ex-
tend to connect interior plaquettes that are
not in close proximity, the energy minimum of
that homotopy class is necessarily higher than
the global ground state energy. To relax to the
ground state, the system requires a topological
change in the strings to change the homotopy
class. Such a process, in which a string changes
in such a way that the system transits from
one homotopy class to another, is called “topo-
logical surgery” (20, 21). The specific process of
topological surgery in SFI occurs when two or
more strings merge and then separate into a
new configuration. This concept of topolog-
ical surgery—the connection and separation
of proximate 1D objects—can also be used to
describe diverse phenomena such as chromo-
some meiosis, DNA recombination, magnetic
flux reconnection in plasmas, vortex line fusion
and reconnection in classical and quantum liq-
uids, and dislocation lines in metallurgy (22–26).
SFI provides an unusual opportunity for ex-
perimentally tracking this process in real time
in a thermal system.
On the basis of the topological character of
the strings in SFI, we considered the kinetics
of our string configurations. We analyzed these
kinetics by comparing sequential XMCD-PEEM
images (27, 28) and recording the motions (we
use the term “motion” to describe any change to
a string, including the creation and annihilation
of loops). We classified string motions into two
categories: trivial and nontrivial. These two cat-
egories correspond to motions that continuously
deform a string or to motions that include a
change in string topology, respectively. In other
words, trivial motions do not change the ho-
motopy class of the string configuration in the
system, whereas nontrivial changes do. The
distinction is demonstrated in Fig. 2, where
we show simple examples of each, and in Fig. 3,
where we give a taxonomy of trivial and non-
trivial motions.
The trivial motions of strings correspond to
continuous deformations of strings that retain
the same interior plaquettes as endpoints. The
different variations of trivial motions are shown
in the top schematic of Fig. 3 and are labeled
“wiggle” [red (T1), dark blue (T2), and purple
(T3)]; “grow” and “shrink” [orange (T4)];
“loop”: “wiggle” [green (T5)]; “loop”: “grow”
and “shrink” [light blue (T6)]; and “loop”: “cre-
ation” and “annihilation” [pink (T7)]. The loop
motions are included as trivial because loops
are contractible to zero, and thus loop crea-
tion and annihilation are topologically trivial.
The nontrivial motions of strings are those
changes that represent a change in the homo-
topy class of the system—that is, a change in
how the interior plaquette endpoints are con-
nected by strings. The different variations of
nontrivial motions are shown in the bottom
schematic of Fig. 3 and are labeled “merge”
and “split” [magenta/red (N1/N2) and light
Zhang et al., Science 380, 526–531 (2023)
5 May 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
A
D
Fig. 2. Trivial and nontrivial string
motions. Red dots indicate unhappy vertices.
(A to C) Examples of trivial string motions.
Between configurations (A) and (B), there is
a “wiggle” motion with no energy change.
Between configurations (B) and (C), there is
a “grow” motion with energy change because
(C) is longer. (D to F) Examples of nontrivial
string motions. Configurations (D) and (F) have
pairs of strings that have different pairs of
endpoints, and configuration (E) has a merged
crossed string with four endpoints. Each of
the three configurations represents
topologically distinct states, that is, different
homotopy classes.
Fig. 3. Illustration and taxonomy of various
string motions. Island moments are shown
as open arrows, and the flipped island
moments (solid arrows) correspond to
the strings. (Top) Trivial string motions.
(Bottom) Nontrivial string motions.
The tables list the motions illustrated
in the schematics.
Trivial
B
E
Trivial
C
F
Nontrivial
Nontrivial
Topologically Trivial String Motions
T4
T2
T3
T4
T5
T2
T5
TT77
T1
T6
T3
Topologically Nontrivial String Motions
Topologically trivial mo ons
Wiggle
Grow (T4)
Shrink (T4)
Loop
Energy increase (T1)
Energy decrease (T1)
No energy change (T2,T3)
Wiggle (T5)
Grow (T6)
Shrink (T6)
Crea on (T7)
Annihila on (T7)
N1
N2
N3
N8
N9
N4
N5
N7
N6
Topologically nontrivial mo ons
Merge (N1,N2 N3)
Split (N3 N1,N2)
Reconnec on (N4,N5 N6,N7)
Adjacent reconnec on (N8)
Strings on adjacent
interior plaque e
Crea on (N9)
Annihila on (N9)
T1
T6
N8
purple (N3)]; “reconnection” [green/light blue
(N4/N5) and olive/dark blue (N6/N7)]; “adja-
cent reconnection” [purple (N8)]; and “strings
on adjacent interior plaquettes”: “creation”
and “annihilation” [orange (N9)]. Merge and
split are necessarily intermediate steps of a
reconnection, but because of the limited time
resolution, we labeled a full reconnection dis-
tinctly, when we could witness a full re-
connection between two consecutive frames.
Adjacent reconnection and interior creation
and annihilation are in a separate category in
that they involve only adjacent interior pla-
quettes with a shared side and the flip of a
single moment along that side. We therefore
focus on the topologically nontrivial kinetics
associated with strings that traverse the pe-
ripheral plaquettes.
In addition to the trivial and nontrivial string
motions that we classified, there are strings
that naturally do not change between sequen-
tial images (that have no motion). There are
also complex motions that involve multiple
nontrivial changes in moment orientations
between two sequential XMCD-PEEM images,
as a consequence of the finite time resolution
of image acquisition. Furthermore, there are
other trivial and nontrivial motions associated
with a small number of nanoislands whose
magnetizations could not be determined from
our XMCD-PEEM images (<0.1%) or that lie
on the edges of those images; these motions
resulted in incomplete strings in our image
recognition program. These other motions
and complex motions are not listed in the
Fig. 3 tables.
Quantifying string motion
To quantitatively and unambiguously charac-
terize the string motions from our sequential
experimental maps of the moments, we used
an analysis program that labels each string
with two sets: One set comprises the interior
plaquettes at which the string ends, and the
other set comprises the unhappy vertices it
connects. By comparing the elements in these
two sets between sequential images, the algo-
rithm can therefore collect the temperature-
Zhang et al., Science 380, 526–531 (2023)
5 May 2023
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
dependent prevalence of each type of string
motion (16). We quantified the string-motion
prevalence as N, the average number of string
motions between subsequent XMCD-PEEM
images per unit cell. N effectively gives the
rate of string motions because the images are
taken at 1-s intervals, and it can be separated
by the type of motion involved.
We then turned to the temperature de-
pendence of our data. We first examined the
total vertex energy of the system as a function
of temperature, for example, the sum of the
magnetostatic energy at each vertex, averaged
over all images at a given temperature, as de-
termined by micromagnetic calculations (16, 29).
The energy, quantified as the excess above the
ground state, has a clear crossover tempera-
ture of around TX = 330 K (Fig. 4A), which is
reflected in the kinetics of the strings and is
also visible in our previously published data on
average string length in the system (14).
In Fig. 4, B to D, we show the statistics of
motions as a function of temperature, with
the temperature regime below TX indicated
by shading. Below TX, trivial string motions
dominate, whereas nontrivial string motions
dominate above TX (the number of trivial mo-
tions at higher temperatures is suppressed by a
fraction of them being subsumed within the non-
trivial motions). The system energy shown in
Fig. 4A is nearly flat below TX; the system does
not substantially change its energy as a func-
tion of temperature (even though string motions
retain considerable temperature dependence).
Fig. 4. Temperature-dependent string
properties. The shaded regions indicate the
low-temperature regime below TX. (A) The
excess energy per unit cell versus temperature.
(B) The temperature dependence of the
average number of string motions per XMCD-
PEEM image. We also tracked the number
of strings that did not change between
subsequent images; we plot the total number
of these strings as “no motion.” (C) The
temperature dependence of the average num-
ber of string motions per XMCD-PEEM image
for all types of nontrivial motions. (D) An
expanded plot of the low-temperature regime
from (C). All data are shown per unit cell
and are for a = 600 nm SFI; error bars are the
SEM from all XMCD-PEEM images taken at
the given temperature.
A
C
This temperature independence is consistent
with the observed crossover in the predominant
types of string motions from nontrivial to trivial
when crossing through TX. Substantial reduc-
tions in system energy require topological changes
in the strings—that is, nontrivial motions—and
therefore the system is unable to relax to a lower-
energy overall configuration.
Among the nontrivial motions above TX, the
complex motions are predominant (Fig. 4C).
Because these represent multiple updates of
the moments among XMCD-PEEM images at
higher temperature, they also include most
other topologically nontrivial motions. By
contrast, local processes among adjacent in-
terior plaquettes dominate below TX (Fig. 4D).
This is consistent with the system’s topological
constraints because such motions require only a
single moment flip and make a minimal change
to the system’s topology.
We next considered the trivial kinetics. As
shown in Fig. 5A, the loop processes (light green
curves) and wiggles (red curves) are the high-
est contributors. The grow (blue) and shrink
(violet) curves always follow each other, as ex-
pected from detailed balance. The same can be
said in general for all reversible inelastic pro-
cesses, as seen also in wiggles corresponding
to energy increase and decrease (Fig. 5B) (the
increase in energy being caused by the strings
occupying a vertex of coordination 4, whose
excitations have slightly higher energy than
vertices of coordination 3) and loop creation
and annihilation (Fig. 5C). Most of the wig-
gles, however, do not change the total system
energy (Fig. 5B), and thus they dominate
(Fig. 5A, red curve).
e(cid:2)ES
Last, in Fig. 5D we show a log-linear plot of
the number of trivial processes, Ntrivial, versus
the reciprocal temperature. Within our fairly
narrow accessible temperature range below
TX, the data display apparent thermally activ-
ated behavior: Ntrivial Tð Þ ¼ 1
T , where to is
to
the inverse attempt frequency for the thermal
processes and ES is an energetic barrier for
changing states. This apparent activated behav-
ior is also observed for the other lattice spacings
(16). Fits of that form to the data, however, de-
pend very much on our choice of fitting range,
with to ~ 10−5 seconds and ES ~ 3000 K for the
a = 600 nm sample. We can qualitatively under-
stand this behavior in that the motion of strings,
even those with no net energy change, requires
moments to flip in direction and overcome the
energy barrier associated with the island’s shape
anisotropy. The energy barrier to flipping a mo-
ment, as calculated with micromagnetics at
zero temperature, is considerably higher than
ES (16), but we expect that the energy barrier
for flipping the moment in a real system will
be affected by the magnetic texture of the
island at elevated temperature (30). Similarly,
the effective attempt frequency, 1=to , is small
compared with the typical spin wave frequency,
which is in the gigahertz regime. Again, this is
perhaps not surprising in that the string mo-
tions, even almost all of the trivial ones, are
associated with changes in multiple islands.
regime
regime
B
D
Zhang et al., Science 380, 526–531 (2023)
5 May 2023
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
A
C
B
D
Fig. 5. Temperature dependence of trivial string motions. Shown are the average number of string motions per XMCD-PEEM image for trivial motions. Shading
indicates the low-temperature regime in which trivial string motions dominate. (A) All types of trivial motions. (B) String-wiggle subtypes of trivial motions.
(C) Loop subtypes of trivial motions. (D) The number of total trivial motions as a function of inverse temperature with fit. All data are shown per unit cell and for
the a = 600 nm sample. The error bars are the SEM from all XMCD-PEEM images taken at the given temperature.
Discussion and outlook
Our observed crossover leads to a consistent
picture of how string kinetics evolve. Below
TX, the system does not change its energy, but
it is kinetically active through topologically
trivial motions (bending and stretching modes).
In this regime, the system has a low likelihood of
undergoing topological surgery and changing
between homotopy classes. In other words,
the system’s limited ability to explore homo-
topy classes likely prevents full ergodicity on
experimental timescales at low temperatures,
and thus the system is impeded from further
reducing its global energy to fully realize the
ground state. The trivial kinetic processes are
thermally equilibrated below TX, but within a
limited phase space of possible string config-
urations. This state is evidenced by the close
tracking of energy-increasing and -decreasing
motions (Fig. 5, A to C), which indicates that
detailed balance holds, and by the activated
behavior shown in Fig. 5D.
By contrast, at high temperature, above TX,
the topologically nontrivial motions become
predominant, and the system appears to be
fully ergodic. This allows the total energy of
the system to change substantially because
the different homotopy classes can have radi-
cally different energies; this is the regime in
which the activated string length was demon-
strated previously (14).
The observation of a clear crossover be-
tween the two regimes correlates somewhat to
glass transitions and other dynamic slowdowns,
but here it is driven by topology rather than by
a disordered potential landscape. We can broad-
ly understand the crossover in that the homo-
topy class–changing nontrivial motions require
substantially longer strings and thus more mo-
ments to reverse directions than do the trivial
motions, and therefore are kinetically sup-
pressed with decreasing temperature. Although
our present data does not allow us to identify
the crossover with a dynamic phase transition
and associated nonanalytic behavior, the relative
sharpness of the crossover is notable relative to
most glass transitions, suggesting future studies
with frequency-sensitive techniques (31–33).
Our findings should be generalizable in de-
fining and characterizing kinetic crossovers to
nonergodicity beyond systems with topolog-
ical constraints (34). Comparing experimen-
tal results with existing theoretical methods to
ascertain ergodicity could elucidate slow relax-
ation after quenches and memory effects and
aid in exploring the relation with other kinetic
crossovers (8, 34–36). Such studies could also
have relevance for new forms of computing (37)
and possibly for quantum tunneling through
homotopy classes within qubit realizations (38)
of SFI or similar structures. Our work might
also prove pertinent to systems such as struc-
tural glasses, for which dislocation lines ending
in vertices proved to be a useful description
(39–42). The advantage of our physical realiza-
tion of SFI is that it allows direct experimental
measurements of these phenomena, suggest-
ing future examinations of other bespoke to-
pologically interesting structures.
REFERENCES AND NOTES
1. K. Huang, Statistical Mechanics (John Wiley and Sons, 2000).
2. D. Weaire, N. Rivier, Contemp. Phys. 25, 59–99 (1984).
3. P. Goldbart, N. Goldenfeld, Phys. Rev. Lett. 58, 2676–2679
(1987).
4. T. Aste, D. Sherrington, J. Phys. Math. Gen. 32, 7049–7056
(1999).
J.-Y. Xiong et al., J. Phys. Chem. B 109, 5638–5643 (2005).
5. F. Ritort, P. Sollich, Adv. Phys. 52, 219–342 (2003).
6.
7. P. Oikonomou, P. Cluzel, Nat. Phys. 2, 532–536 (2006).
8. R. G. Palmer, Adv. Phys. 31, 669–735 (1982).
9. C. L. Henley, J. Phys. Condens. Matter 23, 164212 (2011).
10. C. Nisoli, Europhys. Lett. 132, 47005 (2020).
11. S. H. Skjærvø, C. H. Marrows, R. L. Stamps, L. J. Heyderman,
Nat. Rev. Phys. 2, 13–28 (2020).
12. P. Schiffer, C. Nisoli, Appl. Phys. Lett. 118, 110501 (2021).
13. M. J. Morrison, T. R. Nelson, C. Nisoli, New J. Phys. 15, 045009
(2013).
14. X. Zhang et al., Nat. Commun. 12, 6514 (2021).
15. R. L. Stamps, Nat. Phys. 10, 623–624 (2014).
16. Materials and methods are available as supplementary
materials.
Zhang et al., Science 380, 526–531 (2023)
5 May 2023
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
17. M. D. Frank-Kamenetskii, A. V. Lukashin, A. V. Vologodskii,
37. A. Lucas, Front. Phys. (Lausanne) doi:10.3389/fphy.2014.
Nature 258, 398–402 (1975).
00005 (2014).
18. D. Michieletto, M. S. Turner, Phys. World 27, 28–31 (2014).
19. D. Michieletto, M. S. Turner, Proc. Natl. Acad. Sci. U.S.A. 113,
38. A. D. King, C. Nisoli, E. D. Dahl, G. Poulin-Lamarre,
A. Lopez-Bezanilla, Science 373, 576–580 (2021).
5195–5200 (2016).
20. J. W. Milnor, in Proceedings of Symposia in Pure Mathematics,
vol. 3, Differential Geometry, C. B. Allendorfer, Ed. (American
Mathematical Society, 1961), pp. 39–55.
21. S. Antoniou, S. Lambropoulou, PLOS ONE 12, e0183993 (2017).
22. M. A. Berger, G. B. Field, J. Fluid Mech. 147, 133 (1984).
23. J. Koplik, H. Levine, Phys. Rev. Lett. 71, 1375–1378 (1993).
24. S. Kida, M. Takaoka, Annu. Rev. Fluid Mech. 26, 169–177 (1994).
25. U. F. Kocks, H. Mecking, Prog. Mater. Sci. 48, 171–273 (2003).
26. M. Yamada, R. Kulsrud, H. Ji, Rev. Mod. Phys. 82, 603–664
(2010).
27. A. Farhan et al., Phys. Rev. Lett. 111, 057204 (2013).
28. I. Gilbert et al., Nat. Phys. 12, 162–165 (2016).
29. A. Vansteenkiste et al., AIP Adv. 4, 107133 (2014).
30. S. D. Slöetjes, B. Hjörvarsson, V. Kapaklis, Appl. Phys. Lett. 118,
142407 (2021).
31. M. Pohlit et al., Phys. Rev. B 101, 134404 (2020).
32. M. Goryca et al., Phys. Rev. X 11, 011042 (2021).
33. M. Goryca et al., Phys. Rev. B 105, 094406 (2022).
34. G. Miloshevich, T. Dauxois, R. Khomeriki, S. Ruffo, Europhys.
Lett. 104, 17011 (2013).
35. F. Borgonovi, G. L. Celardo, M. Maianti, E. Pedersoli,
J. Stat. Phys. 116, 1435–1447 (2004).
36. D. Mukamel, S. Ruffo, N. Schreiber, Phys. Rev. Lett. 95, 240604
(2005).
39. D. R. Nelson, M. Rubinstein, F. Spaepen, Philos. Mag. A Phys.
Condens. Matter Defects Mech. Prop. 46, 105–126 (1982).
40. D. R. Nelson, M. Widom, Nucl. Phys. B 240, 113–139 (1984).
41. S. Sachdev, D. R. Nelson, Phys. Rev. Lett. 53, 1947–1950
(1984).
42. D. R. Nelson, in Applications of Field Theory to Statistical
Mechanics, L. Garrido, Ed. (Springer, 1985), pp. 13–30.
43. X. Zhang et al., Topological kinetic crossover in a nanomagnet
array, Dryad (2023); https://doi.org/10.5061/dryad.w0vt4b8w9.
44. X. Zhang et al., Topological kinetic crossover in a nanomagnet
array, version 1, Zenodo (2023); https://doi.org/10.5281/
zenodo.7641526.
ACKN OWLED GMEN TS
We thank C. Castelnovo, N. Goldenfeld, and Y. Shokef for helpful
feedback on the manuscript. We also thank R. Chopdekar and
R. Koch for support at the PEEM-3 endstation at beamline 11.0.1.1
of the Advanced Light Source during our data acquisition.
Funding: Work at Yale University was funded by the US
Department of Energy (DOE) Office of Basic Energy Sciences,
Materials Sciences and Engineering Division, under grant DE-
SC0020162. This research used resources of the Advanced Light
Source, a DOE Office of Science user facility, under contract
DE-AC02-05CH11231. Work at the University of Minnesota was
supported by the NSF through grant DMR-2103711. Work at Los
Alamos National Laboratory was carried out under the auspices of
the DOE through Los Alamos National Laboratory, operated by
Triad National Security, (contract 892333218NCA000001), and
financed by DOE LDRD. Author contributions: J.R. performed film
depositions under the guidance of C.L. X.Z. and N.S.B. oversaw
the lithography. X.Z., N.S.B., H.S., and I.-A.C. performed the XMCD-
PEEM characterization of the thermally active samples. X.Z., G.F.,
and S.S. analyzed the string structures. C.N. developed the theory
and wrote the first draft. P.S. supervised the entire project.
All authors contributed to the discussion of results and to the
finalization of the manuscript. Competing interests: The authors
declare that they have no competing interests. Data and materials
availability: Additional experimental data and underlying data
from the plots generated in this study are available at Dryad (43)
and Zenodo (44). License information: Copyright © 2023 the
authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.science.org/about/science-
licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add6575
Materials and Methods
Supplementary Text
Figs. S1 to S6
Tables S1 and S2
Submitted 25 June 2022; accepted 31 March 2023
10.1126/science.add6575
Zhang et al., Science 380, 526–531 (2023)
5 May 2023
6 of 6
|
10.1126_science.add5574
|
RES EARCH
2D MATERIALS
Correlated insulator of excitons in WSe2/WS2
moiré superlattices
Richen Xiong1, Jacob H. Nie1, Samuel L. Brantly1, Patrick Hays2, Renee Sailus2, Kenji Watanabe3,
Takashi Taniguchi4, Sefaattin Tongay2, Chenhao Jin1*
A panoply of unconventional electronic states has been observed in moiré superlattices. Engineering
similar bosonic phases remains, however, largely unexplored. We report the observation of a
bosonic correlated insulator in tungsten diselenide/tungsten disulfide (WSe2/WS2) moiré
superlattices composed of excitons, that is, tightly bound electron-hole pairs. We develop a pump
probe spectroscopy method that we use to observe an exciton incompressible state at exciton filling
nex = 1 and charge neutrality, indicating a correlated insulator of excitons. With varying charge
density, the bosonic correlated insulator continuously transitions into an electron correlated
insulator at charge filling ne = 1, suggesting a mixed correlated insulating state between the two
limits. Our studies establish semiconducting moiré superlattices as an intriguing platform for
engineering bosonic phases.
S trongly correlated phases can emerge in
flat-band systems when many-body in-
teractions dominate over kinetic energy
(1, 2). Semiconducting transition-metal
dichalcogenide (TMDC) moiré superlat-
tices offer a distinctive platform where both
fermionic and bosonic quasiparticles—charges
(3–12) and excitons (13–21)—occupy flat bands.
Previous studies have primarily focused on the
fermionic sector with exciton filling nex = 0,
such as Mott and Wigner crystal states (3–8),
stripe phase (9), continuous Mott transition
(10, 11), and quantum anomalous Hall insula-
tors (12), along with single-exciton behavior
in the nex→0 limit, such as moiré excitons
(13–16) and trions (20, 21). Recently, increas-
ing efforts have been put into an intermediate
exciton density regime of nex ~ 0.1. Examples
include studies of exciton-mediated ferromag-
netism (22) as well as excitonic insulators
(23, 24)—insulators for charge but “metals” for
excitons, where electron behavior is affected
by the coexisting excitons. However, in these
prior studies, excitons themselves are in a com-
pressible fluid state; strongly correlated phases
of bosons remain elusive.
Here we explore the high–exciton density
regime of nex ~ 1. We create and identify a
bosonic correlated insulator consisting of in-
terlayer excitons in a 60°-aligned WSe2/WS2
moiré superlattice. Each interlayer exciton
contains an electron in the WS2 layer and a
hole in WSe2 with a large binding energy of
hundreds of milli–electron volts (25), which
1Department of Physics, University of California at Santa
Barbara, Santa Barbara, CA 93106, USA. 2School for
Engineering of Matter, Transport, and Energy, Arizona State
University, Tempe, AZ 85287, USA. 3Research Center for
Functional Materials, National Institute for Materials Science,
1-1 Namiki, Tsukuba 305-0044, Japan. 4International Center
for Materials Nanoarchitectonics, National Institute for
Materials Science, 1-1 Namiki, Tsukuba 305-0044, Japan.
*Corresponding author. Email: jinchenhao@ucsb.edu
is the ground state exciton configuration of a
type II heterostructure (Fig. 1A). We find an ex-
citon incompressible state at nex = 1, i.e., one
exciton per moiré site, a hallmark of a bosonic
correlated insulator. We further study the
phase diagram spanned by ne and nex and ob-
serve a mixed correlated insulator along the
path of ntot = nex + ne = 1. Owing to the large
exciton binding energy and strong correlations
in this system, the bosonic correlated insulator
persists to above 30 K, orders of magnitude
higher than in previous studies in cold atoms
and quantum wells (26, 27). Our results high-
light semiconducting moiré superlattices as
an attractive platform for engineering new
bosonic phases at high temperature, such as
valley pseudospin order and pseudospin liq-
uid, bosonic Mott-superfluidity transition (28),
exciton-mediated superconductivity (29), and
topological excitons (18, 19), as well as for ex-
ploring the many-body physics of interacting
fermions and bosons.
Direct measurement of exciton
compressibility
One major challenge in identifying a corre-
lated insulator of excitons is to distinguish it
from disordered excitons without a lattice. Sev-
eral works have reported power-dependent
photoluminescence (PL) spectra in TMDC
systems (30, 31) and emergence of additional
high-energy peaks under strong excitation.
However, this only indicates the existence
of local exciton-exciton interactions while pro-
viding no information on an ordered exciton
lattice. The key evidence of a correlated in-
sulator, as widely recognized in electrical
measurements, is an incompressible state at
particular lattice fillings (32). Optical mea-
surements such as PL, by contrast, typically col-
lect responses from all excitons in the system
and cannot obtain compressibility information.
Here we develop a pump-probe spectroscopy
method that directly measures exciton com-
pressibility (Fig. 1B); this is an optical analog
of electrical capacitance measurements (33)
and distinct from conventional optical pump-
probe configurations. In an electronic capaci-
tance measurement, a DC “pump” gate voltage
tunes the background charge density, and a
small AC “probe” voltage slightly modulates
the charge density. Similarly, here a relatively
strong pump light tunes the background ex-
citon density, and the weak probe light injects
a small number of additional excitons and de-
tects their response. In both cases, the charge
and exciton compressibility can be directly ob-
tained from the minimum energy it takes to
add one more particle on top of a given back-
ground particle density. Notably, a DC “pump”
is necessary to maintain a stable background
particle density and a well-defined ground
state, whereas the “probe” needs to be AC
modulated to isolate the responses of particles
created by the probe (34).
Correlated insulator of bosons
With the capability of tuning charge and exciton
density through electrostatic gating and pump
light, respectively, we fully explore the phase
diagram spanned by the charge filling ne and
exciton filling nex. We start with the axes, i.e.,
nex = 0 or ne = 0. Figure 1, C and E, show the PL
and absorption spectra, respectively, of a 60°-
aligned moiré bilayer (device D1) at nex = 0
(zero pump intensity) and ne ≥ 0 (see fig. S2 for
complete doping dependence). At charge neu-
trality, the PL features a single peak at 1.43 eV
from interlayer exciton emission (6, 8), whereas
the absorption shows three peaks from moiré
intralayer excitons (13). At approximately ne =
1 and 2 [ne = 1 corresponds to the moiré density
n0 = 2.1 × 1012 cm−2 (34)], the emission peak
blueshifts suddenly, and the absorption peaks
show a kink. These features originate from the
emergence of insulating states at integer elec-
tron fillings, which has been independently
confirmed by capacitance and microwave im-
pedance measurements (3, 6). The PL energy
jump at ne = 1 can be intuitively understood
from the emergence of a fermionic correlated
insulator state in which all available sites are
occupied by one electron; any additional exci-
tons injected are therefore forced into a higher-
energy state. We will discuss the mechanisms
of these spectral changes in more depth in the
discussion section.
We now turn to the case of ne = 0 by fixing
the gate voltage Vg at charge-neutral −0.5 V.
Figure 1, D and F, show the dependence of
pump-probe PL and absorption spectra on
the pump light intensity, which effectively con-
trols nex. To account for the nonlinear depen-
dence of nex on pump intensity, we also show
on the right axes dipolar-interaction–induced
interlayer exciton energy shift Ddipole (the ener-
gy shift of peak I relative to zero pump intensity;
Xiong et al., Science 380, 860–864 (2023)
26 May 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
A
WSe2
WS2
B
C
)
V
(
t
e
a
G
3
2
1
0
PL Intensity
(a.u.)
E
600
300
0
)
V
(
t
e
a
G
2
e
r
o
t
c
a
F
g
1
n
i
l
l
i
F
3
2
1
0
1.40
1.45
Photon Energy (eV)
0
1.50
1.6
1.7
Photon Energy (eV)
1.6
1.2
0.8
0.4
D
)
2
/
m
µ
W
µ
(
y
t
i
s
n
e
t
n
I
p
m
u
P
PL Intensity
(a.u.)
8
600
300
0
)
V
e
m
4
(
l
e
o
p
d
i
1.6
1.2
0.8
0.4
F
)
2
/
m
µ
W
µ
(
y
t
i
s
n
e
t
n
I
p
m
u
P
R/R
0.2
-0.05
-0.3
R/R
0.2
-0.05
-0.3
2
e
r
o
t
c
a
F
g
1
n
i
l
l
i
F
0
1.8
8
)
V
e
m
4
(
l
e
o
p
d
i
0
1.40
1.45
Photon Energy (eV)
0
1.50
0
1.6
1.7
Photon Energy (eV)
0
1.8
Fig. 1. Bosonic correlated insulator. (A) Illustration of a bosonic correlated
insulator consisting of interlayer excitons. Magenta spheres indicate holes
and cyan spheres, electrons. (Inset) Type II band alignment of WSe2/WS2
heterostructure. (B) Schematics of continuous-wave pump probe spectroscopy.
The exciton and electron density are independently controlled by pump light
and electrostatic gate. Red and green shading correspond to wide-field pump
light and focused probe light, respectively. (C and E) Gate-dependent PL (C)
and absorption (E) spectra of a 60°-aligned WSe2/WS2 moiré bilayer (device D1)
at zero pump intensity. The PL peak shows a sudden blue shift at electron
filling ne= 1 and 2 (yellow arrows), where the absorption spectrum shows kinks
and splitting. (D and F) Pump intensity–dependent PL (D) and absorption (F)
spectra of device D1 at charge neutrality. Right axes show dipolar-interaction–
induced interlayer exciton energy shift Ddipole, which is approximately
proportional to nex. The dominant PL peak in (D) at low and high pump intensity
are labeled as peak I and II, respectively. All measurements are performed
at a base temperature of 1.65 K.
see Fig. 1D), which is approximately propor-
tional to nex (34–36). Notably, a jump in the
PL energy is observed here as well. This jump
is well reproduced in two other 60°-aligned
moiré bilayers and one 0°-aligned moiré bi-
layer but is absent in a slightly misaligned bi-
layer (34), indicating that its origin is from
correlation effects. The qualitative similarity
to the gate-dependent PL suggests a similar
origin behind the exciton energy jump: the
emergence of a particle lattice that occupies
all available sites. The most natural candidates
are interlayer excitons, which are the immedi-
ate products of pump light absorption, and
interlayer charge transfer in a type II hetero-
junction. To elucidate the nature of the lattice,
we compare its effects on PL and absorption
spectra to those from an electron lattice. PL
spectra probe interlayer exciton responses. Al-
though both the gate-induced electron lattice
(Fig. 1C) and pump-induced lattice (Fig. 1D)
lead to a jump in interlayer exciton energy in
PL, the amplitudes are quite different: 35 meV
and 15 meV, respectively. A more prominent dif-
ference is observed in the absorption spectra,
which examine how intralayer excitons re-
spond to the induced lattices. Electrons induce
rich features, such as shifting, merging, and
splitting of exciton resonances (Fig. 1E). By con-
trast, the pump-injected particles have rather
weak effects on intralayer excitons, only slight-
ly decreasing their oscillator strength (Fig. 1F).
These distinctive behaviors indicate that the
pump-induced lattice is not formed by electrons.
To further confirm the exciton nature of the
pump-induced lattice, we move away from the
axes and investigate the phase diagram at nex >
0 and ne > 0. Figure 2A shows a set of pump
intensity-dependent PL spectra at different
gate voltages, corresponding to increasing elec-
tron density in the WS2 layer. All plots have
qualitatively similar behaviors—interlayer ex-
citon energy shows a jump. Upon increasing
ne, the jump occurs earlier and earlier, even-
tually appearing at zero pump intensity when
ne = 1 (Vg = 0.8 V), consistent with the gate-
injected electrons having already occupied all
sites at ne = 1. On the basis of these observa-
tions, the pump-injected particles will occupy
sites previously available to free electrons in
WS2 and form a mixed lattice together with
free electrons. Besides an electron itself, which
has been already excluded, the only other can-
didate for the pump-injected particle is the
interlayer exciton composed of an electron in
the WS2 layer and a hole in the WSe2 layer.
We, therefore, conclude that the pump light
is creating interlayer excitons and tuning nex.
At charge neutrality, the observed jump in PL
spectra then corresponds to a sudden increase
in the interlayer exciton energy when increas-
ing its density, i.e., an incompressible state
of the interlayer exciton. Moreover, this state
connects smoothly into the ne= 1 electron
correlated insulator, indicating one particle
per moiré site along the entire path and nex=
1 in the limit of sole occupation by excitons
[we have also independently calibrated the
exciton density (34)]. Our observation of an
incompressible exciton state at nex= 1 (half-
filling of an exciton band considering the
valley degeneracy) and charge neutrality is a
hallmark of a correlated insulator purely made
of excitons. This correlated insulator could be
a bosonic Mott insulator (28, 37), a generalized
Wigner crystal (38, 39), or a charge transfer
insulator (40). We unambiguously exclude the
charge transfer insulator scenario through val-
ley (flavor)–resolved measurements (34). Esti-
mations of rs (ratio of Coulomb interaction to
kinetic energy) and the Mott criterion suggest
Xiong et al., Science 380, 860–864 (2023)
26 May 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
0
200
400
A
600
PL (a.u.)
-0.3
B
0.2
-0.4V
-0.2V
0V
0.2V
0.4V
0.6V
0.8V
0µW/µm2
0.05µW/µm2
0.1µW/µm2
0.2µW/µm2
1.6
0.6µW/µm2
1.7
Photon
Energy (eV)
C
0.04
0.03
)
V
e
(
x
e
0.02
0.01
0.8V
0.6V
0.5V
0.4V
0.3V
0.1V
-0.1V
-0.3V
-0.5V
tot=1
0.00
0.0
0.5
1.0
1.5
Pump Intensity (µW/µm2)
0µW/µm2
0.05µW/µm2
0.1µW/µm2
0.2µW/µm2
D
0.15
0.00
V
d
/
)
R
R
/
(
d
-0.15
0.6µW/µm2
0.8µW/µm2
1.6µW/µm2
e=1
1.4
1.45
Photon
Energy (eV)
1.5
0
0.8
Pump Intensity (µW/µm2)
1.6
1.8
0
1.6µW/µm2
1
Gate (V)
2
0
1
Gate (V)
2
Fig. 2. Mixed correlated insulator. (A) Pump intensity–dependent PL spectra at
gate voltages Vg from −0.4 V (near charge neutrality) to 0.8 V (electron-one-filling).
(B) Gate-dependent absorption spectra at pump intensities from 0 to 1.6 mW/mm2.
Blue arrows mark the kink and splitting in absorption peaks at ne = 1. (C) Power-
dependent interlayer exciton energy change DΕex at representative gate voltages.
Green triangles mark middles of the transitions, which appear at smaller pump
intensity (nex) with increasing gate voltage (ne), consistent with a mixed correlated
insulator state at ntot = 1. (D) First-order derivative of absorption spectra with respect
to gate voltage at 1.76 eV under different pump intensities. Blue triangles denote
ne = 1. All measurements are performed at a base temperature of 1.65 K in device D1.
a bosonic Mott insulator nature of our obser-
vation (34, 38, 39).
Mixed correlated insulator
½
ð
ð
Þ= I1 þ I2
To quantify the jump in interlayer exciton en-
ergy, we fit PL spectra with two Lorentzian
peaks (34) and compute the exciton energy
change DEex ¼ E1I1 þ E2I2
(cid:3),
Þ (cid:2) E1
which can be considered as an effective exci-
ton chemical potential. Here, E1 (E2) and I1 (I2)
correspond to the energy and amplitude of the
PL peak I and II, respectively (Fig. 1D). Figure
2C summarizes the evolution of DΕex with pump
intensity at representative gating, where a tran-
sition in exciton energy is clearly observed at
all gate voltages. The transition appears not
particularly sharp around charge neutrality,
largely because of the nonlinear dependence
of exciton density on pump intensity (Fig. 1D).
In addition, the transition is broadened by
spatial inhomogeneity in the exciton density
that is expected to be much larger than the
charge case (34). At charge neutrality Vg = −0.5 V,
we determine the position of nex = 1 from the
middle point of the DΕex transition. We also
independently determine the exciton density at
this point to be (2 T 0.2) × 1012 cm−2 from time-
resolved measurements (34), which matches
well with the expected density of moiré cells
n0 = 2.1 × 1012 cm−2. At finite electron den-
sity, the transition happens when electrons
and excitons cooperate to occupy all availa-
ble sites. The transition points therefore cor-
respond to ntot = ne + nex = 1 (green triangles
in Fig. 2C) and keep shifting to a lower pump
intensity until reaching nex = 0 when ne = 1.
To separate ne and nex better, we also mea-
sure absorption spectra of the moiré super-
lattice while varying gate voltage and pump
intensity. Because intralayer excitons are only
sensitive to ne but not nex (Fig. 1, E and F), we
use spectral changes in absorption to indepen-
dently determine ne. Figure 2B shows the gate-
dependent absorption spectra of the moiré
bilayer at different pump intensities. All plots
show similar behaviors except for a slight shift
in gate voltage. This can be seen more clearly
in the gate-differentiated absorption spectra at
representative pump intensity (Fig. 2D) (34).
Blue triangles in Fig. 2, B and D, label the
position of “kinks” that correspond to ne = 1.
With both ntot and ne extracted, Fig. 3A sum-
marizes the phase diagram spanned by ne and
nex at a base temperature of 1.65 K. The color
scale represents DΕex; the boundaries of ntot =
1 and ne = 1 are overlaid on the phase di-
agrams. The ne = 1 line shifts slightly toward
lower gate voltage at high pump intensity, pre-
sumably from photocarriers generated during
the relaxation of excitons. However, this effect
is rather weak, and the ne = 1 line deviates far
from the ntot = 1 line where transitions in DΕex
are observed. This further confirms that the
transition at ntot = 1 is not from an electron
lattice. Instead, it originates from a mixed lat-
tice composed of both excitons and electrons.
Our observations, therefore, suggest a mixed
correlated insulator along the ntot = 1 line,
which smoothly connects into a bosonic (fer-
mionic) correlated insulator at the end point
of nex = 1 (ne = 1).
We further obtain the phase diagrams at 15
and 30 K (Fig. 3, B and C; see fig. S9 to S12 for
details). A sharp transition in DΕex is always
observed at ne = 1, consistent with the strong
charge correlation and a high melting temper-
ature of charge correlated insulator in WSe2/
WS2 moiré superlattices (3, 4). Similarly, the
behavior at nex = 1 remains largely unchanged
with temperature, indicating survival of the
bosonic correlated insulator to above 30 K. By
Xiong et al., Science 380, 860–864 (2023)
26 May 2023
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
1.5
1.65K
e=1
tot=1
1.0
0.5
A
)
2
/
m
µ
W
µ
(
y
t
i
s
n
e
t
n
I
p
m
u
P
B
)
2
/
m
µ
W
µ
(
y
t
i
s
n
e
n
t
I
p
m
u
P
0.0
-0.5
D
0.03
)
V
e
(
x
e
E
0.02
0.01
0.00
0.0
Gate (V)
0.5
1.65K
-0.5V
-0.1V
0.0
0.5
1.0
1.5
Pump Intensity (µW/µm2)
15K
1.5
e=1
1.0
0.5
0.0
-0.5
0.0
Gate (V)
0.5
C
)
2
/
m
µ
W
µ
(
y
t
i
s
n
e
t
n
I
p
m
u
P
30K
2
1
0
-0.5
E
ex (eV)
0.03
e=1
0.02
0.01
0
0.0
Gate (V)
0.5
E
0.03
)
V
e
(
x
e
E
0.02
0.01
0.00
15K
-0.5V
-0.1V
30K
-0.5V
-0.1V
F
0.03
)
V
e
(
x
e
E
0.02
0.01
0.0
0.5
1.0
1.5
Pump Intensity (µW/µm2)
0.00
0.0
0.5
1.0
2.0
Pump Intensity (µW/µm2)
1.5
2.5
Fig. 3. Phase diagram. (A to C) DΕex with respect to the gate voltage and pump
intensity at 1.65 K (A), 15 K (B), and 30 K (C) measured on device D1. The
boundaries of ntot = 1 and ne = 1, as determined from the pump-probe PL and
absorption measurements, respectively, are highlighted with green and blue triangles.
The clear separation between the two boundaries confirms that the transition of
DΕex at ntot = 1 is not from a charge correlated insulator state until close to ne = 1.
Regions of high pump intensity and/or high gating are not shown because peak I
has already disappeared and cannot be fitted reliably (34). (D to F) Vertical line-
cuts of the phase diagrams for Vg = −0.5 V (charge neutrality) and −0.1 V (electron
doped) at 1.65 K (D), 15 K (E), and 30 K (F). Whereas at 1.65 K, DΕex rises faster
for Vg = −0.1 V than −0.5 V, at 30 K it rises slower for Vg = −0.1 V than −0.5 V,
suggesting partial melting and lower stability of the mixed correlated insulator state.
Electron Density ne (1012 cm-2)
1.0
0.5
1.5
vtot=1
0.0
1.0
A
x
e
ν
Fig. 4. Mixed Hubbard model. (A) DΕex
phase diagram measured on another 60°-aligned
moiré bilayer (device D2) at 1.65 K with
calibrated nex and ne. Green (ntot = 1) and blue
(ne = 1) dashed lines are boundaries predicted by
a mixed Hubbard model, where DΕex is expected
to jump. Green triangles label the experimental
50% transition point in DΕex, which matches well
with the prediction. Error bars denote the range
between 20 and 80% transition in DΕex. (B to
D) Schematics showing the energy required to
add one interlayer exciton (red) into the system
for region I (B), II (C), and III (D) in the phase
diagram. Additional energy cost of DEex ¼ Uex(cid:2)ex
and Ue(cid:2)ex is required in region II and III, respectively,
from on-site repulsion between excitons and
between electron-exciton.
g
n
i
l
l
i
F
n
o
t
i
c
x
E
0.5
0.0
0.0
E
ex (eV)
0.03
0.02
0.01
0.00
B
C
D
E
ex = 0
ex = U
E
ex–ex
2.0
2.0
1.5
E
x
c
i
t
o
n
D
e
n
s
i
t
y
1.0
n
e
x
(
1
0
1
2
c
m
-
2
)
0.5
0.0
0.5
Electron Filling νe
1.0
ex = U
E
e–ex
contrast, for the regions in between, the tran-
sition in DΕex becomes much slower than at
base temperature. This can be seen clearly by
comparing the DΕex evolution at charge neu-
tral (Vg = −0.5 V) and finite electron density
(Vg = −0.1 V) (Fig. 3, D to F). At base tem-
perature the finite ne case shows a sharper rise
than charge neutral (Fig. 3D), which can be
naturally explained by the smaller nex needed
to reach ntot = 1 and therefore less inhomoge-
neous broadening from exciton density. Notably,
at 30 K the rise in DΕex becomes smoother at Vg =
−0.1 V compared to charge neutral, indicating
increasing exciton compressibility and partial
melting of the mixed correlated insulator. This
observation suggests that the mixed correlated in-
sulator is less stable than both components (34).
Discussion
To allow direct comparison with theory, we
calibrate nex throughout the phase diagram by
Xiong et al., Science 380, 860–864 (2023)
26 May 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
time-resolved measurements (34) (Fig. 4A).
The DΕex jumps (green triangles) indeed occur
at nex + ne = 1 (green dashed line), further sup-
porting the mixed correlated insulator origin.
The experimental phase diagram matches very
well with predictions from a two-component
Hubbard model with both fermionic and bo-
sonic species (34). Along the horizontal axis
(nex = 0), DΕex shows a sudden jump at ne = 1.
This can be understood because excitons always
avoid electron-occupied sites owing to the on-
site electron-exciton repulsion [Fig. 4B; see
discussions in (34)] until all sites are electron-
occupied at ne = 1, after which adding an
exciton necessarily pays the additional energy
cost of Ue-ex (Fig. 4D). We thereby determine
Ue-ex to be 32 to 40 meV (varies between sam-
ples). Such doping dependence is distinctively
different from that in monolayer TMDC or
TMDC bilayers where strong correlation is not
observed, such as WSe2/MoSe2 (20, 21). In
these systems, electron doping immediately
leads to the formation of trion and a corre-
sponding emission peak in PL at lower energy
(20, 21). In WSe2/WS2, by contrast, the emis-
sion spectrum remains largely unchanged until
ne = 1, after which the PL shows a blueshift of
~35 meV and ~20 meV in 60° and 0° twisted
moiré superlattices, respectively (6, 8). Our
model provides a natural explanation to this
widely observed yet not fully understood PL
doping dependence.
Similarly, DΕex remains 0 along the vertical
axis until nex = 1, after which adding an exciton
requires an additional energy cost of Uex-ex
(Fig. 4C). We thereby determine Uex-ex to be
14 to 20 meV (varies between samples). Away
from the axes, DΕex remains 0 for nex + ne < 1
(region I in Fig. 4A) because added excitons
can find an empty site to avoid on-site repul-
sion. For nex + ne ≥ 1 but ne < 1 (region II), added
excitons will prefer to stay on an exciton site
because Uex-ex < Ue-ex, and the additional en-
ergy cost is DΕex =Uex-ex. For ne ≥1 (region III),
all sites are already occupied by electrons and
added excitons can only reside on an electron
site; therefore DΕex = Ue-ex.
The above analysis ignores the hopping
term or the two pseudospins of excitons from
the K and K′ valleys (25). Nevertheless, it cap-
tures all salient features of the experiments.
Several interesting theoretical questions re-
main to fully understand the phase diagram,
such as on the effects of finite hopping, as
well as the microscopic mechanism of exciton-
exciton interaction and its dependence on
exciton pseudospin, for which our results
provide a valuable experimental reference.
RE FERENCES AND NOTES
1. D. M. Kennes et al., Nat. Phys. 17, 155–163 (2021).
2. Y. Cao et al., Nature 556, 43–50 (2018).
3. E. C. Regan et al., Nature 579, 359–363 (2020).
4. Y. Tang et al., Nature 579, 353–358 (2020).
5. Y. Shimazaki et al., Nature 580, 472–477 (2020).
6. S. Miao et al., Nat. Commun. 12, 3608 (2021).
7. X. Huang et al., Nat. Phys. 17, 715–719 (2021).
8. E. Liu et al., Phys. Rev. Lett. 127, 037402 (2021).
9. C. Jin et al., Nat. Mater. 20, 940–944 (2021).
10. A. Ghiotto et al., Nature 597, 345–349 (2021).
11. T. Li et al., Nature 597, 350–354 (2021).
12. T. Li et al., Nature 600, 641–646 (2021).
13. C. Jin et al., Nature 567, 76–80 (2019).
14. K. L. Seyler et al., Nature 567, 66–70 (2019).
15. K. Tran et al., Nature 567, 71–75 (2019).
16. E. M. Alexeev et al., Nature 567, 81–86 (2019).
17. L. Zhang et al., Nature 591, 61–65 (2021).
18. H. Yu, G.-B. Liu, J. Tang, X. Xu, W. Yao, Sci. Adv. 3,
e1701696 (2017).
19. F. Wu, T. Lovorn, A. H. MacDonald, Phys. Rev. Lett. 118,
147401 (2017).
20. X. Wang et al., Nat. Nanotechnol. 16, 1208–1213 (2021).
21. E. Liu et al., Nature 594, 46–50 (2021).
22. X. Wang et al., Nature 604, 468–473 (2022).
23. J. Gu et al., Nat. Phys. 18, 395–400 (2022).
24. Z. Zhang et al., Nat. Phys. 18, 1214–1220 (2022).
25. P. Rivera et al., Nat. Nanotechnol. 13, 1004–1015
(2018).
26. S. Sugawa et al., Nat. Phys. 7, 642–648 (2011).
27. C. Lagoin, S. Suffit, K. Baldwin, L. Pfeiffer, F. Dubin, Nat. Phys.
18, 149–153 (2022).
28. M. P. A. Fisher, P. B. Weichman, G. Grinstein, D. S. Fisher,
Phys. Rev. B Condens. Matter 40, 546–570 (1989).
29. F. P. Laussy, A. V. Kavokin, I. A. Shelykh, Phys. Rev. Lett. 104,
106402 (2010).
30. P. Rivera et al., Nat. Commun. 6, 6242 (2015).
31. W. Li, X. Lu, S. Dubey, L. Devenica, A. Srivastava, Nat. Mater.
19, 624–629 (2020).
32. Y. Cao et al., Nature 556, 80–84 (2018).
33. S. Luryi, Appl. Phys. Lett. 52, 501–503 (1988).
34. See additional text and data in the supplementary materials.
35. P. Nagler et al., Mater 4, 025112 (2017).
36. Z. Wang, Y. H. Chiu, K. Honz, K. F. Mak, J. Shan, Nano Lett. 18,
137–143 (2018).
37. M. Greiner, O. Mandel, T. Esslinger, T. W. Hänsch, I. Bloch,
Nature 415, 39–44 (2002).
38. B. Padhi, C. Setty, P. W. Phillips, Nano Lett. 18, 6175–6180
(2018).
39. B. Padhi, R. Chitra, P. W. Phillips, Phys. Rev. B 103, 125146
(2021).
40. Y. Zhang, N. F. Q. Yuan, L. Fu, Phys. Rev. B 102, 201115
(2020).
41. R. Xiong, Data repository for: Correlated insulator of excitons
in WSe2/WS2 moire superlattices, Open Science Framework
(2023).
AC KNOWLED GME NTS
R.X. thanks T. Xie, C. Tschirhart, A. Potts, Y. Guo, W. Gong, and
W. Y. Xing for experimental help. Funding: C.J. acknowledges
support from Air Force Office of Scientific Research under award
FA9550-23-1-0117. R.X. acknowledges support from the UC
Santa Barbara NSF Quantum Foundry funded via the Q-AMASE-i
program under award DMR-1906325. S.T. acknowledges
support from DOE-SC0020653 (materials synthesis), NSF
DMR-2206987 (elemental purification), NSF ECCS 2052527
(electrical characterization for crystal optimization), DMR 2111812
(excitonic characterization during growth optimization), and
CMMI 2129412 (large size crystal development). K.W. and T.T.
acknowledge support from the JSPS KAKENHI (grant nos.
19H05790, 20H00354, and 21H05233). Author contributions:
C.J. conceived and supervised the project. R.X. and J.H.N.
fabricated the devices. R.X. and S.L.B performed the optical
measurements and analyzed the data. P.H., R.S., and S.T. grew
the WSe2 and WS2 crystals. K.W. and T.T. grew the hBN crystals.
C.J. and R.X. wrote the manuscript with the input from all the
authors. Competing interests: The authors declare no competing
interests. Data and materials availability: All data in the
main text and supplementary materials, as well as the code for
peak fitting, are available from Open Science Framework (41).
License information: Copyright © 2023 the authors, some
rights reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government
works. https://www.sciencemag.org/about/science-licenses-
journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add5574
Materials and Methods
Supplementary Text
Figs. S1 to S16
References (42–63)
Submitted 24 June 2022; resubmitted 26 September 2022
Accepted 27 April 2023
Published online 11 May 2023
10.1126/science.add5574
Xiong et al., Science 380, 860–864 (2023)
26 May 2023
5 of 5
|
10.1126_science.add7150
|
RES EARCH
NEUROSCIENCE
Orphan receptor GPR158 serves as a metabotropic
glycine receptor: mGlyR
Thibaut Laboute1, Stefano Zucca1, Matthew Holcomb2, Dipak N. Patil1†, Chris Garza2,
Brittany A. Wheatley3, Raktim N. Roy3, Stefano Forli2, Kirill A. Martemyanov1*
Glycine is a major neurotransmitter involved in several fundamental neuronal processes. The identity
of the metabotropic receptor mediating slow neuromodulatory effects of glycine is unknown. We identified an
orphan G protein–coupled receptor, GPR158, as a metabotropic glycine receptor (mGlyR). Glycine and a related
modulator, taurine, directly bind to a Cache domain of GPR158, and this event inhibits the activity of the
intracellular signaling complex regulator of G protein signaling 7–G protein b5 (RGS7-Gb5), which is associated
with the receptor. Glycine signals through mGlyR to inhibit production of the second messenger adenosine
3′,5′-monophosphate. We further show that glycine, but not taurine, acts through mGlyR to regulate neuronal
excitability in cortical neurons. These results identify a major neuromodulatory system involved in mediating
metabotropic effects of glycine, with implications for understanding cognition and affective states.
GPR158 is one of the most abundant orphan
GPCRs in the brain that transduces signals by
coupling to RGS proteins (25, 26). In neurons,
it regulates signaling to the second messenger
adenosine 3′,5′-monophosphate (cAMP) and
controls key ion channels, kinases, and neuro-
trophic factors involved in neuronal excitability
and synaptic transmission (25, 27). Accord-
ingly, GPR158 has been heavily implicated in
cognition and affective states (25, 28, 29). Ge-
netic suppression of GPR158 in mice results
in a prominent antidepressant phenotype and
stress resiliency, making GPR158 an attract-
ive target for development of new antidepres-
sants (25).
The endogenous ligand for GPR158 remains
unknown. Recent structures of GPR158 revealed
the presence of an extracellular Cache domain,
a putative ligand-binding module (30, 31).
G lycine is the simplest amino acid ubiqui-
tously present in all mammalian tissues.
Glycine serves as an inhibitory neuro-
transmitter, but it can be excitatory in
developing neurons (1, 2). Glycinergic
neurons are distributed across the brain; how-
ever, glycine can also be released by glial cells
(3). Known receptors for glycine belong to the
family of pentameric ligand-gated ion chan-
nels (4). Glycine also serves as a coagonist of
N-methyl-D-aspartate (NMDA) receptors (5).
Metabotropic neuromodulatory effects of gly-
cine have been observed (6, 7), but no recep-
tors mediating these actions have been found.
Glycine has distinct effects on neural circuits
(3), and glycinergic transmission has been im-
plicated in pathological conditions, including
depression (8–10).
Metabotropic neuromodulation in the ner-
vous system is mediated mainly by heterotrimeric
GTP-binding protein (G protein)–coupled re-
ceptors (GPCRs). GPCRs play essential roles
in neuronal physiology and pathology and
present targets for drug development (11).
Canonically, GPCRs transduce their signals by
activating heterotrimeric G proteins (12, 13).
However, G protein–independent modes of
signal transduction triggered by the recruit-
ment of b-arrestins and other scaffolds to
activated GPCRs have also been described
(14–16). G protein signaling is controlled by
regulator of G protein signaling (RGS) pro-
teins, which facilitate their deactivation (17).
RGS proteins also interact with several GPCRs
(18–22).
1Department of Neuroscience, UF Scripps Biomedical
Research, Jupiter, FL 33458, USA. 2Department of Integrative
Structural and Computational Biology, The Scripps Research
Institute, La Jolla, CA 92037 USA. 3Department of Integrative
Structural and Computational Biology, UF Scripps Biomedical
Research, Jupiter, FL 33458, USA.
*Corresponding author. Email: kmartemyanov@ufl.edu
†Present address: Lilly Biotechnology Center, Eli Lilly and
Company, 10290 Campus Point Dr., San Diego, CA 92121, USA.
GPCRs mediate the effects of all major
neurotransmitters except glycine and taurine.
However, many GPCRs still have no identified
endogenous ligands. Orphan GPCRs may have
potential for obtaining insights into physiology
and for drug development (23, 24).
Results
Glycine signals through GPR158 to regulate cAMP
The structure of GPR158 revealed the pres-
ence of a Cache domain, which serves as a
ubiquitous ligand-binding module in bac-
terial chemoreceptors (30). We found that
Fig. 1. Identifica-
tion of glycine
as GPR158
ligand. (A) Three-
dimensional model
of the GPR158
Cache domain
(cyan) with puta-
tive ligand-binding
pocket (orange).
(B) Schematic of
the screening
assay design.
(C) Quantification
of cAMP changes
mediated by
GPR158. BRET
signal in control
cells is subtracted
from the signal
from cells
expressing
GPR158, and the
difference is
plotted. Dotted
lines denote 2× SD
confidence inter-
val. Data represent
mean ± SEM
determined from
three independent
experiments per-
formed in tripli-
cate. (D) Time course of cAMP change induced with 10 mM forskolin in U87 MG glioblastoma cells in response
to glycine (100 mM) application (arrow). (E) Quantification of maximal amplitudes of responses to glycine
in (D). Data represent mean ± SEM determined from three independent experiments performed in triplicate.
**p < 0.01, one-way analysis of variance (ANOVA).
Laboute et al., Science 379, 1352–1358 (2023)
31 March 2023
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
GPR158 Cache domain had a small pocket
with organization similar to that of the amino
acid binding pocket in other Cache domains
(Fig. 1A). We hypothesized that GPR158 may
have an amino acid ligand. We screened a
library of amino acids for their ability to
alter GPR158-mediated signaling. Because
GPR158 has been linked to regulation of cAMP
in the brain (25, 27), we used a bioluminescence
resonance energy transfer (BRET)–based cAMP
biosensor (32) (Fig. 1B). Out of all amino acids
tested, only glycine showed significant decrease
in cAMP when applied to human embryonic
kidney (HEK) 293 cells expressing GPR158
relative to nontransfected cells (Fig. 1C).
To study this effect in more detail, we ana-
lyzed the individual responses to glycine in a
kinetic mode. We found that glycine appli-
cation to U87 glioblastoma cells expressing
GPR158 resulted in cAMP decrease. No glycine-
induced changes in cAMP were observed in
cells lacking GPR158 (Fig. 1, D and E). This
inhibitory effect of GPR158 was further po-
tentiated by coexpressing RGS7-G protein b5
(RGS7-Gb5), suggesting that GPR158 signals
by means of this protein complex to affect
cAMP levels (Fig. 1, D and E).
We further tested the effect of taurine, a
compound closely related to glycine, which
binds to several common receptors (33), in-
cluding ionotropic glycine receptors (34). Tau-
rine caused a significant decrease in cAMP
levels only in HEK293 cells expressing GPR158
(fig. S1, A and B). Again, this effect was poten-
tiated by coexpressing RGS7-Gb5, suggesting
that these proteins act in complex with GPR158
in mediating the effects of taurine. However,
when compared directly, the effect of taurine
on GPR158-mediated suppression of cAMP
was weaker than the effect of glycine (fig. S1C).
Glycine inhibits modulation of RGS7-Gb5
by GPR158
To understand how glycine action on GPR158
regulates intracellular cAMP, we focused on
GPR158 interaction with RGS7-Gb5, an es-
tablished guanosine triphosphatase (GTPase)–
activating protein (GAP) for the Gai/o proteins
(35) known to regulate cAMP production (26).
We used a cell-based assay to monitor GAP ac-
tivity by following kinetics of G protein deacti-
vation (36) (Fig. 2A). In this assay, activation
of G proteins by GPCR stimulation generates
the BRET signal upon interaction of liberated
Venus-Gbg subunits with the masGRK3CT-
Nluc reporter. This signal is quenched when
Ga deactivation is triggered by GPCR antag-
onism and recombines with Venus-Gbg to form
inactive heterotrimer. As previously reported
(22), we found that introduction of RGS7-Gb5
accelerated deactivation of its substrate, Gao
(Fig. 2, B and D). Application of glycine had no
effect on either baseline Gao deactivation or
the RGS7-Gb5-assisted process (Fig. 2, B and
Fig. 2. Glycine and taurine slow deactivation of Gao by GPR158-RGS7-Gb5 complex. (A) Schematics of
the BRET-based GAP assay. G proteins are activated at t = 0 s by stimulating GPCR (dopamine DR receptor, 0.1 mM).
After reaching steady state, the GPCR activity is terminated by injection of haloperidol (0.1 mM) at t = 15 s (arrow).
G protein deactivation is then monitored by following quenching of the BRET signal. (B and C) Traces of BRET
signal showing Gao activation and deactivation time course with or without glycine or taurine (100 mM) treatment in
cells without GPR158 (B) or cells transfected with GPR158 (C). (D) Quantification of deactivation time constant
of the reactions presented in (B) and (C). 1/t is calculated from deactivation curves of n = 5 independent
experiments conducted in triplicate from each cell transfection group. Data represent mean ± SEM. ****p < 0.0001,
ns (not significant) = p > 0.05, two-way ANOVA. (E) Dose–response profile of changes in GAP activity (KGAP)
calculated by subtracting the baseline deactivation rate (1/t) from the rate of the reaction in the presence of
GPR158-RGS7-Gb5. Data represent mean ± SEM of n = 4 independent experiments conducted in triplicate.
D). However, when GPR158 was coexpressed
together with RGS7-Gb5, glycine significantly
decelerated Gao deactivation (Fig. 2, C and D),
suggesting that it specifically inhibited the
GAP activity of RGS7-Gb5 by engaging GPR158.
Dose–response studies showed that the me-
dian inhibitory concentration (IC50) of glycine
on GPR158 is ~3 mM (Fig. 2E). Taurine dis-
played a similar inhibitory effect on Gao de-
activation only in cells coexpressing GPR158
Laboute et al., Science 379, 1352–1358 (2023)
31 March 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Direct interaction of glycine with GPR158. (A) Schematics of assay design for
detecting glycine binding to GPR158 by flow cytometry. (B) Flow cytometry histogram
showing distribution of cellular populations after sorting. (C) Quantification of FITC-glycine
binding detected in flow cytometry experiments. The median of fluorescence (MFI) is
quantified and plotted. Error bars indicate SEM, n = 3, **p < 0.01, ****p < 0.0001, two-way
ANOVA. (D) Schematics of the radioligand binding assay. (E) Quantification of [3H]glycine
binding to membrane expressing GPR158. Data show mean of four independent
experiments; error bars indicate SEM, n = 4. (F) Scatchard plot of the [3H]glycine radio-
ligand binding assay. Data show mean of four independent experiments; error bars indicate
SEM, n = 4. (G) Schematics of the assay design detecting glycine binding to GPR158
by isothermal titration calorimetry (ITC) with purified protein. (H) ITC binding profile
showing glycine binding to GPR158 in the initial run with fresh sample. (I) Quantification of
binding determined by fitting the integrated isotherm to an independent binding model.
Data show mean of four experimental runs; error bars indicate SEM.
with RGS7-Gb5, but with a lower IC50 of ~6 mM
(Fig. 2, B to E).
We further tested whether glycine or taurine
could induce GPR158 to activate G proteins as
canonical GPCRs do (fig. S2A). We observed no
significant activation of any G proteins tested
with either glycine or taurine (fig. S2, B to I).
We also tested whether glycine could induce
b-arrestin recruitment to GPR158 using a BRET
assay and obtained no significant response (fig. S3).
GPR158 directly binds glycine
To confirm that GPR158 is a direct target of gly-
cine, we used several strategies. First, we de-
vised a flow cytometry–based assay to monitor
binding of fluorescein isothiocyanate (FITC)–
Laboute et al., Science 379, 1352–1358 (2023)
31 March 2023
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Probing Cache
domain of GPR158 as a li-
gand binding site. (A) Com-
putational docking of glycine
(teal) into putative ligand-
binding pocket on GPR158
Cache domain (green). Gly-
cine and directly interacting
residues are shown as sticks.
Hydrogen bonds (teal) and
van der Waals interactions
(orange) are shown as
dotted lines. (B) Diagram
showing interactions in the
docked model of glycine
against GPR158 Cache
domain. Hydrogen bonds are
shown as dashed lines,
van der Waals interactions
are shown as intersecting
semicircles, and the
secondary structural
context of Ser266 and
Tyr269 is shown as an arc.
Dagger indicates a residue
implicated in glycine binding
in (A). (C) Radioligand
binding assay of [3H]glycine
in cells expressing GPR158
mutants. Data show mean ±
SEM of three independent
experiments, *p < 0.05, ns =
p > 0.05, one-way ANOVA.
(D) Functional evaluation of
GPR158 mutants in GAP
BRET assay. Error bars indicate
mean ± SEM of three
independent experiments con-
ducted in triplicate. *p < 0.05,
**p < 0.01, ***p < 0.001,
****p < 0.0001, ns = p > 0.05, two-way ANOVA. (E) Quantification of glycine inhibitory effect on KGAP normalized to the effect seen with wild-type (WT) receptor. Error bars indicate
mean ± SEM of three independent experiments conducted in triplicate. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, ns = p > 0.05, one-way ANOVA. (F and G) Traces
of Gao deactivation time course upon glycine addition. Single-letter abbreviations for the amino acid residues are as follows: D, Asp; E, Glu; K, Lys; R, Arg; S, Ser; T, Thr; and Y, Tyr.
conjugated glycine to cells expressing GPR158
(Fig. 3A). When HEK293 cells expressing GPR158
were incubated with FITC-glycine, we observed
labeling of a significant population of cells (Fig.
3B). No such labeling was evident when FITC-
glycine was incubated with cells not trans-
fected with GPR158. Dose–response studies
further confirmed this binding and its selec-
tivity across the ranges of glycine used (Fig. 3C).
Next, we performed radioligand binding
assays examining binding of [3H]-labeled
glycine to HEK293 cells expressing GPR158
(fig. S4A). We detected significant binding of
[3H]glycine to GPR158-expressing cells across
concentrations (Fig. S4B). We isolated cellular
membranes and conducted classical radio-
ligand titration experiments (Fig. 3D). We
detected saturable [3H]glycine binding to
membranes containing GPR158 in substantial
excess over linear nonspecific binding to mem-
branes devoid of GPR158 (Fig. 3E). Scatchard
analysis (Fig. 3F) estimated the dissociation
constant, KD, of GPR158 for glycine to be ~3 mM.
Binding competition experiments directly com-
paring the ability of glycine and taurine to
displace [3H]glycine bound to GPR158 (fig. S5)
confirmed the specificity of glycine and tau-
rine binding to GPR158 and also revealed a
twofold lower affinity of taurine relative to
glycine (IC50: ~3 mM versus ~6 mM).
In addition, we examined binding of non-
labeled glycine directly to purified GPR158
using isothermal titration calorimetry (ITC)
(Fig. 3G). Titration experiments showed satu-
ration of the heat released upon glycine addi-
tion to GPR158 yielding KD ranging from 2 to
16 mM across experiments conducted first
with a fresh sample (Fig. 3, H and I) and sub-
sequently rerun after removal of glycine and
detergent (fig. S6). The affinity of glycine ob-
tained in direct binding experiments is in
good agreement with the affinity measured
in the functional GAP assays, indicating that
binding to glycine is responsible for changes
in GPR158 activity.
Glycine binds to Cache domain of GPR158 and
modulates GAP activity of RGS7-Gb5 complex
We performed molecular docking experiments
fitting glycine into a model of the putative
ligand-binding pocket in the Cache domain of
GPR158, built by supplementing experimental
structure (30) with a missing loop taken from
the AlphaFold2 prediction (Fig. 4A and table
S1). Although another structure of GPR158 is
available (31), it did not resolve side-chain
conformations and thus was not considered
Laboute et al., Science 379, 1352–1358 (2023)
31 March 2023
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. Effect of glycine on
neuronal excitability.
(A) Schematic of the
electrophysiological recordings
in slice preparation targeting
mPFC neurons of layer II and III.
Experiments were conducted
with picrotoxin (100 mM)
(blockade of GABAA receptors),
strychnine (1 mM) (antagonist
of glycine and acetylcholine
receptors), CNQX (6-cyano-7-
nitroquinoxaline-2,3-dione)
(20 mM) (AMPA receptor
antagonist), and APV (D,L-2-
amino-5-phosphonovaleric
acid) (50 mM) (NMDA receptor
antagonist). (B) Traces of
voltage responses to a
200-picoampere (pA) current
ramp injection under control
conditions and after bath
application of glycine (1 mM).
(C) Quantification of changes in excitability by number of action potentials fired in response to 200-pA current ramp
(n = eight neurons from five mice). (D) Quantification of changes in excitability by rheobase current under control
condition and glycine application (n = eight neurons from five mice). (E) Traces of voltage responses to a 200-pA
current ramp injection obtained from layer II and III pyramidal neurons in Gpr158 KO mice under control conditions and
during bath application of glycine (1 mM). (F) Quantification of changes in excitability by number of action potentials
(AP) fired in response to 200-pA current ramp (n = four neurons from three mice). (G) Quantification of changes in
excitability by rheobase current under control condition and glycine application (n = four neurons from three mice).
(H) Schematic representation of the proposed mechanism of glycine effects on mGlyR. In all graphs, nonparametric
t test; Wilcoxon test was used for statistical analysis, ns = p > 0.05, **p < 0.01.
as a source of alternate receptor conforma-
tions for docking. For the best-scored glycine
pose, glycine could be well accommodated in
a pocket where it is stabilized by a network
of hydrogen-bonding interactions with S172,
R173, E271, and D307 side chains, with the
charged side chains ideally positioned to sta-
bilize the carboxylate and amine moieties of
the zwitterion. These residues are embedded
in a web of other hydrophilic residues located
in a close vicinity (e.g., K264, S266, Y269, T284,
and K305) lining the pocket (Fig. 4B). Dock-
ing studies performed with taurine found a
cluster of poses that overall matched the
putative binding mode of glycine, retaining
the features described for glycine (fig. S7).
The size of the pocket is spatially constrained,
particularly by L282, in such a way that other
amino acids cannot be easily accommodated
without steric clashes with side chains of
residues lining the pocket, which provides a
possible explanation for the selectivity of the
recognition (fig. S8).
To test the role of the residues forming the
putative glycine pocket in the GPR158 Cache
domain, we performed site-directed mutagen-
esis. In radioligand binding assays, the R173A,
E271A, and Y269A mutants showed near com-
plete loss of [3H]glycine binding, confirming
the essential role of these residues in ligand
coordination (Fig. 4B). We then tested each
of the mutants in functional assays (Fig. 4, C
and D). Each of the mutants defective in gly-
cine binding also lost an ability to inhibit the
GAP activity of RGS7-Gb5 (Fig. 4E). The activ-
ity of the S266A mutant, which normally binds
glycine, was not regulated by it, suggesting
that some of the residues in the binding pocket
are involved in conformational transitions
triggered by ligand interaction (37). The mech-
anism by which mutating E271 residue resulted
in loss of glycine responsiveness also deviated
for that of other mutations. This mutant ex-
hibited a much slower deactivation kinetics in
the absence of glycine, generating a constitu-
tively inhibited receptor.
Glycine modulates neuronal excitability
through GPR158
Lastly, we assessed the impact of glycine mod-
ulation of GPR158 on neuronal activity. We
examined the intrinsic properties of layer II
and III neurons in the prelimbic cortex, where
GPR158 is prominently expressed (25) and
regulates neuronal excitability (27). The meta-
botropic effects of glycine are not well charac-
terized across the nervous system. Therefore,
we started by defining the effects of glycine on
layer II and III neurons. To isolate metabo-
tropic actions, we antagonized excitatory and
inhibitory synaptic drive with pharmacologi-
cal blockade and measured the current-voltage
relation in response to a depolarizing current
ramp. Application of glycine significantly in-
creased the number of action potentials while
decreasing the amount of current necessary to
elicit the first action potential (Fig. 5, A to C)
without changes in the resting membrane
potential (fig. S9). This excitatory effect of
glycine is distinct from its canonical inhibi-
tory action mediated by glycine receptor (GlyR)
ion channels. Interestingly, glycine application
did not produce any changes in the intrinsic
excitability of layer V neurons (fig. S9), which
do not express GPR158 (27).
To confirm the involvement of GPR158 in
the effects of glycine, we studied Gpr158 knock-
out (Gpr158 KO) mice. Glycine application
failed to alter excitability of layer II and III
neurons in prefrontal cortex of Gpr158 KO
mice (Fig. 5, D to F). We also tested the effect
of taurine on the excitability of layer II and
III neurons (fig. S11). These experiments re-
vealed small effects on neuronal firing in the
same direction as glycine. However, these ef-
fects did not reach the criteria for statistical
significance, possibly because of the lower
efficacy of taurine.
Discussion
In this study, we demonstrated that GPR158
serves as a metabotropic receptor for glycine.
We also report that GPR158 can be modulated
by taurine, which acts as a partial agonist for
this receptor. This finding was enabled by
recently obtained high-resolution structure
of the receptor, which revealed the presence
of a ligand-binding module: the Cache domain.
Cache domains are well-known receptors for
amino acids and other related small molecules
ubiquitously used by bacterial chemorecep-
tors. Only two GPCRs contain them, includ-
ing GPR158 and night-blindness associated
receptor GPR179 (38), whose ligand remains
to be established. We present evidence that
glycine acts as a bona fide ligand on GPR158,
Laboute et al., Science 379, 1352–1358 (2023)
31 March 2023
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
including direct binding and resultant change
in receptor activity eliciting cellular response.
This puts GPR158 in line with other class C
GPCRs, many of which are amino acid sensors,
such as the metabotropic glutamate receptors
(mGluRs) and the receptor for g-aminobutyric
acid (GABA), GABAB. Thus, we propose a ge-
neric name for GPR158 to be metabotropic
glycine receptor, or mGlyR. We did not ob-
serve significant GPR158-mediated neuronal
responses to taurine in cortical neurons, con-
sistent with weaker effects of taurine on GPR158
relative to glycine. However, it remains possible
that GPR158 may still mediate the effects of
taurine in other neuronal populations or under
certain conditions, possibly making GPR158 a
receptor for both glycine and taurine.
The mechanism by which mGlyR (GPR158)
signals upon glycine or taurine binding devi-
ates from canonical actions of GPCRs. Instead
of activating G proteins, mGlyR recruits a RGS7-
Gb5 complex, docking it into the intracellular
pocket that canonical GPCRs use for interact-
ing with G proteins and relaying changes in
seven-transmembrane architecture upon li-
gand binding into conformational changes in
Ga, triggering nucleotide exchange. Thus, in
the model we propose (Fig. 5H), glycine bind-
ing to the Cache domain of mGlyR changes
the conformation of the intracellular surface,
which in turn affects conformation of RGS7-Gb5.
This change reduces the ability of RGS7-Gb5
to stimulate Ga GTPase, likely by disfavoring
its orientation toward the membrane. In this
sense, glycine serves as an antagonist of the
GPR158-RGS7-Gb5 complex by reducing its
activity. Because RGS7-Gb5 is a selective GAP
for the inhibitory Gi/o proteins, which regulate
cAMP production (39, 40), inhibition of RGS7-
Gb5 activity via GPR158 influences cAMP levels.
The direction of the effect on the cAMP pro-
duction is likely determined by the identity of
the adenylyl cyclases present in a particular
cell, as they are known to be differentially
regulated by Gai and Gao (via Gbg) (41). Thus,
glycine signals via mGlyR by inhibiting inhib-
itory G protein regulation, thereby generating
an excitatory influence. This regulation endows
the metabotropic glycinergic system with a dis-
tinct feature that makes the degree of its in-
fluence scale with the extent of Gi/o activation
by other GPCR cascades, with its influence
increasing upon the increase in Gi/o inputs.
The discovery of mGlyR also opens many
interesting avenues for exploring the metab-
otropic influence of glycine and its role in
nervous system physiology. Indeed, metabo-
tropic effects of glycine have been anecdotally
noted (6, 7, 42), but molecular and circuit dis-
section of this influence have been limited. The
relatively high affinity of mGlyR for glycine
(~3 mM) should allow it to signal without con-
comitant engagement of GlyRs, which have an
order-of-magnitude-lower affinity for glycine,
creating an independent neuromodulatory
channel (6). The mGlyR effects on neurons
that we observe are also excitatory, contrast-
ing with the largely inhibitory influence of
ionotropic GlyR receptors (9, 43). The two sys-
tems likely overlap and are involved in auto-
tuning and homeostatic feedback, as has
been noted for other pairs of ionotropic and
metabotropic systems. Thus, in the context of
intact neural circuitry, glycine likely triggers
more complex responses that may involve
interplay between ionotropic and metabo-
tropic systems.
Furthermore, we think that glycinergic sig-
naling by means of mGlyR has implications
for understanding mood disorders and for the
development of new pharmacological strat-
egies. mGlyR is prominently expressed in the
medial prefrontal cortex (mPFC) (25), a region
critically involved in depression (44). Glycine
and its transporters are also colocalized in the
mPFC (45–47). Both taurine and glycine have
been heavily implicated in the pathophysiol-
ogy of depression (8–10, 48) and are dysregu-
lated in plasma of humans diagnosed with
major depressive disorder (10). Furthermore,
taurine has an antidepressant effect on stress-
induced depressive rats (49). Because these
amino acids inhibit mGlyR, and because knock-
out of mGlyR in mice also results in an anti-
depressive phenotype and stress resilience (25),
it seems possible that antidepressant properties
of glycine and taurine may be mediated by
mGlyR. The ubiquitous nature and multitude
of the effects limit the potential of glycine and
taurine to be used as medications. However,
identification of mGlyR presents a new target
for the development of antidepressants that
we postulate to be small molecules that selec-
tively inhibit this receptor to avoid possibly
related receptors, such as GPR179 in the eye.
RE FERENCES AND NOTES
1. M. S. Hernandes, L. R. P. Troncone, J. Neural Transm. 116,
1551–1560 (2009).
2. P. Legendre, Cell. Mol. Life Sci. 58, 760–793 (2001).
3. L. G. Harsing Jr., P. Matyus, Brain Res. Bull. 93, 110–119 (2013).
4. H. Zhu, E. Gouaux, Nature 599, 513–517 (2021).
5.
I. Pérez-Torres, A. M. Zuniga-Munoz, V. Guarner-Lans, Mini Rev.
Med. Chem. 17, 15–32 (2017).
6. Y. Han, J. Zhang, M. M. Slaughter, J. Neurosci. 17, 3392–3400
(1997).
7. M. Hou, L. Duan, M. M. Slaughter, J. Physiol. 586, 2913–2926 (2008).
8. C. C. Huang et al., Biol. Psychiatry 74, 734–741 (2013).
9. W. Li et al., Neuropharmacology 157, 107688 (2019).
10. C. Altamura, M. Maes, J. Dai, H. Y. Meltzer, Eur.
Neuropsychopharmacol. 5 (suppl. 1), 71–75 (1995).
11. N. Wettschureck, S. Offermanns, Physiol. Rev. 85, 1159–1204 (2005).
12. E. J. Neer, Cell 80, 249–257 (1995).
13. D. Yang et al., Signal Transduct. Target. Ther. 6, 7 (2021).
14. E. V. Gurevich, V. V. Gurevich, Genome Biol. 7, 236 (2006).
15. A. K. Shukla, K. Xiao, R. J. Lefkowitz, Trends Biochem. Sci. 36,
457–469 (2011).
16. E. Hermans, Pharmacol. Ther. 99, 25–44 (2003).
I. Masuho et al., Cell 183, 503–521.e19 (2020).
17.
18. M. Abramow-Newerly, A. A. Roy, C. Nunn, P. Chidiac,
Cell. Signal. 18, 579–591 (2006).
19. A. Fajardo-Serrano et al., Hippocampus 23, 1231–1245 (2013).
20. J. Celver, M. Sharma, A. Kovoor, J. Neurochem. 120, 56–69 (2012).
21. K. Psifogeorgou et al., J. Neurosci. 31, 5617–5624 (2011).
22. C. Orlandi et al., J. Cell Biol. 197, 711–719 (2012).
23. J. A. Stockert, L. A. Devi, Front. Pharmacol. 6, 100 (2015).
24. C. Laschet, N. Dupuis, J. Hanson, Biochem. Pharmacol. 153,
62–74 (2018).
25. L. P. Sutton et al., eLife 7, e33273 (2018).
26. C. Orlandi, L. P. Sutton, B. S. Muntean, C. Song, K. A. Martemyanov,
Neuropsychopharmacology 44, 642–653 (2019).
27. C. Song, C. Orlandi, L. P. Sutton, K. A. Martemyanov, J. Biol.
Chem. 294, 13145–13157 (2019).
28. D. Çetereisi et al., Front. Cell. Neurosci. 13, 465 (2019).
29. L. Khrimian et al., J. Exp. Med. 214, 2859–2873 (2017).
30. D. N. Patil et al., Science 375, 86–91 (2022).
31. E. Jeong, Y. Kim, J. Jeong, Y. Cho, Nat. Commun. 12, 6805 (2021).
32. L. I. Jiang et al., J. Biol. Chem. 282, 10576–10584 (2007).
33. O. Kletke, G. Gisselmann, A. May, H. Hatt, O. A. Sergeeva,
PLOS ONE 8, e61733 (2013).
34. N. Hussy, C. Deleuze, A. Pantaloni, M. G. Desarménien, F. Moos,
J. Physiol. 502, 609–621 (1997).
35. I. Masuho, K. Xie, K. A. Martemyanov, J. Biol. Chem. 288,
25129–25142 (2013).
36. I. Masuho, K. A. Martemyanov, N. A. Lambert, G Protein-
Coupled Receptors in Drug Discovery: Methods and Protocols,
M. Filizola, Ed. (Springer, 2015), pp. 107–113.
37. B. W. H. Liauw, H. S. Afsari, R. Vafabakhsh, Nat. Chem. Biol. 17,
291–297 (2021).
38. T. A. Ray et al., J. Neurosci. 34, 6334–6343 (2014).
39. B. S. Muntean et al., Cell Rep. 34, 108718 (2021).
40. P. Sassone-Corsi, Cold Spring Harb. Perspect. Biol. 4,
a011148–a011148 (2012).
41. R. Sadana, C. W. Dessauer, Neurosignals 17, 5–22 (2009).
42. E. Albiñana, S. Sacristán, R. Martín del Río, J. M. Solís,
J. M. Hernández-Guijo, Cell. Mol. Neurobiol. 30, 1225–1233
(2010).
43. M. H. Aprison, N. S. Nadi, Amino Acids as Chemical Transmitters,
F. Fonnum, Ed. (Springer US, 1978), pp. 531–570.
44. B. D. Hare, R. S. Duman, Mol. Psychiatry 25, 2742–2758 (2020).
45. H. U. Zeilhofer et al., J. Comp. Neurol. 482, 123–141 (2005).
46. N. Kawai et al., Amino Acids 42, 2129–2137 (2012).
47. R. J. Harvey, B. K. Yee, Nat. Rev. Drug Discov. 12, 866–885 (2013).
48. S. Kim, K. B. Hong, S. Kim, H. J. Suh, K. Jo, Sci. Rep. 10, 11370
(2020).
49. G. F. Wu et al., Sci. Rep. 7, 4989 (2017).
50. T. Laboute et al., Orphan receptor GPR158 serves as a
metabotropic glycine receptor: mGlyR, Zenodo (2023);
https://doi.org/10.5281/zenodo.7631532.
AC KNOWLED GME NTS
We thank N. Martemyanova for help with mouse husbandry, X. Li
for experimental help, and members of the Martemyanov Lab
for helpful discussions. We thank Servier Medical Art for templates
used in graphics. Funding: This work was supported by NIH
grants MH105482 (to K.A.M.) and GM069832 (to S.F.). Author
contributions: T.L. and K.A.M. conceived the project; T.L. performed
all functional experiments; S.Z. performed electrophysiology
experiments; C.G., D.P., S.F., and M.H. performed and analyzed
computational docking and structural modeling; B.A.W. and R.N.R.
performed and analyzed ITC experiments; T.L. and K.A.M. wrote the
manuscript with input from all other authors; and K.A.M. supervised
the project. Competing interests: T.L. and K.A.M. are listed as
inventors on a patent application describing methods to study
GPR158 activity; K.A.M. is a cofounder of Blueshield Therapeutics, a
startup company pursuing GPR158 as a drug target. Data and
materials availability: All data that support the findings are
available at Zenodo (50). All plasmids generated during this study are
freely available upon request. License information: Copyright ©
2023 the authors, some rights reserved; exclusive licensee
American Association for the Advancement of Science. No claim to
original US government works. https://www.science.org/about/
science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add7150
Materials and Methods
Figs. S1 to S10
Table S1
References (51–65)
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
Submitted 29 June 2022; resubmitted 4 November 2022
Accepted 3 March 2023
10.1126/science.add7150
Laboute et al., Science 379, 1352–1358 (2023)
31 March 2023
6 of 6
|
10.1126_science.abq1414
|
RES EARCH
SURFACE CHEMISTRY
Quantum effects in thermal reaction rates at
metal surfaces
Dmitriy Borodin1,2*, Nils Hertl1,2, G. Barratt Park1,2,3, Michael Schwarzer1, Jan Fingerhut1,
Yingqi Wang4, Junxiang Zuo4, Florian Nitz1, Georgios Skoulatakis2, Alexander Kandratsenka2,
Daniel J. Auerbach2, Dirk Schwarzer2, Hua Guo4, Theofanis N. Kitsopoulos1,2,5,6, Alec M. Wodtke1,2*
There is wide interest in developing accurate theories for predicting rates of chemical reactions
that occur at metal surfaces, especially for applications in industrial catalysis. Conventional methods
contain many approximations that lack experimental validation. In practice, there are few reactions
where sufficiently accurate experimental data exist to even allow meaningful comparisons to theory.
Here, we present experimentally derived thermal rate constants for hydrogen atom recombination
on platinum single-crystal surfaces, which are accurate enough to test established theoretical
approximations. A quantum rate model is also presented, making possible a direct evaluation of
the accuracy of commonly used approximations to adsorbate entropy. We find that neglecting the
wave nature of adsorbed hydrogen atoms and their electronic spin degeneracy leads to a 10× to
1000× overestimation of the rate constant for temperatures relevant to heterogeneous catalysis.
These quantum effects are also found to be important for nanoparticle catalysts.
E normous effort has gone into developing
predictive theories of thermal reaction
rates (1), with one goal being accurate
kinetic models of heterogeneous catal-
ysis, an industrial cornerstone of modern
society (2). Modeling real catalytic reactors
presents technical problems because they often
involve networks of reactions (3, 4), compli-
cating meaningful comparisons to experiment
that could test a theory’s assumptions. A pos-
sible solution is to compare experiment and
theory using simplified model systems that
involve only a single elementary reaction.
Unfortunately, even this comparison is seldom
achieved because accurate measurements of
elementary reaction rates are rare in surface
chemistry (5).
Illustrative of these problems is the thermal
recombination of H atoms on transition metals,
leading to H2 formation. Being perhaps the
simplest reaction for theoretical modeling and
omnipresent as an elementary step in indus-
trial catalysis [e.g., hydrogenation of unsaturated
fats (6), ammonia synthesis (7), and electro-
chemical hydrogen production (8)], it is an
obvious starting point for the development of
accurate rate theories in surface chemistry.
Unfortunately, large uncertainties in the ex-
perimentally derived second-order rate con-
stants arise because of difficulties in obtaining
1Institute for Physical Chemistry, University of Göttingen,
Tammannstraße 6, 37077 Göttingen, Germany. 2Department
of Dynamics at Surfaces, Max Planck Institute for
Multidisciplinary Sciences, am Faßberg 11, 37077 Göttingen,
Germany. 3Department of Chemistry and Biochemistry,
Texas Tech University, Lubbock, TX 79409-1061, USA.
4Department of Chemistry and Chemical Biology, University
of New Mexico, Albuquerque, NM 87131, USA. 5Department
of Chemistry, University of Crete, 71003 Heraklion, Greece.
6Institute of Electronic Structure and Laser, FORTH, 71110
Heraklion, Greece.
*Corresponding author. Email: dborodi@gwdg.de (D.B.);
alec.wodtke@mpinat.mpg.de (A.M.W.)
accurate initial concentrations (9). If these and
other experimental problems could be overcome,
this reaction would provide an ideal system for
benchmarking rate theory, especially for testing
approximate treatments of quantum effects.
From the study of gas-phase reactions, exact
treatments of nuclear quantum effects are often
considered to be unnecessary above ~500 K (10),
and, because most catalytic reactors operate
at increased temperatures, one might conclude
that a classical approximation (11) or approx-
imate ad hoc quantum treatments, like har-
monic transition-state theory (hTST) (12, 13),
would be sufficient to model surface chemis-
try. But the need to go beyond hTST has been
pointed out recently (14) and new methods
were reported, although they also lack valida-
tion from experiment. Electron spin is another
important quantum effect on surface reactions;
for example, in H atom recombination, only
one out of four electron spin combinations
yields a stable H2 molecule. However, the spin-
degeneracy of reactants and products has, to
our knowledge, never been included in calcu-
lations of reaction rates at metal surfaces.
This paper reports kinetic data for H atom
recombination on both the Pt(111) and Pt(332)
surfaces obtained with velocity-resolved kinetics
(VRK), which was previously used only to study
first-order and pseudo–first-order reactions on
model catalysts (15, 16). For this work, we have
extended VRK to the measurement of rate
constants for second-order reactions by mea-
suring the absolute reactant flux, which, when
combined with known sticking probabilities
(17, 18), provides accurate initial concentra-
tions [H]0 and eliminates the main source of
error found in previous work.
To understand the kinetics more deeply, we
also constructed a quantum rate model (QRM)
that accurately reproduced experimental rate
constants over 12 orders of magnitude for
temperatures between 250 and 1000 K with no
adjustable parameters. Comparison to a cor-
responding classical rate model (CRM) revealed
how large and crucially important quantum
effects are; the classical reaction rate constants
were ~20 times larger than quantum rate
constants even at 1000 K, with an increasing
deviation at lower temperatures. For reactions
at stepped surfaces, the errors were even higher.
This dramatic quantum reduction of the reac-
tion rate resulted from both the delocalization of
the adsorbed H* nuclei as well as the influence
of electron spin degeneracy.
Results
The experiments are described in detail in
the supplementary materials (SM). Briefly, a
pulsed molecular beam with a controlled mix-
ture of H2 and D2 illuminated either a Pt(111)
or Pt(332) crystal facet, with step densities
of 0.1 to 0.6% and 16.7%, respectively. The
transient rates of HD formation were then
recorded using VRK, where pulsed laser-
ionization, time-of-flight mass spectrometry
reports the product’s mass-to-charge ratio (m/Z)
and its density as a function of delay between
the pulsed molecular and laser beams. Because
the ions were detected with slice imaging
(19, 20) yielding product velocity, we could
accurately compute the transient product
flux as a function of reaction time at the
surface. Initial reactant concentrations are
needed to obtain second-order rate constants
(SM section S2a). These values were obtained
from the absolute flux profiles of the incident
molecular beams (SM section S2b) and known
sticking coefficients (17, 18) (SM section S3).
Finally, VRK data obtained at m/Z = 2, 3, and 4
led to isotopic branching fractions (SM section
S4), from which we obtained isotope-specific
rate constants.
The QRM developed in this work is an exact
formulation of a thermal rate constant. It
yields accurate isotope-specific thermal rates
as long as one has accurate isotope-specific
thermal sticking probabilities SH2;HD;D2
adsorption energies EH2;HD;D2
, and reactant
0
QH(cid:2);D(cid:2) and product QH2;HD;D2 partition functions:
Tð Þ,
E
D
0
kH2
(cid:2)
Tð Þ ¼
(cid:3)
Tð Þ
SH2
0
s
ffiffiffiffiffiffiffiffiffiffiffiffiffiffi
kBT
2pmH2
(cid:5)
Þ2 exp (cid:3)
QH2
=V
QH(cid:2) =A
ð
EH2
0
kBT
(cid:6)
ð1Þ
where T is temperature, kB is the Boltzmann
constant, V is reference volume, and A is refer-
ence area for the partition function evaluation.
The QRM rate constant for H2 desorption by
the recombination of two adsorbed H atoms,
Tð Þ, given by Eq. 1, is derived from the
kH2
principle of detailed balance provided in SM
Borodin et al., Science 377, 394–398 (2022)
22 July 2022
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
section S5a. Expressions for kHD Tð Þ and kD2
Tð Þ
were easily obtained by analogy. Accurate values
D
and SH2;HD;D2
for EH2;HD;D2
0
0
Tð Þ can be ob-
E
tained from prior experiments (SM sections S3
and S6). These measured quantities allowed
us to avoid errors associated with the theoret-
ical determination of the thermal dissociative
adsorption rates and density functional theory
(DFT) calculations of adsorption energies,
which can be highly dependent on the choice
of exchange-correlation functional (21). In ad-
dition, the partition function for the hydrogen
molecule in the gas phase QH2 is well known.
The adsorbate partition functions QH(cid:2);D(cid:2) are
crucial inputs to the QRM and were computed
with a quantum potential energy sampling
(QPES) method, where the nuclear part of the
partition function is obtained by a direct state
count. States and energies were obtained by
solving the nuclear Schrödinger equation with
DFT interaction potentials computed with two
different functionals and assuming a static Pt
surface (SM section S1b). This procedure was
performed for H interacting with both Pt(111)
and Pt(332). We found that QH(cid:2) is weakly depen-
dent on the choice of DFT functional (SM sec-
tion S5e). The electronic contribution to QH(cid:2) ,
which accounts for the twofold spin degeneracy
of the H-Pt system, was explicitly included.
Figure 1 presents the experimentally ob-
tained HD formation rates for reactions on
Pt(111) and Pt(332) and compares them with a
simulation of the experiment. The simulations
used rate constants from the QRM for all three
isotopologs (fig. S9) and accounted for the
temporal profile of the dosing pulse and its
spatial inhomogeneity, f (t, r) (Fig. 2A), where
t is time and r is radial distance, as well as
reactant diffusion. This aspect of the data
analysis goes beyond past work and is essential
because the rates of second-order reactions
are sensitive to surface concentration dis-
tributions and gradients. The full diffusion-
reaction model is described in SM section S7
and accounts for well-known diffusion effects
on surface reaction rates discussed in previous
work (22). Figure 2B shows that the isotope
effect at these temperatures is small and well
described by the QRM. Inspection of Figs.
1 and 2B clearly shows that the QRM, which
has no adjustable parameters, reproduces ex-
perimental data for reactions on both Pt(111)
and Pt(332).
Figure 3 shows the VRK-derived H* recom-
bination rate constants (black circles) compared
with those of previous work (light red trapezoids)
for reactions on Pt(111). Previous studies used
temperature programmed desorption (TPD)
for T < 400 K (23–26) and molecular beam
relaxation spectrometry (MBRS) for T > 400 K
(27, 28). The uncertainty in the previously
reported rate constants spans three orders of
magnitude. We note that previous work studied
Fig. 1. VRK of H atom recombination on Pt(111) and Pt(332). Measured HD formation rates for Pt(111)
(○) and Pt(332) (+) are compared with the results of the QRM (dashed and solid lines). The temperature
dependence and the transient rate of the measurements are quantitatively captured by the model for both
facets. The shaded regions of the top three panels indicate 2s uncertainty, mainly associated with the
absolute reactant flux measurement (~30%) and the dissociative adsorption energies. The excellent
agreement between VRK and QRM is achieved without adjustable parameters. a.u., arbitrary units.
different isotopic recombination reactions
(fig. S8); however, given the small isotope effect
found in the present study (Fig. 2B and fig. S9),
these differences between experiments cannot
explain the large range of reported values.
The VRK measurements clearly distinguish
the accuracy of two previous MBRS measure-
ments that fall within the uncertainty range
of (27) but differ by two orders of magnitude
from rate constants reported in (28).
A hallmark of a fundamentally correct
model is its ability to reproduce accurate ex-
perimental data over a broad temperature
range. The QRM uses a fundamentally correct
ab initio adsorbate partition function that leads
to excellent agreement with experiment over
a large temperature range. The performance
of the QRM for Pt(111) at temperatures between
650 and 950 K is demonstrated by comparison
to VRK-derived rate constants, whereas low-
temperature comparisons rely on TPD. Un-
certainties in the TPD-derived rate constants
arise from questionable approximations used
to derive rate constants from the data, neglect
of the coverage dependence of adsorption
energies (26), dubious estimations of prefac-
tors (25), and neglect of the influence of steps
(23). To make the most meaningful compar-
ison, we used the QRM to directly simulate
TPD spectra from (29), where the influence
of steps was carefully identified and removed
(SM section S8). Here, we also accounted for
the previously reported coverage dependence
of the adsorption energy (24, 30) (fig. S11 and
SM section S6). The comparison is shown in
the top-right inset of Fig. 3A. The solid black
lines of the QRM are in excellent agreement
with the TPD spectra [broad gray lines, from
(29)] for three initial coverages.
Discussion
The aforementioned comparisons to kinetics
experiments carried out between 250 and 950 K
demonstrated the validity of the QRM rate
constants over 12 orders of magnitude and for
H atom coverages up to 0.3 monolayer (ML).
Within the context of the principle of detailed
balance as implemented in the QRM, this
coherent picture demonstrates the quantitative
consistency of previously reported sticking co-
efficients and binding energies with the kinetics
measurements of this work. The agreement over
such a wide range of rates provides confidence
in the QRM rate constants, making H recombi-
nation on Pt(111) a reliable benchmark for ap-
proximate rate theories in surface chemistry.
Borodin et al., Science 377, 394–398 (2022)
22 July 2022
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
A
B
Fig. 2. Calibration of the molecular beam and isotopic branching. (A) The
space-dependent H2 and D2 dosing profiles used to determine the absolute initial
concentration of H* and D*. These results were obtained from laser-based
calibration of the molecular beam flux and are required to accurately determine
the recombination rate constants. The shaded regions indicate the 2s uncertainty
range. The inset shows the temporal profile of the molecular beam pulse.
(B) Isotopic branching fraction from VRK experiments (symbols) and QRM (lines).
The agreement shows that QRM correctly predicts the isotope effect. The error
bars and the gray-shaded region reflect 2s uncertainty in the experiment and model,
respectively. Note that some symbols have been shifted by ±5 K for clarity.
A
B
Fig. 3. Rate constants for H atom recombination on Pt(111). (A) Light red
trapezoids show the temperature-range and rate-constant uncertainties of
previous work (23–28). Shown are experimental results from this work (○) with
2s error bars compared with the results of the QRM (black solid line), hTST
(green dotted line), CRM (blue dash-dotted line), and QRM neglecting electron
spin (black dashed line). The inset at the bottom left shows an expanded
view. The inset at the top right compares TPD spectra (broad gray lines) from
(29) with the predictions of the QRM model, QRM neglecting spin, the CPES
model, and the hTST model for three initial H* coverages of 0.1, 0.2,
and 0.3 ML. The gray-shaded region and the horizontal error bar on one of
the modeled TPD spectra reflects the uncertainty of the experimental H2
chemisorption energy. The ability of the QRM rate constants to quantitatively
reproduce experimental data demonstrates the importance of both nuclear
and electronic quantum effects. (B) Comparison of the approximate
predictions of three rate models to QRM rate constants. Neglecting spin
degeneracy, using a fully classical approximation or a commonly adopted
approximate quantum model both introduce large errors even at high
temperatures. Similar errors are seen for recombination rates on the stepped
Pt(332) surface (see fig. S14). See fig. S12 for a detailed decomposition of the
errors observed from hTST and adsorbate entropy approximations.
It is worth noting that the QRM as imple-
mented in this work is semiempirical because
it relies on experimental values of thermal
sticking coefficients and adsorption energies.
However, it also provides a path to an ab initio
theory of thermal reaction rates, if these quan-
tities can be accurately calculated from first
principles.
The framework of the QRM allows us to
critically test the quality of predictions based
on approximations that are commonly used
in kinetic modeling of heterogeneous catal-
ysis. The results of this analysis are shown in
Fig. 3. The most widely used model for rate
constants (hTST) introduces quantum effects
in an approximate way, where nuclear parti-
tion functions are computed assuming separable
motion of contributing degrees of freedom
Borodin et al., Science 377, 394–398 (2022)
22 July 2022
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
A
B
Fig. 4. The influence of steps on H atom recombination on Pt. (A) Rate
constants derived from VRK experiments (symbols) for H atom recombination on
Pt(111) and Pt(332) are compared with QRM predictions (solid lines). The 2s
uncertainty of the rate constants of Pt(332) QRM is shown as a red-shaded
region. The inset is a magnification of the area enclosed by the dotted rectangle.
(B) Entropies obtained experimentally at 598 K (symbols with 2s error bars) and
from CPES (blue dash-dotted line) for H* bound to Pt nanoparticles from (11).
Also shown are QPES entropies for H* bound to Pt(111) (solid black line) and
Pt(332) (solid red line) that were obtained in this work (see SM section S9). The
nuclear quantum effect contribution is 12 J mol−1 K−1, and the contribution of
electron spin is 6 J mol−1 K−1. The comparison suggests that the nanoparticle-
size dependence of the H* entropy is determined by the concentration of steps.
that can each be approximated as a harmonic
oscillator. By definition, recrossing correc-
tions are not included in hTST (12, 13). The
hTST rate constants, calculated by placing the
dividing surface far above the surface, are shown
as a green dotted line in Fig. 3. This approxi-
mation overestimates the experimental reac-
tion rate constant by two to three orders of
magnitude at all temperatures between 200
and 1200 K. The major source of errors in hTST
arise from the harmonic simplifications made
to the H-Pt interaction potential (resulting in
errors of a factor 5 to 25) and the neglect of
recrossing corrections to TST (with errors of
a factor 5 to 10) (see fig. S12A for details).
The next, more sophisticated level of rate
theory uses the complete potential energy
sampling (CPES) method to characterize en-
tropy associated with the in-plane degrees of
freedom of H*. CPES is considered by many to
provide the most accurate adsorbate partition
function (31), and it has been applied to char-
acterize H interaction at metals (11). It ac-
counts for anharmonicity by using a semiclassical
partition function computed from the adsorbate
potential energy surface, which may be obtained
with DFT (11, 14, 31). To evaluate this approach,
we modified the QRM, replacing the QPES
by the CPES adsorbate partition function but
retaining the other parameters in Eq. 1. This
substitution serves to illustrate the classical
counterpart of the QRM, which we hereafter
denote as the CRM. The rate constants predicted
by the CRM are shown as blue dash-dotted lines
in Fig. 3. The CRM performed better than hTST
but nevertheless overestimated the rate constant
by a factor of 20, even at temperatures as high as
1000 K. The error is more than 100-fold at 300 K,
a temperature typical for electrochemical appli-
cations. Although our detailed analysis is focused
on Pt(111), the errors introduced by hTST and
the CRM are similar for reactions on Pt(332)
(fig. S14).
A major source of error in the CPES method
arises from the classical description of the
adsorbate’s in-plane motion. This can be under-
stood by considering that the in-plane zero-point
energy of H* on Pt(111) (58 meV) is almost
equal to the classical diffusion barrier (60 meV)
(see fig. S13). Thus, classical and quantum de-
scriptions of H* motion on the surface lead
to very different results. CPES excludes H* from
classically forbidden regions of space, whereas
quantum mechanically, there is a substantial
probability to populate these regions. Further-
more, CPES does not account for the uncer-
tainty principle, which prevents localization
of H* at the classical energy minimum at low
temperature. The surface area explored by the
H atom is underestimated by CPES and thus
so too is the adsorbate entropy. This results in
an overestimate of the corresponding rate con-
stant. Our results underscore the importance of
quantum delocalization and help explain why
the deviations of CRM become more severe at
low temperatures. Quantum delocalization is
also the reason why hTST fails. All quantum
states above the ground state exhibit prob-
ability maxima at positions far from the po-
tential energy minimum (fig. S13).
We may investigate other sources of error in
the CRM by using the QPES partition function
but neglecting electron spin. Figure 3A shows
rate constants predicted on this basis, and in
Fig. 3B, one can see that neglect of electron
spin degeneracy led to a 4× overestimate of
the rate constant at all temperatures. This
result can be understood intuitively if we con-
sider that when two H* atoms attempt to react,
they must approach one another in one of four
degenerate states with either parallel (triplet)
or antiparallel (singlet) spins. Only the singlet
state correlates with the formation of gas-
phase singlet H2 products; hence, including spin
degeneracy reduces the reaction rate by a factor
of four. This result should not come as a surprise
to those familiar with rate calculations for gas-
phase reactions where spin degeneracy is rou-
tinely included (1). Nonetheless, to the best of
our knowledge, this is the first demonstration
that the rates of thermal reactions at metal
surfaces depend on adsorbate spin.
The QRM developed in this work also de-
scribes the reaction rate on stepped surfaces.
In Fig. 4A, we compare the predicted rate con-
stants of the QRM for Pt(111) and Pt(332) sur-
faces to those derived from VRK experiments.
For details of the QRM treatment of the re-
action on Pt(332), see SM sections S1b and S6.
Experiment shows that near 700 K, the rate
constants for reaction on the (332) facet is
larger than that on the (111) facet, an effect that
is quantitatively captured by the QRM. This
result may appear surprising, because the H
atom’s binding energy is larger at steps than
at terraces (30, 32). A naïve view of Eq. 1 suggests
that this leads to a lower rate constant. However,
careful analysis of the thermally populated
Borodin et al., Science 377, 394–398 (2022)
22 July 2022
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
quantum states used in the QPES partition
functions showed that at these temperatures,
H atoms on the (332) facet tend to remain lo-
calized near step sites (fig. S15). This fact re-
duces their in-plane translational entropy and
leads to an increase in the rate constant because
the effect of entropy is larger than that produced
by a larger step binding energy.
This observation reflects how changing tem-
perature alters the relative influence of energy
and entropy on the rate constant. In past work,
similarities in TPD spectra of H2 desorbing
from Pt(111) and a B-type stepped Pt surface
at T ~ 350 K were taken as evidence for a lack
of preferential step binding (26). Inspection
of QRM rate constants in Fig. 4A reveals that
at 350 K, the similarity in desorption rate
constants arises from compensation between
energetic and entropic contributions (see SM
section S8 for details). Our work supports
conclusions derived from He-atom and ion-
scattering experiments that H binds more
strongly to B-type steps (30, 32). Only at much
lower temperatures does the energetic prefer-
ence for H binding at steps cause the rate
constant on the stepped surface to drop below
that on (111) terraces.
Because the QRM approach developed in
this work provides an accurate determination
of rate constants on stepped surfaces, we expect
that QPES entropies used in the QRM would
help us understand experiments performed
on size-selected Pt nanoparticles (11), because
smaller nanoparticles exhibit higher step con-
centrations. Figure 4B shows measured H*
entropies for Pt nanoparticles of various sizes
reproduced from (11)—the entropy increases
with the nanoparticle size, consistent with ob-
servations presented above that Pt steps reduce
H* entropy. Also shown are CPES entropies
reported in (11), which fail to describe entropies
derived from experiment. Notably, the entropies
found using the QPES method for H* bound
to Pt(111) (see Fig. 4B) are in good agreement
with entropies for the largest particle sizes.
Note that surfaces of large nanoparticles are
primarily composed of the (111) facets (11).
Figure 4B also shows QPES entropies for H*
on the (332) facet, which compare well to ex-
perimentally obtained entropies on small nano-
particles. This comparison further supports our
hypothesis that the H* entropy decreases as
nanoparticle size decreases and the relative
importance of step defects increases. This re-
sult points out the importance of quantum
effects even for the description of thermody-
namic state functions in advanced catalytic
materials.
Conclusion
As this work has shown, H* recombination on
Pt surfaces exhibits large quantum effects even
at increased temperatures relevant to cataly-
sis. These quantum effects in the reaction rates
and in the thermodynamic properties of the
adsorbed H atoms arise in part from the H atom’s
light mass, where a careful treatment of its wave
properties is required to obtain accurate results.
Such nuclear quantum effects will diminish in
importance for heavier adsorbates. However,
the effect of spin degeneracy demonstrated
here will remain of general importance for a
host of reactions of heavier species involved in
real-world catalysis. At present, it is not easily
possible to determine the lowest-energy spin
state for metal surfaces with DFT. Developing
theoretical and experimental methods that are
able to probe the general influence of spin on
reaction rates presents the next challenge on the
way toward fully predictive surface chemistry at
metal catalysts.
RE FERENCES AND NOTES
1.
J. L. Bao, D. G. Truhlar, Chem. Soc. Rev. 46, 7548–7596
(2017).
2. F. Studt, Front. Catal. 1, 658965 (2021).
3. A. H. Motagamwala, J. A. Dumesic, Chem. Rev. 121, 1049–1076
4.
(2021).
J. E. Sutton, W. Guo, M. A. Katsoulakis, D. G. Vlachos, Nat.
Chem. 8, 331–337 (2016).
5. G. B. Park et al., Nat. Rev. Chem. 3, 723–732 (2019).
6. P. Sabatier, Catalysis in Organic Chemistry (D. Van Nostrand
Company, 1922).
7. G. Ertl, Angew. Chem. Int. Ed. 47, 3524–3535 (2008).
8. A. Godula-Jopek, Hydrogen Production by Electrolysis (Wiley-VCH,
2015).
9. M. P. D’Evelyn, R. J. Madix, Surf. Sci. Rep. 3, 413–495 (1983).
10. Y. V. Suleimanov, F. J. Aoiz, H. Guo, J. Phys. Chem. A 120,
8488–8502 (2016).
11. M. García-Diéguez, D. D. Hibbitts, E. Iglesia, J. Phys. Chem. C
123, 8447–8462 (2019).
12. S. Bhandari, S. Rangarajan, C. T. Maravelias, J. A. Dumesic,
M. Mavrikakis, ACS Catal. 10, 4112–4126 (2020).
13. M. Jørgensen, H. Grönbeck, ACS Catal. 6, 6730–6738 (2016).
14. A. Bajpai, P. Mehta, K. Frey, A. M. Lehmer, W. F. Schneider, ACS
Catal. 8, 1945–1954 (2018).
15. D. Borodin et al., J. Am. Chem. Soc. 143, 18305–18316
(2021).
16. J. Neugebohren et al., Nature 558, 280–283 (2018).
17. R. van Lent et al., Science 363, 155–157 (2019).
18. A. C. Luntz, J. K. Brown, M. D. Williams, J. Chem. Phys. 93,
5240–5246 (1990).
19. K. Tonokura, T. Suzuki, Chem. Phys. Lett. 224, 1–6 (1994).
20. C. R. Gebhardt, T. P. Rakitzis, P. C. Samartzis, V. Ladopoulos,
T. N. Kitsopoulos, Rev. Sci. Instrum. 72, 3848–3853 (2001).
21. G. J. Kroes, Phys. Chem. Chem. Phys. 23, 8962–9048 (2021).
22. D. R. Olander, J. Colloid Interface Sci. 58, 169–183 (1977).
23. K. Christmann, G. Ertl, T. Pignet, Surf. Sci. 54, 365–392 (1976).
24. B. Poelsema, K. Lenz, G. Comsa, J. Phys. Condens. Matter 22,
304006 (2010).
25. P. Samson, A. Nesbitt, B. E. Koel, A. Hodgson, J. Chem. Phys.
109, 3255–3264 (1998).
26. M. J. van der Niet, A. den Dunnen, L. B. Juurlink, M. T. Koper,
J. Chem. Phys. 132, 174705 (2010).
27. G. E. Gdowski, J. A. Fair, R. J. Madix, Surf. Sci. 127, 541–554
(1983).
28. M. Salmerón, R. J. Gale, G. A. Somorjai, J. Chem. Phys. 70,
2807–2818 (1979).
29. S. K. Jo, Surf. Sci. 635, 99–107 (2015).
30. B. J. J. Koeleman, S. T. de Zwart, A. L. Boers, B. Poelsema,
L. K. Verhey, Nucl. Instrum. Methods Phys. Res. 218, 225–229
(1983).
31. M. Jørgensen, H. Grönbeck, J. Phys. Chem. C 121, 7199–7207
(2017).
32. B. Poelsema, G. Mechtersheimer, G. Comsa, Surf. Sci. 111,
519–544 (1981).
33. D. Borodin, Hydrogen atom recombination on Pt(111) and
Pt(332). Zenodo (2022); doi:10.5281/zenodo.6565004.
AC KNOWLED GME NTS
We thank J. C. Tully for helpful discussions. Funding: D.B. and M.S.
thank the BENCh graduate school, funded by the DFG
(389479699/GRK2455). T.N.K., G.S., A.K., M.S., and J.F.
acknowledge support from the European Research Council under
the European Union’s Horizon 2020 research and innovation
program (grant agreement no. 833404). Y.W., J.Z., and H.G.
acknowledge the US National Science Foundation (grant. no. CHE-
1951328), and H.G. thanks the Alexander von Humboldt Foundation
for a Humboldt Research Award. The calculations were partially
performed at the Center for Advanced Research Computing
(CARC) at the University of New Mexico and at the National Energy
Research Scientific Computing (NERSC) Center. Author
contributions: D.B., M.S., and J.F. conducted the transient kinetics
experiments. Flux calibration procedures were developed by D.B.,
G.B.P., M.S., F.N., D.J.A., and T.N.K. D.B. and D.S. developed the
reaction-diffusion analysis. D.B. and A.M.W. developed the
quantum rate model. N.H., Y.W., J.Z., and H.G. conducted DFT
calculations. D.B., N.H., A.K., and H.G. analyzed DFT calculations
and developed methods for description of nuclear partition
functions. M.S., J.F., G.S., T.N.K., D.J.A., D.S., and A.K. participated
in discussion of the results. D.B., D.J.A., H.G., and A.M.W. wrote
the manuscript and the supporting material. All authors
contributed to the reviews of the manuscript and the supporting
material. Competing interests: None declared. Data and
material availability: All data needed to evaluate the conclusions
in the paper are present in the paper or the supplementary
materials and are publicly available in the Zenodo repository
(33). License information: Copyright © 2022 the authors, some
rights reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government
works. https://www.science.org/about/science-licenses-journal-
article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abq1414
Materials and Methods
Supplementary Text
Figs. S1 to S15
References (34–71)
Submitted 21 March 2022; accepted 16 June 2022
10.1126/science.abq1414
Borodin et al., Science 377, 394–398 (2022)
22 July 2022
5 of 5
|
10.1126_science.add7795
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
ENZYMOLOGY
Visualizing the DNA repair process by a photolyase
at atomic resolution
Manuel Maestre-Reyna*, Po-Hsun Wang, Eriko Nango, Yuhei Hosokawa, Martin Saft, Antonia Furrer,
Cheng-Han Yang, Eka Putra Gusti Ngurah Putu, Wen-Jin Wu, Hans-Joachim Emmerich, Nicolas Caramello,
Sophie Franz-Badur, Chao Yang, Sylvain Engilberge, Maximilian Wranik, Hannah Louise Glover,
Tobias Weinert, Hsiang-Yi Wu, Cheng-Chung Lee, Wei-Cheng Huang, Kai-Fa Huang, Yao-Kai Chang,
Jiahn-Haur Liao, Jui-Hung Weng, Wael Gad, Chiung-Wen Chang, Allan H. Pang, Kai-Chun Yang,
Wei-Ting Lin, Yu-Chen Chang, Dardan Gashi, Emma Beale, Dmitry Ozerov, Karol Nass, Gregor Knopp,
Philip J. M. Johnson, Claudio Cirelli, Chris Milne, Camila Bacellar, Michihiro Sugahara, Shigeki Owada,
Yasumasa Joti, Ayumi Yamashita, Rie Tanaka, Tomoyuki Tanaka, Fangjia Luo, Kensuke Tono,
Wiktoria Zarzycka, Pavel Müller, Maisa Alkheder Alahmad, Filipp Bezold, Valerie Fuchs, Petra Gnau,
Stephan Kiontke, Lukas Korf, Viktoria Reithofer, Christian Joshua Rosner, Elisa Marie Seiler,
Mohamed Watad, Laura Werel, Roberta Spadaccini, Junpei Yamamoto, So Iwata, Dongping Zhong,
Jörg Standfuss, Antoine Royant, Yoshitaka Bessho*, Lars-Oliver Essen*, Ming-Daw Tsai*
INTRODUCTION: Dimerization of thymine bases
to form a cyclobutane-pyrimidine dimer (CPD) is
a common ultraviolet light–induced DNA lesion.
Photolyases catalyze light-triggered repair of CPD-
DNA, thus contributing to genome stability in
many organisms. Combining time-resolved crys-
tallography and computational analyses, we report
an atomic visualization of the photolyase-catalyzed
DNA repair process. We captured electron transfer
at low picoseconds, chemical steps at picoseconds
to nanoseconds, active-site recovery at nanosec-
onds to microseconds, and reannealing of the
double-stranded DNA (dsDNA) at submillisec-
onds, forging new ground in DNA repair, struc-
tural biology, and enzymology.
RATIONALE: Mechanistic models from previous
spectroscopic studies set a framework for using
time-resolved serial femtosecond crystallog-
raphy (TR-SFX) to visualize this catalytic pro-
cess. This technique provides a concise view of
not only the repair chemistry but also hitherto
unknown postrepair events.
RESULTS: Two series of TR-SFX experiments
were performed, one from picoseconds to nano-
seconds and the other from nanoseconds to
microseconds (see the figure). Our visualization
of the repair of a CPD begins at 100 ps, with
Arg256 (R256) becoming dynamic and moving
to stabilize the CPD, which suggests the initia-
tion of the forward electron transfer from the
reduced, anionic hydroquinone state of flavin
adenine dinucleotide (FADH−) to the CPD. At
650 ps, the C5–C5′ bond of the CPD is predom-
inantly split, and at 1 ns, the C6–C6′ is likewise
split. Recovery of R256, a five-water cluster
(5WC), and the FADH− coenzyme occurs dur-
ing the next 500 ns, returning to their re-
spective resting-state conformations. The
repaired thymine bases remain in the active
site during this time and then start to return
to reanneal with the dsDNA in the microsec-
ond range. The 200-ms structures show the
coexistence of a back-flipping intermediate
and the reannealed product before its final
release from the enzyme.
CONCLUSION: Our results uncover the atomic
mechanism of how DNA photolyases repair
DNA in real time. These data reveal an
ordered breaking of the covalent C–C bonds
and opening of the cyclobutane ring, as well
as the concomitant conformational changes
of the photolyase and its FAD coenzyme. De-
fined intermediates were also captured when
the enzyme-product complex was recovering
and repaired product bases were departing
from the active site to pair with their com-
plementary bases in the dsDNA.▪
The list of author affiliations is available in the full article online.
*Corresponding author. Email: mmaestre@ntu.edu.tw (M.M.-R.);
bessho@spring8.or.jp (Y.B.); essen@chemie.uni-marburg.de
(L.-O.E.); mdtsai@gate.sinica.edu.tw (M.-D.T.)
Cite this article as M. Maestre-Reyna et al., Science 382,
eadd7795 (2023). DOI: 10.1126/science.add7795
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.add7795
Active site
CPD
FAD
FAD
5WC
CPD
T8
C5′
C5
T7
C6-C6′
R256
(Gray: Dark)
C5
C6
C5′
C6′
5WC
R256
3′
CPD
5′
(Gray: dsDNA)
3′
5′
Forward electron transfer
R256 in action; CPD still intact
C5–C5′
bond broken
C6–C6′
bond broken
Product complex
Active site recovered
Thymine bases
flipping back
DNA reannealed and
ready to dissociate
100 ps
650 ps
1 ns
500 ns
200 µs
200 µs
Dark
Time
0.1
DNA repair
1
Active-site recovery
TT return to dsDNA
10
100
1000
10,000
100,000 (ns)
Elucidation of the main processes and key intermediates in the DNA repair reaction catalyzed by photolyase. Selected intermediates (cyan) of the repair process
are overlaid with the structure of the dark state (gray) to illustrate structural changes during catalysis. T7 and T8 are the thymine-7 and thymine-8, the damaged 5′-
and 3′-thymines of the CPD lesion in the DNA strand. TT refers to the two thymines together.
Maestre-Reyna et al., Science 382, 1014 (2023)
1 December 2023
1 of 1
RES EARCH
R E S E A R C H A R T I C L E
◥
ENZYMOLOGY
Visualizing the DNA repair process by a photolyase
at atomic resolution
Manuel Maestre-Reyna1,2*, Po-Hsun Wang1, Eriko Nango3,4, Yuhei Hosokawa1,2,5, Martin Saft6,
Antonia Furrer7, Cheng-Han Yang1, Eka Putra Gusti Ngurah Putu1, Wen-Jin Wu1,
Hans-Joachim Emmerich6, Nicolas Caramello8,9, Sophie Franz-Badur6, Chao Yang10,
Sylvain Engilberge8,11, Maximilian Wranik7, Hannah Louise Glover7, Tobias Weinert7, Hsiang-Yi Wu1,
Cheng-Chung Lee1, Wei-Cheng Huang1, Kai-Fa Huang1, Yao-Kai Chang1, Jiahn-Haur Liao1,
Jui-Hung Weng1†, Wael Gad1, Chiung-Wen Chang1, Allan H. Pang1, Kai-Chun Yang2, Wei-Ting Lin2,
Yu-Chen Chang2, Dardan Gashi7, Emma Beale7, Dmitry Ozerov7, Karol Nass7, Gregor Knopp7,
Philip J. M. Johnson7, Claudio Cirelli7, Chris Milne7, Camila Bacellar7, Michihiro Sugahara3,
Shigeki Owada3,12, Yasumasa Joti3,12, Ayumi Yamashita3,13, Rie Tanaka3,13, Tomoyuki Tanaka3,13,
Fangjia Luo12, Kensuke Tono3,12, Wiktoria Zarzycka14, Pavel Müller14, Maisa Alkheder Alahmad6,
Filipp Bezold6‡, Valerie Fuchs6, Petra Gnau6, Stephan Kiontke6, Lukas Korf6, Viktoria Reithofer6,
Christian Joshua Rosner6, Elisa Marie Seiler6, Mohamed Watad6, Laura Werel6,
Roberta Spadaccini6,15, Junpei Yamamoto5, So Iwata3,13, Dongping Zhong10,16,17, Jörg Standfuss7,
Antoine Royant8,11, Yoshitaka Bessho1,3*, Lars-Oliver Essen6*, Ming-Daw Tsai1,18*
Photolyases, a ubiquitous class of flavoproteins, use blue light to repair DNA photolesions. In this work, we
determined the structural mechanism of the photolyase-catalyzed repair of a cyclobutane pyrimidine dimer
(CPD) lesion using time-resolved serial femtosecond crystallography (TR-SFX). We obtained 18 snapshots
that show time-dependent changes in four reaction loci. We used these results to create a movie that
depicts the repair of CPD lesions in the picosecond-to-nanosecond range, followed by the recovery of the
enzymatic moieties involved in catalysis, completing the formation of the fully reduced enzyme-product
complex at 500 nanoseconds. Finally, back-flip intermediates of the thymine bases to reanneal the DNA were
captured at 25 to 200 microseconds. Our data cover the complete molecular mechanism of a photolyase and,
importantly, its chemistry and enzymatic catalysis at work across a wide timescale and at atomic resolution.
P hotolyases catalyze the repair of damaged
DNA that contains ultraviolet (UV) light–
induced lesions, including cyclobutane
pyrimidine dimers (CPDs) and the 6-4
photoproduct (6-4PP) (1, 2). CPD lesions
are the most abundant type of DNA damage
that occurs in nature upon UV-B exposure and
are a major cause of skin cancer in humans (3).
Catalysis by CPD photolyases involves multi-
ple redox reactions and a multistep splitting
of the cyclobutane ring (4). The enzyme is first
activated through photoreduction of its coen-
zyme flavin adenine dinucleotide (FAD) from the
oxidized (FADox) to the reduced state (FADH–)
(5). This process requires two single-electron
transfer steps mediated by an electron transfer
chain that involves three tryptophan residues,
yielding, respectively, the radical semiquinone
intermediate (FAD(cid:129)– or its protonated form
FADH(cid:129)) and the reduced, anionic hydroquinone
state (FADH–) (6). Only the latter is capable of
catalyzing light-induced DNA repair. The mech-
anisms of photoreduction and DNA repair have
been elucidated by extensive spectroscopic and
theoretical studies over the past three decades
(1, 5, 7–11). However, relatively limited informa-
tion from high-resolution crystal structures is
available, and, apart from the apo form (12–14),
all such structures were obtained using synchro-
tron x-ray radiation that induced both photo-
reduction (15–17) and DNA repair reactions (18).
Recently, we used time-resolved serial femto-
second crystallography (TR-SFX) to elucidate
the structural mechanism of the photoreduc-
tion processes in the Methanosarcina mazei
class II CPD photolyase (MmCPDII) (12). We
identified an Asn/Arg-Asp redox sensor triad
that regulates FAD rehybridization and proto-
nation and observed buckling and twisting of
the isoalloxazine ring of the coenzyme FAD,
which occurred in the submicrosecond regime
after the light-triggered electron transfer step.
The TR-SFX study of the photoreduction pro-
cess set the stage for us to study the main func-
tion of the MmCPDII photolyase—the repair
of CPD lesions, which we present here.
On the basis of extensive spectroscopic and
computational analyses (1, 2, 4, 5, 19, 20), the
repair of the cyclobutane-bridged TT dimer
(T<>T, 1 in Fig. 1A) by photolyases was sug-
gested to go through three intermediates. First,
a forward electron transfer (FET) from the re-
duced coenzyme FADH– produces (T<>T)•– (2),
where the excess electron may be delocalized
between the two pyrimidines. The resulting
CPD radical anion then undergoes cleavage of
its C5–C5′ bond to produce (T_T)•– (3) and fur-
ther cleavage of the C6–C6′ bond to give (T+T)•–
(4). The latter then transfers an electron back
to the flavin coenzyme to produce the repaired
thymine bases (T+T, the product, designated
as 5 in Fig. 1A). Although such a repair mech-
anism is chemically feasible and well supported
by spectroscopic data, structures of its reaction
intermediates await elucidation in the enzyme-
bound form. Furthermore, given the moderate
quantum yields for DNA repair, at least in class II
photolyases, the underlying interactions between
reaction intermediates, the isoalloxazine ring
of the coenzyme, and active-site residues of the
photolyase remain to be revealed at high reso-
lution. The first goal of this TR-SFX study was
hence to identify and characterize five reaction
intermediates that correspond to the chemical in-
termediates shown in Fig. 1A in the DNA-enzyme
complex, Int1 to Int5. Their assignment was based
on observed changes of the coenzyme’s geom-
etry, conformations of active-site residues, and
the chemical state of the CPD during its repair.
The chemical reactions and electron trans-
fers shown in Fig. 1A usually occur at the ps-ns
timescale, but we are interested in the com-
plete catalytic cycle of the enzyme, which addi-
tionally involves slower conformational steps
in the ns-ms range. When complexed to photo-
lyases, the TT dimer in UV-damaged DNA, which
retains base pairing with adenine bases in
the unbound state, though somewhat twisted
(21), flips out of double-stranded DNA (dsDNA)
1Institute of Biological Chemistry, Academia Sinica, 128 Academia Rd. Sec. 2, Nankang, Taipei 115, Taiwan. 2Department of Chemistry, National Taiwan University, 1, Roosevelt Rd. Sec. 4, Taipei
106, Taiwan. 3RIKEN SPring-8 Center, 1-1-1 Kouto, Sayo, Hyogo 679-5148, Japan. 4Institute of Multidisciplinary Research for Advanced Materials, Tohoku University, 2-1-1 Katahira, Aoba-ku,
Sendai 980-8577, Japan. 5Division of Chemistry, Graduate School of Engineering Science, Osaka University, 1-3 Machikaneyama, Toyonaka, Osaka 560-8531, Japan. 6Department of Chemistry,
Philipps University Marburg, Hans-Meerwein Strasse 4, Marburg 35032, Germany. 7Paul Scherrer Institute, Forschungstrasse 111, 5232 Villigen PSI, Switzerland. 8European Synchrotron Radiation
Facility, 38043 Grenoble, France. 9Hamburg Centre for Ultrafast Imaging, Universität Hamburg, 22761 Hamburg, Germany. 10Department of Physics, The Ohio State University, Columbus, OH
43210, USA. 11Université Grenoble Alpes, CNRS, CEA, Institut de Biologie Structurale (IBS), 38044 Grenoble, France. 12Japan Synchrotron Radiation Research Institute, 1-1-1 Kouto, Sayo, Hyogo
679-5198, Japan. 13Department of Cell Biology, Graduate School of Medicine, Kyoto University, Yoshidakonoe-cho, Sakyo-ku, Kyoto 606-8501, Japan. 14Université Paris-Saclay, CEA, CNRS,
Institute for Integrative Biology of the Cell (I2BC), 91198 Gif-sur-Yvette, France. 15Dipartimento di Scienze e tecnologie, Universita degli studi del Sannio, Benevento, Italy. 16Department of
Chemistry and Biochemistry, The Ohio State University, Columbus, OH 43210, USA. 17Center for Ultrafast Science and Technology, Shanghai Jiao Tong University, Shanghai 200240, China.
18Institute of Biochemical Sciences, National Taiwan University, 1, Roosevelt Rd. Sec. 4, Taipei 106, Taiwan.
*Corresponding author. Email: mmaestre@ntu.edu.tw (M.M.-R.); bessho@spring8.or.jp (Y.B.); essen@chemie.uni-marburg.de (L.-O.E.); mdtsai@gate.sinica.edu.tw (M.-D.T.)
†Present address: Department of Pharmacology, University of California, San Diego, La Jolla, CA 92093, USA. ‡Present address: CSL Behring Innovation GmbH, 35041 Marburg, Germany.
Maestre-Reyna et al., Science 382, eadd7795 (2023)
1 December 2023
1 of 14
RES EARCH | R E S E A R C H A R T I C L E
A
8
O
9
9'
O
8'
1
HN
3
2
7
O
5
6
4
1
N
NH
3'
2'
5'
6'
4'
1'
N
7'
O
hν
T<>T
e-
FADH-
FADH -*
FADH(cid:129)
O
3'T
N
HN
O
5
NH
O
5'T
N
O
T+T
e-
O
O
HN
-
(cid:129)
2
NH
O-
O
HN
(cid:129)
3
NH
O-
O
HN
(cid:129)
4
NH
O
N
N
O
O
N
N
O
O
N
N
O
(T<>T)(cid:129)-
(T_T)(cid:129)-
(T+T)(cid:129)-
B
C
O
3
2
4
4a
10a
1
N-
O
6
H
5
N
5a
9a
10
9
N
R
7
8
CH3
CH3
Fig. 1. Proposed intermediates during repair of T<>T lesions as catalyzed by DNA photolyases.
(A) Canonical scheme of DNA repair by CPD photolyases. 3′-T, 3′-thymine; 5′-T, 5′ thymine. (B) The CPD
lesion–containing dsDNA used in this work, which contains the cis-syn CPD with its phosphodiester linkage
(22). (C) Illustration of the rC and rN dihedral angles of the FADH– isoalloxazine moiety. As shown by
TR-SFX of photoreduction, the isoalloxazine moiety of FAD in MmCPDII undergoes symmetrical buckling or
bending (rC = rN) or asymmetrical twisting (rC ≠ rN) upon redox change (12). Most importantly, the
transition from one geometry to another is orders of magnitude slower than electron transfer itself.
to interact with the FAD coenzyme and active-
site residues. This double flip causes a strong
kink in the canonical B-DNA structure and
creates an unpaired bubble that is stabilized
by the photolyase’s bubble-intruding region
(BIR) (22). Accordingly, a second goal of this
TR-SFX study was to identify the conformational
intermediates (Con) in the inverse conforma-
tional changes that are required for dsDNA re-
annealing and dissolution of the DNA-photolyase
complex after completion of the chemical CPD
repair. This analysis may be particularly reward-
ing given that very little dynamic information
is available for the late part of the reaction cy-
cle and that photolyases bind already repaired
dsDNA >4 orders of magnitude more weakly
than they bind UV-damaged dsDNA (23).
Results
Tailoring TR-SFX to uncover the
photolyase mechanism
Our DNA repair study depended on cocrystals
of fully reduced MmCPDII and CPD-containing
DNA (Fig. 1B) (22), which were grown within
12 hours before use directly at x-ray free elec-
tron laser (XFEL) sites using anaerobic tents
and safety-light conditions. Two sets of TR-SFX
data were collected: The first covers the CPD
repair reaction itself from 100 ps to 10 ns (the
ps-ns series), as performed at SwissFEL, where
crystals were excited by a 0.98-ps-long, 10-mJ,
400-nm pump laser pulse with a 50-mm diam-
eter focal spot. For the second time series,
which covers relaxation of the photolyase-pro-
duct complex and DNA release [10 ns to 200 ms,
the ns-ms series, SACLA (SPring-8 Ångstrom
Compact free-electron Laser)], samples were
excited by a 3-ns-long, 150-mJ, 408-nm pump
laser pulse with a 100-mm focal spot diameter.
Although illumination at ~400 nm likely gen-
erates a mixture of S1 and S2 excited FADH–*
species, the S2→S1 decay has been noted to
occur within a few ps, considerably faster than
FET (24).
For photolyase-mediated DNA repair, TR-
SFX conditions were optimized for achieving
anticipated DNA repair yields, as detailed in
supplementary text S1 and figs. S1 and S2. A
brief summary is provided here: (i) We confirmed
that the oxidized form of MmCPDII was unable
to initiate FET toward CPD and failed to cause
any effects, including repair of the CPD under
very high laser powers, both under ps-ns and
ns-ms conditions (fig. S1, A and B). (ii) Before per-
forming TR-SFX experiments that addressed
CPD repair itself, we validated that difference
density signatures of the 10-ns TR-SFX structure
(fig. S1E) coincided with those initially reported
from x-ray radiation-triggered repair (fig. S1C)
(17, 18) and low-dosage light-emitting diode (LED)–
induced repair (fig. S1D). (iii) The reduction of
laser power from 10 to 5 mJ in the ps-ns series
caused diminished signature signals (fig. S2B).
(iv) Control experiments for light contamination
assured clean dark datasets (fig. S2, A and B).
In TR-SFX experiments, the power and ener-
gies of laser excitation pulses generally exceed
one absorbed photon per chromophore, thus
implying potential multiphoton effects in light-
triggered systems, which may affect the photo-
chemical reaction course and dynamics (25–30).
Additionally, multiphoton absorption can result
in sample heating due to chromophore relaxa-
tion from high-energy states. We found that not
more than ~68% of incident light is accessible
for microcrystals embedded in grease because
of light scattering (supplementary text S2 and
fig. S3). Accordingly, calculation of photon dos-
ages as absorbed by the enzyme must take this
and other factors such as the quantum yields into
account. Nevertheless, our results indicated that
up to 10 and 30 photons are potentially absorbed
per FAD chromophore per pulse for the ps-ns
and ns-ms series, respectively (supplementary
text S3), numbers that are high enough to cause
activation of the FADH– coenzyme to levels
higher than the S1 state. However, electronic
relaxation of FADH–* to S1 from higher states
than S2 is expected to proceed even faster than
the S2→S1 transition by internal conversion (24),
that is, in the sub-ps range. Given the time-
resolution of our TR-SFX experiments (Table 1),
multiply excited FADH– species should have
decayed before the first catalytic event, that is,
electron-transfer to the CPD lesion.
Analyzing TR-SFX data for
photolyase-mediated CPD-DNA repair
Given the complex set of TR-SFX data, a brief
explanation of the methods we used for struc-
tural analysis and the nomenclature is given
first. For structure factor extrapolation and
refinement, we followed previously described
guidelines (31–34). For both time series, the
basic conditions and main structural features
are summarized in Table 1, whereas the de-
tailed structural parameters are listed in tables
S1 and S2. Each structure is given a name in
the first column of Table 1, for example, F10ns
and N10ns for the 10-ns snapshots, where the
one-letter code indicates the type of pumping
laser (F for femtosecond laser pump in the ps-ns
series and N for nanosecond pump in the ns-
ms series), followed by the delay time after
Maestre-Reyna et al., Science 382, eadd7795 (2023)
1 December 2023
2 of 14
RES EARCH | R E S E A R C H A R T I C L E
Table 1. Properties of the static and time-resolved structures of MmCPDII complexes with damaged DNA. A “–” indicates that no value was assigned.
N/A, not applicable; n.d., not determined; PDB, Protein Data Bank.
Structure*
Intermediate* TT state†
FAD state
rC /rN (°)
5WC/R256§
BIR¶
D428-W431-
R441
C–C peak
distance
C5/C6 (Å)#
Negative
densities
C5/C6
(e–/Å3)††
PDB
code
Resolution (Å)
Oxidized photolyase:DNA complex
............................................................................................................................................................................................................................................................................................................................................
Dynamic/
CPD§
Int1
T<>T
FADox
2.0/2.1
ox/ss
............................................................................................................................................................................................................................................................................................................................................
ps-ns time series (SwissFEL)
............................................................................................................................................................................................................................................................................................................................................
Fdark
............................................................................................................................................................................................................................................................................................................................................
F100ps
............................................................................................................................................................................................................................................................................................................................................
F250ps
............................................................................................................................................................................................................................................................................................................................................
F450ps
............................................................................................................................................................................................................................................................................................................................................
F650ps
............................................................................................................................................................................................................................................................................................................................................
F1ns
............................................................................................................................................................................................................................................................................................................................................
FADH–
FET‡ FADH–*/
FADH•
FADH•
FADH•
T<>T
FET‡ T<>T/
(T<>T)•–
(T<>T)•–
(T_T)•–
(T+T)•–
20.1/20.0
4.6/6.3
−8.9/−5.3
4.0/7.6
7.1/10.1
3.7/9.1
−/−
−/−
−/−
1.5/−
1.9/−
1.9/1.9
0.19/0.00
0.18/0.02
0.21/0.03
0.19/0.03
7YC7
0.00/0.00 7YCM
Ordered/5WC
Ordered/CPD
7YCP
7YCR
7YD6
7YD7
Int2/3
Int3
Int3/4
2.08
2.15
2.15
2.25
Dynamic/CPD
1.95
2.00
Int1*/2
Locked
7YE0
2.75
0/0
0/0
Locked
−/−
Dynamic/CPD
BET‡ (T+T)•–/
T+T
Int4/5
Int4/5
F2ns
............................................................................................................................................................................................................................................................................................................................................
F3.35ns
............................................................................................................................................................................................................................................................................................................................................
F6ns
............................................................................................................................................................................................................................................................................................................................................
F10ns
............................................................................................................................................................................................................................................................................................................................................
ns-μs series (SACLA)
............................................................................................................................................................................................................................................................................................................................................
Ndark
............................................................................................................................................................................................................................................................................................................................................
FADH– recovering
–
FADH
recovering
20.0/20.0 Ordered/5WC
0.25/0.08
0.25/0.24
0.19/0.18
11.0/14.2
Int5
Int5
2.7/2.3
6.4/1.4
2.3/1.9
2.3/1.9
7.1/8.3
T+T
T+T
T<>T
7YDZ
7YEB
7YEC
7YEE
−/−
2.20
2.20
2.23
2.15
0.21/0.08
7YD8
2.3/2.7
2.15
−6.1/8.2
T+T
7YEI
2.70
Con1
T+T
T+T
1.9/1.9
Locked
10.9/13.4
Dynamic/CPD
Dynamic/5WC
1.9/2.3
2.3/2.7
T/T
T/T
-T-T-
FADH–
recovered
20.0/20.0 Ordered/5WC
Con2a
Con2b
Con3a
Con3b/4
Con4/3b
N10ns
............................................................................................................................................................................................................................................................................................................................................
N100ns
............................................................................................................................................................................................................................................................................................................................................
N500ns
............................................................................................................................................................................................................................................................................................................................................
N25μs
............................................................................................................................................................................................................................................................................................................................................
N200μsA*
............................................................................................................................................................................................................................................................................................................................................
N200μsB*
............................................................................................................................................................................................................................................................................................................................................
*The definitions of structure names and intermediate numbers are explained in the text. Each structure has a unique PDB code, except for N200μsA and N200μsB, which share the same code
(7YEM) because they correspond in N200μs to the two independent complexes, A and B, of the asymmetric unit. When multiple intermediates coexist in a structure, the bolded number designates
the major form. F250ps may also contain a contribution by Int3 based on negative density at C5–C5′ only.
†The TT state defines the linkage between the 5′-thymine and the 3′-thymine as the
following: T<>T, linked by cyclobutane; T_T, linked by the C6–C6′ bond; T+T, not linked, but both thymines are in the active site; T/T, not linked, and only the 3′-thymine is in the active site; -T-T-,
‡Here, no dominant CPD, TT, and FAD intermediates can be defined. Hence, these structures should be
both thymines have flipped back and are reannealing to adenine counterbases.
§These are the conformations of active-site moieties 5WC and the R256 side chain. 5WC is designated as ordered (intact) or dynamic. R256 is designated as pointing
considered as mixtures.
toward 5WC or CPD. N/A is noted for 200 μs because the thymine bases have moved out of the active site.
¶The BIR composed of amino acids D428, W431, and R441 acts as a lock. In its
locked state, BIR stabilizes the unpaired bubble and prevents CPD flip-back before repair. Upon completion of CPD repair, flip-back of the repaired bases unlocks the BIR, displaces it, and initiates
#C–C peak distances corresponding to the distance between the maxima of the characteristic CPD repair signals, that is, the cyclobutane negative density and the 5′-
product release.
thymine positive density, along the C5–C5′ and the C6–C6′ axes. The position of the maximum was calculated as described in the Materials and methods summary and fig. S5.
**The thymine bases have started to flip back to pair with adenine bases.
atoms, calculated as described in supplementary text S4 and fig. S13B.
††These values are the average integrated negative electron densities from difference maps around the C5 and C6
7YEJ
7YEK
7YEL
7YEM
7YEM
2.55
2.40
2.55
2.60
2.60
Unlocking
Unlocked
n.d./n.d.
−/−**
N/A§
FADH•/
FADH•/
FADH–
FADH–
FADH–
FADH–
recovering
illumination (with “dark” designating a non-
illuminated time-resolved control structure).
Exceptions are the steady-state structures in the
oxidized (ox/ss) and fully reduced states (Fdark,
Ndark). Isomorphous difference electron density
maps (DFo) (35) were calculated to decipher
structural differences, which are designated as
the difference between two states, DFo(Y-X).
Major properties of each structural snap-
shot are depicted in Table 1, which include
features from the four main reaction loci (Fig.
2). These properties are considered in the in-
terpretation of each structural snapshot. For
the ps-ns series, the focus was on CPD repair. The
structural snapshots were hence assigned to
match the TT status of the proposed interme-
diates 1 to 5 (Fig. 1A) and designated as Int1 to
Int5 of the photolyase-DNA complexes (Table 1).
For the ns-ms series with its already completed
CPD repair reaction, structural snapshots re-
flecting postrepair reactions are designated as
conformational intermediates (Con1 to Con4),
with close variants being further subdivided
by a, b, and so on.
Structural analysis of the TR-SFX data was
conducted in three rounds: First, on the basis
of mainly structural changes at CPD and FADH–
(two of the four reaction loci; Fig. 2), each time
point was assigned an intermediate or inter-
mediates that play a role in the catalytic cycle.
Next, structural changes in the other two re-
action loci were included to fine-tune the in-
terpretation of each time point. Finally, all
datasets and interpretations were considered
together along with additional supporting evi-
dence derived from electron density analyses,
singular value decomposition (SVD) analysis,
and quantum mechanical (QM) calculations.
The four main reaction loci of CPD-DNA repair
by MmCPDII
Global structures of the MmCPDII-DNA com-
plex, as determined by SFX, in both the oxi-
dized and fully reduced states (ox/ss and Fdark
in Fig. 2A and Ndark) agree well with prior
synchrotron structures of the DNA complex
in the oxidized states (16, 22) and MmCPDII
SFX structures without bound DNA in all
redox states (12). However, difference den-
sity peaks are found at four main loci: the
CPD structure (Fig. 2B), the FAD coenzyme
(Fig. 2C), the active-site moieties R256 (R, Arg)
and five-water cluster (5WC) (Fig. 2D), and the
class II photolyase–defining BIR (Fig. 2E).
These loci also show all prominent difference
density map features after light-triggered DNA
repair, though the general protein fold itself is
not affected (fig. S4). In this section, we survey
Maestre-Reyna et al., Science 382, eadd7795 (2023)
1 December 2023
3 of 14
RES EARCH | R E S E A R C H A R T I C L E
A
D
B
C
E
Fig. 2. Static structures of MmCPDII complexes with damaged DNA and
illustration of the four main reaction loci. Structure of Fdark (reduced, green)
superposed with ox/ss (steady-state oxidized, gray) are shown along with
3s-contoured difference maps between the dark and oxidized states, DFo(dark-ox/ss)
(positive peaks in cyan, negative peaks in magenta). (A) Global structures showing
the overall fold, with the bound DNA backbone in orange. (B) Structures of the
bound CPD. (C) Structures of the isoalloxazine ring in the oxidized (FADox, gray) and
fully reduced form (FADH–, gold), showing increased buckling of the FAD in the
reduced state. (D) The 5WC/R256 locus at the active site. The interactions between
CPD; FAD and active-site residues R256, W305, and W421; and the 5WC that only
appears in the Fdark state are also shown. In the presence of the 5WC, R256
undergoes a conformational change, with its guanidinium moiety becoming the final
vertex in the bipyramidal 5WC. Some interatomic distances are highlighted by
dashed lines, with values in angstroms nearby. (E) The BIR with its D428/W431/
R441 locking triad, as well as R429, which mostly interacts with the unpaired bases
dA7′ and dA8′ complementary to the thymine dimer. The lack of substantial
difference electron density here suggests little conformational change related to the
flavin’s redox state for this region.
the four main reaction loci and explain how
the relevant data in Table 1 were obtained.
1) CPD locus. To track CPD repair itself, we
used two key indicators based on the charac-
teristic CPD difference density features. First,
we determined the C5–C5′ and C6–C6′ peak-to-
peak distances (designated as C–C peak dis-
tances) between the maxima of the difference
map peaks along the axes of the C5–C5′ and
C6–C6′ bonds (Table 1, C-C peak distance C5/C6
and footnote ¶; Materials and methods sum-
mary; and fig. S5). Secondly, we calculated the
Maestre-Reyna et al., Science 382, eadd7795 (2023)
1 December 2023
4 of 14
RES EARCH | R E S E A R C H A R T I C L E
average integrated negative electron den-
sities (designated as negative densities) from
difference maps around C5 and C6 atoms as a
function of time (Table 1, negative densities
C5/C6 and footnote **; and supplementary
text S4), which is a good indicator of the accu-
mulation of repair intermediates where the
corresponding bond had been broken. The as-
signed linkage and charge state of the 5′-
thymine and the 3′-thymine is shown in Table
1 (TT-state) using a notation (T<>T, T_T, T+T,
T/T, -T-T-) that is defined in the corresponding
footnote.
2) FAD locus. Subtle redox-dependent changes
around the flavin site, including increased buckl-
ing of the FAD upon reduction (Fig. 2C), are
consistent with our recent studies of MmCPDII
photoreduction (12). In that work, we showed
by monitoring the rC and rN dihedral angles,
as defined in Fig. 1C, that the isoalloxazine
ring of the FAD coenzyme undergoes a sequence
of buckling and twisting motions during the
process of photoreduction and protonation.
In the oxidized state FADox, the isoalloxazine
moiety was only very mildly buckled (rC, rN:
2.0°, 2.0°), whereas in the catalytic FADH–
state, the isoalloxazine moiety was strongly
buckled (rC, rN: 14.3°, 14.5°). In the semiquinone
FADH• state, it moved closer to planarity (rC, rN:
4.6°, 4.8°) (12). Because FADH– is transiently
oxidized into FADH• upon FET toward CPD
and then recovered upon reaction completion,
the rC and rN dihedral angles may hence well
reflect the electron flow during CPD repair
(Table 1, FAD state and rC/rN).
3) 5WC/R256 locus. In the active site, the
substrate interacts with several key features of
the enzyme (Fig. 2D), including the adenine
moiety of FAD, the highly conserved residue
R256, and the structurally conserved 5WC.
In the fully reduced state, that is, Fdark and
Ndark, the 5WC bridges the bisphosphate
backbone of FAD and R256 so that it con-
tributes to the electrostatic stabilization of the
anionic FADH–. Upon FET, R256 moves to
stabilize the CPD radical anion and the 5WC
becomes disordered or dynamic. Changes in
5WC/R256 during repair are monitored in Table
1 (5WC/R256 and footnote ‡).
4) BIR locus. In the BIR (Fig. 2E), the D428-
W431-R441 triad (D, Asp; W, Trp) acts as a lock,
both stabilizing the unpaired bubble and firmly
maintaining the protein-DNA complex (22),
whereas R429 is more flexible and capable of
dynamically interacting with both CPD com-
plementary bases, dA7′ and dA8′. Here, post-
repair base flip-back induced the opening of
the BIR lock, as shown in Table 1 (BIR D428-
W431-R441 and footnote §).
Given our structural snapshots, the entire
process of photolyase-mediated repair of dsDNA
can be assigned to three periods, that is, early,
middle, and late events, which correspond to
CPD repair, active-site recovery, and thymine
back-flip/DNA release, respectively. The loci of
CPD structure and FAD geometry will be con-
sidered first to assess the primary repair process.
The remaining two reaction loci will then be
separately addressed. Considering these results
together, we then used them to assemble a mo-
lecular movie covering DNA repair and ordered
product release by photolyases.
Early events: Electron transfer–driven repair of
the CPD lesion
The refined structures and difference maps
from the ps-ns series monitor both the ring-
opening steps of the CPD repair reaction and
the concomitant FAD isoalloxazine moiety dy-
namics (Fig. 3A). The C5–C5′ and C6–C6′ peak
distances as well as the integrated negative
electron densities (Fig. 3B) and the flavin di-
hedral angles rC and rN (Fig. 3C) are plotted
over time. The structure of each time point in
Fig. 3A represents the predominantly accu-
mulated species of a snapshot during CPD repair.
Although the structural changes shown here are
subtle, the DFc(Y-dark) of each individual struc-
ture is always maximally correlated to its cor-
responding DFo(Y-dark) (fig. S6, A and B), which
supports the reliability and significance of the
refined structural changes and yields a clear
view of the process of DNA repair.
Accordingly, we have observed all proposed
CPD repair intermediates 1 to 5 (Fig. 1A) in the
enzyme bound form as Int1/Int1* to Int5 (Table 1),
where the Fdark structure represents Int1 (T<>T/
FADH–). At 650 ps, the magnitude of the negative
density feature around the C5–C5′ bond is much
higher than that around the C6–C6′ bond. In good
agreement, the C5–C5′ peak distance is 1.9 Å
(Fig. 3, A and B, and Table 1), which implies a
broken C5–C5′ and an intact C6–C6′ bond, sug-
gesting that the CPD ring-opening intermediate
(T_T)•–, Int3, is dominant in the F650ps
snapshot. At F1ns, with the negative density
around C5–C5′ being strong and that around
C6–C6′ growing, both C5–C5′ and C6–C6′
difference density peak distances reach 1.9 Å.
Here, the dominant species corresponds to the
repaired CPD intermediate (T+T)•–, Int4. Inter-
estingly, these trends continue to the F2ns snap-
shot, suggesting a continuing buildup of Int4.
This in turn suggests that F1ns may still con-
tain a contribution by Int3.
The first structure after photon absorption
corresponds to the activated coenzyme, FADH–*.
The F100ps snapshot (Fig. 3A) might represent
this state because its CPD structure is unaltered,
whereas its isoalloxazine ring has greatly
changed. Nevertheless, the R256 side chain
has moved toward CPD, suggesting that FET
from FADH–* has already begun at this point.
In the F250ps snapshot, the isoalloxazine ring
continued to swing, as shown by sign changes
of its rC and rN angles (Fig. 3C). Although the
CPD structure still remains unchanged, the
nearby active-site moieties R256/5WC have
already begun to move, as described later. These
changes suggest that FET is in progress during
both the F100ps and F250ps snapshots, which
represent a mixture of species, T<>T/FADH–*
(Int1*) and (T<>T)•–/FADH• (Int2). This elec-
tron transfer–dependent reorganization is in
good agreement with the proposal that such
ultrafast redox transitions are stabilized by rapid
isolloxazine fluttering (36) and harmonic mo-
tions of other light-gathering cofactors after
absorption (28).
In the F450ps snapshot, a C5–C5′ peak dis-
tance could be measured for the first time
because C5–C5′ negative density has accumu-
lated. Nevertheless, in the refined structure,
both the C5–C5′ and C6–C6′ bonds remained
intact. Hence, this snapshot can be assigned to
an Int2/3 mixture, in which Int2 is the major
species. Interestingly, from F450ps up to F1ns,
structures showed relatively planar FAD iso-
alloxazine geometries compared with the dark,
fully reduced state (Fig. 3C), which is expected
for the semiquinoid FADH• form (12). This indi-
cates that FET is mostly finished in the F450ps
snapshot and that the radical anionic CPD
species dominate these snapshots. However, im-
mediately after the 1-ns delay (at 2 and 3.35 ns),
there are notable oscillatory motions of the
FAD’s isoalloxazine ring, as indicated by the rC
and rN dihedral angles, which correlate well
with fluttering that is associated with rapid
relaxation after sudden redox changes (1). This
fluttering suggests back electron transfer (BET)
from the thymine radical anion to the coen-
zyme and is in good agreement with the BET
time constant of ~700 ps that was determined
during time-resolved spectroscopic studies
of CPD repair by other photolyases (37, 38).
Accordingly, we assign the F2ns and F3.35ns
snapshots as representatives for the BET re-
action, which should contain varying mixtures
of (T+T)•–/FADH• (Int4) and (T+T)/FADH–
(Int5). In the F6ns snapshot, the isoalloxazine
oscillatory motion appears to have receded,
and therefore we propose that F6ns consists
predominantly of the product after completion
of the CPD repair, Int5. Because the structural
properties of the next snapshot, F10ns, are very
similar to those of F6ns (Fig. 3A and Table 1),
F10ns represents further accumulation of the
Int5 product.
Middle events: Base stacking and FAD
site recovery
The ns-ms series of TR-SFX data addresses
conformational changes of the repaired thy-
mines and enzymatic moieties during product
release. Throughout this series, the status of
the intact thymines (T+T) and the coenzyme
FADH– corresponds chemically to Int5. Never-
theless, given that the T+T and FADH– con-
formations and active-site arrangements also
vary with time, we describe them as distinct
conformational intermediates. Notably, the
Maestre-Reyna et al., Science 382, eadd7795 (2023)
1 December 2023
5 of 14
RES EARCH | R E S E A R C H A R T I C L E
A
B
C
Fig. 3. Bond breaking and ring opening during cleavage of the cyclobutane
ring. (A) Structures of the CPD moiety (T<>T) of the bound DNA as well as the FAD
isoalloxazine moiety from the first TR-SFX series (orange), from F100ps to F10ns,
each superposed with the reduced dark structure (Fdark in green and FADH–
in gold). Each panel also shows 3s-contoured DFo(Y-dark) maps, with positive peaks
in cyan and negative ones in magenta. (B) Plots of C5–C5′ (green) and C6–C6′
(orange) peak distances (full lines) and average integrated negative electron
densities (dashed lines) over time. Because both the negative cyclobutane and
the positive 5′-thymine peaks are necessary to calculate peak distances, when
either was missing, the distance could not be calculated and no value is plotted
(100 to 250 ps for C5–C5′ and 100 to 650 ps for C6–C6′). (C) Plots of rC (blue)
and rN (red) dihedral angles over time. ET, electron transfer.
N10ns snapshot (Con1, the first structure in
Fig. 4A) and the F6ns/F10ns snapshots (Int5;
Fig. 3A) coincide closely, thus providing fur-
ther evidence for the adequacy of the experi-
mental parameters used in both series. To our
surprise, before the onset of product release,
the intact thymines stay fully stacked in the ac-
tive site for up to 500 ns (Table 1 and Fig. 4A).
Apparently, the active site undergoes relaxation
before the two repaired thymine bases start to
flip back out of the active site. In the N100ns
and N500ns snapshots, the isoalloxazine geom-
etry of the FADH– coenzyme returns to the ini-
tial fully reduced state (Fig. 4, A and C). These
two structures, assigned as Con2a and Con2b
Maestre-Reyna et al., Science 382, eadd7795 (2023)
1 December 2023
6 of 14
B
C
RES EARCH | R E S E A R C H A R T I C L E
A
D
E
Fig. 4. Postrepair events that lead to thymine-base return and DNA release.
(A) Structures of the newly repaired thymines as well as the isoalloxazine moiety
from the second TR-SFX series (orange), from N10ns, N100ns, and N500ns, each
superposed over the reduced dark structure (Ndark CPD is shown in green, whereas
its isoalloxazine moiety is depicted in gold). Each panel also shows 3s-contoured
DFo(Y-dark) maps, with positive peaks in cyan and negative ones in magenta. (B and
C) Peak distance plots along the C5–C5′ (green) and C6–C6′ (yellow) axes (B) and
rC (blue) and rN (red) angles plots (C) over time for the second TR-SFX time series.
(D) Structures of the newly repaired thymines from the second TR-SFX series
(orange) during their return to the dsDNA, including N25ms and N200ms, each
superposed over the reduced dark structure (Ndark in green). Because the two
MmCPDII-DNA complexes in N200ms are at two different DNA release stages,
they are both shown here. Each panel also shows 3s-contoured DFo(Y-dark)
maps as in (A). (E) Difference maps at the dsDNA, from left to right: N10ns,
N500ns, N25ms, and complex B from N200ms. For all complexes, 3s-contoured
DFo(Y-dark) maps are superposed over the entire DNA molecule.
Maestre-Reyna et al., Science 382, eadd7795 (2023)
1 December 2023
7 of 14
RES EARCH | R E S E A R C H A R T I C L E
intermediates, respectively, differ only slightly for
the 5WC/R256 locus (see below). Interestingly
during photoreduction of MmCPDII without
bound DNA, the isoalloxazine ring similarly
adopts its full buckling at about 300 ns after its
light-driven FADH•→FADH– transition (12).
An interesting observation from these data
is the rapid decrease of isoalloxazine dihedral
angles during FET but slow increase and re-
covery after BET (Fig. 3C and Table 1). A possible
explanation is that FET occurs between CPD
and the light-excited FADH–* in its S1 state
(39). The oxidation of FADH–* into FADH• is a
downhill transition between an unstable, high-
energy state to a stable, low-energy state, where-
as reduction of FADH• to FADH– represents a
transition between two ground states (D0 to S0).
Although BET transiently perturbs the FADH
structure, as evidenced by strong torsion and
fluttering between 1 and 3 ns (Fig. 3C), the accu-
mulated energy within the isoalloxazine moiety
is only slowly dispersed over hundreds of nano-
seconds, as seen both here and previously for the
MmCPDII photoreduction (12).
Late events: Base back-flip and onset of
reannealing of dsDNA
The isoalloxazine geometry no longer changes
in the 25 and 200 ms snapshots (Fig. 4D). Here,
the repaired thymine bases finally move out of
the active site and flip back toward the dsDNA.
In a continuation from prior time points, where
postrepair movement concentrated around
the 5′-thymine (Figs. 3A and 4A), it is the 5′-
thymine (dT7) that first starts to exit the active
site during the N25ms snapshot. However, this
back-flipping base is too disordered to be refined
because only the pentose is well resolved (Fig.
4D). Because the two thymine bases are no
longer stacked and the 5′-thymine flips back,
this structure can be considered as a “thymine
back-flipping intermediate,” which we define
as TT status T/T (Con3a), reflecting the non-
stacking relationship between the two thymine
bases. Meanwhile, the 3′-thymine dT8 is tilted
in such a way that it occupies the full space
within the active site (Fig. 4D). The N200ms
snapshot is particularly interesting because the
two thymine bases continued the back-flipping
process but were captured at different stages
in the two complexes in the asymmetric unit.
In complex A, the dT7 has further moved and
its density has now become well defined (in-
termediate Con3b, TT status: T/T). In com-
plex B, both thymine bases have flipped back
and undergo base pairing with their respective
complementary bases. This complex can be de-
signated as the “product-reannealing complex”
(Con4), which corresponds to regular dsDNA
and is denoted as the “-T-T-” TT status. There
are extra difference signals in complex A that
can be fit to Con4 (30 to 50%) and likewise in
complex B that correspond to Con3b (30 to
40%) (fig. S7). This population shift between
complexes A and B in the N200ms snapshot
likely derives from crucial crystal contacts made
by complex A but not by complex B (fig. S8). In
terms of thymine back-flipping, complex B is
hence less constrained by the crystal lattice and
can restore the integrity of the duplex DNA in a
shorter time frame. Considering that repaired
dsDNA is known to dissociate from the photo-
lyase, it is reasonable to consider Con4 as a
transient intermediate captured just before
final complex dissociation. Unlike all structures
up to 500 ns that show only a few difference
density peaks in the DNA region (Fig. 4E, left),
the N200ms complex B reveals extensive dif-
ference density features along the entire dsDNA
(Fig. 4E, right).
Active-site dynamics during CPD repair: The
5WC/R256 locus
Concomitant to the reaction course of CPD
repair as outlined before, further DFo peaks
nearby hint at conformational changes in two
highly conserved features of the class II photo-
lyase active site: the 5WC and the side chain of
R256 (Fig. 2D). Only in the catalytically active
fully reduced state (Int1), but not the oxidized
state (ox/ss), does the 5WC interact with one of
the Nh atoms of R256 to form a square bi-
pyramidal structure that connects the FAD phos-
phate backbone, the CPD 3′-thymine, and R256
(Fig. 2D). Upon reaction initiation, that is, at
100 ps, R256 quickly retracts from the 5WC
toward the CPD, with the side chain still being
mobile and showing two alternative confor-
mations (Fig. 5A). During the further reaction,
the 5WC loses its order, whereas the R256-CPD
interaction persists, as shown for Int3 of the
F650ps snapshot as an example (Fig. 5B). Based
on these structures and the corresponding struc-
tures of all time points (fig. S9), we suggest that
the 5WC/R256 locus allows fast reorganization
during the electron transfer reaction. In the
resting, that is, fully reduced, state, 5WC/R256
electrostatically stabilizes the negatively charged
FADH–. Upon FET, R256 preferably electro-
statically stabilizes the CPD•– anion radical
rather than the FADH• radical and destabilizes
the ordered 5WC. In support of this hypothesis,
the 5WC could not be found in either the ox/ss
(Fig. 2D) or in any of the room-temperature apo
structures of MmCPDII (fig. S10) (12), indicating
that for the formation of a square bipyramidal
structure of 5WC, anionic FAD and a bound CPD
substrate are necessary. In MmCPDII-CPD-DNA
complexes obtained under cryogenic conditions
at the synchrotron, a highly similar six-water
cluster (6WC), in which the sixth water replaces
R256, can be observed (16, 22) (fig. S10). Given
that such a 6WC transiently appears during the
BET, that is, in the F3.35ns snapshot (fig. S9),
we propose that, for the synchrotron structures,
the 6WC mimics an intermediate derived from
either cryo-trapping or x-ray induced electron
transfer.
Our results on the electrostatic stabilization
of the CPD radical anion by R256 are corrob-
orated by previous mutational analyses (37, 40).
These studies showed that the corresponding
R342A mutation (A, Ala) in Escherichia coli
photolyase caused impaired binding for photo-
damaged DNA and a diminished quantum yield
for CPD repair (37). Accordingly, the 5WC/R256
locus of class II photolyases may act as a fail-
safe device for FADH–, priming it for FET ac-
tivity only in the presence of a bound CPD lesion.
The 5WC is reordered at 500 ns (Fig. 5C),
when the isoalloxazine ring of FADH– has
returned to its relaxed fully reduced state.
Accordingly, the Con2b intermediate can be
understood as the “enzyme-recovered product
complex” before initiation of product release.
In the N100ns snapshot (Con2a), the 5WC adopts
a slightly different geometry when compared
with both the dark and N500ns states, which
apparently represents a structural snapshot of
the process toward the enzyme-recovered product
complex. The concept of enzyme recycling is
well known, but it is notable that the enzyme
does not rush to release the product before
resetting the conformations of its active-site
moieties that have been altered during catalysis.
In this way, the enzyme may drive product
release while being primed for cognate substrate
recognition and the next reaction cycle.
Role of the BIR during product release
by photolyases
Whereas the 5WC/R256 locus is involved in
fine-steering the active site’s affinity for cognate
CPD substrates and repaired products, the BIR
acts in stabilizing the unpaired bubble of the
bound dsDNA, that is, the DNA region vacated
by CPD flip-out upon binding to MmCPDII
(Fig. 2E) (22). Unlike the 5WC/R256 locus, the
BIR locus remains essentially unchanged dur-
ing the first 500 ns (Fig. 5D and fig. S11) but
begins showing slight changes for the N25ms
snapshot (Con3a) and then, as expected, large-
scale changes in both complexes of the N200ms
snapshot, where the two thymines have partially
entered the unpaired bubble (Con3b and Con4
in Fig. 5, E and F; and figs. S7 and S11). As the
5′-thymine first flips back toward the unpaired
bubble (Figs. 4D and 5E), the resulting rota-
tion of the intra-CPD phosphate causes BIR
residue R441 to swivel away from the phos-
phate by changing its center of mass distance
from 4.4 to 7.2 Å (Fig. 5E). This opens the BIR
“lock” and allows both thymines to return se-
quentially to the unpaired bubble. Once filling
the unpaired bubble, the two thymine bases
stack well with the preceding dC6 base and
interact with their complementary bases dA7′
and dA8′, essentially displacing D428 and R429,
while R441 moves back closer to the dT8 phos-
phate, stabilizing it in its new orientation (Fig.
5F). However, the remaining BIR element W431
continues interacting with dC9 and the thymine
Maestre-Reyna et al., Science 382, eadd7795 (2023)
1 December 2023
8 of 14
RES EARCH | R E S E A R C H A R T I C L E
A
D
B
E
C
F
Fig. 5. Roles of the 5WC/R256 locus during DNA repair and the BIR locus
during DNA release. (A to C) Detailed structures of the 5WC/R256 locus
[orange, F100ps for (A), F650ps for (B), and N500ns for (C)] superposed with
the corresponding dark structure (green). The FAD coenzyme is depicted in gold
for both structures in each panel. The R256 side chain is dynamic, with two
alternative conformations at 100 ps. 3s-contoured DFo(Y-dark) maps are shown
as well, with positive peaks in cyan and negative ones in magenta. The characteristic
dark 5WC is shown as green spheres, whereas the remaining water molecules for
each specific time-resolved snapshot are in red. Here, blue dashed lines highlight the
5WC arrangement in each time-resolved snapshot. Notably, already at 100 ps (A),
the square bipyramidal interaction between 5WC and R256 is broken, resulting in a
square pyramidal arrangement. R256 and the CPD are shown as stick figures, with
interatomic distances between the center of mass of the R256 guanidinium moiety
and the O2 carbonyl of each of the CPD thymines highlighted with black dahsed
lines. Distances in angstroms are provided nearby and are color-coded to match the
structures. (D to F) Detailed view of the unpaired bubble and the BIR residues D428,
R429, W431, and R441 with Ndark (green) superposed onto three individual time-
resolved snapshots [N500ns (D), N200ms complex A (E), and N200ms complex B
(F)]. dT7 and dT8, corresponding to the CPD bases, as well as all BIR residues are
shown as stick models, whereas all other bases and amino acids are shown as
cartoon representations. 3s-contoured DFo(Y-dark) maps as above are also shown.
To highlight BIR lock opening, marked by the movement of R441, the R441 to CPD
phosphate center-of-mass distance is shown as a red dashed line, with color-coded
distances in angstroms provided nearby.
bases, which rationalizes why the repaired DNA,
after losing so many interactions at the active
site, still sticks to the enzyme before full
release and why the bound and still-kinked
DNA at this point is highly dynamic, as shown
by the large number of diffuse difference den-
sity peaks around the DNA backbone (right
structure in Fig. 4E). Indeed, clustering anal-
ysis of steered molecular dynamics simulations
restrained by the DFo(200ms-dark) electron
density map revealed that dsDNA in complex
A adopts a single major conformation simi-
lar to the starting structure (fig. S12). By con-
trast, the dsDNA of complex B showcases two
major conformations with different degrees
of deformation in the unpaired bubble re-
gion, supporting that DNA release starts at the
unpaired bubble and extends from it toward
the upstream and downstream regions of the
DNA (fig. S12).
Upon full DNA release, even this last protein-
DNA interaction will be broken, and dC9 and -T-T-
co-stack again as in canonical dsDNA. It is
feasible that CPD-DNA binding by MmCPDII
follows the reverse sequence of events: First,
W431 intercalates between the 3′-thymine of the
CPD lesion and its downstream base, enforcing
suboptimal base stacking on both sides of the
CPD lesion. Then, an onset of interactions be-
tween R429 and the counterbases dA7′ and dA8′
(16) supplies further stabilization. Finally,
the R441-D428 lock facilitates CPD phosphate
rotation, thus causing the CPD to flip into the
enzyme’s active site.
Complementary approaches to validate the
TR-SFX–derived photolyase mechanism
To provide further support for our proposed
structural mechanism, we performed three
additional analyses on the TR-SFX data as a
whole, with details provided in supplementary
texts S4 to S6. First, we developed a kinetic
model (supplementary text S4 and fig. S13A)
based on the accumulation of negative den-
sities around the two affected bonds, C5–C5′
and C6–C6′, during the first 3.35 ns of the re-
action (Fig. 3B and fig. S13B). By numerical
integration of this model, we could deter-
mine that indeed Int3, that is, (T_T)•–, was the
dominant species between 450 and 650 ps,
whereas a mixture of Int4 and Int5 became
dominant after 1 ns (fig. S13C), as described by
our molecular structures (Fig. 3A). These re-
sults show that the kinetic data derived from
the negative density accumulation correlate
well with the derived intermediate structures
and thus produce a valid reaction mechanism
(fig. S13D).
Second, we performed SVD analysis of both
the ps-ns and ns-ms time series as described in
supplementary text S5 and figs. S14 and S15.
For the ns-ps series, the first main component
(lSV1) acted by symmetrically separating the
CPD bases, whereas the second one (lSV2)
dampened the effect of the first one in the vi-
cinity of C6–C6′, effectively delaying their sepa-
ration (fig. S14, A and B). Meanwhile, SVD
results for the ns-ms series agreed very well
with our molecular structures (Fig. 4 and fig.
S15) and supported a three-state mechanism
for DNA release: (i) local CPD changes prev-
alent at 100 to 500 ns (lSV1), (ii) global DNA
Maestre-Reyna et al., Science 382, eadd7795 (2023)
1 December 2023
9 of 14
RES EARCH | R E S E A R C H A R T I C L E
changes during complex release at 200 ms (lSV2),
and (iii) base flip-back between 25 and 200 ms
(lSV3). Because SVD is performed over the en-
tire asymmetric unit, including complexes A
and B, we performed further structural analy-
ses to show that complexes A and B of 200 ms
involve shifts of populations (fig. S7).
Third, we sought to validate our interpreta-
tions of time-dependent structural changes with
QM computational analysis by using the density
functional theory method to calculate the highest
occupied molecular orbitals of the TR-SFX–
derived atomic coordinates of CPD repair inter-
mediates (supplementary text S6 and fig. S16A).
Our data showed overlapping orbitals up to
650 ps (Int3) and formation of a node at 1 ns
(Int4; fig. S16B), which supports complete
rupture of the cyclobutane ring between 650 ps
and 1 ns. Furthermore, when the calculation
was repeated with addition of a negative
charge to CPD, node formation then occurred
at 3.35 ns, that is, BET involving Int4 and Int5
(fig. S16C), which reflects the extra time needed
for BET.
Discussion
Comparing TR-SFX data with kinetic data from
spectroscopic studies of photolyases
Even though this work focused on structural
details of the intermediates and processes during
catalysis of DNA repair by a class II photolyase,
the kinetics and order of occurrence of inter-
mediate species should resemble that from spec-
troscopic studies in solution, though detailed
data can differ because of different experimental
conditions and sample sources. Overall, our re-
sults are in close agreement with results of
previous ultrafast time-resolved spectroscopic
studies, wherever a comparison can be made,
such as the chemical structures of the inter-
mediates of cyclobutane repair (1, 4, 38), and
the roles of active-site residues such as R256
and hydration dynamics (37). The role of primary
hydration dynamics in stabilizing and mediating
the charge transfer reactions, which had been
previously hypothesized based on ultrafast spec-
troscopic analysis of the DNA repair reaction (4),
has been now given a molecular and structural
basis. Here, the R256-5WC locus plays a key role
in stabilizing the negative charge of the electron
that travels back and forth between the FADH–
coenzyme and the CPD lesion.
Furthermore, we can also address some of
the differences in the kinetics and quantum
yields reported previously for different photo-
lyases. For example, our data clearly indicate that
the occupancy of activated complexes during
CPD repair stays constant from 100 ps to 10 ns
(table S1). This result signifies that once FET has
occurred, the reaction proceeds with almost
100% efficiency and, conversely, that the ob-
served quantum yields for repair depend on
nonproductive deexcitation of FADH–*. Another
possible culprit for lowered quantum yields is
the proposed intramolecular electron transfer
(iET) between the FADH– isoalloxazine and
adenine moieties (20, 41). We believe that the
latter is a plausible explanation, although our
data showed no obvious time-dependent con-
formational changes around the adenine moiety.
iET is part of the FET pathway in class II
photolyase–mediated DNA repair but is not
predominant in class I photolyase–mediated
repair (20). This hypothesis, which also implies
that iET is completed within the overall FET
time frame of about 100 to 250 ps, fits well
with the comparatively low quantum yield of
MmCPDII (~25%; fig. S18) compared with yields
of 45 to 100% reported for class I photolyases
(19, 42). These results suggest that different
photolyases, despite high structural homology,
differ in their kinetic schemes, which is also
reflected by the variations of kinetic constants
and quantum yields depending on class, species,
substrate, and reaction conditions (20, 38, 41, 43).
Because our TR-SFX analysis of MmCPDII
simultaneously examined multiple reaction loci,
as well as chemical and conformational prop-
erties, the order of events found in this study
should reflect the events across a realistic
timescale, though it was not our intent to
determine detailed kinetic constants. Never-
theless, our finding that BET occurs 1 to 2 ns
after photon absorption and FET (Fig. 3C) is
comparable to findings of previous reports
about class II photolyases, where (T+T)•– ap-
peared ~600 ps after photon absorption and
BET was completed 850 ps later (20).
Prior spectroscopic data had suggested that
bond breaking occurred sequentially, with
C5–C5′ breaking first, followed by C6–C6′ (Fig.
1A) (1, 2, 4, 7, 40). Our results have validated
this mechanism and characterized the chem-
ical structures of the ring-opening interme-
diates in the enzyme-bound form.
The issue of multiphoton effects on DNA
repair by photolyases
Although it is conceivable that FAD dynamics,
and possibly the resulting FET kinetics, is
affected by multiphoton excitation, the focus of
this work is the repair of DNA that contains
CPD lesions, which should be minimally af-
fected, if at all, by multiphoton effects on the
basis of the following mechanistic and experi-
mental considerations: First, the structural moiety
being repaired, a CPD, does not interact at
400 nm with the excitation light. Initiation of
its repair only requires an electron through FET
from the excited FADH–*. Our results show that
FET is completed in ~250 ps, whereas CPD
repair takes place mainly from 0.45 to 1.0 ns,
and the repair mechanism itself should be un-
affected by the mechanism and kinetics of FET.
This conclusion is further supported by relevant
literature and additional control experiments:
Previous spectroscopic studies showed that elec-
tronic relaxation of free FADH–* proceeds within
5 to 10 ps but takes about 1.8 ns when bound to
free photolyases (6), indicating that there is
sufficient time to generate higher states than
S1 by excitation of FADH–*. However, relaxation
of FADH–* from S2 and higher states to S1
occurs within a few ps (24) or even faster by
internal conversion. Even if multiple photon
absorption by FADH– may cause high-energy
states, which eject a free electron by reaching
the ionization continuum, this process might
be a biologically feasible FET mechanism. In the
presence of the 8-hydroxy-deazaflavin antenna
chromophore, light energy absorbed by the
antenna is predicted to cause free-electron
ejection from FADH– by a process called inter-
molecular coulombic decay (44). Although the
effects on quantum yield cannot be discounted,
effects on the FET speed appear rather unlikely
according to our TR-SFX data, because previ-
ously published FET time constants from time-
resolved spectroscopy range between 200 and
600 ps, for example, 209 ps for a class I and
565 ps for a class II photolyase (20). Thermal
effects caused by vibrational cooling upon back-
conversion of high-energy states of FADH–* are
also unlikely to affect catalysis because we did
not observe substantial difference map peaks
in the CPD environment of the photolyase-DNA
complex in its oxidized state under multiphoton
conditions (fig. S1, A and B).
Second, the interpretation of our results is
based on multiple active-site moieties (Fig. 2, A
to E) that all display different time-dependent
changes. As mentioned above, CPD repair is by
orders of magnitude slower than multiphoton
relaxation. Based on Fig. 6A, the recovery of
active-site residue conformations (500 ns) and
the thymine back-flipping and product return
(ms range) are even slower and thus unlikely to
be affected by any multiphoton effect.
Third, it is interesting to note the high degree
of correlation between the CPD DFo features in
the F10ns and N10ns snapshots (see Materials
and methods summary). Because simultaneous
multiphoton excitation may occur in the ps-ns
series (pulse duration: 0.98 ps) but not in the
ns-ms series because of long pulse duration
(3 ns), as addressed in supplementary text S3, we
would expect the F10ns and N10ns snapshots
and their difference maps to differ considerably
if multiphoton excitation affected repair kinetics
in the ps-ns series.
Finally, as described in the beginning of the
Results section, the power titration analysis
suggested that the power used in our experi-
ments is just enough, in contrast to the cal-
culated values that suggest excessive photons.
This disparity has also been reported recent-
ly (45), which suggests that far fewer pho-
tons than the nominal values are absorbed
for reasons that require further investigation.
An alternative scenario, especially for proteins
optimized for light-driven electron transfer, is
that “proteins may have evolved to direct all
Maestre-Reyna et al., Science 382, eadd7795 (2023)
1 December 2023
10 of 14
RES EARCH | R E S E A R C H A R T I C L E
Fig. 6. The complete timeline for major events and video presentation.
Evolution of the average C5–C5′ and C6–C6′ atomic distances of the CPD and
dihedral angles of the isoalloxazine ring of FAD, as derived from our structural
models. The intermediate number or numbers for each time point, and the catalytic
hallmarks of different time periods, are indicated. The intermediates up to 10 ns
are reaction intermediates for CPD repair (Int1 to Int5, indicated as I1 to I5), whereas
the intermediates from 10 ns upward are conformational intermediates (Con1
to Con4, indicated as C1 to C4) in the processes of active-site recovery and
product release. Importantly, because the refined structural models may represent an
average structure derived from the composite of different intermediates that are
present at any given time point, the evolving C5–C5′ and C6–C6′ distances should
not be simply considered as bond elongation but rather as shifts in the populations
of bond split-unsplit intermediates. In any case, these values visualize the progress
along the reaction coordinate for CPD cleavage over time.
deposited energy toward functional outcomes”
(25, 29).
The issue of identifiable reaction intermediates
from TR-SFX experiments
Many TR-SFX reports have focused on dynamic
features of a specific chemical or physical step
(28, 46–48), and others have used kinetic mod-
els from spectroscopic studies to guide their
search for reaction intermediates (49). By con-
trast, our work deals with an entire catalytic
cycle with many elementary steps not assigned
previously, such as ordered product release.
Intermediates of chemical or enzymatic reac-
tions are commonly considered stable when
detectable by kinetic or biophysical methods.
These methods are usually limited by their
capacity to detect characteristic signals for
each intermediate moiety and thus provide
mostly local information about the reactants
themselves. Notably, the International Union
of Pure and Applied Chemistry (IUPAC) Gold
Book (50) defines an intermediate as “a mo-
lecular entity with a lifetime appreciably longer
than a molecular vibration that is formed from
the reactants and reacts further to give the
products of a chemical reaction.” Because TR-SFX
captures global information about all atoms
within the enzyme-substrate complex, in our
case, a photolyase bound to the biopolymeric
CPD-DNA, it should be able to unravel multi-
ple intermediates, or molecular entities, in a
chemical step or a process (e.g., conformational
change, product release), particularly because
they can be stabilized by the enzyme.
Conversely, because each time point may
consist of multiple intermediates with vary-
ing populations, each snapshot listed in
Table 1 represents either the predominant
species of that particular snapshot or a mix-
ture of species. Importantly, we based our
assignments not only on the refined structural
models but also on the features of the DFo
maps, such as C–C peak distances and the
accumulation of negative densities along the
relevant bonds (Fig. 3B). These assignments
were then further validated by other relevant
structural changes at the active site, such as
the status of the dihedral angles of the FAD
and the structural movements of the 5WC/
R256 locus (Table 1), and by different experi-
mental approaches (see supplementary texts
S4 to S6).
The reliability and importance of the refined
structural changes are supported by correla-
tion coefficient analysis (fig. S6), which quan-
tifies the similarity between the observed and
calculated difference electron density maps.
Especially within the initial 2 ns, where prior
solution spectroscopic studies predicted that
most chemical changes would occur (5, 20),
individual structures strongly correlate with
their corresponding density maps but not with
those of their neighboring time points.
A conformational change is commonly con-
sidered as a simple two-state process, where a
time-dependent structural change corresponds
to a shift of population between two states. In
our view, a conformational change could be a
continuous or multistep process, as demonstrated
recently by a temperature-resolved cryo–electron
microscopy study of a ligand-induced conforma-
tional change of another enzyme (51). In this
work, we found that the active site of the en-
zyme goes through a multistep process (10 to
500 ns) to recover from catalysis, and we iden-
tified two variants, Con2a and Con2b, in the
process. For the product release, we also iden-
tified two variants (Con3a and Con3b) in the
process of back-flipping of the repaired thymines
and characterized the reannealed product Con4.
Increased conformational freedom around the
DNA also leads to the shift of population be-
tween Con3b and Con4 (figs. S7 and S8).
For the CPD repair in the ps-ns series, the
structure from a snapshot often consists of
multiple intermediates. Those with distinct
structures in a predominant state can be well
characterized, such as the ring-opening inter-
mediates in the F650ps (C5–C5′ cleaved) and
F1ns (C6–C6′ also cleaved) snapshots. Those of
the electron transfer processes (FET and BET)
are not directly resolvable by TR-SFX before
the occurrence of the next step. However, the
electron transfer processes can be indirectly
identified from changes in the coenzyme and
active-site residues.
Molecular movie of photolyase catalysis
Our results described in this work are sum-
marized in an overall timeline (Fig. 6) and in a
molecular movie (Movie 1) that enables visu-
alization of the structural basis of (photo)
enzymatic catalysis for this multistep reaction.
The FET reaction leads to the accumulation of
ring-opening intermediates, followed by BET,
postrepair recovery of enzyme and coenzyme,
and thymine back-flip and DNA release. The
events of CPD repair are mediated by struc-
tural fluctuations of the FAD coenzyme and
the surrounding active-site moieties, including
the 5WC/R256 locus, whereas thymine back-
flip and the onset of DNA release proceed along
the BIR region of the unpaired bubble. We also
observed the directionality of DNA repair, that
is, the 5′-thymine moves away from the 3′-thymine
after the opening of the CPD ring so that the
5′-thymine leaves the active site first.
Materials and methods summary
Sample preparation
MmCPDII-DNA cocrystals were grown accord-
ing to previously published conditions (16, 22)
Maestre-Reyna et al., Science 382, eadd7795 (2023)
1 December 2023
11 of 14
RES EARCH | R E S E A R C H A R T I C L E
Movie 1. 3D molecular movie of photolyase-catalyzed DNA repair. Key structural components of all
intermediates are highlighted in Fig. 6 and Table 1. As also explained in Fig. 6, the evolving C5–C5′ and C6–C6′
distances reflect shifts in the populations of the intermediates in the opening of the cyclobutane ring, not
the elongation of bonds.
after applying further optimizations for enzyme
activity and large-scale microcrystal produc-
tion. Here, the MmCPDII protein solution was
directly activated before crystallization by photo-
reduction in the presence of dithiothreitol
(DTT), white light, and anaerobic conditions,
followed by mixing with CPD-containing DNA.
After an incubation period of ~2 hours at 23°C,
crystals were harvested, mixed with grease, and
loaded into injectors according to previously
established protocols (52). To preserve enzy-
matic activity, all steps were performed under
anaerobic conditions in a Coy Labs vinyl
anaerobic chamber. Further, to prevent acciden-
tal DNA repair, all experimental steps after
enzyme photoreduction were performed under
safety light (640 nm). To confirm the enzyme
status, in-solution reoxidation was followed
by UV-Vis absorption spectroscopy (fig. S17A),
which showed that MmCPDII remained in the
fully reduced state under experimental condi-
tions for at least 24 hours after photoreduc-
tion. Although the presence of photodamaged
DNA is known to further stabilize FADH– in
photolyases, we further confirmed MmCPDII in
crystallo status by in crystallo UV-Vis absorption
spectroscopy of 20-hour-old crystals (fig. S17B).
No crystals older than 20 hours were used for
any of the experiments presented in this work.
Further details about sample preparation are
provided in the supplementary methods.
Data collection
At SACLA (53), data were collected at the BL2
beamline using a 30-Hz pulse frequency and
10-keV x-ray with a pulse duration of <10 fs and
focused to a focal spot with a diameter of 1.5 mm.
For time-resolved experiments, a 15-Hz, 150-mJ,
408-nm optical parametric oscillator (OPO)
pump laser with a 3-ns pulse length with a
100-mm diameter focal spot was used. During
all experiments, the DAPHNIS (54) chamber
was flooded with a 98% helium atmosphere to
increase the signal-to-noise ratio and to pre-
vent oxidation of the sample. Samples were
extruded at 2.6 ml/min through a 75-mm nozzle
(55). Images were collected on a short-working
distance octal multiport detector with a
50-mm sample-to-detector distance (56). Ini-
tial light and dark splitting as well as initial
processing was performed by the Cheetah
pipeline (57).
At SwissFEL (58), the XFEL was also 1.5 mm
in diameter with a ~20-fs pulse duration but was
run at 100 Hz. For time-resolved data collection,
the in-house laser was set for a pulse duration
of 0.98 ps, with a pulse energy of 10 mJ and a
focal spot diameter of 50 mm running in a 3:1
setup. Accordingly, for every three light images,
one dark image was collected. Sample was ex-
truded at 5 ml/min through a 75-mm nozzle
while the chamber was maintained at 200 mbar
after flooding and evacuating three times with
helium.
The average crystal size was 50 mm by 50 mm
by 50 mm, with a 1/e penetration depth of
103 mm at 400 nm. Further details and control
experiments for light contamination and laser
power can be found in the supplementary
methods, supplementary texts 1 to 3, and
figs. S1 to S3.
Processing, refinement, and analysis
Data processing with CrystFEL (59, 60) and
refinement procedures are described in detail
in the supplementary methods. Briefly, we fol-
lowed previously described guidelines (31–34)
to perform structure factor extrapolation fol-
lowed by structural refinement (tables S1 and
S2 for refinement statistics). Additionally, two
features were given particular attention, namely
the exact structure of the CPD cyclobutane ring
(Fig. 1A) and the dihedral angles rC and rN of
the FAD isoalloxazine moiety (Fig. 1C). To ac-
curately determine the crucial CPD and FAD
geometries, we performed real-space correlation
coefficient–based refinement of both features.
Here, we searched for maximum correlation
between a DFo(Y-X) map and an equivalent
calculated difference map [DFc(Y-X)] based on
a fully refined static structure (X) and the time-
resolved structure to be refined (Y). This approach
produces a quantitative quality parameter, cor-
relation coefficient, which can be compared for
a given structure versus all experimental data
to quantify how well small structural changes
can be monitored in our time courses (fig. S6).
The occupancies we achieved by light-triggering
(tables S1 and S2) are close to the experimental
quantum yield of 24 ± 1% for DNA repair by
MmCPDII (fig. S18 and table S3) for the ns-ms
series (20 to 22%) and about half of that for
the ps-ns series (11 to 13%).
The asymmetric unit contains two complexes.
Unless otherwise stated, the maps shown in the
figures are always from complex A, because
those of complex B are much noisier than those
of complex A (I/s values of 3 or below).
Consistency between F series and N series
To control for consistency between the ps-ns
and ns-ms series, we collected the 10-ns time
point both at SwissFEL and SACLA. Two im-
portant factors need to be considered when
comparing these two 10-ns snapshots. The first
one is resolution. Given the lower resolution of
N10ns (2.7 Å), atomic positions are less well
defined than those in the high-resolution
counterpart F10ns (2.15 Å). Secondly, and more
importantly, is pulse duration. For the F10ns
snapshot, each crystal is illuminated for 0.98 ps,
whereas for N10ns, illumination is for 3 ns.
Accordingly, reaction coherence is higher in
F10ns than in N10ns. For example, the variance
for the ps-ns series is expected to be ±0.001 ns,
that is, the pump pulse duration. Meanwhile, in
the ns-ms series, the variance could amount up
to ±3 ns, that is, an error of 30% relative to
N10ns, which decreases for longer delay times.
Accordingly, the N10ns snapshot corresponds
more to a mixture of the F6ns and F10ns
snapshots, as implied by the finding that their
correlation coefficients around the CPD are
about the same (N10ns/F6ns: 0.62; N10ns/
F10ns: 0.61, F6ns/F10ns: 0.67).
Calculation of C–C peak distances
Bias-free monitoring of distances between peaks
from difference maps is more reliable than
atom-atom distances from refined structural
models in tracking structural changes by TR-SFX.
In our study, only the 5′-thymine, and not the
Maestre-Reyna et al., Science 382, eadd7795 (2023)
1 December 2023
12 of 14
RES EARCH | R E S E A R C H A R T I C L E
3′-thymine, appears to move upon opening of
the cyclobutane, leading to mainly negative
peaks for the cyclobutane and positive peaks at
the 5′-thymine side. To determine the distance
between the negative cyclobutane and the pos-
itive 5′-thymine difference density peaks along
the C5–C5′ and C6–C6′ bonds, DFo maps were
generated with the phenix isomorphous dif-
ference tool as described in the supplementary
materials and as shown in Figs. 3A and 4. The
CCP4 tool fft was then used to extract electron
densities at a 2.5s contour level. To ensure that
an identical number of voxels existed in each
map even at different resolutions, the grid was
set to 1/4 times the maximum resolution. Next,
a Python script was used first to position a
dummy atom on the C5 and C6 positions and
then to move them along the C5–C5′ and the
C6–C6′ bond axes in steps of 0.385 Å (half the
atomic radius of carbon) for 11 steps (4.235 Å)
toward, and beyond, the 5′-thymine. In both
cases, the dummy atom was then used as a mask
to extract the integrated electron density at each
position and plotted versus distance (fig. S5). The
distance between the minimum (cyclobutane
peak) and maximum (5′-thymine peak) along
each axis was then extracted from the plots.
Importantly, if either a minimum or a max-
imum could not be determined, no peak distance
could be calculated and therefore no value was
assigned [noted as a dash (–) in Table 1] or
plotted (Figs. 3B and 4B).
RE FE RENCES AND N OT ES
1. M. Zhang, L. Wang, D. Zhong, Photolyase: Dynamics and electron-
transfer mechanisms of DNA repair. Arch. Biochem. Biophys. 632,
158–174 (2017). doi: 10.1016/j.abb.2017.08.007; pmid: 28802828
2. L. O. Essen, T. Klar, Light-driven DNA repair by photolyases.
Cell. Mol. Life Sci. 63, 1266–1277 (2006). doi: 10.1007/
s00018-005-5447-y; pmid: 16699813
3. P. J. Rochette et al., Influence of cytosine methylation on
ultraviolet-induced cyclobutane pyrimidine dimer formation
in genomic DNA. Mutat. Res. 665, 7–13 (2009). doi: 10.1016/
j.mrfmmm.2009.02.008; pmid: 19427505
4. Y.-T. Kao, C. Saxena, L. Wang, A. Sancar, D. Zhong, Direct
observation of thymine dimer repair in DNA by photolyase.
Proc. Natl. Acad. Sci. U.S.A. 102, 16128–16132 (2005).
doi: 10.1073/pnas.0506586102; pmid: 16169906
5. K. Brettel, M. Byrdin, Reaction mechanisms of DNA photolyase.
Curr. Opin. Struct. Biol. 20, 693–701 (2010). doi: 10.1016/
j.sbi.2010.07.003; pmid: 20705454
6. Y. T. Kao et al., Ultrafast dynamics of flavins in five redox
states. J. Am. Chem. Soc. 130, 13132–13139 (2008).
doi: 10.1021/ja8045469; pmid: 18767842
7. A. Sancar, Mechanisms of DNA repair by photolyase and
excision nuclease (Nobel Lecture). Angew. Chem. Int. Ed. 55,
8502–8527 (2016). doi: 10.1002/anie.201601524;
pmid: 27337655
8. A. Sancar, Structure and function of DNA photolyase.
9.
Biochemistry 33, 2–9 (1994). doi: 10.1021/bi00167a001;
pmid: 8286340
I. Chaves et al., The cryptochromes: Blue light photoreceptors
in plants and animals. Annu. Rev. Plant Biol. 62, 335–364
(2011). doi: 10.1146/annurev-arplant-042110-103759;
pmid: 21526969
10. C. Aubert, M. H. Vos, P. Mathis, A. P. M. Eker, K. Brettel,
Intraprotein radical transfer during photoactivation of DNA
photolyase. Nature 405, 586–590 (2000). doi: 10.1038/
35014644; pmid: 10850720
11. A. Sancar, D. Zhong, It is chemistry but not your grandfather’s
chemistry. Biochemistry 56, 1–2 (2017). doi: 10.1021/acs.
biochem.6b01262; pmid: 28068777
12. M. Maestre-Reyna et al., Serial crystallography captures
32. J. H. Yun et al., Early-stage dynamics of chloride ion-pumping
dynamic control of sequential electron and proton transfer
events in a flavoenzyme. Nat. Chem. 14, 677–685 (2022).
doi: 10.1038/s41557-022-00922-3; pmid: 35393554
13. A. Cellini et al., The three-dimensional structure of Drosophila
melanogaster (6-4) photolyase at room temperature. Acta
Crystallogr. D Struct. Biol. 77, 1001–1009 (2021). doi: 10.1107/
S2059798321005830; pmid: 34342273
14. A. Cellini et al., Structural basis of the radical pair state in
photolyases and cryptochromes. Chem. Commun. 58,
4889–4892 (2022). doi: 10.1039/D2CC00376G; pmid: 35352724
rhodopsin revealed by a femtosecond x-ray laser. Proc. Natl.
Acad. Sci. U.S.A. 118, e2020486118 (2021). doi: 10.1073/
pnas.2020486118; pmid: 33753488
33. M. Schmidt, Time-resolved macromolecular crystallography at
pulsed x-ray sources. Int. J. Mol. Sci. 20, 1401 (2019).
doi: 10.3390/ijms20061401; pmid: 30897736
34. E. Claesson et al., The primary structural photoresponse of
phytochrome proteins captured by a femtosecond x-ray laser.
eLife 9, e53514 (2020). doi: 10.7554/eLife.53514;
pmid: 32228856
15. R. Kort, H. Komori, S. Adachi, K. Miki, A. Eker, DNA
35. M. A. Rould, C. W. Carter Jr., Isomorphous difference methods.
apophotolyase from Anacystis nidulans: 1.8 Å structure, 8-HDF
reconstitution and x-ray-induced FAD reduction. Acta
Crystallogr. D Biol. Crystallogr. 60, 1205–1213 (2004).
doi: 10.1107/S0907444904009321; pmid: 15213381
16. S. Kiontke et al., Crystal structures of an archaeal class II DNA
photolyase and its complex with UV-damaged duplex DNA.
EMBO J. 30, 4437–4449 (2011). doi: 10.1038/emboj.2011.313;
pmid: 21892138
17. R. Pokorny et al., Recognition and repair of UV lesions in loop
structures of duplex DNA by DASH-type cryptochrome. Proc.
Natl. Acad. Sci. U.S.A. 105, 21023–21027 (2008). doi: 10.1073/
pnas.0805830106; pmid: 19074258
18. A. Mees et al., Crystal structure of a photolyase bound to a
CPD-like DNA lesion after in situ repair. Science 306,
1789–1793 (2004). doi: 10.1126/science.1101598;
pmid: 15576622
19. Z. Liu et al., Dynamics and mechanism of cyclobutane
pyrimidine dimer repair by DNA photolyase. Proc. Natl. Acad.
Sci. U.S.A. 108, 14831–14836 (2011). doi: 10.1073/
pnas.1110927108; pmid: 21804035
20. M. Zhang, L. Wang, S. Shu, A. Sancar, D. Zhong, Bifurcating
electron-transfer pathways in DNA photolyases determine the
repair quantum yield. Science 354, 209–213 (2016).
doi: 10.1126/science.aah6071; pmid: 27738168
21. H. Park et al., Crystal structure of a DNA decamer containing a
cis-syn thymine dimer. Proc. Natl. Acad. Sci. U.S.A. 99,
15965–15970 (2002). doi: 10.1073/pnas.242422699;
pmid: 12456887
22. M. Maestre-Reyna et al., Twist and turn: A revised structural
view on the unpaired bubble of class II CPD photolyase in
complex with damaged DNA. IUCrJ 5, 608–618 (2018).
doi: 10.1107/S205225251800996X; pmid: 30224964
23. S. Weber, Light-driven enzymatic catalysis of DNA repair:
A review of recent biophysical studies on photolyase.
Biochim. Biophys. Acta 1707, 1–23 (2005). doi: 10.1016/
j.bbabio.2004.02.010; pmid: 15721603
24. R. F. PauszekIII, G. Kodali, M. S. U. Siddiqui, R. J. Stanley,
Overlapping electronic states with nearly parallel transition
dipole moments in reduced anionic flavin can distort
photobiological dynamics. J. Am. Chem. Soc. 138,
14880–14889 (2016). doi: 10.1021/jacs.6b06449;
pmid: 27686753
25. G. Brändén, R. Neutze, Advances and challenges in time-
resolved macromolecular crystallography. Science 373,
eaba0954 (2021). doi: 10.1126/science.aba0954;
pmid: 34446579
26. E. C. Schulz, B. A. Yorke, A. R. Pearson, P. Mehrabi, Best
practices for time-resolved serial synchrotron crystallography.
Acta Crystallogr. D Struct. Biol. 78, 14–29 (2022). doi: 10.1107/
S2059798321011621; pmid: 34981758
27. A. Gorel, I. Schlichting, T. R. M. Barends, Discerning best
practices in XFEL-based biological crystallography – standards
for nonstandard experiments. IUCrJ 8, 532–543 (2021).
doi: 10.1107/S205225252100467X; pmid: 34258002
28. G. Nass Kovacs et al., Three-dimensional view of ultrafast
dynamics in photoexcited bacteriorhodopsin. Nat. Commun.
10, 3177 (2019). doi: 10.1038/s41467-019-10758-0;
pmid: 31320619
29. R. J. D. Miller, O. Paré-Labrosse, A. Sarracini, J. E. Besaw,
Three-dimensional view of ultrafast dynamics in photoexcited
bacteriorhodopsin in the multiphoton regime and biological
relevance. Nat. Commun. 11, 1240 (2020). doi: 10.1038/
s41467-020-14971-0; pmid: 32144255
30. M. L. Grünbein et al., Illumination guidelines for ultrafast pump-
probe experiments by serial femtosecond crystallography. Nat.
Methods 17, 681–684 (2020). doi: 10.1038/s41592-020-0847-3;
pmid: 32451477
31. M. Carrillo et al., High-resolution crystal structures of transient
intermediates in the phytochrome photocycle. Structure 29,
743–754.e4 (2021). doi: 10.1016/j.str.2021.03.004;
pmid: 33756101
Methods Enzymol. 374, 145–163 (2003). doi: 10.1016/
S0076-6879(03)74007-5; pmid: 14696372
36. Y.-T. Kao et al., Ultrafast dynamics and anionic active states of
the flavin cofactor in cryptochrome and photolyase. J. Am.
Chem. Soc. 130, 7695–7701 (2008). doi: 10.1021/ja801152h;
pmid: 18500802
37. C. Tan et al., The molecular origin of high DNA-repair efficiency
by photolyase. Nat. Commun. 6, 7302 (2015). doi: 10.1038/
ncomms8302; pmid: 26065359
38. V. Thiagarajan, M. Byrdin, A. P. M. Eker, P. Müller, K. Brettel,
Kinetics of cyclobutane thymine dimer splitting by DNA
photolyase directly monitored in the UV. Proc. Natl. Acad. Sci.
U.S.A. 108, 9402–9407 (2011). doi: 10.1073/pnas.1101026108;
pmid: 21606324
39. T. R. Prytkova, D. N. Beratan, S. S. Skourtis, Photoselected
electron transfer pathways in DNA photolyase. Proc. Natl.
Acad. Sci. U.S.A. 104, 802–807 (2007). doi: 10.1073/
pnas.0605319104; pmid: 17209014
40. Z. Liu, L. Wang, D. Zhong, Dynamics and mechanisms of DNA
repair by photolyase. Phys. Chem. Chem. Phys. 17, 11933–11949
(2015). doi: 10.1039/C4CP05286B; pmid: 25870862
41. B. J. G. Rousseau, S. Shafei, A. Migliore, R. J. Stanley,
D. N. Beratan, Determinants of photolyase’s DNA repair
mechanism in mesophiles and extremophiles. J. Am. Chem.
Soc. 140, 2853–2861 (2018). doi: 10.1021/jacs.7b11926;
pmid: 29401372
42. V. Thiagarajan et al., DNA repair by photolyase: A novel
substrate with low background absorption around
265 nm for transient absorption studies in the UV.
Biochemistry 49, 297–303 (2010). doi: 10.1021/bi901562a;
pmid: 20000331
43. K. Yang, R. J. Stanley, The extent of DNA deformation in DNA
photolyase-substrate complexes: A solution state fluorescence
study. Photochem. Photobiol. 84, 741–749 (2008).
doi: 10.1111/j.1751-1097.2007.00251.x; pmid: 18086248
44. P. H. Harbach, M. Schneider, S. Faraji, A. Dreuw, Intermolecular
Coulombic decay in biology: The initial electron detachment
from FADH− in DNA photolyases. J. Phys. Chem. Lett. 4,
943–949 (2013). doi: 10.1021/jz400104h; pmid: 26291360
45. T. Gruhl et al., Ultrafast structural changes direct
the first molecular events of vision. Nature 615,
939–944 (2023). doi: 10.1038/s41586-023-05863-6;
pmid: 36949205
46. P. Nogly et al., Retinal isomerization in bacteriorhodopsin
captured by a femtosecond x-ray laser. Science 361,
eaat0094 (2018). doi: 10.1126/science.aat0094;
pmid: 29903883
47. N. Coquelle et al., Chromophore twisting in the excited
state of a photoswitchable fluorescent protein captured
by time-resolved serial femtosecond crystallography.
Nat. Chem. 10, 31–37 (2018). doi: 10.1038/nchem.2853;
pmid: 29256511
48. K. Pande et al., Femtosecond structural dynamics drives the
trans/cis isomerization in photoactive yellow protein. Science
352, 725–729 (2016). doi: 10.1126/science.aad5081;
pmid: 27151871
49. D. Sorigué et al., Mechanism and dynamics of fatty acid
photodecarboxylase. Science 372, eabd5687 (2021).
doi: 10.1126/science.abd5687; pmid: 33833098
50. A. D. McNaught, A. Wilkinson, Eds., International Union of Pure
and Applied Chemistry (IUPAC) Compendium of Chemical
Terminology (the “Gold Book”) (Blackwell Scientific
Publications, ed. 2, 1997).
51. C.-Y. Chen, Y.-C. Chang, B.-L. Lin, C.-H. Huang, M.-D. Tsai,
Temperature-resolved cryo-EM uncovers structural bases of
temperature-dependent enzyme functions. J. Am. Chem. Soc.
141, 19983–19987 (2019). doi: 10.1021/jacs.9b10687;
pmid: 31829582
52. M. Sugahara et al., Grease matrix as a versatile carrier of
proteins for serial crystallography. Nat. Methods 12, 61–63
(2015). doi: 10.1038/nmeth.3172; pmid: 25384243
Maestre-Reyna et al., Science 382, eadd7795 (2023)
1 December 2023
13 of 14
RES EARCH | R E S E A R C H A R T I C L E
53. T. Ishikawa et al., A compact x-ray free-electron laser emitting
in the sub-ångström region. Nat. Photonics 6, 540–544 (2012).
doi: 10.1038/nphoton.2012.141
54. K. Tono et al., Diverse application platform for hard x-ray
diffraction in SACLA (DAPHNIS): Application to serial protein
crystallography using an x-ray free-electron laser. J.
Synchrotron Radiat. 22, 532–537 (2015). doi: 10.1107/
S1600577515004464; pmid: 25931065
55. Y. Shimazu et al., High-viscosity sample-injection device for
serial femtosecond crystallography at atmospheric pressure.
J. Appl. Cryst. 52, 1280–1288 (2019). doi: 10.1107/
S1600576719012846; pmid: 31798359
56. T. Kameshima et al., Development of an x-ray pixel detector
with multi-port charge-coupled device for x-ray free-electron
laser experiments. Rev. Sci. Instrum. 85, 033110 (2014).
doi: 10.1063/1.4867668; pmid: 24689567
57. T. Nakane et al., Data processing pipeline for serial femtosecond
crystallography at SACLA. J. Appl. Cryst. 49, 1035–1041 (2016).
doi: 10.1107/S1600576716005720; pmid: 27275146
58. E. Prat et al., A compact and cost-effective hard x-ray free-
electron laser driven by a high-brightness and low-energy
electron beam. Nat. Photonics 14, 748–754 (2020).
doi: 10.1038/s41566-020-00712-8
59. T. A. White, Processing serial crystallography data with
CrystFEL: A step-by-step guide. Acta Crystallogr. D Struct. Biol.
75, 219–233 (2019). doi: 10.1107/S205979831801238X;
pmid: 30821710
60. T. A. White et al., CrystFEL: A software suite for snapshot serial
crystallography. J. Appl. Cryst. 45, 335–341 (2012).
doi: 10.1107/S0021889812002312
ACKN OW LEDG MEN TS
M.-D.T. wishes to dedicate this work to his former PhD adviser
H. G. Floss (1934–2022), and L.-O.E. wishes to dedicate this work
to his former mentor D. Oesterhelt (1940–2022). We thank all staff
members of the TPS05A beamline, National Synchrotron Radiation
Research Center (NSRRC), a national user facility supported by
the Ministry of Science and Technology (MOST), Republic of China;
and, in particular, C.-C. Tseng and C.-K. Chou for their help in
setting up nonstandard conditions for crystal testing. We also
thank all staff members of the BL32XU beamline, SPring-8, Japan,
for their help in setting up SACLA-similar beam conditions for
testing our crystals before SACLA beamtime. Also, H.-L. Shr
(Crystallization Facility of the Institute of Biological Chemistry,
Academia Sinica) provided the facility for crystallization under
nonstandard conditions, which were initially tested with equipment
kindly provided by S.-G. Shyu. Finally, we thank G. Schertler for
very stimulating discussions on the analytical value of quantifying
negative electron density within difference density maps, and
J. C. Essen for the delicate transport of crystals and logistical
support at the SwissFEL site. The graphics for the summary figure
were created by Y. Wang. Special thanks are extended to three
anonymous reviewers, whose insightful suggestions have been
incorporated into the published version. Funding: This work was
supported by Academia Sinica and the Taiwan Protein Project
funded by the MOST (grant no. AS-KPQ-105-TPP and AS-KPQ-109-
TPP2 to M.-D.T.) and in part by Japan Society for the Promotion
of Science (JSPS) KAKENHI (16K01942) to Y.B.; the Air Force
Office of Scientific Research (AFOSR) (grant no. FA9550-14-1-
0409) and German Research Foundation (DFG) (grant no. ES152/
18) to L.-O.E.; the National Science and Technology Council (NSTC)
(111-2113-M-002-029-MY3) and the National Taiwan University
grant (NTU112L894305) to M.M.-R.; the Ministry of Education,
Culture, Sports, Science and Technology of Japan (Grants-in-Aid
for Scientific Research on Innovative Areas “Molecular movie,”
JP20H05442) and Japan Science and Technology Agency (JST)
FOREST (JPMJFR2057) to J.Y.; the Platform Project for Supporting
Drug Discovery and Life Science Research [Basis for Supporting
Innovative Drug Discovery and Life Science Research (BINDS)]
from Japan Agency for Medical Research and Development
(AMED) to S.I. and E.N.; and JSPS KAKENHI(JP19H05776) to S.I.
The XFEL experiments were performed at the BL2 of SACLA with
the approval of the Japan Synchrotron Radiation Research Institute
(JASRI) (proposal nos. 2017A8019, 2017B8052, 2018A8008,
2018B8031, 2019A8014, and 2019B8005). SwissFEL experiments
were performed at the Alvra beamline (proposal nos. 20191801 and
20190089). Synchrotron experiments were performed at SPring8
(Super Photon Ring - 8 GeV) at beamline BL32XU with the
approval of JASRI (proposal nos. 2018A2514, 2019A2534, and
2020A2585). This work used the icOS platform of the Grenoble
Instruct-ERIC center (ISBG; UAR 3518 CNRS-CEA-UGA-EMBL)
within the Grenoble Partnership for Structural Biology (PSB),
supported by FRISBI (ANR-10-INBS-0005-02) and GRAL, financed
within the University Grenoble Alpes graduate school (Ecoles
Universitaires de Recherche) CBH-EUR-GS (ANR-17-EURE-0003).
We thank T. Arima, Y. Matsuura, H. Naitow, N. Kunishima,
M. Nishiyama, M. Yamamoto, T. Kin, and the members of the
Engineering Support Team of SACLA for help during our x-ray
experiments, as well as T. Nakane for his introduction to CrystFEL.
Author contributions: M.M.-R., A.R., J.Y., S.I., J.S., Y.B., L.-O.E.,
and M.-D.T. conceived the research and designed experiments.
M.M.-R., P.-H.W., E.N., Y.H., M.Sa., A.F., C.-H.Y., E.P.G.N.P., W.-J.W.,
H.-J.E., S.F.-B., C.Y., S.E., N.C., M.Wr., H.L.G., T.W., H.-Y.W., C.-C.L.,
W.-C.H., K.-F.H., Y.-K.C., J.-H.L., J.-H.W., W.G., C.-W.C., A.H.P., K.-C.Y.,
W.-T.L., Y.-C.C., D.G., E.B., G.K., C.C., C.M., C.B., M.Su., Y.J., A.Y.,
R.T., T.T., F.L., K.T., M.A.A., F.B., V.F., P.G., S.K., L.K., V.R., C.J.R.,
E.M.S., M.Wa., L.W., R.S., A.R., J.Y., Y.B., and L.-O.E. performed
experiments. M.M.-R., D.O., K.N., R.S., A.R., and L.-O.E. were
responsible for on-site data processing at SwissFEL; M.M.-R., C.-H.Y.,
Y.J., R.S., A.R., and L.-O.E. fulfilled the same role at SACLA.
P.J.M.J. was responsible for pump laser setup at SwissFEL, and
S.O. fulfilled the same role at SACLA. M.M.-R., P.-H.W., E.N., C.-H.Y.,
S.K., R.S., A.R., J.Y., Y.B., L.-O.E., and M.-D.T. analyzed the data.
M.M.-R., P.-H.W., E.P.G.N.P., L.-O.E., and A.R. performed in
crystallo spectroscopy. P.-H.W. performed in-solution
spectroscopy. W.Z. and P.M. determined quantum yield
spectroscopically. N.C. and A.R. performed SVD analyses. C.-H.Y.
performed QM calculation. M.M.-R., P.-H.W., and C.-H.Y.
established and analyzed the refinement protocol. D.Z. provided
critical insight on the early events of dimer repair. M.M.-R.,
C.-H.Y., L.-O.E., and M.-D.T. wrote the manuscript. Competing
interests: The authors declare no competing interests. Data and
materials availability: The preparation of protein and damaged
DNA are based on previously reported procedures, with references
provided. The structural coordinates are deposited in the Protein
Data Bank under accession codes 8KCM (low dosage LED
synchrotron structure), 7YE0 (oxidized photolyase:DNA complex);
7YC7, 7YCM, 7YCP, 7YCR, 7YD6, 7YD7, 7YD8, 7YEB, 7YEC, and
7YEE (ps-ns time series); and 7YDZ, 7YEI, 7YEJ, 7YEK, 7YEL, 7YEM,
and 7YEM (ns-ms time series). See Table 1 and tables S1, S2,
and S4 for details. License information: Copyright © 2023 the
authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.science.org/about/science-
licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add7795
Supplementary Methods
Supplementary Text
Figs. S1 to S18
Tables S1 to S4
References (61–93)
MDAR Reproducibility Checklist
Submitted 12 July 2022; resubmitted 23 August 2023
Accepted 5 October 2023
10.1126/science.add7795
Maestre-Reyna et al., Science 382, eadd7795 (2023)
1 December 2023
14 of 14
|
10.1126_science.ade1499
|
RES EARCH
ELECTROCHEMISTRY
La- and Mn-doped cobalt spinel oxygen evolution
catalyst for proton exchange membrane electrolysis
Lina Chong1, Guoping Gao2, Jianguo Wen3, Haixia Li2, Haiping Xu1, Zach Green4, Joshua D. Sugar5,
A. Jeremy Kropf1, Wenqian Xu6, Xiao-Min Lin3, Hui Xu4, Lin-Wang Wang2, Di-Jia Liu1,7*
Discovery of earth-abundant electrocatalysts to replace iridium for the oxygen evolution reaction (OER)
in a proton exchange membrane water electrolyzer (PEMWE) represents a critical step in reducing the
cost for green hydrogen production. We report a nanofibrous cobalt spinel catalyst codoped with
lanthanum (La) and manganese (Mn) prepared from a zeolitic imidazolate framework embedded in
electrospun polymer fiber. The catalyst demonstrated a low overpotential of 353 millivolts at 10
milliamperes per square centimeter and a low degradation for OER over 360 hours in acidic electrolyte.
A PEMWE containing this catalyst at the anode demonstrated a current density of 2000 milliamperes
per square centimeter at 2.47 volts (Nafion 115 membrane) or 4000 milliamperes per square centimeter
at 3.00 volts (Nafion 212 membrane) and low degradation in an accelerated stress test.
L ow-temperature water electrolysis can
rapidly produce environmentally sus-
tainable, or green, hydrogen and is a
prospective means of storing energy from
renewable but intermittent power sources,
such as wind and solar, in future clean energy
infrastructure (1–4). Commercial systems use
either liquid alkaline electrolyte or proton ex-
change membrane electrolyte (1). Compared
with the alkaline counterpart, a proton ex-
change membrane water electrolyzer (PEMWE)
offers the advantages of higher current density,
higher H2 purity, lower resistance losses, and
more compact design, rendering it a preferred
technology where high efficiency and small
footprint are essential (1, 2). Working under
the acidic and oxidative environment, how-
ever, adds substantial challenges to the catalyst
activity and stability profile. This is particularly
the case for the anode catalyst because of the
high overpotential for oxygen evolution reac-
tion (OER) (3). At present, OER catalysts for
PEMWE are primarily restricted to the plat-
inum group metal (PGM) materials, such as
IrOx (1). Their high cost and limited reserve,
however, pose substantial barriers to the wide-
spread implementation of PEMWE. Low-cost
transition metals and their oxides are known
to be active toward OER in alkaline electrolyte
(4–6), but their demonstration in acidic elec-
trolyte is very limited (1, 7, 8).
Cobalt molecular and oxide compounds have
emerged as promising OER catalysts for water
1Chemical Science and Engineering Division, Argonne
National Laboratory, Lemont, IL 60439, USA. 2Material
Science Division, Lawrence Berkeley National Laboratory,
Berkeley, CA 94720, USA. 3Center for Nanoscale Materials,
Argonne National Laboratory, Lemont, IL 60439, USA.
4Giner Inc., Auburndale, MA 02466, USA. 5Sandia National
Laboratory, Livermore, CA 94550, USA. 6X-ray Science
Division, Advanced Photon Source, Argonne National
Laboratory, Lemont, IL 60439, USA. 7Pritzker School of
Molecular Engineering, University of Chicago, Chicago,
IL 60637, USA.
*Corresponding author. Email: djliu@anl.gov
splitting in recent years (9). Kanan and Nocera
systematically investigated water oxidation
using a catalyst deposited from Co2+ solution
in pH 7 phosphate buffer (10). Gerken et al.
performed a comprehensive mechanistic study
of cobalt-catalyzed water oxidation from homo-
geneous to heterogeneous phases in electrolyte
over a broad pH range of 0 to 14 (11). Those
studies have provided profound understanding
of electrocatalytic oxygen evolution by cobalt.
More recently, thin film spinel-type Co3O4 was
found to be active toward OER and stable at
low overpotential in acid (12). The activity and
stability of Co oxide were improved substan-
tially when modified with iron, manganese,
antimonite, and PbOx (13–15). In addition to
activity and stability, the inherent conductivity
represents an essential requirement for ef-
ficient OER electrocatalysis to overcome the
insulating properties of most transition metal
oxides in their crystalline form. Conventional
carbon supports, used to facilitate the elec-
tron conductivity, are not stable against oxi-
dation to CO2 at the PEMWE anode under
OER operating potential. For example, a Co-
polyoxometalate composited with carbon paste
achieved a lower OER onset potential than
IrO2 in 1 M sulfuric acid but decayed quickly,
presumably as a result of oxidative corrosion
of the carbon (16). Self-conductive oxide cat-
alysts can also enhance the active -site volu-
metric density without being diluted by a
secondary nonactive support. A recently devel-
oped self-healing OER catalyst has demon-
strated excellent activity and durability in
acidic electrolyte (14). The approach, however,
requires the presence of metal precursors in
the aqueous electrolyte, hindering integration
into a PEMWE. Most of the aforementioned
studies were carried out either in half-cells or
aqueous electrolyzers, where the demands for
OER catalyst stability and conductivity are
different from those in a PEMWE. For exam-
ple, the dissolved transition metal concentra-
tion must be minimized to avoid poisoning the
proton exchange membrane in the PEMWE.
Effective OER for PEMWE requires optimal
interfacial properties, microporosity, and sur-
face catalytic activity (1), all of which need to be
validated in the operating electrolyzer.
In this work, we present a cobalt spinel–
based OER catalyst derived from a zeolitic
methyl-imidazolate framework (Co-ZIF) and
processed by electrospinning. The catalyst
demonstrated excellent OER activity ben-
efiting from its high specific surface area,
porous interconnected nanonetwork structure,
and high conductivity.
Design of an acid-stable cobalt OER catalyst
Our design concept of an efficient cobalt
spinel–based OER catalyst for PEMWE anodes
is based on the following rationale: To en-
hance OER activity in acid, an oversized and
more stable second element can be selectively
introduced to the cobalt oxide surface to
generate strain, oxygen vacancy (Vo), and acid
tolerance (17); to improve the oxide electronic
conductivity, a third element with similar
charge and dimension to cobalt may be in-
corporated inside the lattice to bridge the
Fermi bandgap through d orbital partial oc-
cupation of the third element induced by its
d-electron delocalization. Advancing further
from the rotating disk electrode (RDE) or half-
cell study to a membrane electrode assembly
(MEA) demonstration, the catalyst should
have a high porosity and surface area easily
accessible to the reactant (H2O). Meanwhile,
the electrode layer should be effective in trans-
porting H2O and releasing O2 without blocking
the water-catalyst interface. Furthermore, the
oxide catalyst should be self-conductive with-
out the need for another conductive support,
such as carbon, that is unstable under high
OER operating potential and current density.
Finally, the metal oxide should be stable against
oxidative and acidic (pH 2 to 4) corrosion in
the PEMWE environment.
We selected Co-ZIF as the precursor for the
catalyst preparation because of its high intrin-
sic porosity and reticular structure. It has re-
cently been used to prepare a nanoplate oxide
with excellent OER activity tested in strong
alkaline electrolyte (1 M KOH) using the RDE
method (18). Our catalyst preparation through
low-temperature oxidation partially retained
the porosity and morphology of Co-ZIFs after
their conversion into interconnected hollow
metal oxide particles, providing an excellent
platform for enhanced charge and mass transfer.
Among different elements that we screened,
we selected lanthanum (La3+) as the second
element because of its much larger radius
compared with that of Co2+ (1.06 Å versus
0.72 Å) along with its strong affinity to bind
−OH groups at the surface of cobalt oxide. We
also added manganese ions (Mn2+) of similar
Chong et al., Science 380, 609–616 (2023)
12 May 2023
1 of 8
RES EARCH | R E S E A R C H A R T I C L E
A
Co-ZIF
+ Mn(II)
+ La(III)
B
F
E-spin
Activation
Electrospinning
ZIF in Polymer Nanofiber
Interconnected Porous Oxide
C
G
D
H
E
I
La
Mn
Fig. 1. Synthesis, morphology, and structure of LMCF. (A) Schematics of
LMCF synthesis including formation of Co-ZIF embedded PAN polymer fiber by
electrospinning and thermal oxidative activation to produce interconnected
porous cobalt oxide particles after removing all the organics. The background
SEM images show a cross-linked, ZIF-containing fiber nanonetwork produced by
electrospinning and the interconnected porous oxide after the activation.
(B) SEM image (scale bar, 1 mm). (C) HAADF-STEM image (scale bar, 500 nm).
(D) TEM image (scale bar, 200 nm). (E) HRTEM image (scale bar, 5 nm).
(F to H) STEM image and the corresponding La and Mn distributions
(scale bar, 2 nm). The color bars show the element counts. The maximal counts are
321 for La and 142 for Mn. (I) HRTEM image (scale bar, 0.5 nm). The green dots
represent atomic columns of tetrahedral (T) and octahedral (O) oxygens, and the red
dots represent the cobalt atomic columns simulated based on bulk phase inside
lattice. The dotted yellow ellipses (on surface) show a different orientation compared
with the solid yellow ellipses from the simulation (in bulk), suggesting a shift of
oxygen position as a result of lattice relaxation.
radius to Co2+ (0.8 Å versus 0.72 Å) during the
Co-ZIF synthesis, which would be oxidized to
Mn3+ (0.72 Å) during the oxidative conversion
and uniformly distributed inside the cobalt
spinel lattice to promote conductivity (via
bandgap) and OER activity (via affinity OH
or H group). The catalyst synthesis scheme is
shown in Fig. 1A. Briefly, La- and Mn-doped
Co-ZIF, LaMn@Co-ZIF, was prepared in solu-
tion. Powder x-ray diffraction (XRD) combined
with scanning electron microscopy (SEM)–
energy dispersive x-ray spectroscopy (EDX)
elemental mapping confirmed the successful
incorporation of atomic La and Mn into the
structure and cavity of the Co-ZIF (fig. S1).
The LaMn@Co-ZIF was then suspended in a
polyacrylonitrile (PAN) polymer slurry (LaMn@
Co-ZIF/PAN) (19), which was subsequently
electrospun into a fibrous mat. The nanofiber
embedded with individual LaMn@Co-ZIF was
activated in flowing air at 360°C for 6 hours to
remove the organic components, forming La-
and Mn-codoped porous cobalt spinel fibers
(denoted LMCF). Thermogravimetric analysis
(TGA) confirmed the removal of carbon and
nitrogen in LMCF after oxidative activation
(fig. S2). The SEM image of LMCF shows an
interconnected nanofibrous network mor-
phology with ample macropores in between
the entwined nanofibers (Fig. 1B). High-angle
annular dark-field scanning transmission
electron microscopy (HAADF-STEM) revealed
that individual LMCF particles retained the
original Co-ZIF’s rhombic dodecahedral shape,
aligned and fused together in strings after
oxidation (Fig. 1C). The ZIF-shaped LMCF
particle is highly porous with a hollow struc-
ture composed of nanopores, as shown by
transmission electron microscopy (TEM) (Fig.
1D). Each particle is composed of aggregates of
Co3O4 nanocrystallites, with an average size of
~3.5 nm, determined by aberration-corrected
high-resolution transmission electron micros-
copy (HRTEM) (Fig. 1E and fig. S3). Nitrogen
adsorption measurement of LMCF at 77 K
Chong et al., Science 380, 609–616 (2023)
12 May 2023
2 of 8
RES EARCH | R E S E A R C H A R T I C L E
A
D
)
4
-
(
B
E
)
4
-
(
C
F
|
)
R
(
|
)
R
(
|
|
Fig. 2. XAS study of LMCF. (A to C) Fluorescence XANES spectra collected at
Co K-edge (A) (inset: enlarged pre-edge 1s → 3d transition), Mn K-edge (B),
and La LIII-edge (C) under ex situ (LMCF) and operando conditions at different
potentials. Co3O4, CoO, Mn2O3, MnO2, and La2O3 are used as the references.
(D and E) R-space EXAFS spectra at Co K-edge (D) and Mn K-edge (E) of the
same samples. (F) Co-O bond distances (blue line) and DWF (red line) surround
Co derived from EXAFS data taken at different cell potentials. Error bars represent
the uncertainty of Fourier transformation of the experimental EXAFS data.
provided a Brunauer–Emmett–Teller (BET)
specific surface area (SSA) of 197 m2 g−1 and
a pore volume of 0.463 cm3 g−1 (fig. S4). The
high porosity of individual LMCF particle com-
bined with nanofibrous morphology improves
accessibility of reactants to the catalytic sites
and facilitates water-oxygen mass transport in
and out of the catalyst structure—an essential
attribute for high-OER current density. XRD
and Raman spectroscopy confirmed that the
catalyst exhibited a spinel Co3O4 structure of
slightly expanded lattice and higher Co2+/Co3+
ratio (fig. S5C and table S1). STEM images
showed that the individual crystal surface is
dominated by (111) facets, and electron energy-
loss spectroscopy (EELS) elemental mapping
revealed La localization on the surface, with
Mn distributed mainly in the bulk (Fig. 1, F to
H, and fig. S6). Low-magnification EDX ele-
mental mapping disclosed a uniform distribution
of Co, Mn, La, and O in LMCF, and inductively
coupled plasma optical emission spectrome-
try (ICP-OES) determined an atomic ratio of
Co:Mn:La of 80:12:8 (fig. S7 and table S3).
HRTEM imaging revealed that the LMCF lat-
tice surface was terminated by oxygen atoms
in a relaxed state, with positions shifted from
those inside the crystallite (20) due to Vo (Fig.
1I and fig. S8), an important attribute in low-
ering the OER activation energy and stab-
ilizing the intermediate during the reaction
(17, 21). We also measured LMCF conductivity
using the four-probe van der Pauw method for
comparison with IrOx and commercial Co3O4.
The LMCF conductivity was found to be 8.6
times as high as that of commercial Co3O4 and
about two-thirds that of IrOx (fig. S9).
The electronic configuration and the coor-
dination structure of Co, La, and Mn in LMCF
were investigated using x-ray photoelectron
spectroscopy (XPS) (fig. S10A and table S4)
and synchrotron x-ray absorption spectroscopy
(XAS). The high-resolution O 1s XPS spectrum
confirmed the presence of high Vo concentra-
tion in LMCF (fig. S11A). The Co 2p XPS spec-
trum revealed a higher ratio of Co2+:Co3+ in
LMCF compared with that in Co3O4 (fig. S11B).
The x-ray absorption near-edge structure
(XANES) spectrum at the Co K-edge (LMCF)
shows a very similar spectral pattern to that
of Co3O4 with a slightly red-shifted absorp-
tion energy and a decreased white line (WL)
intensity (Fig. 2A), indicating a lower aver-
age oxidation state and a smaller O coordina-
tion number (CN) to cobalt in LMCF, which
agrees well with the Raman and XPS results.
We also observed an enhanced 1s → 3d tran-
sition peak intensity, which reveals that the
cobalt in LMCF is in a less centrosymmetric
coordination environment than that in Co3O4,
suggesting a distorted Co oxide lattice by Vo.
Compared with Co3O4, the LMCF extended
x-ray absorption fine structure (EXAFS) shows
lower peak intensities and CNs corresponding
to Co-O and Co-Co shells (Fig. 2D, fig. S12, and
table S5), which further supports a less de-
veloped lattice with a high concentration of Vo
from smaller particle size and higher Co2+
(tetrahedral O-coordinated) fraction in LMCF.
XANES at the Mn K-edge indicates that the
average manganese oxidation state is between
+3 and +4 (Fig. 2B). Both K-space and R-space
spectra extracted from EXAFS exhibit signifi-
cant differences from that in Mn oxide ref-
erences (Mn3O4, Mn2O3, and MnO2) (fig. S13,
B and C). R-space fitting determined the CNs
of Mn to O and Mn to Co to be 5.5 ± 0.4 and
7.2 ± 0.3, respectively (table S5). Most no-
ticeably, the first and second shell radii and
CNs are close to those of Co-O and Co-Co(oh)
in Co3O4 (Fig. 2E and tables S5 and S6)
instead of Mn-O and Mn-Mn paths in Mn
references, including Mn3O4, Mn2O3, and
MnO2 (fig. S13B). These observations provide
convincing evidence that the Mn substitutes
for the Co3+ at the edge-sharing octahedral
site and is embedded inside the cobalt oxide
matrix in LMCF, in agreement with the HAADF-
STEM result. The Mn3+ in the lattice is known
to enhance OER activity of the oxide in acid
(22). The XANES spectrum of La in LMCF
shows significantly higher WL intensity than
that of the La2O3 reference (Fig. 2C). Given
that the WL intensity for oxides is generally
proportional to the number of coordinated
oxygens, R-space fitting determined the CN
Chong et al., Science 380, 609–616 (2023)
12 May 2023
3 of 8
RES EARCH | R E S E A R C H A R T I C L E
of La to oxygen in LMCF to be 8.3, which is in
between that of lanthanum oxide (CN = 6)
and hydroxide (CN = 9) (fig. S14).
Electrocatalytic activity in solution
We first evaluated the OER catalytic activity of
LMCF using the RDE method in 0.1 M HClO4
electrolyte (pH = 1). To better understand the
effects of the second and third elemental
doping, we also prepared Co3O4 fiber (CF)
and La-doped Co3O4 fiber (LCF) using a sim-
ilar Co-ZIF–electrospinning method. Commer-
cial Co3O4 and Ir black were also studied as
benchmarks. Figure 3A shows a progressive
improvement of OER activity measured by
cyclic voltammogram (CV) through the addi-
tion of Mn and La in CF. The performance of
LMCF also significantly exceeds that of com-
mercial Co3O4 and approaches that of Ir
black. Figure 3B presents the mass activity
(MA) Tafel plot for LMCF together with the
benchmark samples. LMCF again exhibits a
high intrinsic catalytic activity. For example,
LMCF catalyst shows a bulk MA of 126 ± 20 A g−1
at an overpotential of 370 mV, which is higher
than that of commercial Ir black (table S7).
This is possibly because of a higher grav-
imetric catalytic site density of LMCF due
to its lower molecular weight compared
with IrOx. LMCF evaluated by linear sweep
CF
LCF
LMCF
Ir black
Commercial Co3O4
A
)
2
-
m
c
A
m
(
y
t
i
s
n
e
d
t
n
e
r
r
u
C
80
70
60
50
40
30
20
10
0
B
E
H
R
s
v
)
V
(
E
1.65
1.62
1.59
1.56
1.53
1.0
1.2
1.4
1.6
1.8
2.0
0.1
1
10
E (V) vs. RHE
i = 10 mA cm-2
C
30
)
2
-
m
c
A
m
(
y
t
i
s
n
e
d
t
n
e
r
r
u
C
LMCF@1st
LMCF@14Kth
20
10
0
1.2
1.3
1.4
1.6
1.5
E (V) vs. RHE
1.7
1.8
1.0
0
50
100
150
200
Time (h)
250
300
350
1.8
1.7
1.6
1.5
1.4
1.3
10s
10s
0 10 20 30 40 50 60 70
Time (second)
)
V
(
l
a
i
t
n
e
t
o
p
l
l
e
C
1.9
1.8
1.7
1.6
1.5
1.4
1.3
0.00
0.10
0.20
0.30
0.40
Current Density (A cm- )
LMCF
Reference
0
2000
6000
4000
Voltage Cycles
8000
10000
0.0
I
300
)
2
-
m
c
A
m
(
y
t
i
s
n
e
d
t
n
e
r
r
u
C
200
100
0
0
LMCF
LMCF iR corrected
Ir 0.4 mg cm-2
Ir 2 mg cm-2
0.5
1.0
1.5
Current density (A cm -2)
2.0
@ 1.65 V
10 20 30 40 50 60 70 80 90 100
Time (hour)
Fig. 3. Electrocatalytic performance of LMCF. (A) CVs of LMCF, LCF, CF, Ir
black, and commercial Co3O4 in O2-saturated 0.1 M HClO4 (PGM-free catalyst
loadings = ~260 ± 30 mg cm−2, Ir black loading = ~230 ± 30 mg cm−2). (B) Tafel
plots of LMCF, LCF, CF, Ir black, and commercial Co3O4, with error bars of
one standard deviation over four experimental replicates. (C) LSVs of LMCF
measured by RDE before and after 14,000 voltage cycles in O2-saturated
0.1 M HClO4 (with 95% iR correction). (D) Chronopotentiometric response at
10 mA cm−2 with LMCF catalyst loading of 0.9 mg cm−2 over 353-hour test.
(E) S number calculated for LMCF after different hours onstream compared
with selected benchmark Ir-based catalysts. (F) Current-voltage polarizations
(iR corrected and uncorrected) of the PEMWE cell with LMCF anodic catalyst
compared with that of Ir black catalysts with different loadings at 80°C. The
polarization plot for LMCF represents an average of three MEA measurements
with one standard deviation. (G) Current-voltage polarizations of the PEMWE
after selected cycle numbers during a multiple-voltage cycling AST. The inset
shows the stepwise voltage swing between 1.4 V and 1.7 V with 10-s dwell time
at each potential. (H) PEMWE cell voltage measured at different current
densities after selected voltage cycles during the AST. (I) Potentiostatic
measurement of PEMWE at the cell potential of 1.65 V for LMCF anodic catalyst
over 100 hours. Test conditions: anode LMCF catalyst loading, 1 to 2 mg cm−2;
cathode Pt loading, 0.4 mgPt cm−2; 60° or 80°C cell temperature, unless
otherwise specified; 5-cm2 active electrode area; Nafion 115 membrane for
(F), (G), and (H), and Nafion 212 membrane for (I); DI water at flow rate
of 10 ml min−1.
Chong et al., Science 380, 609–616 (2023)
12 May 2023
4 of 8
Current density (mA mg
-1
2
O
r
l
h
0
1
6
O
r
l
r
P
2
a
B
x
O
r
l
h
1
l
a
i
t
i
n
I
100
)
-
3
O
1
.
0
r
I
9
.
0
o
C
r
S
F
)
V
(
e
g
a
t
l
o
V
l
l
e
C
2.5
2.0
1.5
1.0
0.5
0.0
E
107
106
r
e
b
m
u
n
-
S
105
104
103
102
1.90
1.85
1.80
1.75
1.70
1.65
1.60
1.55
H
)
V
(
e
g
a
t
l
o
V
l
l
e
C
D
2.0
1.8
1.6
1.4
1.2
E
H
R
.
s
v
)
V
(
E
G
)
V
(
e
g
a
t
l
o
V
l
l
e
C
RES EARCH | R E S E A R C H A R T I C L E
voltammetry (LSV) (Fig. 3C, with 95% iR cor-
rection) shows an onset potential of 1.28 V
measured at 0.46 mA cm−2 and an over-
potential of 353 ± 30 mV at 10 mA cm−2. The
LMCF catalyst activity was also measured in
0.5 M H2SO4 (pH = 0), and the overpotential
was reduced to 335 ± 30 mV at 10 mA cm−2
(fig. S15). These results placed LMCF among
the most active PGM-free catalysts reported
in aqueous acid (23) (table S2). We estimated
the electrochemical surface areas (ECSA) of
the catalysts by measuring the double-layer
capacitance from the CV curves in the non-
Faradaic region (fig. S16) and produced Tafel
plots of the ECSA-based specific current den-
sities (fig. S17). The intrinsic activities of LMCF
were further assessed based on turnover
frequencies (TOFs) at different overpotentials
(320 mV, 370 mV, and 650 mV), which are
among the highest when compared with rep-
resentative PGM-free and PGM OER catalysts
tested in various acidic media (table S8). For
example, the TOF of LMCF is calculated to be
0.079 ± 0.011 s−1 at an overpotential of 370 mV
based on the total loading mass, which in-
creases to 0.87 ± 0.09 s−1 when calculated based
on ECSA (table S7). We also analyzed O2 prod-
uced during OER over LMCF in a H cell using
in situ gas chromatography (GC) and calculated
the Faradaic efficiencies (FEs) for the oxygen
produced at the current densities of 10 mA cm−2,
20 mA cm−2, 30 mA cm−2, and 50 mA cm−2
(fig. S18, A to C). An average FE of 99.3 ± 5%
was obtained, indicating that the oxygen for-
mation through the four-electron transfer
during water splitting is the only electrochem-
ical reaction over LMCF. Ir was measured as
reference, for which the FE was 97.0 ± 5%
(fig. S18C).
LMCF was subjected to an accelerated aging
test through voltage cycling between 1.4 V and
2.0 V versus reversible hydrogen electrode
(RHE) by RDE in 0.1 M HClO4 electrolyte. A
mere ~20-mV potential loss at 10 mA cm−2
was observed after 14,000 CV cycles (Fig. 3C).
Such stability was found to be better than
that of commercial Co3O4 or Ir black at com-
parable catalyst weight loadings (fig. S19).
The morphology, composition, and electronic
states of the LMCF after the voltage cycling
were found to be nearly unchanged from the
pristine state (fig. S20, A to E). Similarly, no
appreciable changes in the XANES spectra of
Co, Mn, and La were observed after the volt-
age cycling (fig. S20, F to H). We further tested
LMCF galvanostatic stability by holding the
electrode current density at 10 mA cm−2 for an
extended operation period, following a test
protocol (23, 24) (Fig. 3D). The change of
electrode potential in acidic electrolyte was
monitored over 353 hours, and a slow deg-
radation at an average rate of 0.28 mV hour−1
was observed. Transient potential dips in Fig. 3D
were the result of pauses of the measurement
while the electrolyte was refreshed. We also
checked the metal contents in the electrolyte
from acid leaching by performing ICP-OES
under OER electrocatalysis. A mild loss of Co,
Mn, and La ions by ~1.9 wt %, ~2.8 wt %, and
~1.1 wt % over their stoichiometric loadings
in the LMCF catalyst, respectively, in a period
of 80 hours was observed during the chrono-
potentiometry at 10 mA cm−2 (fig. S21A). After
an initial jump at the first hour, the metal
dissolution rate decreased markedly afterward.
Meanwhile, the catalytic activity remained
nearly the same, which suggests that the
dissolved metals could be attributed mainly
to some loosely bound oxides at the surface.
We also calculated the stability number (S
number) for LMCF based on the amount of
Co dissolved in the electrolyte at different
testing times using the framework proposed
by Geiger et al. (25) as well as the activity-
stability factor (ASF) by Kim et al. (26). These
values are compared with some of the Ir-
based benchmarks in the literature, and the
LMCF stability was found to be comparable
to some less-stable Ir materials (Fig. 3E and
fig. S21B).
Electrocatalytic activity in a PEMWE
The ultimate test of an OER catalyst is its
performance in the operating PEMWE. Key
properties, such as porosity, stability, and
conductivity, become more important for the
catalyst performance under the electrolyzer
working conditions in comparison with the
less-strenuous RDE measurements. The LMCF
was assembled into the anode of a PEMWE
single cell and tested using deionized (DI)
water as the feed. Figure 3F shows composite
current-voltage polarization curves derived from
three measurements in the PEMWE. The elec-
trolyzer reached a current density of 2000 mA
cm−2 at a voltage of 2.47 V (2.20 V after cell iR
correction), which could be further reduced to
2.30 V by switching the membrane from Nafion
115 to Nafion 212, and reached a current den-
sity of 4000 ± 200 mA cm−2 at 3.0 V (fig. S22).
In comparison, PEMWEs with Ir loading at
0.4 mgIr cm−2 and 2.0 mgIr cm−2 at the anode
displayed cell voltages of 1.93 V and 1.78 V at
2000 mA cm−2 before iR correction, respec-
tively. Comparisons with other reported anodic
catalysts are given in table S2 (14–16).
The MEA with LCMF was also subjected to
accelerated stress tests (ASTs) using voltage
cycling, galvanostatic, and potentiostatic meth-
ods. In the voltage cycling test, the PEMWE cell
voltage was swept between 1.4 V and 1.7 V with
a 10-s dwell time at each voltage (square-wave)
up to 10,000 cycles. Current-voltage polariza-
tion was recorded periodically after designated
voltage cycles. Figure 3G shows the polarization
curves recorded after the selected voltage cy-
cles. Figure 3H demonstrates the iR-corrected
PEMWE cell voltages at four different current
densities (100 mA cm−2, 200 mA cm−2, 300 mA cm−2,
and 400 mA cm−2) after each designated cycle
number taken from the tests in Fig. 3G. Only
small fluctuations in the cell voltage were
observed between the first and the 10,000th
cycles, suggesting excellent catalyst stabil-
ity under such cycling conditions. After the
voltage cycling, we conducted a galvanostatic
test on the same MEA at incrementally increased
current densities of 50 mA cm−2, 100 mA cm−2,
200 mA cm−2, and 300 mA cm−2 over a 90-hour
time span (fig. S23). The cell voltages re-
mained nearly constant after each incremen-
tal increase of the cell current until the last
10 hours, when an increasing cell voltage was
observed after the current density was raised
to 300 mA cm−2. We also conducted a separate
potentiostatic test over another PEMWE with
LMCF anodic catalyst at a cell potential of
1.65 V for 100 hours (Fig. 3I). A constant cell
current density of ~210 mA cm−2 was observed,
and the anodic effluent from the PEMWE was
analyzed by ICP-MS. The cobalt dissolution
gradually leveled off during the first 10 hours
and became negligible thereafter (fig. S24).
The S number calculated based on the disso-
lution rate was found to be two orders of
magnitude higher than that obtained from
RDE (fig. S25A). The difference may be ascribed
to the higher acidity of the electrolyte used in the
RDE (pH = 1) relative to that at the MEA anode
layer (pH = 2 to 4). A similar phenomenon was
found in the study of RuO2 catalyzed water
oxidation (25). The lifetime determined by
the S number suggests excellent durability of
LMCF operated at the PEMWE (fig. S25B).
The morphology and the surface structure of
the catalyst were preserved after AST by RDE
as well as by combined voltage cycling and
galvanostatic tests in the PEMWE (fig. S26).
Active -site analysis
To understand the nature of the active site
and the impact of the second and third metal
doping, we performed in situ XAS of LMCF in
an O2-saturated electrolysis cell (0.1 M HClO4)
at the Co, Mn K-edge, and La-LIII edge. Figure
2A shows the XANES spectra at the Co K-edge
under different OER operating potentials. At
the onset potential of 1.23 V when OER has yet
to commence, XANES already shows a decrease
in WL intensity and an increase in pre-edge
1s → 3d peak intensity compared with the
spectrum of as-prepared LMCF, accompanied
by a decrease of CN of the first Co-O shell (Fig.
2D, fig. S12, and table S5). This signals oxygen
loss and structural change as a result of in-
teraction with the acidic electrolyte under the
electric field, altering cobalt’s centrosymmetric
coordination (symmetry breakdown) and elec-
tronic structure (27). The CN is significantly
lower than that in Co3O4, indicating a higher
fraction of tetrahedral coordinated Co2+ combined
with higher concentration of Vo, which serves
Chong et al., Science 380, 609–616 (2023)
12 May 2023
5 of 8
RES EARCH | R E S E A R C H A R T I C L E
as the nucleophilic site in promoting O–O
bond formation (28). As the cell potential
increases to initiate OER, the 1s → 3d tran-
sition becomes more intense (Fig. 2A, inset).
The average valence state of cobalt in LMCF
was maintained at a lower value than that of
Co3O4 (+2.67) under all the test potentials
from +2.51 at 1.02 V to +2.32 at 1.70 V (29)
(fig. S27). Simultaneously, Fourier transforma-
tion of EXAFS spectra shows the reduction of
Co-O and Co-Co shell intensities (Fig. 2D). At
a potential of 1.7 V when the OER reaction
proceeds much more rapidly, XAS analysis
shows shortening of the Co-O bond length
(Fig. 2F and table S5), indicating a high degree
of covalency contraction, which positively af-
fects the catalytic activity of the nanoparticles
(28). A similar phenomenon was observed in
an MnO2 film, where the Mn-O bond was
shortened concomitant with accumulation of
Mn3+ and Vo under applied potential in acid
(22). Shortened Co-O bonds during OER were
also found in the amorphous cobalt oxide
OER catalyst (CoCat) (30), except that Co was
in the +4 state, distinct from our catalyst. Ac-
companying the rapid acceleration of OER at
1.7 V, a clear increase of the Debye–Waller
factor (DWF) s2 was also observed. DWF per-
tains to the motion of the coordinated atoms.
In this case, it indicates increased Co-O vi-
bration as the adsorbed H2O is converted to
O2, possibly involving the shuffling lattice
oxygen in LMCF, as was recently observed in
a perovskite OER catalyst (31). The participa-
tion of surface and lattice oxygen creates
anisotropic displacement of the ligation to
cobalt, causing an increase of 1s → 3d tran-
sition intensity. The Co XANES and EXAFS
spectra were nearly fully restored to their
original intensities at the onset voltage (1.23 V)
when the cell potential was returned to 1.02 V,
suggesting the reversibility of the active -site
restructuring in LMCF. By contrast, XANES
spectra at the Mn K-edge and La LIII-edges
show little changes under different cell poten-
tials (Fig. 2, B and C). Compared with Co, the
EXAFS spectrum of Mn only showed a minor
reduction of shell peak intensity at 1.23 V after
the catalyst was immersed in the electrolyte. No
apparent changes in shell structure and DWF
were observed for Mn during OER (Fig. 2E and
fig. S13). XANES and EXAFS analyses indicate
that Mn and La do not participate in the electro-
catalysis directly. Rather, their presence modifies
the structure and activity of the cobalt site. To
this end, we also investigated the in situ XAS at
different potentials over the fibrous cobalt oxide
catalyst prepared by the same method but in the
absence of Mn and La (CF). The changes in
XANES and EXAFS are significantly less dom-
inant than those found in LMCF (fig. S28).
Comparison with theory
The experimentally observed atomic and elec-
tronic configurations in LMCF correspond
well with density functional theory (DFT)
A
)
V
(
l
a
i
t
n
e
t
o
P
3
2
1
0
-1
-2
-3
II
V: Lao3O_CoTO*
IV: bri3OH*
III: bri3OH_H*
I: disLao_H* + 2La3+
C
)
V
e
(
y
g
r
e
n
E
1
0
-1
up
down
0
1
2
3
4
5
6
7
pH
8
9
10
11
12
13 14
G
A
H
K
G
M
L
H
B
I
II
III
IV
V
D
Fig. 4. Computational Pourbaix diagram and Fermi band structure of
LMCF. (A) Surface Pourbaix diagram for the La-doped Co3O4 (111) facet obtained
from the DFT + U calculations. (B) Possible intermediate state configurations.
Blue, gray, red, and white balls denote the octahedral Co3+, tetrahedral Co2+,
oxygen, and hydrogen ions. An asterisk denotes the pure surface; H* and _H*
denote the configurations in which the surface oxygen atoms are covered
by H; CoO denotes the octahedral coordinated cobalt, and CoT denotes the
tetrahedral coordinated cobalt; the nO/nOH/nOOH refer to the numbers of
O/OH/OOH groups covering over each Co atom; configurations denoted
“disCo/La” are generated after the Co/La dissolution; and the prefix “bri-”
means that the covered groups act as a bridge connecting LaO and CoT.
The potential is relative to the standard hydrogen electrode (SHE).
(C) Calculated Fermi band structure of LMCF by replacing Co3+ with
Mn3+ in the Co3O4 lattice. (D) Charge density distribution at the Fermi level
upon Mn substitution in LMCF. Yellow refers to the charge density contour.
Blue, gray, and red balls indicate octahedral Co3+, tetrahedral Co2+, and
oxygen, respectively. Mn ions are behind yellow contour and circled by
violet dotted line.
Chong et al., Science 380, 609–616 (2023)
12 May 2023
6 of 8
RES EARCH | R E S E A R C H A R T I C L E
calculations. For example, our calculation re-
veals that in the bulk, in terms of total energy,
both Mn and La preferentially replace Co in
octahedral rather than tetrahedral sites, with
stability differences of 0.25 eV and 1.28 eV, re-
spectively. This is in agreement with the above
XAS observation of the increased Co2+ per-
centage due to the Mn and La doping. The
stronger preference for La is because of its
most stable +3 oxidation state, whereas Mn
could be present in either a +2 or +3 state. Our
calculation also shows that Mn preferentially
remains in the bulk of the Co3O4, whereas La
is extruded to the surface because of its large
size. We compared the relaxed atomic struc-
tures of Mn and La at the surface layers and
in the bulk of Co3O4 (fig. S29). For Mn, the
surface layer structure is merely 0.1 eV higher
than that in the bulk, suggesting that it can
displace Co anywhere in the system. For La,
the energy is 3.09 eV lower at the surface than
in the bulk. This agrees well with the STEM
measurement of Mn and La distributions, as
shown in Fig. 1, G and H, affirming their roles
as dopants in enhancing oxide conductivity
and surface defect or oxygen vacancies for
better OER activity according to our design
concept (32).
For cobalt spinel–based OER catalysts oper-
able in a PEMWE, the most imposing chal-
lenge is stability in the acidic media. To better
understand the enhanced acid tolerance of
LMCF under electrolytic condition, we calcu-
lated the surface Pourbaix diagram of the La-
decorated Co3O4 (111) facet (Fig. 4A) because it
represents the dominant facet in the LMCF
observed by STEM (Fig. 1, F, G, and I, and fig.
S6, D and E). The Pourbaix diagram consists
of five regions—I: disLao_H* + 2La3+, II:
disLao_CoTOOH* + 2La3+, III: bri3OH_H*,
IV: bri3OH*, and V: Lao3O_CoTO*—with their
intermediate state configuration shown in
Fig. 4B. Among these five phases, phases I
and II contain ionic La3+ and therefore are
unstable and soluble, whereas phases III,
IV, and V contain the surfaces of Co2+, Co3+,
and La3+ bridged by OH*, OH_H*, and O*
and exhibit relative stability against Co dis-
solution. Under low pH and low or negative
potential, the surface La in the La-doped
Co3O4 (111) facet dissolves to form La3+ (phase
I). At potentials between 0.86 V and 1.1 V
(RHE), the cation on the surface is oxidized
and stabilized by OH*, and the oxygen on the
surface is covered by H* to form the stable
bri3OH_H* structure (phase III). As the po-
tential further increases to ≥1.23 V (RHE), the
surface is protected by OH* to form a bri3OH*
stable structure (phase IV). Under high pH
and high potential, the surface of the La-
doped Co3O4 (111) facet tends to be oxidized by
water, and the surface cations La and Co are
protected by O* by forming a stable structure
Lao3O_CoTO* (phase V).
The calculated Pourbaix diagram demon-
strates that the stability of the LMCF catalyst
is defined by the combination of cell potential
and pH. It also agrees well with our experi-
mental observations. For example, under no
electric field (potential = 0) at pH < 7, the
catalyst is in phase I and is thermodynami-
cally unstable in the acidic solution. During
our RDE study in 0.1 M HClO4 (pH = 1) at cell
potentials between 1.5 V and 2 V, only a very
small fraction of tests at low cell potential will
overlap with phase II, which could explain
why the dissolution rate was higher in the
RDE test. In our PEMWE test, the actual cell
voltage runs from 1.5 V to 3.0 V (between the
two red dashed lines) within the anodic pH
between 2 and 4 (defined by the two vertical
green dashed lines) in Fig. 4A. Therefore, the
actual PEMWE anode operates in a window
within these two boundaries, as marked by
the area covered by the green diagonal stripes.
This operating window falls within phase V,
Lao3O_CoTO*, which is stable, thus offering a
preliminary explanation of the stability of the
LMCF catalyst in the PEMWE.
For comparison, we also calculated the
Pourbaix diagram of the pure Co3O4 (111)
facet and confirmed its weaker stability com-
pared with the La-doped surface (fig. S30).
We furthermore calculated Pourbaix diagrams
of the LMCF (110) facet (fig. S31) and (100)
facet (fig. S32). Both facets offer good stability
under PEMWE operating conditions accord-
ing to the calculation. Our study revealed the
critical role of surface La in stabilizing mul-
tiple Co3O4 facets against corrosion under
working PEMWE condition.
Another critical issue in electrocatalysis is
the inherent electron conductivity of the oxide
itself. Our calculation of LMCF shows that
substituting a low concentration of Co3+ ions
with uniformly distributed Mn3+ ions in the
Co3O4 lattice induces two partially occupied
defect states in the midbandgap (Fig. 4, C
and D). The Mn-induced electron wave func-
tion overlaps significantly with neighboring
Co ions, causing obvious dispersion and hence
good electron mobility. This provides a direct
enhancement of bulk-based electron conduc-
tivity, which, combined with the connectivity
of the nanofibrous oxide network, offers an
overall high conductivity value for LMCF. The
improved electronic conductivity was further
confirmed experimentally by the four-probe
van der Pauw method (fig. S9).
Outlook
This study offers prospective directions and
design insight for the future development of
PGM-free OER catalysts for hydrogen produc-
tion using PEMWE technology. For example,
the catalytic activity enhancement can be fur-
ther explored by increasing the surface func-
tional group density through elemental doping,
primary size control, and morphology innova-
tion. The PEMWE durability can be improved
by removing the electrochemically unattached
oxide, therefore limiting metal ion dissolution
because of the lack of electro-potential stabili-
zation. Fundamental understanding of the OER
mechanism with respect to mononuclear versus
binuclear reaction intermediates and catalytic
pathways will help to guide the precursor and
catalyst designs for lower overpotential and bet-
ter acid tolerance (33, 34). These improvements
offer paths to the next-generation, PGM-free
OER catalysts as viable replacements for pre-
cious metals, such as iridium.
REFERENCES AND NOTES
1. K. Ayers et al., Annu. Rev. Chem. Biomol. Eng. 10, 219–239
(2019).
2. M. Carmo, D. L. Fritz, J. Mergel, D. Stolten, Int. J. Hydrogen
Energy 38, 4901–4934 (2013).
3. M. T. M. Koper, J. Electroanal. Chem. 660, 254–260 (2011).
4. R. D. L. Smith et al., Science 340, 60–63 (2013).
5. B. Zhang et al., Science 352, 333–337 (2016).
6. Z.-F. Huang et al., Nat. Energy 4, 329–338 (2019).
7. T. Reier, H. N. Nong, D. Teschner, R. Schlögl, P. Strasser,
Adv. Energy Mater. 7, 1601275 (2017).
8. Z. W. Seh et al., Science 355, eaad4998 (2017).
9. V. Artero, M. Chavarot-Kerlidou, M. Fontecave, Angew. Chem.
Int. Ed. 50, 7238–7266 (2011).
10. M. W. Kanan, D. G. Nocera, Science 321, 1072–1075 (2008).
11. J. B. Gerken et al., J. Am. Chem. Soc. 133, 14431–14442 (2011).
12. J. S. Mondschein et al., Chem. Mater. 29, 950–957 (2017).
13. B. Rodríguez-García et al., Sustain. Energy Fuels 2, 589–597
(2018).
14. M. Chatti et al., Nat. Catal. 2, 457–465 (2019).
15.
I. A. Moreno-Hernandez et al., Energy Environ. Sci. 10,
2103–2108 (2017).
16. M. Blasco-Ahicart, J. Soriano-López, J. J. Carbó, J. M. Poblet,
J. R. Galan-Mascaros, Nat. Chem. 10, 24–30 (2018).
17. J. Bao et al., Angew. Chem. Int. Ed. 54, 7399–7404 (2015).
18. S. L. Zhang et al., Adv. Mater. 32, 2002235 (2020).
19. J. Shui, C. Chen, L. Grabstanowicz, D. Zhao, D. J. Liu,
Proc. Natl. Acad. Sci. U.S.A. 112, 10629–10634 (2015).
20. W. Wang, A. Janotti, C. G. Van de Walle, J. Mater. Chem. C 4,
6641–6648 (2016).
21. Y. Wang et al., Adv. Energy Mater. 4, 1400696 (2014).
22. Z. Morgan Chan et al., Proc. Natl. Acad. Sci. U.S.A. 115,
E5261–E5268 (2018).
23. L. C. Seitz et al., Science 353, 1011–1014 (2016).
24. C. C. McCrory et al., J. Am. Chem. Soc. 137, 4347–4357 (2015).
25. S. Geiger et al., Nat. Catal. 1, 508–515 (2018).
26. Y. T. Kim et al., Nat. Commun. 8, 1449 (2017).
27. C. H. van Oversteeg, H. Q. Doan, F. M. de Groot, T. Cuk, Chem.
Soc. Rev. 46, 102–125 (2017).
28. H. N. Nong et al., Nat. Catal. 1, 841–851 (2018).
29. H. Dau, P. Liebisch, M. Haumann, Anal. Bioanal. Chem. 376,
562–583 (2003).
30. M. Risch et al., Energy Environ. Sci. 8, 661–674 (2015).
31. A. Grimaud et al., Nat. Chem. 9, 457–465 (2017).
32. O. Diaz-Morales et al., Nat. Commun. 7, 12363 (2016).
33. I. C. Man et al., ChemCatChem 3, 1159–1165 (2011).
34. M. Busch, Curr. Opin. Electrochem. 9, 278–284 (2018).
35. W. Xu et al., La and Mn-doped cobalt spinel oxygen evolution
catalyst for proton exchange membrane electrolysis, dataset,
Dryad (2023); https://doi.org/10.5061/dryad.76hdr7t1v.
AC KNOWLED GME NTS
We thank J. Bareno, C. Yang, B. Fisher, M. S. Ferrandon, D. J. Myers,
D. Abraham, and J. Wang of Argonne National Laboratory and
S. Kabir of Giner Inc. for experimental assistance and comments
on the manuscript. Funding: This work is supported by the US
Department of Energy (DOE), Office of Energy Efficiency and
Renewable Energy, Hydrogen and Fuel Cell Technologies Office
(D. Peterson, project manager), and by Laboratory Directed
Research and Development (LDRD) funding of Argonne National
Laboratory, provided by the Director, Office of Science, of the US
DOE under contract no. DEAC02-06CH11357 through a Maria
Goeppert Mayer Fellowship to L.C. Work performed at the Center
for Nanoscale Materials and Advanced Photon Source, both US
Chong et al., Science 380, 609–616 (2023)
12 May 2023
7 of 8
RES EARCH | R E S E A R C H A R T I C L E
DOE Office of Science User Facilities, was supported by the US
DOE, Office of Basic Energy Sciences, under contract no. DE-AC02-
06CH11357. The work at Lawrence Berkeley National Laboratory
was supported by the Assistant Secretary for Energy Efficiency
and Renewable Energy of the US DOE under the Hydrogen
Generation program. Sandia National Laboratories is a
multimission laboratory managed and operated by National
Technology and Engineering Solutions of Sandia, LLC, a wholly
owned subsidiary of Honeywell International, Inc., for the US
DOE’s National Nuclear Security Administration under contract
no. DE-NA0003525. This paper describes objective technical
results and analysis. The views and opinions of the authors
expressed herein do not necessarily state or reflect those of the
United States government or any agency thereof; neither the
United States government nor any agency thereof, nor any of
their employees, makes any warranty, expressed or implied, or
assumes any legal liability or responsibility for the accuracy,
completeness, nor usefulness of any information, apparatus,
product, or process disclosed, or represents that its use would
not infringe privately owned rights. Author contributions: D.-J.L.
designed and supervised the experiment. L.C. designed and
synthesized the catalyst and conducted electrochemical
measurement and data analysis. L.C., Ha.X., D.-J.L., Z.G., and
Hu.X. conducted MEA preparation and PEMWE measurements.
J.W. and J.D.S. performed electron microscopy imaging and
analysis. L.C., Ha.X., A.J.K., W.X., X.-M.L., and D.-J.L. conducted
spectroscopic and catalyst structural investigation and data
analysis. G.G., H.L., and L.-W.W. performed DFT calculation.
L.C., L.-W.W., and D.-J.L. wrote the manuscript. Competing
interests: A US patent (USP 11,033,888) on the nanofibrous
catalyst for OER with D.-J.L. and L.C. as the coinventors was
granted to UCHICAGO ARGONNE, LLC. The authors declare no
other competing interests. Data and materials availability: Data
used for Pourbaix diagram calculations on La-doped cobalt spinel
are available from Dryad (35). Other data are available in the main
text and the supplementary materials. License information:
Copyright © 2023 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science. No
claim to original US government works. https://www.science.org/
about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade1499
Materials and Methods
Figs. S1 to S32
Tables S1 to S8
References (36–70)
Submitted 31 July 2022; resubmitted 13 February 2023
Accepted 12 April 2023
10.1126/science.ade1499
Chong et al., Science 380, 609–616 (2023)
12 May 2023
8 of 8
|
10.1126_science.abq1416
|
RES EARCH
R E S E A R C H A R T I C L E
◥
ENZYME EVOLUTION
Evolution of increased complexity and specificity at
the dawn of form I Rubiscos
Luca Schulz1, Zhijun Guo2, Jan Zarzycki1, Wieland Steinchen3,4, Jan M. Schuller3,4,
Thomas Heimerl3,5, Simone Prinz6, Oliver Mueller-Cajar2, Tobias J. Erb1,3*, Georg K. A. Hochberg3,4,7*
The evolution of ribulose-1,5-bisphosphate carboxylase/oxygenases (Rubiscos) that discriminate strongly
between their substrate carbon dioxide and the undesired side substrate dioxygen was an important
event for photosynthetic organisms adapting to an oxygenated environment. We use ancestral sequence
reconstruction to recapitulate this event. We show that Rubisco increased its specificity and carboxylation
efficiency through the gain of an accessory subunit before atmospheric oxygen was present. Using structural
and biochemical approaches, we retrace how this subunit was gained and became essential. Our work
illuminates the emergence of an adaptation to rising ambient oxygen levels, provides a template
for investigating the function of interactions that have remained elusive because of their essentiality,
and sheds light on the determinants of specificity in Rubisco.
R ibulose-1,5-bisphosphate carboxylase/
oxygenases (Rubiscos) that are paired
with oxygenic photosynthesis are re-
sponsible for most inorganic carbon as-
similation on Earth today (1, 2). Rubisco
ancestrally evolved in anaerobic environments,
predating the emergence of oxygenic photosyn-
thesis (3, 4). As oxygenic photosynthesis evolved,
Rubisco faced molecular oxygen, which acts as
an undesired side substrate during catalysis. Re-
action with O2 produces 2-phosphoglycolate
(2PG), a metabolite that inhibits carbon me-
tabolism and causes a loss of carbon from me-
tabolism (5, 6). To reduce the buildup of 2PG
and enable its conversion, several mitigation
strategies evolved: recycling of 2PG through
photorespiration (6, 7), carbon-concentrating
mechanisms that concentrate CO2 around
Rubisco (8), and the use of Rubiscos with higher
specificity for CO2 (9–11). Although the exact
strategies used for photorespiration and car-
bon concentration vary, all aerobic phototrophs,
in particular algae and plants, use high-specificity
Rubiscos (9, 11). Their evolution was an impor-
tant ingredient in the rise of oxygenic photo-
synthesis. Yet, how and when Rubisco evolved
high specificity remain unknown despite exten-
sive research and speculation (4, 12–14).
1Department of Biochemistry and Synthetic Metabolism,
Max Planck Institute for Terrestrial Microbiology, 35043
Marburg, Germany. 2School of Biological Sciences, Nanyang
Technological University, Singapore 637551, Singapore.
3Center for Synthetic Microbiology (SYNMIKRO), Philipps
University Marburg, 35043 Marburg, Germany. 4Department
of Chemistry, Philipps University Marburg, 35043 Marburg,
Germany. 5Department of Biology, Philipps University Marburg,
35043 Marburg, Germany. 6Central Electron Microscopy
Facility, Max Planck Institute of Biophysics, 60438 Frankfurt am
Main, Germany. 7Evolutionary Biochemistry Group, Max Planck
Institute for Terrestrial Microbiology, 35043 Marburg, Germany.
*Corresponding author. Email: toerb@mpi-marburg.mpg.de (T.J.E.);
georg.hochberg@mpi-marburg.mpg.de (G.K.A.H.)
The defining structural feature of all known
high-specificity form I Rubiscos is their as-
sembly into a complex of eight catalytic large
subunits (LSUs) and eight noncatalytic small
subunits (SSUs) (total stoichiometry: L8S8)
(1). This stoichiometry evolved from simpler
ancestors that did not interact with SSUs, as
evidenced by modern-day SSU-independent
form I′, II, and III Rubiscos (3, 13). Because the
SSU was previously shown to influence Rubisco
catalysis (15–20) and because it is the most
obvious structural difference between form I
and other Rubiscos, it is thought to be at least
partly responsible for increased specificity
toward CO2 (10, 17). However, it has so far not
been possible to directly test this hypothesis
because the SSU is essential for both cata-
lytic activity and solubility of form I Rubiscos
(15, 21, 22). Studies on its effect have therefore
been limited to in silico calculations (23), homo-
log shuffling experiments (16–19), and small-
scale mutational perturbations (17, 20), which
have not been conclusive and have shown
relatively small, mostly deleterious effects on
specificity.
In this work, we overcome this challenge
using ancestral sequence reconstruction (24)
to recapitulate the evolution of form I Rubiscos.
We identify the genetic and structural causes
for the recruitment of the SSU, show that the
SSU was indirectly responsible for an increase
in specificity, and reveal how the solubility of
form I Rubiscos became dependent on the SSU.
Notably, our experiments suggest that form I
Rubiscos increased their specificity for CO2 over
O2 even before oxygen was abundant (25–28).
SSU gain coincided with rising specificity
To determine when Rubisco gained the SSU,
we inferred maximum-likelihood phylogenies
of Rubisco’s LSU and SSU (Fig. 1A and figs. S1
and S2). On the LSU phylogeny, three clades of
orthologs branch on the stem lineage toward
form I Rubiscos (form I-a, form I′, and form I′′)
(13, 29). These orthologs were identified from
metagenome-assembled genomes (MAGs), which
do not encode SSUs (fig. S3A). We also identi-
fied a fourth clade of true form I Rubiscos from
MAGs that do encode SSUs (form I Anaero).
This clade branches close to the last common
ancestor (LCA) of known cyanobacterial and
plant form I AB and proteobacterial and red
algal form I CD Rubiscos (see the supplemen-
tary text).
These Rubisco-encoding MAGs were sam-
pled in hot environments and belong to or-
ganisms related to anaerobic, thermophilic
Chloroflexaeota (30) and Firmicutes (fig. S3,
A, B, and C). For further characterization, we
purified Rubiscos of clades branching close
to the SSU’s appearance and measured their
oligomerization state using mass photometry
(MP). Form I-a Rubiscos assembled into di-
mers, form I′ Rubiscos into octamers (13), and
form I Anaero variants into L8S8 hexadecamers
(Fig. 1B and fig. S3D). Furthermore, form I
Anaero Rubiscos exhibited high thermal sta-
bility (>80°C), a high catalysis temperature
optimum of 55° to 70°C (kcatC ~ 5 s−1), and low
specificity for CO2 over O2 for form I Rubiscos
(specificity < 27; fig. S3, E, F, and G, and Table 1).
Together, the stoichiometry measurements
and the absence of SSU genes in the MAGs
of form I′′–encoding organisms suggest that
Rubisco began to interact with the SSU after
branching of the form I′′ clade. Furthermore,
the MAG’s origin in hot, anaerobic environ-
ments and the thermophilicity of form I Anaero
Rubiscos suggest that the SSU first evolved
in thermophilic anaerobes that existed before
crown cyanobacteria diverged and thus be-
fore atmospheric oxygen levels rose substan-
tially during the Great Oxygenation Event (GOE)
(25–28). We cannot rule out that yet-unsampled
organisms carry forms of Rubisco that conflict
with these inferences. However, the phyloge-
netic interval in question is bracketed by three
separate Rubisco clades with similar character-
istics that derive from multiple sampling sites
and sequencing projects. An anaerobic, thermo-
philic origin of form I Rubiscos is thus the most
parsimonious scenario based on current data.
We next investigated how the evolution of
Rubisco’s L8S8 assembly influenced its func-
tion. To do this, we inferred sequences of an-
cestral Rubisco LSUs that existed before (AncL)
and after (AncLS) the gain of the SSU (Fig. 1A
and fig. S4, A and B). We also resurrected an
ancestral small subunit (AncSSU). Because the
SSU phylogeny has no outgroup, we recon-
structed a deep node within the form I Anaero
SSUs as a surrogate that existed close to the
presumed root. MP of purified proteins showed
that AncL formed a homo-octamer and that
AncLS formed an L8S8 heterocomplex with
Schulz et al., Science 378, 155–160 (2022)
14 October 2022
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
AncSSU (Fig. 1C). Both complexes were able to
bind the active site inhibitor carboxyarabinitol-
1,5-bisphosphate (CABP) and were stable to
>75°C (fig. S4, C and D).
We quantified AncL and AncLS+AncSSU’s
carboxylation with ribulose-1,5-bisphosphate
(RuBP) as well as their specificities for CO2 over
the side-reaction substrate O2 (SC/O) at 25°C.
From AncL to AncLS+AncSSU, the Km(CO2)
(the Michaelis constant for CO2) decreased
from 568 to 69 mM, the Km(RuBP) increased
from 92 to 544 mM, and the maximal rate of
carboxylation decreased from 0.61 to 0.29
(Table 1). Overall, this resulted in an increased
catalytic efficiency for the carboxylation reac-
tion (1.1 × 103 to 4.2 × 103) as well as a faster
carboxylation at physiologically relevant dis-
solved CO2 concentrations <500 mM and sat-
urating RuBP concentrations in AncLS (fig.
S4E). Lastly, the SC/O increased from 21.9 to
47.3, which is comparable to the change in
specificity from a form II to a cyanobacterial
form I Rubisco (Table 1). Thus, the carbox-
ylation efficiency and specificity of Rubisco
improved when the canonical form I L8S8
assembly first evolved. Subsequently, extant
form I Anaero Rubiscos appear to have reverted
to low specificities (SC/O values between 12 and
26), consistent with their anaerobic habitat.
This suggests that binding of the SSU alone
is not the sole determinant of specificity.
The SSU quickly became essential
Our next aim was to determine to what extent
the SSU was responsible for the functional
changes that we observed. We sought to iden-
tify a Rubisco that can bind the SSU but does
not yet depend on it for solubility and activ-
ity. We first tested whether AncL was already
able to bind AncSSU. We produced AncL with
untagged AncSSU and vice versa. In either case,
the untagged protein did not coelute (fig. S4,
Fig. 1. The evolution of form I Rubiscos. (A) Reduced phylogeny of form I
and related Rubiscos. Arrows depict operon structure. Complete phylogeny is
shown in fig. S2. (B) MP spectra of metagenomic Rubiscos. Color scheme
is according to (A). Inferred stoichiometries are depicted as cartoons. Representative
spectra of two technical replicates. (C) MP spectra of ancestral Rubiscos AncL and
AncLS+AncSSU. Representative spectra of more than five technical replicates.
(D) MP spectrum of AncL after incubation with a fourfold molar excess of purified
AncSSU. (E) Coomassie-stained SDS-PAGE of crude and soluble lysates of AncL and
AncLS produced with or without AncSSU coproduction. MM, molecular mass; sol.,
soluble. Representative gel of more than five biological replicates. See fig. S5A
for full gel. (F) Native PAGE and Western blot analyses of lysates from cultures
producing AncL or AncLS with or without AncSSU coproduction. Loading
control is shown in fig. S5B. HRP, horseradish peroxidase. (G) Relative Rubisco
activities in soluble lysates of cultures producing AncL and AncLS with and without
coproduction of AncSSU. Data are relative to measurements of AncL without
AncSSU coproduction. For SDS-PAGE analysis of biological triplicates, see fig. S5C.
N = 9 technical replicates for AncL, and N = 6 for AncLS. Error bars depict
standard deviations (SDs).
Table 1. Kinetic characterization of ancestral and related extant Rubisco at 25°C. AncL+7 derives from AncL and contains seven substitutions at the
LSU-SSU interface to enable SSU binding (Fig. 2). kcatC is the maximal rate of carboxylation under saturating substrate concentrations. Km(CO2) and Km(RuBP)
are the Michaelis constants for CO2 and RuBP, respectively. SC/O = [kcatC/Km(CO2)]/[kcatO/Km(O2)]. R. rubrum, Rhodospirillum rubrum; A. ferrooxidans,
Acidithiobacillus ferrooxidans; Syn., Synechococcus; PCC, Pasteur Culture Collection. Values are means ± 95% confidence intervals with the number of
technical replicates (N) indicated in parentheses. Extant form I Rubiscos were produced and measured with their native SSUs. Kinetic curves are shown in
fig. S17. The Km(CO2) measurement for RME08239 Rubisco was performed at 1 mM RuBP as opposed to 2.2 mM. n.d., not determined.
Rubisco name or identifier
kcatC
(s−1)
Km(CO2)
(mM)
Km(RuBP)
(mM)
Specificity
(SC/O)
AncL
............................................................................................................................................................................................................................................................................................................................................
AncL+7
............................................................................................................................................................................................................................................................................................................................................
AncL+7 + 4x AncSSU
............................................................................................................................................................................................................................................................................................................................................
AncL+7 s437W
............................................................................................................................................................................................................................................................................................................................................
AncL+7 s437W + 2x AncSSU
............................................................................................................................................................................................................................................................................................................................................
AncL+7 e170N
............................................................................................................................................................................................................................................................................................................................................
AncL+7 e170N + 4x AncSSU
............................................................................................................................................................................................................................................................................................................................................
AncL+7 e170N s437W + 4x AncSSU
............................................................................................................................................................................................................................................................................................................................................
AncLS + AncSSU
............................................................................................................................................................................................................................................................................................................................................
RME08239 + SSU (form I Anaero)
............................................................................................................................................................................................................................................................................................................................................
RMG64267 + SSU (form I Anaero)
............................................................................................................................................................................................................................................................................................................................................
R. rubrum (form II)
............................................................................................................................................................................................................................................................................................................................................
A. ferrooxidans (form II)
............................................................................................................................................................................................................................................................................................................................................
Syn. PCC 6301 + SSU (form I)
............................................................................................................................................................................................................................................................................................................................................
0.61 ± 0.10 (2)
0.63 ± 0.11 (2)
0.41 ± 0.02 (2)
0.61 ± 0.11 (2)
0.49 ± 0.02 (2)
0.55 ± 0.16 (2)
0.47 ± 0.03 (4)
0.49 ± 0.02 (2)
0.29 ± 0.02 (2)
1.1 ± 0.12 (2)
1.94 ± 0.28 (2)
6.31 ± 0.56 (2)
n.d.
n.d.
21.9 ± 1.1 (9)
25.3 ± 1.9 (5)
29.6 ± 1.7 (4)
25.6 ± 1.4 (9)
25.8 ± 2.2 (5)
25.6 ± 2.3 (6)
38.7 ± 2.9 (6)
40.7 ± 2.8 (4)
47.3 ± 2.1 (9)
12.2 ± 2.5 (2)
26.4 (1)
12.6 ± 0.5 (15)
17.9 ± 0.5 (6)
45.7 ± 1.7 (9)
568 ± 177 (2)
499 ± 167 (2)
148 ± 21 (2)
460 ± 158 (2)
162 ± 24 (2)
569 ± 300 (2)
113 ± 29 (4)
119 ± 17 (2)
69 ± 16 (2)
200 ± 47 (2)
88 ± 53 (2)
116 ± 38 (2)
n.d.
n.d.
92 (1)
58 ± 18 (3)
331 ± 76 (3)
63 (1)
483 (1)
28 (1)
203 (1)
n.d.
544 (1)
n.d.
67 (1)
9 (1)
n.d.
n.d.
kcatC/Km(CO2)
(M−1 s−1)
1.1 × 103
1.3 × 103
2.8 × 103
1.3 × 103
3.0 × 103
1.0 × 103
4.2 × 103
4.1 × 103
4.2 × 103
5.5 × 103
2.2 × 104
5.4 × 104
n.d.
n.d.
kcatO/Km(O2)
(M−1 s−1)
0.5 × 102
0.5 × 102
0.9 × 102
0.5 × 102
1.2 × 102
0.4 × 102
1.1 × 102
1.0 × 102
0.9 × 102
4.5 × 102
8.4 × 102
4.3 × 103
n.d.
n.d.
Schulz et al., Science 378, 155–160 (2022)
14 October 2022
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
F and G). Additionally, we purified AncL and
AncSSU separately and used MP to probe the
interaction, but we could not detect binding of
AncSSU to AncL (Fig. 1D).
In parallel, we tested whether AncLS could
still function without the SSU. We produced
both ancestors with and without coproduc-
tion of AncSSU and quantified solubility using
SDS–polyacrylamide gel electrophoresis (SDS-
PAGE), Western blots, and Rubisco activities
in clarified lysates. The yield of soluble AncLS
was markedly reduced without AncSSU, where-
as AncL titers were unchanged (Fig. 1, E, F,
and G). The interaction with the SSU and a
total dependence on it thus evolved in quick
succession between AncL and AncLS. Inter-
actions, dependencies, and catalytic proper-
ties of AncL and AncLS were qualitatively
unchanged when validating their robustness
to statistical uncertainty (31). Moreover, infer-
ences of AncLS’s stoichiometry and SSU de-
pendence were unchanged when AncLS was
reconstructed using an alternative phylog-
eny (29) (figs. S6, S7, and S8 and supplemen-
tary text).
Genetic basis of LSU-SSU interaction
Because AncL cannot yet interact with the
SSU, and AncLS already cannot function with-
out it, historical substitutions that separate
the two constructs must be responsible for both
creating the LSU-SSU interaction and making
it essential. We reasoned that we might be able
to construct a nonobligate LSU-SSU interaction
by introducing only those substitutions into
AncL that create the interaction.
To pinpoint relevant substitutions, we solved
x-ray crystal structures of inhibitor-bound AncL
and AncLS+AncSSU to 2.1- and 1.8-Å resolu-
tion, respectively (Fig. 2, A and B). AncL and
AncLS differ by 95 substitutions, 14 of which
are close to the SSU in the AncLS structure,
and three insertions (Fig. 2, C and D, and fig.
S4B). The 14 substituted sites make up only
22% of all interface sites and 25% of the total
buried interface area in AncLS.
Next, we created AncL constructs contain-
ing single, double, or triple combinations of
the 14 substitutions in proximity to the SSU
(8, 28, and 10 variants, respectively). At least
three substitutions were required to create a
Fig. 2. A small set of substitutions can create the LSU-SSU interface. (A) Structure of CABP-bound
AncL solved to 2.1 Å. (B) Structure of CABP-bound AncLS cocrystallized with AncSSU solved to 1.8 Å. LSU is
shown in light and dark gray, and AncSSU is shown in beige. (C) Schematic phylogeny highlighting the
differences between AncL and AncLS. subs, substitutions; ins, insertions; IF, interface. (D) Substitutions
separating AncL from AncLS that occurred in proximity to the SSU. Small letters denote ancestral states, and
capitalized letters denote derived states. Purple side chains depict substitutions that are transplanted into
AncL+7. The SSU surface is depicted in beige. Heteroatoms within 3.8 Å of the substitutions are highlighted in
red (oxygen) and blue (nitrogen). Water molecule is shown as a red sphere. (E) MP spectra of AncL+7 in
isolation (bottom) or with a fourfold molar excess of purified AncSSU (top). Empirical masses and inferred
stoichiometries are depicted in the histogram. Theoretical L8 mass: 428.0 kDa; theoretical L8S8 mass: 527.6 kDa.
Data are from three technical replicates. Single-letter abbreviations for the amino acid residues are as
follows: A, Ala; C, Cys; D, Asp; E, Glu; F, Phe; G, Gly; H, His; I, Ile; K, Lys; L, Leu; M, Met; N, Asn; P, Pro; Q, Gln;
R, Arg; S, Ser; T, Thr; V, Val; W, Trp; and Y, Tyr.
MP-detectable LSU-SSU interaction (fig. S9),
although this interaction remained weak. To
achieve tighter binding, we simultaneously
introduced seven interface substitutions into
AncL (Fig. 2D). This variant (AncL+7) was sol-
uble without AncSSU and rapidly reconstituted
an L8S8 heterocomplex upon mixing with pu-
rified AncSSU (Fig. 2E).
SSU improved catalysis and specificity
We were now able to directly test how the in-
teraction with the SSU affected Rubisco. To do
this, we measured AncL+7’s kinetic parame-
ters in the presence or absence of AncSSU (the
L8 and L8S8 forms of AncL+7, respectively).
The carboxylation rate of AncL+7 decreased
slightly when interacting with the SSU (kcatC
values of 0.63 and 0.41, respectively), and the
Km(RuBP) increased sixfold. Notably, how-
ever, the Km(CO2) was decreased threefold,
which, in total, leads to a 2.2-fold improvement
in carboxylation efficiency. Carboxylation is
up to twofold faster at physiologically rele-
vant dissolved CO2 concentrations <500 mM
in the L8S8 form at saturating RuBP concen-
trations (Fig. 3, A and B, and Table 1). We
verified robustness of these inferences with
an alternative SSU reconstruction, correspond-
ing to a deeper but worse-reconstructed node
on our SSU phylogeny (fig. S10).
Besides increasing carboxylation efficiency,
the addition of AncSSU to AncL+7 markedly
increased the temperature tolerance of catal-
ysis from ~40°C up to 50° to 70°C, which results
in an up to 12.5-fold improved carboxylation
rate in the L8S8 form at 70°C (Fig. 3C). Yet,
unexpectedly, AncSSU did not increase overall
stability of AncL+7 toward denaturation: We
found no difference in stability between the
L8 and L8S8 forms upon exposure to elevated
temperatures (Fig. 3D).
Next, we quantified how the SSU affected
CO2/O2 specificity. The seven substitutions in-
troduced at the LSU-SSU interface to create
AncL+7 already caused a statistically signifi-
cant increase in specificity from 21.9 to 25.3
(P = 0.0013, two-tailed t test; Fig. 3E and
Table 1). This increase was amplified to 29.6
when the SSU was added. We could further
improve specificity by introducing one more
substitution at the LSU-SSU interface. Intro-
ducing e170N (where small letters denote an-
cestral states and capitalized letters denote
derived states) into AncL+7 boosted speci-
ficity to 38.7 in the presence of AncSSU but
did not improve AncL+7’s specificity without
AncSSU (Fig. 3E and Table 1). Together, eight
substitutions and the presence of AncSSU
almost doubled the specificity of AncL from
21.9 to 38.7, whereas no changes in or close to
the active site were required. The SSU is thus
directly and indirectly responsible for the key
functional improvements associated with form I
Rubiscos: The interaction directly improves
Schulz et al., Science 378, 155–160 (2022)
14 October 2022
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
catalytic efficiency and catalysis at high tem-
peratures. Its effect on specificity is (at least
partly) indirect because it epistatically enhances
the effect of substitution e170N, which has no
influence on specificity without the SSU.
To elucidate the structural basis of the SSUs’
effects on catalysis, we solved x-ray crystal struc-
tures of inhibitor-bound AncL+7 as well as
AncL+7 e170N in their L8 and L8S8 forms
(Fig. 3F and fig. S11A). In both complexes,
individual monomers are virtually identical
between the L8 and L8S8 forms [average root
mean square deviation (RMSD) < 0.3 Å; fig.
S11B]. The only observable structural differ-
ence is that the LSU octamer is slightly more
compact in the L8S8 forms, but this does not
translate into observable active site rearrange-
ments (Fig. 3G and fig. S11, C and D). The high
specificity–conferring e170N substitution at the
LSU-SSU interface is far away from the active
site, where it engages in a hydrogen-bonding
interaction with Y83 and Q88 on the SSU (fig.
S11E). The SSU thus improves catalysis through
allosteric effects on either the open, non–
inhibitor-bound state or protein dynamics.
A single substitution causes SSU dependence
All characterized form I Rubiscos depend on
the SSU for solubility or catalysis (15, 21, 22),
and our results indicate that this trait had
already evolved by the time of AncLS. We
therefore sought to discover how this depen-
dence evolved. Drawing on prior work on how
Fig. 3. The SSU allosterically enhances catalysis.
(A and B) Kinetic characterization of AncL+7 with or
without AncSSU (L8 and L8S8 forms, respectively). Data
are the means of two (A) or three (B) technical replicates.
Error bars indicate SDs. (C) Carboxylation rates of the
L8 and L8S8 forms of AncL+7 at varying temperatures.
Bars represent the means of two technical replicates.
Activity is relative to the rate of catalysis at 25°C in L8 form.
(D) Remaining activities of the L8 and L8S8 forms of AncL+7
after incubation at elevated temperatures, shown
relative to incubation (Incub.) at 30°C. Curve is a sigmoidal
fit (N = 3 technical replicates; error bars indicate SDs).
(E) Specificity of AncL variants with and without AncSSU.
Significance tested by two-tailed t test. ns, nonsignificant;
**P < 0.01; ***P < 0.001; ****P < 0.0001. N values are
listed in Table 1, and error bars represent 95% confidence
intervals. (F) Structures of CABP-bound AncL+7 crystallized
as a homomer (left) or in complex with AncSSU (right)
solved at 2.65 and 2.0 Å, respectively. LSU monomers are
shown in light and dark purple, and AncSSU is shown in
beige. (G) Active site residues of the L8 (purple) and L8S8
forms (orange) of AncL+7 with coordinated magnesium
(green spheres) and bound CABPs (L8, light gray; L8S8,
dark gray). K*187 marks the carbamylated active site lysine.
Fig. 4. Rubisco evolves to depend on the SSU for
solubility and catalysis. (A) Relative Rubisco activities
and SDS-PAGE analysis of soluble lysates from
strains producing AncL and AncL+14 with or without
AncSSU. LSU band is indicated. Full gel is shown in
fig. S16A. N = 2 technical replicates. (B) Relative
Rubisco activities and SDS-PAGE analysis of soluble
lysates from strains producing AncL+7 and derivates
thereof. Full gel is shown in fig. S16B, and crude lysate
pictures are shown in fig. S16C. N = 2 technical
replicates. (C) A 3.01-Å resolution cryo-EM density map
of AncL+7 s437W fibrils. (D) Structural interactions
across the fiber interface of sites that contain
entrenching states. (E) Coomassie-stained SDS-PAGE
of clarified lysates from cells producing AncLS and
variants thereof. AncSSU was not coproduced. Genotype
of the fiber IF reversion construct is listed in fig. S11H.
Full gel is shown in fig. S16B. (F) 3-Phosphoglycerate
production of fiber interface reversion is SSU-dependent,
even though solubility is recovered. Arrows mark addition
of RuBP and AncSSU (10-fold excess). (G) Relative
active site contents of Rubisco LSU (0.2 nmol) and SSU
(2 nmol) samples, as determined by 14C-CABP binding
assays. FIR, fiber interface reversion construct. Bars
represent the means of two technical replicates.
Schulz et al., Science 378, 155–160 (2022)
14 October 2022
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
interactions can become essential (32, 33), we
hypothesized that entrenching substitutions
occurred near the SSU interface. To test this
idea, we simultaneously transplanted all seven
additional SSU-neighboring substitutions from
AncLS into AncL+7. The resulting AncL+14 was
only soluble when AncSSU was coproduced
and formed an L8S8 heterocomplex (Fig. 4A
and fig. S12A).
To pinpoint which changes cause SSU de-
pendence, we introduced all 14 substitutions
individually into AncL and tested their effect
on solubility (fig. S12B and fig. S13). Substi-
tution s437W reduced solubility of both AncL
and AncL+7 by ~75%, as assessed by SDS-PAGE
and relative Rubisco activities in clarified ly-
sate (Fig. 4B). We examined the remaining six
substitutions separating AncL+7 s437W from
AncL+14 for their effect on solubility (fig. S12,
C and D). Substitution e170N, which was also
responsible for the specificity increase, had
the most severe effect. When introduced into
AncL+7 s437W, it caused a complete loss of sol-
ubility, which could be rescued by coproducing
AncSSU. Notably, e170N alone had no nega-
tive effect on the solubility of AncL+7 (Fig. 4B).
Thus, a single historical substitution (s437W)
could make Rubisco become dependent on
the SSU for solubility.
Introducing s437W had no measurable ef-
fect on most biochemical parameters of AncL+7
or on catalysis and specificity of AncL+7
e170N in the presence of AncSSU (fig. S14 and
Table 1). Our results imply that Rubisco’s de-
pendence on the SSU for solubility is function-
ally neutral and is not required for the catalytic
improvements associated with form I Rubiscos.
This conclusion is further supported by the
fact that most of the functional differences
between AncL and AncLS are recapitulated in
AncL+7 e170N+AncSSU, which does not de-
pend on the SSU (Table 1).
Mechanisms of essentiality
To conclude our investigation, we sought to un-
derstand how these two substitutions make
the LSU require the SSU for solubility. Rubiscos
containing s437W formed higher-order oligo-
mers in MP measurements and stacked into
fibers in negative-stain electron microscopy
images (fig. S12, E, F, and G). This led us to
hypothesize that insolubility stems from LSU
self-assembly into fibers (34), which can be
prevented by the SSU.
To understand the molecular basis of fiber
formation, we solved the structure of AncL+7
s437W fibers to 3.01 Å using cryo–electron mi-
croscopy (cryo-EM) (Fig. 4C and fig. S15). In
fibers, LSU octamers interact at their apices,
forming a helix with a twist of ~21° per oc-
tamer. The fiber-inducing W437 directly en-
gages in a cation-p interaction with R3 of an
adjacent LSU octamer at the fiber interface.
N170, which exacerbates the effect of W437,
seems to have a less direct mechanism of action,
as it is >10 and >25 Å away from either of the
fiber interfaces (Fig. 4D). The fiber interface
partially overlaps with the LSU-SSU interface,
which makes the cavity between interacting
octamers too small to accommodate SSUs. Cap-
ping of LSU octamers by SSUs therefore makes
fiber formation impossible (fig. S12, H and I).
We tested whether substitutions e170N and
s437W remain the sole causes of insolubility in
AncLS or whether additional entrenching sub-
stitutions accumulated at the SSU interface.
Neither N170e nor W437s reversions, alone or
as a pair, fully restored solubility without the
SSU (Fig. 4E). Simultaneously reverting all 14
substitutions that occurred within 5 Å of the
fiber interface fully recovered solubility in the
absence of the SSU (Fig. 4E and fig. S12H).
Hence, the region around the fiber interface
remains the cause of insolubility, and additional
entrenching substitutions beyond e170N and
s437W accumulated after the gain of the SSU.
In the absence of SSUs, this so-called fiber
interface reversion construct assembled into a
conventional L8 complex but was only partially
able to bind the active site inhibitor CABP and
was barely active (~68% active site content;
kcatC < 0.05 s−1). Addition of AncSSU quickly
reconstituted an L8S8 complex, increased ac-
tive site content to levels comparable to those
of AncLS, and restored catalytic activity (Fig. 4,
F and G, and fig. S12J). This indicated that
even though we reverted the dependence on
SSU for solubility, some of the remaining 81
substitutions separating AncL and the fiber
interface reversion construct made Rubisco
depend on the SSU for catalysis. This catalytic
dependence is also present in the only form I
Rubisco that is soluble without the SSU (15).
Conclusions
We show that the SSU modified the functional
sequence space available to Rubisco. Conse-
quently, certain substitutions became ad-
vantageous in the presence of the SSU that
otherwise would not have been. However, it
also enabled the accumulation of substitutions
that created a dependence on the interaction.
Rubisco is a slow-evolving (35), notoriously
constrained enzyme (36, 37), and the recruit-
ment of an epistatic modifier like the SSU may
have been the only way for it to access high
specificity.
All features of modern form I Rubiscos—
interaction with and dependence on the SSU,
as well as high specificity—had already evolved
by the time of AncLS. Notably, this timeline
implies that high-specificity Rubiscos predate
the GOE and may even predate the evolution
of photosystem II, although the exact timing
of this event is still contested (25–28, 38). One
possibility is that specificity first improved as a
by-product of selection for decreasing Km(CO2)
alongside a drop in atmospheric CO2 concen-
trations before the GOE (39). This would have
fortuitously prepared Rubisco for the subse-
quent rise of atmospheric O2 concentrations.
However, there is also tentative evidence that
O2 (and, by implication, oxygenic photosyn-
thesis) may have been present before the GOE
(28, 38, 40–42), which could have directly se-
lected for improved specificity. In either case,
the evolution of the SSU facilitated Rubisco’s
coexistence with oxygenic photosynthesis,
which helped set the stage for complex life in
aerobic environments.
REFERENCES AND NOTES
1.
I. Andersson, A. Backlund, Plant Physiol. Biochem. 46, 275–291
(2008).
2. Y. M. Bar-On, R. Milo, Proc. Natl. Acad. Sci. U.S.A. 116,
4738–4743 (2019).
3. F. R. Tabita, S. Satagopan, T. E. Hanson, N. E. Kreel, S. S. Scott,
J. Exp. Bot. 59, 1515–1524 (2008).
4. T. J. Erb, J. Zarzycki, Curr. Opin. Biotechnol. 49, 100–107 (2018).
5. W. W. Cleland, T. J. Andrews, S. Gutteridge, F. C. Hartman,
G. H. Lorimer, Chem. Rev. 98, 549–562 (1998).
6. F. Flügel et al., Plant Cell 29, 2537–2551 (2017).
7. H. Bauwe, M. Hagemann, A. R. Fernie, Trends Plant Sci. 15,
8.
330–336 (2010).
J. A. Raven, C. S. Cockell, C. L. De La Rocha, Phil. Trans. R. Soc. B
363, 2641–2650 (2008).
9. A. I. Flamholz et al., Biochemistry 58, 3365–3376 (2019).
10. R. J. Spreitzer, Arch. Biochem. Biophys. 414, 141–149 (2003).
11. C. Iñiguez et al., Plant J. 101, 897–918 (2020).
12. B. Kacar, V. Hanson-Smith, Z. R. Adam, N. Boekelheide,
Geobiology 15, 628–640 (2017).
13. D. M. Banda et al., Nat. Plants 6, 1158–1166 (2020).
14. P. M. Shih et al., Nat. Commun. 7, 10382 (2016).
15. T. J. Andrews, J. Biol. Chem. 263, 12213–12219 (1988).
16. T. J. Andrews, D. M. Greenwood, D. Yellowlees, Arch. Biochem.
Biophys. 234, 313–317 (1984).
17. S. Karkehabadi et al., Biochemistry 44, 9851–9861 (2005).
18. E. Martin-Avila et al., Plant Cell 32, 2898–2916 (2020).
19. M. T. Lin, W. D. Stone, V. Chaudhari, M. R. Hanson, Nat. Plants
6, 1289–1299 (2020).
20. B. A. Read, F. R. Tabita, Biochemistry 31, 519–525 (1992).
21. C. Liu et al., Nature 463, 197–202 (2010).
22. J. Joshi, O. Mueller-Cajar, Y.-C. C. Tsai, F. U. Hartl,
M. Hayer-Hartl, J. Biol. Chem. 290, 1066–1074 (2015).
23. M. van Lun, J. S. Hub, D. van der Spoel, I. Andersson, J. Am.
Chem. Soc. 136, 3165–3171 (2014).
24. G. K. A. Hochberg, J. W. Thornton, Annu. Rev. Biophys. 46,
247–269 (2017).
25. P. M. Shih, J. Hemp, L. M. Ward, N. J. Matzke, W. W. Fischer,
Geobiology 15, 19–29 (2017).
26. B. E. Schirrmeister, J. M. de Vos, A. Antonelli, H. C. Bagheri,
Proc. Natl. Acad. Sci. U.S.A. 110, 1791–1796 (2013).
27. B. E. Schirrmeister, M. Gugger, P. C. J. Donoghue,
Palaeontology 58, 769–785 (2015).
28. N. J. Planavsky et al., Nat. Geosci. 7, 283–286 (2014).
29. J. A. West-Roberts et al., bioRxiv 2021.08.23.457424 [Preprint]
(2021). https://doi.org/10.1101/2021.08.23.457424.
30. Y. Sekiguchi et al., Int. J. Syst. Evol. Microbiol. 53, 1843–1851
(2003).
31. G. N. Eick, J. T. Bridgham, D. P. Anderson, M. J. Harms,
J. W. Thornton, Mol. Biol. Evol. 34, 247–261 (2017).
32. G. K. A. Hochberg et al., Nature 588, 503–508 (2020).
33. A. Stoltzfus, J. Mol. Evol. 49, 169–181 (1999).
34. H. Garcia-Seisdedos, C. Empereur-Mot, N. Elad, E. D. Levy,
Nature 548, 244–247 (2017).
35. J. W. Bouvier, D. M. Emms, S. Kelly, bioRxiv 2022.07.06.498985
[Preprint] (2022). https://doi.org/10.1101/2022.07.06.498985.
36. G. G. B. Tcherkez, G. D. Farquhar, T. J. Andrews, Proc. Natl.
Acad. Sci. U.S.A. 103, 7246–7251 (2006).
37. J. W. Bouvier et al., Mol. Biol. Evol. 38, 2880–2896 (2021).
38. T. Cardona, J. W. Murray, A. W. Rutherford, Mol. Biol. Evol. 32,
1310–1328 (2015).
39. A. M. Hessler, D. R. Lowe, R. L. Jones, D. K. Bird, Nature 428,
736–738 (2004).
40. A. M. Satkoski, N. J. Beukes, W. Li, B. L. Beard, C. M. Johnson,
Earth Planet. Sci. Lett. 430, 43–53 (2015).
41. J. Jabłońska, D. S. Tawfik, Nat. Ecol. Evol. 5, 442–448 (2021).
42. S. Kumar, G. Stecher, M. Suleski, S. B. Hedges, Mol. Biol. Evol.
34, 1812–1819 (2017).
43. L. Schulz, G. Hochberg, T. Erb, Evolution of increased
complexity and specificity at the dawn of Form I Rubiscos,
Schulz et al., Science 378, 155–160 (2022)
14 October 2022
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
version 1.0, Edmond (Max Planck Digital Library, 2022);
https://doi.org/10.17617/3.F9BWOU.
ACKN OW LEDG MEN TS
The authors thank the Central Electron Microscopy Facility at
the Max Planck Institute of Biophysics for expertise and access to
their instruments. We acknowledge support from the European
Synchrotron Radiation Facility (ESRF, Grenoble, France) and the
EMBL Hamburg at the PETRA III storage ring (DESY, Hamburg,
Germany). Funding: O.M.-C. is thankful for support from a
Ministry of Education (MOE) of Singapore Tier 2 grant (MOE-
T2EP30120-0005). J.M.S. is grateful for an Emmy Noether grant
(SCHU 3364/1-1) from the Deutsche Forschungsgemeinschaft
(DFG). L.S., J.Z., S.P., T.J.E., and G.K.A.H. are grateful for generous
support from the Max Planck Society. L.S. thanks Refeyn Ltd. for a
travel grant that funded the presentation of this work. Author
contributions: L.S., T.J.E., and G.K.A.H. conceived the project,
analyzed data, and planned experiments. L.S. performed molecular
work, phylogenetics, ancestral sequence reconstruction, protein
purification, cryo-EM and x-ray analysis, comparative in vitro
biochemistry, and MP measurements. Z.G. performed enzyme
kinetic analysis and specificity constant measurements. O.M.-C.
supervised enzyme kinetic analysis and specificity constant
measurements. W.S. analyzed x-ray structures. J.Z. collected,
solved, refined, and analyzed x-ray structures and refined the cryo-EM
structure. T.H. performed negative-stain EM. S.P. performed
cryo-EM grid preparation and collected cryo-EM datasets, and
J.M.S. processed the cryo-EM datasets. O.M.-C., T.J.E., and
G.K.A.H. supervised the project. L.S., T.J.E., and G.K.A.H. wrote the
manuscript with contributions and comments from all authors.
Competing interests: O.M.-C. has consulted for FL79 Inc. This
company did not fund and had no role in this research project. The
authors declare no other competing interests. Data and materials
availability: All raw data for MP spectra and kinetic traces as
well as phylogenetic trees, alignments, and ancestral sequences
are deposited on Edmond, the Open Research Data Repository of
the Max Planck Society for public access (43). Atomic structures
reported in this paper are deposited to the Protein Data Bank
under accession codes 7QSV, 7QSW, 7QSX, 7QSY, 7QSZ, 7QT1, and
7QVI. The cryo-EM data were deposited to the Electron Microscopy
Data Bank under EMD-14178. License information: Copyright ©
2022 the authors, some rights reserved; exclusive licensee
American Association for the Advancement of Science. No claim to
original US government works. https://www.science.org/about/
science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abq1416
Materials and Methods
Supplementary Text
Figs. S1 to S18
Tables S1 to S7
References (44–73)
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
Submitted 21 March 2022; accepted 26 August 2022
10.1126/science.abq1416
Schulz et al., Science 378, 155–160 (2022)
14 October 2022
6 of 6
|
10.1126_science.ade0004
|
RES EARCH
CONVERGENT EVOLUTION
High level of novelty under the hood
of convergent evolution
Steven M. Van Belleghem1,2*, Angelo A. Ruggieri1, Carolina Concha1,3, Luca Livraghi3,4,
Laura Hebberecht3,5,6, Edgardo Santiago Rivera1,3,7, James G. Ogilvie3,8, Joseph J. Hanly3,4,
Ian A. Warren6, Silvia Planas9, Yadira Ortiz-Ruiz1,9, Robert Reed10, James J. Lewis11,
Chris D. Jiggins6, Brian A. Counterman8, W. Owen McMillan3, Riccardo Papa1,9,12*
Little is known about the extent to which species use homologous regulatory architectures to achieve
phenotypic convergence. By characterizing chromatin accessibility and gene expression in developing
wing tissues, we compared the regulatory architecture of convergence between a pair of mimetic
butterfly species. Although a handful of color pattern genes are known to be involved in their
convergence, our data suggest that different mutational paths underlie the integration of these genes
into wing pattern development. This is supported by a large fraction of accessible chromatin being
exclusive to each species, including the de novo lineage-specific evolution of a modular optix enhancer.
These findings may be explained by a high level of developmental drift and evolutionary contingency
that occurs during the independent evolution of mimicry.
A s species diverge, mutations accumulate,
and genes, regulatory elements, or path-
ways that are tightly regulated during
development in one species may not be
similarly constrained in the other. These
genetic changes can generate different ge-
nomic environments that still underlie the
same phenotypes, a process called develop-
mental systems drift (1). Cases of convergent
evolution allow us to study how natural selec-
tion can generate biological similarities in
independent lineages despite their different
genomic environments (2). This largely un-
answered question has implications for un-
derstanding the molecular mechanisms that
promote biological diversity.
Studying convergent evolution within Hel-
iconius butterflies (3) and other adaptive ra-
diations such as African cichlids (4) has
provided insight into the link between nat-
ural selection and the genetic variation that
has shaped the appearance of diverse mor-
phologies. Recently, owing to technological
advances in chromatin profiling, we can study
the gene regulatory architecture of these mor-
phological adaptations (5–8). Chromatin re-
1Department of Biology, University of Puerto Rico, Rio
Piedras, Puerto Rico. 2Ecology, Evolution and Conservation
Biology, Biology Department, KU Leuven, Leuven, Belgium.
3Smithsonian Tropical Research Institute, Panama City,
Republic of Panama. 4Department of Biological Sciences,
The George Washington University, Washington, DC, USA.
5School of Biological Sciences, Bristol University, Bristol, UK.
6Department of Zoology, University of Cambridge, Cambridge,
UK. 7Department of Biomaterials, Universität Bayreuth,
Bayreuth, Germany. 8Department of Biological Sciences,
Auburn University, Auburn, Alabama, USA. 9Molecular Sciences
and Research Center, University of Puerto Rico, San Juan,
Puerto Rico. 10Department of Ecology and Evolutionary Biology,
Cornell University, Ithaca, NY, USA. 11Department of Genetics
and Biochemistry, Clemson University, Clemson, South
Carolina, USA. 12Comprehensive Cancer Center, University of
Puerto Rico, San Juan, Puerto Rico.
*Corresponding author. Email: steven.vanbelleghem@kuleuven.be
(S.M.V.B.); rpapa.lab@gmail.com (R.P.)
modeling plays a key role in determining
cellular identity by exposing cis-regulatory
elements (CREs) to transcription factors (TFs)
and regulating gene expression, thus present-
ing an important link between genetic muta-
tions and developmental processes (9). In
this work, we used this approach to study
the degree of regulatory homology in a case
of Müllerian mimicry between two pairs of
Heliconius species and determine how muta-
tional differences have affected the evolution-
ary trajectory toward convergence.
In Heliconius butterflies, divergence and con-
vergence of wing color patterns has largely
been assigned to allelic changes at only a few
genes with major phenotypic effects (10–13).
However, recent studies of accessible chro-
matin have revealed an intricate regulatory
architecture near these genes that modulates
their spatiotemporal expression patterns (14–17).
Whereas one study revealed that indepen-
dent modular CREs at the cortex gene con-
trol the mimetic yellow hindwing bar between
Heliconius erato and Heliconius melpomene
(16), a similar study on the optix gene proposed
that conserved pleiotropic CREs underlie red
color patterns between these comimics (17).
Moreover, a third study on WntA suggested
that divergent regulatory changes could ex-
plain the different melanic wing patterns in-
duced by a CRISPR-Cas9 WntA gene deletion,
or knockout (KO), across three pairs of Heli-
conius mimics (18, 19). Overall, these studies
are suggesting a divergent regulation of mimetic
wing patterning that has evolved from an an-
cestral developmental plan (20).
Our work focuses on a pair of comimetic
Heliconius species from Panama that di-
verged ~11.1 million years ago and converged in
forewing pattern (H. erato, geographic morphs
demophoon and hydara, and H. melpomene,
geographic morphs melpomene and rosina)
(Fig. 1A). To understand the extent to which
convergence in wing color patterns has oc-
curred through a homologous or nonhomol-
ogous regulatory architecture, we combined
differences in chromatin accessibility and
gene expression data with a pangenome ref-
erence approach that accounted for genomic
deletions and insertions (21). Using this strat-
egy, we (i) investigated the level of chromatin
similarities genome-wide, (ii) quantified and
characterized differences in chromatin ac-
cessibility and gene expression in develop-
ing wings and sections of the forewing, and
(iii) used CRISPR-Cas9 to validate a previ-
ously uncharacterized functional CRE near
the red color pattern gene optix that under-
lies this convergent phenotype exclusive to
the H. erato lineage.
Differences in chromatin accessibility suggest
a divergent regulatory architecture
We quantified the magnitude of genome-wide
changes in chromatin accessibility in the two
butterfly species as a function of tissue and
development (Fig. 1B). As expected, we ob-
served highly dynamic chromatin remodeling
over development (22), which represented the
strongest predictor of chromatin accessibil-
ity within species (Fig. 1C). Out of the 152,897
ATAC-seq (assay for transposase-accessible
chromatin with sequencing) peaks identi-
fied across the total dataset in H. erato (all
tissues and time points), a total of 7.02, 4.51,
and 7.92% were differentially more accessi-
ble (i.e., had significantly more ATAC-seq read
counts), respectively, in fifth-instar larvae and
36- and 60-hour pupae. In H. melpomene, out
of a total of 135,296 ATAC-seq peaks, 8.39,
3.08, and 2.55% were differentially more acces-
sible, respectively in fifth-instar larvae and 36-
and 60-hour pupae (fig. S1).
To explore the distinctness of the species’
chromatin landscapes, we compared the posi-
tion and DNA sequence conservation of ATAC-
seq peaks between H. erato and H. melpomene
using a pangenome assembly. We tested for
different overlap [1 base pair (bp) versus 50%
reciprocal overlap] and replication criteria
(i.e., peak present in at least two samples ver-
sus all samples, tissues, or time points) and
consistently found a high number of species-
specific open chromatin regions (table S1). For
example, we found 57% of the total number
of ATAC-seq peaks to be species-specific, with
7277 in H. erato and 10,762 in H. melpomene
when we used our most conservative analyses
(peak present in all samples for a tissue or time
point within species and only 1-bp overlap
between species) (table S1, panel iv). This level
of distinctness was increased to 70.2% when
we used a 50% reciprocal overlap between
peaks from all samples and developmental
time points, with 10,467 in H. erato and 13,952
in H. melpomene. Across all overlap criteria,
Van Belleghem et al., Science 379, 1043–1049 (2023)
10 March 2023
1 of 7
RES EARCH | R E S E A R C H A R T I C L E
the lowest proportion of specific ATAC-seq
peaks observed was 26.1% in H. erato and
28.3% in H. melpomene (table S1). The dis-
tinctness in the chromatin landscape further
increased when we accounted for differen-
tial accessibility among overlapping peaks
(tables S2 and S3). For example, among the
total of 33,678 ATAC-seq peaks that were
identified as having 50% reciprocal overlap
between the two species, 8.1% (2724) to 18.1%
(6084) were significantly differentially ac-
cessible between the same tissues and time
points [table S3, foldchange (FC) > 1, adjusted
P < 0.05]. We found such a distinct chro-
matin architecture between H. erato and
H. melpomene to be equally distributed across
the 21 chromosomes (fig. S2).
Finally, we observed that ATAC-seq peaks with
less overlap between species in the pangenome
alignment generally occurred in less-conserved
genomic regions (Fig. 1D, left column). We
identified that for 11.7% (2347) and 7.9% (758)
of the total ATAC-seq peaks identified, the
sequence was only present (0% sequence sim-
ilarity) in H. erato and H. melpomene, respec-
tively, and up to 41.4% (8332) and 46.5% (4479)
had less than 50% sequence conservation (Fig.
1F, right column). Specific ATAC-seq peaks
(with 0-bp overlap between species) had sim-
ilar fold changes when compared to shared
peaks (Fig. 1E), which suggests that they have
similar changes in accessibility (see supplemen-
tal text and figs. S3 to S5 for details on fold-
change comparisons between species). Overall,
these results highlight the existence of a wide-
spread chromatin divergence, which is strongly
driven by genomic sequence evolution.
Dissimilarities in chromatin profiles of
developing fore- and hindwings
between comimics
To compare the chromatin landscape of devel-
oping wings between species, we first studied
the differences between fore- and hindwing
chromatin in each species-specific genomic
background. Our analysis of ATAC-seq peak
position and sequence similarity highlighted
that highly overlapping peaks can have low
sequence similarity (and vice versa) (Fig. 1D).
For these analyses, we used a less-restrictive
minimum of two samples within-species for the
ATAC-seq peak to be retained in the analysis
and a 50% reciprocal overlap between the spe-
cies for the ATAC-seq peak to be considered
“shared” (table S1, panel i). These criteria
allowed us to also analyze the variable portion
of the ATAC-seq signal within-species and
enforced both physical overlap of ATAC-seq
peaks and sequence similarity between species.
As expected from the shared ontogeny of
wings (23), less than 0.5% of ATAC-seq peaks
across development had significantly different
chromatin accessibility between the fore- and
hindwings in each of the two species (Fig. 2A).
Out of 2535 ATAC-seq peaks subdivided into
1563 and 972 peaks that were significantly
differentially more accessible in one of the two
wings in H. erato and H. melpomene, respec-
tively, only 7.2% (183 regions) were considered
shared and had similar accessibility patterns
in the two comimics. These included peaks
near potentially important wing developmen-
tal genes such as distal-less (Dll), pangolin
(pan), and dachsous (ds) in the forewing and
Ultrabithorax (Ubx), aristaless (al), split ends
(spen), winged eye (wge), and cubitus interruptus
(ci) in the hindwing (Fig. 2A and table S4). Of
the remaining peaks with a species-specific
fore- or hindwing accessibility pattern, 58.8%
(1490) were not identified (distinctly called
peak) in the other species even at different
time points, and 3.8% (96) were identified
in both species but had significantly differ-
ent accessibility (tables S2 and S3). The 183
Fig. 1. Sampling of chromatin accessibility (ATAC-seq) data and archi-
tecture of specific and shared chromatin landscape between H. erato
and H. melpomene. (A) Geographic distribution of red-banded H. erato and
H. melpomene postman morphs used in this study. The populations of H. erato
demophoon and H. melpomene rosina have a red forewing band and a yellow
hindwing bar and admix with, respectively, H. erato hydara and H. melpomene
melpomene that lacks the yellow hindwing bar. Samples come from reared
morphs of Panama indicated with an asterisk (*). (B) Tissue sampling of fifth-
instar head, forewing (FW), and hindwing (HW), and 36-hour pupal (day 1) and
60-hour pupal (day 2) FW sections (FP, FW posterior; FM, FW medial; FD, FW
distal) and HW. (C) Principal components analysis (PCA) of ATAC-seq count
values for peaks with at least 25% overlap between species. (D) Sequence
similarity distribution between H. erato and H. melpomene for shared (left, ≥1-bp
overlap, with overlapping ranges investigated at 25% intervals) and specific
(right, 0-bp overlap) ATAC-seq peaks. Dashed lines indicate density distributions.
(E) Log2-fold changes of shared (colored) and specific (dashed lines) ATAC-seq
peaks that were differentially more accessible at a developmental time point
in H. erato and H. melpomene.
Van Belleghem et al., Science 379, 1043–1049 (2023)
10 March 2023
2 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Forewing and hindwing identity observed from gene expression
and chromatin landscape. (A) Venn diagrams show the differentially accessible
(DA) ATAC-seq peaks between the fore- and hindwings in H. erato and
H. melpomene. Circles connected with dashed lines indicate how many of these
wing-specific ATAC-seq peaks are shared between the two species (50%
reciprocal overlap). (B) ATAC-seq profile near the Ubx gene in fifth-instar
caterpillars. Blue and green shading indicate sequence that is specific to H. erato
and H. melpomene, respectively. Peaks in red are significantly more accessible in
the hindwing compared with forewing near Ubx and indicate the expected
conserved homology at this gene. Asterisks (*) indicate peaks that are shared
between species but significantly differentially accessible. (C) TF motifs enriched
in differentially accessible ATAC-peaks between fore- and hindwing and their RNA
expression levels. Log(e-value) indicates the significance level of the enrichment
signal, with red and blue indicating higher enrichment in the fore- and hindwing,
respectively, and black indicating enrichment in both fore- and hindwing. Log2FC
indicates the expression level relative to the alternative wing. (D) Gene expression
volcano plots with differentially expressed genes that have a differentially accessible
ATAC-seq peak nearby. Red and blue indicate open ATAC-seq peak in fore- or hindwing,
respectively. Upward and downward triangles indicate the enhancing or suppressing
effect of the ATAC-seq peak. Significantly differentially expressed TFs with significant
motif enrichment signal are indicated in gray. The bar plots show the counts of the
enhancing and suppressing ATAC-seq peaks in fore- and hindwing.
Van Belleghem et al., Science 379, 1043–1049 (2023)
10 March 2023
3 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Chromatin accessibility and gene expression in 36-hour pupa
forewing sections. (A) Differentially accessible (DA) ATAC-seq peaks between
forewing sections in H. erato and H. melpomene. ATAC-seq peaks are either
significantly open (black lines) or closed (dark red lines) in FP, FM, FD, or a
gradient + to − (increasing or decreasing accessibility from the proximal to
distal wing section). Green lines indicate ATAC-seq peaks that are considered
shared between H. erato and H. melpomene. For each comparison, we present
the total and shared count numbers. (B) Numbers are differentially accessible
ATAC-seq peaks in the wing sections. In contrast to (A), these numbers are
obtained by pairwise comparisons between wing sections. Numbers at the
boundaries of wing sections indicate peaks with shared differential accessibility
compared to the other wing section. Numbers in the middle of the wings
indicate peaks identified as shared between H. erato and H. melpomene
(50% reciprocal overlap). Wings on the right show the wild-type phenotypes
of H. erato and H. melpomene, with the blue lines indicating the extent of
red scale development (and optix expression) in the WntA CRISPR-Cas9 KO.
Numbers next to the wings represent DA peaks between FP or FM and FD
in H. erato and H. melpomene, respectively. (C) TF motif enrichment (left) for
differentially accessible ATAC-peaks between wing sections and expression
of associated TFs (right). Log(e-value) indicates the significance level of
the enrichment signal, and log2FC indicates the expression level relative to all
other sections.
shared wing identity peaks had an average
sequence similarity of 80.7% (SD = 14.5),
whereas the 2352 distinct wing identity peaks
had an average sequence similarity of 66.6%
(SD = 24.7), with 4.1% (97) being explained by
0% sequence conservation and 30.4% having
less than 50% sequence similarity in the alter-
native species (tables S5 and S6).
Of the ATAC-seq peaks that were more ac-
cessible in the hindwing, many were concen-
trated within 100 kb of the Ubx gene [5.56% (82)
and 2.04% (20) in H. erato and H. melpomene,
respectively], which is known to be a key gene
for insect hindwing specification (23, 24) (Fig.
2B and fig. S6). The Ubx gene was in the only
genomic region where homologous ATAC-seq
peaks were enriched in the hindwing between
the species across all developmental time points
investigated (fig. S7). Although most peaks
had a similar accessibility pattern in both spe-
cies, we also found 36 species-specific ATAC-
seq peaks near Ubx. Sequence conservation
was generally high at these chromatin regions
(83.9%). Nevertheless, one peak at this genomic
Van Belleghem et al., Science 379, 1043–1049 (2023)
10 March 2023
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Key regulatory switch of red forewing band development. (A) Diver-
gence [Fixation index (FST)], phylogenetic association (tree weighting), and ATAC-seq
profile of red FW band near the optix gene. Blue shading indicates sequence that
is specific to H. erato compared with H. melpomene. Red triangles indicate CRISPR-
Cas9 excision targets. The solid red triangle indicates the target for which loss of
red scales and gain of yellow scales in the FM section were observed. (B) Zoom-in on
the only differentially accessible peak near optix associated with red forewing
band. Gray bars and colors indicate aligned nucleotides and single-nucleotide
polymorphisms (SNPs), respectively, whereas horizontal lines represent gaps.
Blue arrows indicate in silico TF binding sites specific to each haplotype. The dashed
lines indicate complete absence of homologous sequence. (C) CRISPR-Cas9 KO
phenotype of key regulatory switch. Because of the mosaicism of CRISPR-Cas9
mutants, the complete color pattern transition is represented by the composite
analysis of the individual mutant wing phenotypes. (D) Examples of geographic
morphs with yellow forewing band phenotypes. (E) Detail of phylogeny of
red (red circles) versus yellow (yellow circles) forewing band phenotypes at
the key regulatory optix switch. The dashed branch for the outgroup species and
H. melpomene indicates complete absence of homologous sequence.
region was completely specific to H. erato and
three peaks were shared between both species
but significantly differentially accessible.
Enrichment analysis of TF-binding motifs
in peaks differentially accessible between the
fore- and hindwing also showed differences
between H. erato and H. melpomene (Fig. 2C).
At the fifth-instar stage, we found similarly en-
riched TF-binding motifs for Ubx, extradenticle
(exd), hunchback (hb), bric a brac 1 (bab1),
Arrowhead (Awh), and Deformed (DFD) in the
forewing and Medea (Med) in the hindwing.
Overrepresented TF-binding motifs specific
to either H. erato or H. melpomene matched
more than 26 additional TF-binding sites (Fig.
2C). The pattern of differentially accessible
ATAC-seq peaks was corroborated by sim-
ilarly highly divergent patterns of differential
gene expression between the fore- and hind-
wings of H. erato and H. melpomene (Fig. 2D),
including the TFs with enriched binding mo-
tifs. These genes showed patterns of activa-
tion or suppression by nearby CREs with a
relative distribution that changes between
wings, development, and species (Fig. 2D).
Our ATAC-seq and gene expression data show
conservation of chromatin accessibility at
the Ubx locus but also a substantial number
of distinct chromatin peaks between the fore-
and hindwings of H. erato and H. melpomene.
Our results thus suggest that regulatory di-
vergence has evolved between the wings of
these comimetic species, which may poten-
tially include some functional changes at the
Ubx locus itself.
Low conservation in forewing patterning
between comimics
To study the developmental architecture (genes
and CREs) of the comimetic red forewing band
pattern, we collected and analyzed ATAC-
seq and RNA sequencing (RNA-seq) data for
Van Belleghem et al., Science 379, 1043–1049 (2023)
10 March 2023
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
forewing sections of 36-hour pupae of both
species (Fig. 3). With this approach, we tested
different possible combinations of wing pat-
terning (wing section and gradient-like ex-
pressed ATAC-seq peaks and genes; Fig. 3A).
As for the differences between whole wings,
the results from the three forewing sections
suggested a distinct architecture of pattern-
ing represented by the divergent chromatin
landscape and gene expression between the
comimics.
Using 50% reciprocal overlap, we identified,
across all of the different patterning tests, a
total of 2239 and 848 differentially accessible
ATAC-seq peaks between sections of the fore-
wing in H. erato and H. melpomene, respec-
tively. Only 3.3% of these were shared between
the two species. Similarly, when comparing
gene expression across wing sections, we iden-
tified 69 and 544 differentially expressed
genes in the forewing sections of H. erato and
H. melpomene, respectively, of which only two
(0.3%) had shared expression patterns (Fig.
3A). The shared ATAC-seq peaks had an av-
erage sequence similarity of 82.4% (SD = 12.6),
whereas the total of 2871 differentially acces-
sible ATAC-seq peaks specific to H. erato and
H. melpomene had an average sequence sim-
ilarity of 64.8% (SD = 25.1), with 3.0% explained
by 0% sequence conservation and 30.2% having
less than 50% sequence similarity in the alter-
native species (tables S5 and S6). Moreover,
our forewing section data provided a molecu-
lar opportunity to investigate the distinct WntA
KO behavior in the two comimics. Loss of func-
tion of the WntA morphogen resulted in the
expansion of red scales and optix expression
in the proximal part of the forewing only in
H. erato and not in H. melpomene (Fig. 3B)
(18). Differential accessibility analyses between
forewing sections within each species resulted
in 247 common chromatin peaks between
the proximal and medial forewing sections
in H. erato but zero between the proximal
and medial forewing sections in H. melpomene.
This result matches the different effect of WntA
CRISPR-Cas9 KOs in H. erato and H. melpomene
(fig. S8), thus reinforcing the existence of a dis-
tinct regulatory architecture of forewing prox-
imal black in the two butterflies.
Apart from Med, bab1, and hb, we found no
patterns of shared TF motif enrichment be-
tween H. erato and H. melpomene in wing
sections (Fig. 3C). We identified 12 TFs and
signaling molecules with nearby wing section–
specific ATAC-seq peaks or differential expres-
sion patterns in both species (tables S7 and
S8). These genes are known to be involved in
developmental processes that include cell
polarity, dorsoventral determination, and
proximodistal axis identity and may repre-
sent important developmental building blocks
around which gene regulatory networks have
diverged. However, these TFs had distinct
patterns of regulation or expression between
the species because they were identified in
different tissue comparisons. Genes such as
engrailed and distal-less eyespots (25) rep-
resent additional genes, apart from the major
known color pattern ones, that may be im-
plicated in Heliconius wing pattern develop-
ment and evolution. From these analyses, it
emerges that a distinct regulatory architec-
ture and gene expression of phenotypically
identical wing patterning has evolved since
their split ~11.1 million years ago.
A species-specific modular enhancer underlies
the independent evolution of mimicry
To further investigate the implication of the
observed widespread divergence in regulatory
architectures between the comimetic butter-
flies on adaptive evolution (summarized in
table S2), we studied the regulation of the
“red” color pattern gene, optix. Our experimen-
tal design allowed us to study black and red
sections of the wings during key developmen-
tal time points when optix expression is active
[12 to 60 hours after pupation (10)]. Our ex-
pectation was thus to identify an open chroma-
tin region that is significantly more accessible
in the red medial forewing (FM) section and
within the respective genomic association in-
terval (13).
Within a 320-kb associated interval around
the optix gene (13), we identified a total of
106 and 93 ATAC-seq peaks in H. erato and
H. melpomene, respectively (Fig. 4A and fig.
S9). Only one of these ATAC-seq peaks (155.5 kb
downstream of the optix gene) was within a
genetic yellow or red association interval hy-
pothesized to be a candidate region for red
forewing band regulation (13) and was signif-
icantly differentially accessible in the FM sec-
tion in H. erato (Fig. 4B). Functional validation
of this candidate CRE with CRISPR-Cas9 re-
sulted in a mutant phenotype in which scale
color–type changed from red to black or yel-
low in the FM of H. erato and did not affect red
color patterns on the ventral side of the wings
(efficiency equal to 65% of emerging adults)
(Fig. 4C., fig. S10, and table S8; see fig. S11 for
validation of excision mutations). Considering
that these mutants may be mosaic because
not all cells are being mutated in the wings, we
generated a composite of the yellow-forewing
mutant phenotypes, which resembled its sister
species H. himera and similar yellow-forewing
bands of other geographic H. erato morphs
(Fig. 4D). Excising two additional candidate
loci near optix, but outside the association
interval, did not affect the red band pheno-
type (Fig. 4A and table S9). By contrast, a re-
cent study proposed a pleiotropic architecture
of the red hindwing rays and basal forewing
pattern (referred to as “dennis”) and suggested
that modular cis-acting enhancers of the gene
optix that are sufficient to activate the pres-
ence of red rays and dennis patches likely do
not exist (17, 26). Our data demonstrate that a
modular CRE near optix is necessary to induce
a red band phenotype.
Phylogenetic analysis of the H. erato optix
CRE clustered H. erato populations or spe-
cies within its lineage according to yellow
or red color phenotypes (Fig. 4E). The se-
quence of this optix CRE was completely absent
in H. melpomene and in butterfly species more
distantly related to the Heliconius genus (Fig.
4E), thus suggesting its appearance at the origin
of the H. erato clade. In silico identification
of TF-binding sites, with the Drosophila data-
base as a reference, identified up to nine poten-
tial TF-binding sites specific to the red band
haplotype and 15 in the yellow haplotype (Fig.
4B). One of these TF-binding sites was for spalt-
related (salr), a transcriptional repressor that,
in Drosophila, mediates most decapentaplegic
(dpp) functions during the development of the
central part of the wing (27). These targets
represent candidates for upstream regula-
tion of optix and red pattern development in
Heliconius. These results reveal the evolution
of an adaptive optix CRE in H. erato, which
demonstrates a distinct regulatory integration
of a wing color pattern gene in the develop-
ment of convergent morphologies.
Conclusion
Morphological characters of an individual re-
quire the organization of spatial and tem-
poral gene expression (28). The integration
of these genes and their products over the
course of development defines a gene regu-
latory network (GRN) in which TFs interact
with CREs of their target genes. There is a
general consensus that gain and loss of CREs
occurs at substantially higher rates than that
of protein-coding genes (29). Despite the im-
portance of CRE changes in the evolution of
form and function in animals (30), the mag-
nitude of CRE evolution and the context and
evolutionary times necessary for CRE func-
tion to diverge or for new ones to evolve are
not well understood and may be faster than
generally described (31).
Convergent evolution provides an opportu-
nity for comparative studies of CRE evolution
and function during adaptive diversifica-
tion. In Heliconius butterflies, mimetic species
have independently evolved virtually identi-
cal wing color patterns through a shared set
of color patterning genes. This has led us to
assume that convergent wing color pattern
evolution in Heliconius was achieved through
a common developmental plan. However, this
view has begun to shift recently (16, 18, 32).
In this light, the highly divergent chromatin
landscapes that we report for H. erato and
H. melpomene suggest low conservation of
CREs in the development of mimetic wing
patterns. Aside from similarities at Ubx, many
Van Belleghem et al., Science 379, 1043–1049 (2023)
10 March 2023
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
genes were distinctively expressed, regulated, or
organized between H. erato and H. melpomene.
Thus, our findings provide a contrast between
the extremely conserved color pattern control
at the level of protein-coding genes, with low
similarity at the level of regulatory sequence.
We show how these highly divergent regula-
tory architectures play out in the evolution of
the red forewing band. A species-specific en-
hancer can switch red scales into yellow on the
forewing of H. erato (Fig. 4C). The composite
of the mosaic CRISPR-Cas9 mutants of the
optix CRE in H. erato resembled H. erato’s
sister species H. himera (33). This suggests that
the modular regulatory changes that underlie
wing color patterns also affect morphological
diversification in the early stages of speciation.
The lineage-specific nature of this CRE indi-
cates that independent genetic changes are
likely to be involved in species diversification
of the erato and melpomene clades.
Over the ~11.1 million years since the H. erato
and H. melpomene lineages split, they have
retained a shared toolkit of genes involved
in wing patterning (e.g., WntA, optix, cortex,
aristaless, distal-less, engrailed, antennape-
dia). However, they evolved nonhomologous
and quite distinct regulation of those genes
throughout development. Although the wing
patterns among mimics are highly similar,
they are not identical, with consistent minor
differences in pattern elements (fig. S12). These
phenotypic differences may be a direct result
of the fixation of independent developmental
alterations (e.g., CRE changes) in the two but-
terfly lineages. Thus, since their split, H. erato
and H. melpomene appear to have indepen-
dently accumulated distinct genetic variations
that modified an initially shared developmen-
tal system (20).
Our work highlights a high flexibility of evolu-
tionary trajectories that could be a widespread
property of any biological system. Whereas
neutral and selected genetic changes in repro-
ductively isolated species can create distinct
genomic environments, a developmental sys-
tem may thus be able to compensate for and
accommodate these context-specific effects
of genetic variation. This may, in turn, result
in apparently similar but ultimately distinc-
tive species-specific developmental solutions,
as demonstrated by a high evolutionary turn-
over of CREs.
RE FERENCES AND NOTES
J. R. True, E. S. Haag, Evol. Dev. 3, 109–119 (2001).
1.
2. Z. D. Blount, R. E. Lenski, J. B. Losos, Science 362, eaam5979
(2018).
3. S. M. Van Belleghem, J. J. Lewis, E. S. Rivera, R. Papa,
Curr. Opin. Genet. Dev. 69, 72–81 (2021).
4. C. F. Kratochwil et al., Science 362, 457–460 (2018).
5.
J. G. Roscito et al., Nat. Commun. 9, 4737 (2018).
6. X. Luo et al., Cell 184, 723–740.e21 (2021).
7. M. Uesaka, S. Kuratani, H. Takeda, N. Irie, Zoological Lett. 5, 33
8.
(2019).
J. Buenrostro, P. Giresi, L. Zaba, H. Chang, W. Greenleaf,
Nat. Methods 10, 1213–1218 (2013).
9. T. Yadav, J. P. Quivy, G. Almouzni, Science 361, 1332–1336
(2018).
10. R. D. Reed et al., Science 333, 1137–1141 (2011).
11. A. Martin et al., Proc. Natl. Acad. Sci. U.S.A. 109, 12632–12637
(2012).
12. N. J. Nadeau et al., Nature 534, 106–110 (2016).
13. S. M. Van Belleghem et al., Nat. Ecol. Evol. 1, 52 (2017).
14. L. Zhang, A. Mazo-Vargas, R. D. Reed, Proc. Natl. Acad. Sci. U.S.A.
114, 10707–10712 (2017).
15. A. Mazo-Vargas et al., Proc. Natl. Acad. Sci. U.S.A. 114,
10701–10706 (2017).
16. L. Livraghi et al., eLife 10, e68549 (2021).
17. J. J. Lewis et al., Proc. Natl. Acad. Sci. U.S.A. 116, 24174–24183
(2019).
18. C. Concha et al., Curr. Biol. 29, 3996–4009.e4 (2019).
19. S. M. Van Belleghem, P. A. Alicea Roman, H. Carbia Gutierrez,
B. A. Counterman, R. Papa, Proc. Biol. Sci. 287, 20201267
(2020).
20. A. Mazo-Vargas et al., Science 378, 304–308 (2022).
21. A. Ruggieri et al., Genome Res. 32, 1862–1875 (2022).
22. D. U. Gorkin et al., Nature 583, 744–751 (2020).
23. Y. Tomoyasu, Curr. Opin. Insect Sci. 19, 8–15 (2017).
24. A. Tendolkar et al., Front. Ecol. Evol. 9, 643661 (2021).
25. S. R. Schachat, J. C. Oliver, A. Monteiro, BMC Evol. Biol. 15, 20
(2015).
26. J. J. Lewis, S. M. Van Belleghem, Front. Ecol. Evol. 8, 261
(2020).
27. J. Sánchez et al., Biochem. J. 438, 437–445 (2011).
28. S. D. Hueber et al., Development 134, 381–392 (2007).
29. M. T. Weirauch et al., Cell 158, 1431–1443 (2014).
30. P. J. Wittkopp, G. Kalay, Nat. Rev. Genet. 13, 59–69 (2011).
31. P. L. Davidson et al., Nat. Ecol. Evol. 6, 1907–1920 (2022).
32. H. E. Bainbridge et al., J. Evol. Biol. 33, 1516–1529 (2020).
33. S. M. Van Belleghem et al., Evolution 75, 2251–2268 (2021).
34. S. M. Van Belleghem, StevenVB12/Genomics_ATAC:
Genomics_ATAC, version 1.1, Zenodo (2022); https://doi.org/
10.5281/zenodo.7368259.
AC KNOWLED GME NTS
We thank T. Mackay and F. Nijhout for providing valuable feedback
on the manuscript. Funding: This work was funded by NSF
EPSCoR RII Track-2 FEC (OIA 1736026) (R.P. and B.A.C.); NSF IOS
1656389 (R.P.); a Puerto Rico Science, Technology & Research
Trust catalyzer award (2020-00142) (S.M.V.B. and R.P.); and a
BBSRC grant (BB/R007500/1) (C.D.J.). S.M.V.B. and A.A.R. were
also supported by a National Institutes of Health 4 NIGMS 604
COBRE Phase 2 Award from the Center for Neuroplasticity at the
University of Puerto Rico (605 1P20GM103642). R.P. was also
supported by the Hispanic Alliance for Clinical and Translational
Research (Alliance) supported by the National Institute of
General Medical Sciences (NIGMS), National Institutes of Health
(U54GM133807). Additional funding to C.C. and W.O.M. was
provided by the Smithsonian Tropical Research Institute and the
Smithsonian’s Scholarly Studies program. For the support of
sequencing and computational resources, we thank the University
of Puerto Rico Sequencing and Genomics Facility INBRE Grant
P20 GM103475 from NIGMS, a component of the NIH, and the
Bioinformatics Research Core of the INBRE. Its contents are solely
the responsibility of the authors and do not necessarily represent
the official view of NIGMS or NIH. Author contributions: S.M.V.B.
and R.P. conceived the study and wrote the manuscript. S.M.V.B.,
A.A.R., L.L., and J.J.L. analyzed the data. C.C. performed CRISPR-Cas9
experiments and J.G.O. performed genotyping of mutants. L.L., L.H.,
E.S.R., J.J.H., and I.A.W. collected ATAC- and RNA-seq data. S.P. and
Y.O.-R. performed sequencing. R.R., J.J.L., C.D.J., B.A.C., and W.O.M.
provided materials and insights for data collection and analyses.
All authors commented on the final manuscript. Competing interests:
All authors declare no conflict of interest. Data and materials
availability: Custom codes, analyses pipelines, and images
of CRISPR-Cas9 KO phenotypes are available through the archived
Zenodo repository (34). ATAC- and RNA-seq sample GenBank
accession numbers are available in the supplementary materials.
License information: Copyright © 2023 the authors, some
rights reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government works.
https://www.science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade0004
Materials and Methods
Supplementary Text
Figs. S1 to S13
Tables S1 to S12
References (35–78)
MDAR Reproducibility Checklist
Submitted 18 July 2022; accepted 8 February 2023
10.1126/science.ade0004
Van Belleghem et al., Science 379, 1043–1049 (2023)
10 March 2023
7 of 7
|
10.1126_science.adj7576
|
RES EARCH
SPINTRONICS
Observation and control of hybrid
spin-wave–Meissner-current transport modes
M. Borst, P. H. Vree, A. Lowther, A. Teepe, S. Kurdi, I. Bertelli, B. G. Simon,
Y. M. Blanter, T. van der Sar*
Superconductors are materials with zero electrical resistivity and the ability to expel magnetic
fields, which is known as the Meissner effect. Their dissipationless diamagnetic response is central
to magnetic levitation and circuits such as quantum interference devices. In this work, we used
superconducting diamagnetism to shape the magnetic environment governing the transport of
spin waves—collective spin excitations in magnets that are promising on-chip signal carriers—in a
thin-film magnet. Using diamond-based magnetic imaging, we observed hybridized spin-wave–
Meissner-current transport modes with strongly altered, temperature-tunable wavelengths and
then demonstrated local control of spin-wave refraction using a focused laser. Our results demonstrate
the versatility of superconductor-manipulated spin-wave transport and have potential applications
in spin-wave gratings, filters, crystals, and cavities.
T he ability to control the transport of spins
and charges with metal electrodes is fun-
damental to information-processing
devices and an indispensable tool in
quantum and condensed-matter physics.
Although devices such as spin valves and tran-
sistors are based on the transport of uncorre-
lated particles (1), the excitations of magnetic
materials, known as spin waves, are emerging
as promising alternative information carriers (2).
These collective spin excitations provide new
opportunities for realizing analog or binary
device functionality based on their wave nature,
nonreciprocal transport properties, and low
intrinsic damping (3).
Control of spin-wave transport is possible by
heavy-metal electrodes that enable modulation
by means of the spin-Hall effect (4–6) or by
auxiliary magnetic materials that modify the
spin-wave spectrum (7, 8). However, metallic
gates can also introduce additional spin-wave
damping because of uncontrolled spin pumping
or spin-wave–induced eddy currents (6, 9, 10).
Furthermore, the diamagnetic response of
normal metals is dominated by ohmic resistance,
precluding effective stray-field control of the
spin-wave spectrum.
An attractive approach for strong, low-damping
spin-wave modulation is to use superconduct-
ing electrodes. Superconductors are materials
with zero electrical resistivity and a strong
diamagnetic response that enables the crea-
tion of magnetic shields, magnetic lenses, and
circuits such as quantum bits and quantum
interference devices (11, 12). Spin-wave spec-
troscopy measurements have demonstrated
that superconducting strips on magnetic films
can alter the spin-wave spectrum through the
backaction of induced currents (13) or the in-
Department of Quantum Nanoscience, Kavli Institute of
Nanoscience, Delft University of Technology, 2628 CJ Delft,
Netherlands.
*Corresponding author. Email: t.vandersar@tudelft.nl
teraction with Abrikosov vortices (14). Recently,
it was proposed to harness the diamagnetism
of a superconductor to create the spin-wave
equivalents of optical mirrors and cavities (15).
Being able to image and control spin waves as
they travel underneath superconducting elec-
trodes would enable insight into the nature of
the spin wave–superconductor interaction and
unlock opportunities to control the propaga-
tion, dispersion, and refraction of spin waves.
In this work, we developed, imaged, and
studied temperature-, field-, and laser-tunable
spin-wave transport enabled by a supercon-
ducting strip on a thin-film magnetic insu-
lator (Fig. 1A). We used magnetic resonance
imaging based on nitrogen-vacancy (NV) spins
in diamond (16–18) to study the spin waves
as they travelled underneath the optically
opaque superconductor.
Imaging hybrid spin-wave–Meissner-current
transport modes using spins in diamond
Our system consisted of a thin film of yttrium
iron garnet (YIG), a magnetic insulator with
low spin-wave damping (2), equipped with gold
microstrips for spin-wave excitation and a
molybdenum-rhenium superconducting strip
for spin-wave modulation (Fig. 1A). To image
the spin waves, we placed a diamond mem-
brane that contained a thin layer of NV sensor
spins on top of the sample (Fig. 1A and fig. S1)
(18, 19). These spins detect the spin waves by
their microwave magnetic stray fields, which
enables imaging through optically opaque
materials (10). The sample was embedded in a
variable temperature cryostat with a base tem-
perature of 5.5 K and free-space optical access
to read out the NV sensor spins.
NV centers are atomic defects in the dia-
mond carbon lattice with an S = 1 electron
spin (16). The sensitivity of the NV spin to
magnetic fields, combined with its optical spin
readout and excellent spin coherence, has en-
abled widespread sensing applications in fields
ranging from condensed-matter science to ge-
ology and biophysics (20–22). Here we used
the sensitivity of the NV spins to microwave
magnetic fields to image the spin waves in
our YIG film (17, 18, 23, 24). When resonant
with an NV electron spin resonance (ESR) fre-
quency, the stray field of the spin waves drives
transitions between the NV spin states, which
we detected through the spin-dependent NV
photoluminescence under green-laser excitation.
We applied a magnetic bias field to tune the
NV ESR frequency into resonance with spin
waves of different spin wavelengths (Fig. 1B).
By orienting the field along one of the four
possible crystallographic NV orientations (Fig.
1A), we split the ESR frequency of this “field-
aligned” NV ensemble (f1) off from that of the
three other NV ensembles (f2) as shown in the
optically detected resonance spectrum of Fig.
1C. Alternatively, we applied the field in plane
along ^z to enable measurements at different
frequencies at a given magnetic field. Because
the applied magnetic fields are much smaller
than the YIG saturation magnetization, the
YIG magnetization lies predominantly in-plane
along ^z for both field orientations (Damon-
Eshbach geometry).
To demonstrate the spin-wave modulation
capabilities of our superconductor, we imaged
the spin-wave transport above and below the
molybdenum-rhenium (MoRe) superconduct-
ing transition temperature Tc = 8.7 K (Fig. 2, A
to D). We generated NV-resonant spin waves
with wavevector k ¼ k^y by applying a micro-
wave current at NV frequency f1 to the gold
microstrip located just left outside of the im-
aging area. The interference between the mi-
crowave magnetic stray field generated by
these spin waves and the direct microstrip
field leads to a spatial standing-wave modula-
tion of the NV ESR contrast (17, 18). Crucial for
our measurements, this interference effect en-
ables a straightforward extraction of the spin
wavelength. The spatial map of the ESR con-
trast C1 at T = 10.7 K (above Tc) shows spin
waves traveling toward and then underneath
the MoRe strip without a change in wavelength
(Fig. 2B). By contrast, the spin-wavelength in-
creased almost twofold when the strip was
cooled into its superconducting state at T =
5.5 K (Fig. 2C). Averaging the maps along ^z
(Fig. 2E) highlights the spatial homogeneity of
the wavelength change.
We explained the superconductor-induced
change of the spin wavelength by developing
an analytical expression for the spin-wave dis-
persion in a magnet-superconductor thin-film
hybrid. In this model, building on the formalism
developed in (15), the spin waves induce AC
Meissner currents that are governed by the
London penetration depth lL of the super-
conductor. These currents, in turn, generate
a magnetic field that acts back on the spin waves.
Borst et al., Science 382, 430–434 (2023)
27 October 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
A
X
B
3
2
1
)
z
H
G
(
f
0
0
C
1
0
L
P
/
L
P
0.98
f 2
f 1
10-2
10-1
1
10
20
30
B
z
(mT)
f 1
f 2
C2
C1
2.6
f
sw
2.8
(GHz)
3
Fig. 1. Magnetic resonance imaging of hybridized spin-wave–Meissner-
current transport modes. (A) Overview of the experiment. A gold (Au)
microstrip excites spin waves with wavevector k ¼ k^y in a 245-nm-thick film
of yttrium iron garnet (YIG). The spin waves travel toward a molybdenum
rhenium (MoRe) superconducting strip (width W = 30 mm, thickness t =
140 nm) where their stray fields induce Meissner currents that act back on
the spin waves, shifting their wave number to kYIG/SC < kYIG (SC, superconductor).
We imaged the waves underneath and next to the superconductor by their
microwave magnetic stray fields using a ~20-nm-thick layer of nitrogen-
vacancy (NV) spins implanted at ~0.07 mm below the bottom surface of a
100 × 100 × 5 mm3 diamond membrane placed on top of the sample. A
magnetic field BNV ¼ Bx^x þ Bz^z applied at q = 54° with respect to the
x axis yields an in-plane YIG magnetization along ^z for the small fields applied
and directional Damon-Eshbach spin-wave excitation. GGG, gadolinium gallium
garnet, SW, spin wave. (Inset) Optical micrograph of NV diamond and Au and
MoRe strips on the YIG. Scale bar, 30 mm. Au thickness, 200 nm. (B) YIG
dispersion (color map) and NV electron spin resonance (ESR) frequencies
(red lines) as a function of the in-plane field component Bz = BNVcosq. f1 and
f2 denote the ESR frequencies of NV spins with zero-field quantization axis
aligned or misaligned with BNV, respectively. The intersection of the ESR
frequencies with the spin-wave dispersion sets the detectable spin-wave
numbers k. (C) Optically detected NV ESR spectrum at Bz = 10 mT, at a location
denoted by the red cross in the inset of (A). The ESR contrast C1 or C2,
where the subscript identifies the field-aligned or misaligned NV ensemble,
respectively, results from interference between the microstrip field and spin-wave
field, enabling spatial mapping of the spin-wave fronts.
Fig. 2. Magnetic resonance imaging of spin waves
above and below the superconducting transition
temperature. (A) Spatial map of the NV photo-
luminescence PL0 in the absence of microwaves, showing
the MoRe strip (between the vertical dashes). Scale
bar, 10 mm. (B and C) Spatial maps of the NV electron
spin resonance contrast C1 above (B) and below (C) the
superconducting transition temperature of Tc = 8.7 K
at T = 10.7 and 5.5 K, respectively. The Au, which
excites spin waves, is located just outside the left
edge of the imaged area. Above Tc, the wavelength is
unaffected by the MoRe strip; below Tc, it is lengthened.
(D) DC resistance R of the MoRe strip as a function of
temperature T, with markers indicating the resistance
of the film during the measurements of (B) triangle and
(C) square. (E) Data from (B) and (C) averaged over
the z direction, with the MoRe strip indicated by yellow
shading. (F) Calculated spin-wave dispersion fYIG(k)
for bare YIG, and fYIG/SC(k, lL) for YIG, covered by a
superconducting film with London penetration depth
lL = 400 nm. The superconductor shifts the dispersion
upwards by fSC(k, lL), which manifests as a reduction in the
wave number at NV frequency f1 from k 1ð Þ
YIG=SC,
indicated by the dashed lines.
YIG to k 1ð Þ
D
200
)
(
R
0
A
PL0 (10 7s-1)
1
0
2
B
C1 (%)
2
C
4
0
C1 (%)
2
0
4
T = 10.7 K
T = 10.7 K
T = 5.5 K
YIG YIG/MoRe
YIG
z
y
z
y
E
4
+0.3
)
%
(
g
v
1a
2
C
0
5
9
13
T (K)
y
f 1
z
F
)
z
H
G
(
f
3
2
1
10.7 K
5.5 K
k(1)
YIG/SC k(1)
YIG
f YIG
f YIG/SC
-20
0
20
40
60
0
0.1
0.2
0.3
0.4
Borst et al., Science 382, 430–434 (2023)
27 October 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
By integrating this field self-consistently into
the Landau-Lifshitz-Gilbert equation, we found
that the spin-wave dispersion shifts upwards
in frequency as
ð
fYIG=SC k; lL
Þ ¼ fYIG kð Þ þ fSC k; lL
ð
Þ
ð1Þ
where fYIG(k) is the bare-YIG spin-wave dis-
persion (SI) and
ð
fSC k; lL
Þ ≈gm0
Msktr
1 (cid:2) e(cid:2)2h=lL
Þ2 (cid:2) klL (cid:2) 1
ð
ð
klL þ 1
Þ2e(cid:2)2h=lL
ð2Þ
is the superconductor-induced shift [supple-
mentary text, section S1 (19)]. Here, Ms is the
YIG saturation magnetization, t = 245 nm is
the YIG thickness, h = 140 nm is the supercon-
YIG
12
8
4
A
)
T
m
(
z
B
B
)
T
m
(
z
B
YIG/MoRe
3
C
0.35
0.3
0.25
)
1
-
0.2
k
0.15
0.1
0.05
)
%
(
1
C
0
2.6
)
%
(
2
C
0
YIG
12
YIG/MoRe
8
4
-20
-10
0
10
20
y
fit Ms
fit L
k(1)
YIG/SC
k(2)
YIG/SC
k(1)
YIG
k(2)
YIG
0.35
0.3
0.25
)
1
-
0.2
0.15
k
0.1
0.05
0
10
Bz (mT)
2.4 2.6 2.8
f (GHz)
Fig. 3. Magnetic field dependence of the spin-wave dispersion in the magnet-superconductor hybrid.
(A and B) Spatial line traces of the NV ESR contrasts C1 and C2 as a function of magnetic field Bz. Spin
waves excited in the bare YIG region (y < 0) by the left Au microstrip (outside the imaged area) travel toward
and then underneath the superconducting strip (y > 0), changing their wavelength. Interference with
secondary spin waves excited at the MoRe strip edge (at y = 0) due to inductive coupling between Au
and MoRe strips yields a beating pattern along Bz for y > 0. T = 5.5 K. The drive frequency is adjusted at each
Bz to maintain resonance with the NV ESR transitions. (C) Spin-wave number as a function of field (left)
and frequency (right), extracted from the data in (A) and (B) by means of Fourier transformation (fig. S2).
The k(i) are the wave numbers measured with the field-aligned (i = 1) and misaligned (i = 2) NV ensembles.
The error bars (indicated by shading) are determined by the inverse of the spatial sampling range in the
y direction. We determined the saturation magnetization Ms by fitting the data in the bare YIG region, and the
London penetration depth lL by fitting the data in the YIG-MoRe region using our YIG/SC model
[supplementary text, section 2 (19)].
A
10
)
K
(
T
8
3.3 mT
6
10
4.9 mT
)
K
(
T
8
6
0
C *
1
(a.u.)
B
0.3
0.2
0.1
Tc
0
5
10
Bz (mT)
C
0.8
data
theory
0.6
0.4
0.2
6
7
T (K)
8
6.5 mT
9.8 mT
10
y
20
0
10
y
20
6
8
T (K)
10
Fig. 4. Temperature tunability of the hybrid spin-wave–Meissner-current dispersion and extraction
of the London penetration depth. (A) Spatial line traces of the NV ESR contrast C(cid:3)
region as a function of temperature, showing the continuous change of the spin wavelength underneath
the MoRe strip, for different in-plane magnetic fields Bz. The data are linearly detrended along y. Above
the superconducting phase transition there is no temperature dependence of the wavelength, indicating the
absence of the Meissner effect. A.u., arbitrary units. (B) Spin-wave numbers k extracted from data in (A)
and from additional data in fig. S4 as a function of temperature. The colors indicate the different magnetic
field values Bz. (C) London penetration depth lL of the MoRe film as a function of temperature, extracted
from the data in (B) through our YIG/SC model. The black line represents the fit of the temperature
dependence of lL(T) from which we extracted Tc = 8.7 K and l0
values of Bz as in (B).
L = 380 nm. The colors indicate the different
1 across the YIG-MoRe
ductor thickness, g = 28 GHz/T is the electron
gyromagnetic ratio, m0 is the vacuum perme-
ability, and r is a dimensionless factor asso-
ciated with the YIG thickness and spin-wave
ellipticity. The approximation holds when the
kinetic inductance dominates the impedance, as
is the case for our superconducting strip [sup-
plementary text, section S1 (19)], and when
≪ 1. A more general expression is given
k2l2
L
in the supplementary text (19). The dispersion
shift fSC(k,lL) is maximal when lL → 0, in which
case the superconductor perfectly screens the
spin-wave stray field. This limit was analyzed
in (25, 26) by considering a magnetic film cov-
ered by a “perfect metal” as defined by a per-
fect magnetic field screening. Indeed, when we
let lL → 0, the shift calculated by our model
approached the shift predicted in (19, 25, 26).
The calculated bare YIG and hybridized YIG-
MoRe spin-wave dispersions are compared in
Fig. 2F. The upwards frequency shift under-
neath the superconductor manifested as a re-
duction in wave number of the NV-resonant spin
waves detected in our experiments (Fig. 2E).
We observed that the ESR contrast just to the
right of the MoRe strip was higher than that in
the MoRe-strip region (Fig. 2C), consistent with
the screening of the spin-wave field by the
Meissner currents. In addition, we observed that
the ESR contrast just to the right of the MoRe
strip exceeded that just to the left of it (Fig. 2E),
which indicated the excitation of additional, sec-
ondary spin waves by the MoRe strip itself. This
enhanced excitation of secondary spin waves is
presumably caused by an additional microwave
current in the MoRe strip that is excited through
the direct geometric inductive coupling with the
gold microstrip when the MoRe impedance
changes as it is cooled below Tc.
Temperature- and field dependence of the
spin-wave dispersion and extraction of the
London penetration depth
We characterized the magnetic-field depen-
dence of the spin-wave dispersion underneath
the superconductor and used it to extract the
London penetration depth lL at the T = 5.5 K
base temperature of our cryostat. Spatial line
traces of the NV ESR contrast across the strip
shows the dependence of the spin wave-
length on the applied magnetic field for the
field-aligned (Fig. 3A) and misaligned (Fig. 3B)
NV ensembles. In both measurements, we ad-
justed the drive frequency at each magnetic
field to maintain resonance with the NV ESR
frequency. We extracted the spin-wave numbers
in the bare YIG and YIG-MoRe regions sepa-
rately by Fourier transformation (fig. S2) and
plotted these as a function of field and frequen-
cy in Fig. 3, C and D. A similar measurement
with the bias field applied in plane along ^z
shows that Meissner screening of the bias field
does not play a significant role in the wave-
length shift (fig. S3).
Borst et al., Science 382, 430–434 (2023)
27 October 2023
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
A
x
B
z
laser
hot spot
NV diamond
MoRe
YIG
GGG
y
kYIG/SC
C
laser
kYIG
kYIG/SC
D
z
z
y
y
y
laser
laser
2
3
4
PL0 (10 6 s-1)
1
2
3
on (%)
C1
1
-1
0
on - C1
C1
off (%)
Fig. 5. Laser-induced spin-wave refraction at
target locations. (A) Schematic illustration of a
laser-induced scattering spot. By shining an
auxiliary 594-nm laser on the sample, we created
a hot spot in the MoRe strip that locally altered
the effective refractive index governing the
spin-wave propagation. (B) Scanning confocal
microscope images of the NV photoluminescence
at T = 5.5 K, with the auxiliary laser focused onto
the MoRe strip at three different locations indicated
by the arrows. Scale bar, 10 mm. (C) Spatial maps
of the NV ESR contrast with auxiliary laser turned
on (Con
1 ) showing spin waves in the YIG-MoRe region
that scatter on the laser spot. (D) Background-
subtracted ESR contrast highlighting the laser-
induced spin-wave scattering obtained by subtracting
the ESR contrast with the auxiliary laser turned
off Coff
1
from the measurements in (C). Bz = 3.3 mT.
From the field dependence of the extracted
spin-wave numbers in the bare YIG region, we
extracted the YIG saturation magnetization
Ms = 194(1) kA/m [supplementary text, sec-
tion S2 (19)] in agreement with previous low-
temperature measurements (27). We then
used Ms as a fixed parameter to fit the field
dependence of the spin-wave numbers under-
neath the superconducting strip using the hy-
bridized YIG-MoRe spin-wave dispersion (Eq. 1).
From this fit, we extracted a London penetration
depth lL = 405(10) nm at T = 5.5 K, which agrees
well with static-field nano–superconducting quan-
tum interference device measurements (28).
The temperature dependence of the London
penetration depth provides a powerful tool for
tuning the spin wavelength. To demonstrate
this, we imaged the spin waves in the YIG-
MoRe region while sweeping through Tc at
different magnetic fields (Fig. 4A and fig. S4).
The extracted spin-wave number k is shown in
Fig. 4B, with the color indicating the in-plane
component of the magnetic field. We observed
that k changes continuously with temperature
over the superconducting phase transition in
the YIG-MoRe region and remains unchanged
in the bare YIG region (fig. S5). We did not
observe global heating of the superconductor
due to our excitation laser (fig. S6). Using our
model, we extracted the London penetration
depth lL(T) for every observed value of k (Fig. 4C)
[supplementary text, section S2 (19)]. We found
that almost all data collapse onto a single curve
TcÞ4(cid:4)(cid:2)1=2 (29),
described by lL ¼ l0
½
L 1 (cid:2) ðT =
¼ 380nm. The excep-
with Tc = 8.7 K and l0
L
tions occurred when the spin wavelength
lSW = 2p/k became comparable to the width
of the MoRe strip. Here, our approximation
of the superconducting strip by an infinite film
breaks down. These results highlight that
imaging the hybridized spin-wave–Meissner-
current transport modes is a powerful tool for
extracting the temperature dependence of the
London penetration depth.
Local control of spin-wave transport by
laser-induced spin-wave refraction
Thus far, we have demonstrated dispersion
engineering through global control of temper-
ature and magnetic field. We now show that
the creation of a hot spot in the superconduc-
tor with a focused laser enables local manip-
ulation of the spin-wave transport by tuning
the effective refractive index (Fig. 5A). To do
so, we coupled an auxiliary, orange laser into
our setup and focused it at target sites on the
superconductor (Fig. 5B). The laser spot is
visible through the locally enhanced NV photo-
luminescence. Spatial measurements of the NV
ESR contrast C on
1 with the auxiliary laser on
(Fig. 5C) show the spin-wave scattering pat-
terns induced by the local hot spot. The reduc-
tion in amplitude behind the hot spot indicates
destructive interference between the scattered
and incident spin waves. Subtracting a refer-
ence measurement with the auxiliary laser
turned off (Fig. 5D) highlights the angular pro-
file of the scattered spin-wave patterns.
The characteristic “caustic” angles observed
in these scattered patterns (dashed lines, Fig.
5D) result from the highly anisotropic dipolar
spin-wave dispersion (30). Tracing the patterns
to their origin shows that the scattering site is
tightly confined to the laser location. Presum-
ably, the laser locally breaks the superconduc-
tivity, inducing a local change in the magnetic
environment seen by the spin waves, leading
to local spin-wave refraction akin to defect-
controlled spin-wave scattering (30). The abil-
ity to optically induce spin-wave refraction at
target sites could be used to create devices
such as gratings or magnonic crystals (31) and
enable spin-wave manipulation through opti-
cal switching of flux-focusing regions in the
superconducting strip.
Conclusions
i
h
≈
(cid:1) (cid:3)
þ 0:5
ln lL
x
We demonstrated local measurements of hy-
bridized spin-wave–Meissner-current transport
modes in a magnetic thin film equipped with
a superconducting gate. The wavelength was
tunable by temperature and field, enabling
efficient phase-shifting of the spin-wavefronts
and a notable in situ visualization and quanti-
tative extraction of the London penetration depth
as a function of temperature. Because MoRe
is a type-2 superconductor with an estimated
lower critical field of Hc1 ¼ F0
4pl2
L
4 mT at T ≈ 5 K (32, 33), where x ≈ 0.01 mm is
the coherence length (34) and F0 the super-
conducting flux quantum, Abrikosov vortices
were expected in measurements such as those
in Fig. 3B with a few-mT out-of-plane field com-
ponent. However, we did not identify vortex-
related effects in these measurements, which
look qualitatively similar to those with purely
in-plane field (fig. S3). Presumably, the pres-
ence and location of vortices is strongly in-
fluenced by our focused excitation laser, as
highlighted by recent magneto-optical (35) and
wide-field NV-imaging experiments (36). In
particular, (35) showed that vortices can be
annihilated by the laser through local heating
above Tc or pushed by the laser to new pinning
sites or border regions of the superconductor.
The presented microwave magnetic imaging
of the spin-wave transport modes in a YIG-
MoRe heterostructure shows the versatility of
superconducting gates for spin-wave manipu-
lation, enables determining the temperature-
dependent London penetration depth, and
opens new opportunities for creating wave-
based circuit elements such as filters, mirrors,
and cavities.
REFERENCES AND NOTES
1. S. A. Wolf et al., Science 294, 1488–1495 (2001).
2. A. V. Chumak, V. I. Vasyuchka, A. A. Serga, B. Hillebrands,
Nat. Phys. 11, 453–461 (2015).
3. A. Barman et al., J. Phys. Condens. Matter 33, 413001 (2021).
4.
J. Sinova, S. O. Valenzuela, J. Wunderlich, C. H. Back,
T. Jungwirth, Rev. Mod. Phys. 87, 1213–1260 (2015).
5. L. J. Cornelissen, J. Liu, R. A. Duine, J. Ben Youssef,
B. J. Van Wees, Nat. Phys. 11, 1022–1026 (2015).
6. A. Hamadeh et al., Phys. Rev. Lett. 113, 197203 (2014).
7. T. Yu, C. Liu, H. Yu, Y. M. Blanter, G. E. W. Bauer, Phys. Rev. B
99, 134424 (2019).
J. Chen et al., Phys. Rev. Lett. 120, 217202 (2018).
8.
9. Y. Sun et al., Phys. Rev. Lett. 111, 106601 (2013).
10. I. Bertelli et al., Adv. Quantum Technol. 4, 2100094 (2021).
11. Z. Y. Zhang et al., Supercond. Sci. Technol. 25, 025009 (2012).
12. A. Blais, A. L. Grimsmo, S. M. Girvin, A. Wallraff, Rev. Mod. Phys. 93,
025005 (2021).
Borst et al., Science 382, 430–434 (2023)
27 October 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
I. A. Golovchanskiy et al., J. Appl. Phys. 127, 093903 (2020).
13.
14. O. V. Dobrovolskiy et al., Nat. Phys. 15, 477–482 (2019).
15. T. Yu, G. E. W. Bauer, Phys. Rev. Lett. 129, 117201 (2022).
16. L. Rondin et al., Rep. Prog. Phys. 77, 056503 (2014).
17. T. X. Zhou et al., Proc. Natl. Acad. Sci. U.S.A. 118, e2019473118
(2021).
I. Bertelli et al., Sci. Adv. 6, eabd3556 (2020).
18.
19. Materials and methods, and supplementary text, figures and
references are available as supplementary materials.
20. F. Casola, T. Van Der Sar, A. Yacoby, Nat. Rev. Mater. 3, 17088
(2018).
21. R. Schirhagl, K. Chang, M. Loretz, C. L. Degen, Annu. Rev.
Phys. Chem. 65, 83–105 (2014).
22. D. R. Glenn et al., Geochem. Geophys. Geosyst. 18, 3254–3267
(2017).
29. M. Tinkham, Introduction to superconductivity (Dover Publications,
2nd edition, 2004).
30. R. Gieniusz et al., Appl. Phys. Lett. 102, 102409 (2013).
31. A. V. Chumak, A. A. Serga, B. Hillebrands, J. Phys. D Appl. Phys.
50, 244001 (2017).
32. W.-K. Kwok et al., Rep. Prog. Phys. 79, 116501 (2016).
33. V. E. Calado et al., Nat. Nanotechnol. 10, 761–764 (2015).
34. K. Makise et al., Mater. Res. Express 5, 096406 (2018).
35. I. S. Veshchunov et al., Nat. Commun. 7, 12801 (2016).
36. S. E. Lillie et al., Nano Lett. 20, 1855–1861 (2020).
37. M. Borst et al., Data for: Observation and control of hybrid
spin-wave–Meissner-current transport modes, version 1,
Zenodo (2023).
placement: A.L. and S.K.; Constructing experimental setup: M.B. and
A.L.; Measurements and software: M.B., P.H.V., A.L., and I.B.; Data
analysis: M.B., P.H.V., and A.T.; Theory: T.S. and Y.M.B.; Figures: M.B.
and P.H.V.; Funding acquisition: T.S.; Project administration: M.B.
and T.S.; Supervision: T.S. and M.B.; Writing – original draft:
M.B., T.S., and P.H.V.; Writing – review and editing: M.B., T.S., P.H.V.,
and Y.M.B. Competing interests: The authors declare that they have
no competing interests. Data and materials availability: All
data presented in this work are available at Zenodo (37). License
information: Copyright © 2023 the authors, some rights reserved;
exclusive licensee American Association for the Advancement of
Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse
ACKN OWLED GMEN TS
23. P. Andrich et al., npj Quantum Inf. 3, 28 (2017).
24. C. S. Wolfe et al., Phys. Rev. B Condens. Matter Mater. Phys. 89,
180406 (2014).
25. A. G. Gurevich, G. A. Melkov, Magnetization Oscillations and
Waves (CRC Press, 1996).
26. S. R. Seshadri, Proc. IEEE 58, 506–507 (1970).
27. S. Knauer et al., J. Appl. Phys. 133, 143905 (2023).
28. A. G. Shishkin et al., Supercond. Sci. Technol. 33, 065005 (2020).
We thank A. F. Otte for commenting on the manuscript and
T. Yu and G.E.W. Bauer for discussions. Funding: This work was
supported by the Dutch Research Council (NWO) under awards
VI.Vidi.193.077, NGF.1582.22.018, and OCENW.XL21.XL21.058. This
work was also supported by the Dutch National Growth Fund (NGF)
as part of the Quantum Delta NL programme. Author contributions:
Conceptualization: T.S. and M.B.; Sample design and fabrication:
M.B.; Diamond membrane fabrication: B.S. and M.B.; Membrane
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adj7576
Materials and Methods
Supplementary Text
Figs. S1 to S6
References (38–42)
Submitted 14 July 2023; accepted 18 September 2023
10.1126/science.adj7576
Borst et al., Science 382, 430–434 (2023)
27 October 2023
5 of 5
|
10.1126_science.add8737
|
Cite as: C. Fischer et al., Science
10.1126/science.adg2821 (2022).
First release: 20 December 2022
www.science.org
(Page numbers not final at time of first release) 1
RESEARCH ARTICLES
Cite as: C. Fischer et al., Science
10.1126/science.add8737 (2022).
Gradual emergence followed by exponential spread
of the SARS-CoV-2 Omicron variant in Africa
Carlo Fischer1†, Tongai Gibson Maponga2†, Anges Yadouleton3†, Nuro Abílio4, Emmanuel Aboce5, Praise
Adewumi3, Pedro Afonso6, Jewelna Akorli7, Soa Fy Andriamandimby8, Latifa Anga9, Yvonne Ashong7, Mohamed
Amine Beloufa10, Aicha Bensalem10, Richard Birtles11,12, Anicet Luc Magloire Boumba13,14, Freddie Bwanga5,15,
Mike Chaponda16, Paradzai Chibukira17, R. Matthew Chico18, Justin Chileshe16, Gershom Chongwe16, Assana
Cissé19, Umberto D’Alessandro20, Xavier Nicolas de Lamballerie21, Joana F. M. de Morais6, Fawzi Derrar10,
Ndongo Dia22, Youssouf Diarra23, Lassina Doumbia23, Christian Drosten1,24, Philippe Dussart8, Richard
Echodu11, Yannik Eggers25,26, Abdelmajid Eloualid9, Ousmane Faye22, Torsten Feldt25,26, Anna Frühauf1, Afiwa
Halatoko27, Pauliana-Vanessa Ilouga28, Nalia Ismael4, Ronan Jambou29, Sheikh Jarju20, Antje Kamprad1, Ben
Katowa30,31, John Kayiwa32, Leonard King’wara33, Ousmane Koita23, Vincent Lacoste8, Adamou Lagare29, Olfert
Landt34, Sonia Etenna Lekana-Douki35, Jean-Bernard Lekana-Douki35, Etuhole Iipumbu36, Hugues Loemba37,14,
Julius Lutwama32, Santou Mamadou29, Issaka Maman27, Brendon Manyisa17, Pedro A. Martinez6, Japhet
Matoba30,31, Lusia Mhuulu36, Andres Moreira-Soto1, Judy Mwangi11,12, Nadine N´dilimabaka35, Charity Angella
Nassuna32, Mamadou Ousmane Ndiath20, Emmanuel Nepolo36, Richard Njouom28, Jalal Nourlil9, Steven Ger
Nyanjom38, Eddy Okoth Odari38, Alfred Okeng5, Jean Bienvenue Ouoba19, Michael Owusu39, Irene Owusu
Donkor7, Karabo Kristen Phadu2, Richard Odame Phillips39, Wolfgang Preiser2,40, Vurayai Ruhanya17, Fortune
Salah27, Sourakatou Salifou41, Amadou Alpha Sall22, Augustina Angelina Sylverken39,42, Paul Alain Tagnouokam-
Ngoupo28, Zekiba Tarnagda19, Francis Olivier Tchikaya14, Tafese Beyene Tufa25,26, Jan Felix Drexler1,24*
1Charité – Universitätsmedizin Berlin, corporate member of Freie Universität Berlin and Humboldt Universität zu Berlin, Institute of Virology, Berlin, Germany. 2Division of
Medical Virology, Stellenbosch University Faculty of Medicine and Health Sciences, Cape Town, South Africa. 3Laboratoire dés fievres hemorragiques virales de Cotonou,
Akpakpa, Benin. 4Instituto Nacional de Saúde, Maputo, Mozambique. 5MBN Clinical Laboratories, Kampala, Uganda. 6Instituto Nacional de Investigação em Saúde (INIS),
Luanda, Angola. 7Noguchi Memorial Institute for Medical Research, University of Ghana, Legon, Ghana. 8Institut Pasteur de Madagascar, Antananarivo, Madagascar.
9Institut Pasteur du Maroc, Casablanca, Morocco. 10Institut Pasteur of Algeria, National Influenza Centre, Sidi-Fredj, Algeria. 11Gulu University Multifunctional Research
Laboratories, Gulu, Uganda. 12School of Science, Engineering and Environment, University of Salford, Salford, UK. 13Faculty of Health Sciences, Marien Ngouabi University,
Pointe-Noire, Congo. 14Molecular Diagnostic Laboratory HDL, Pointe-Noire, Congo. 15Makerere University College of Health Sciences, Kampala, Uganda. 16Tropical Diseases
Research Centre, Ndola Teaching Hospital, Ndola, Zambia. 17National Virology Laboratory, Faculty of Medicine and Health Sciences, University of Zimbabwe, Avondale,
Zimbabwe. 18London School of Hygiene and Tropical Medicine, London, UK. 19Laboratoire National de Référence-Grippes, Ouagadougou, Burkina Faso. 20Medical Research
Council Unit at London School of Hygiene and Tropical Medicine, Banjul, Gambia. 21Unité des Virus Émergents, Aix Marseille Université, Marseille, France. 22Institut Pasteur
de Dakar (IPD), Dakar, Senegal. 23Université des Sciences, des Techniques et des Technologies de Bamako (USTTB), Bamako, Mali. 24German Centre for Infection Research
(DZIF), associated Partner Charité-Universitätsmedizin Berlin, Berlin, Germany. 25Hirsch Institute of Tropical Medicine, Asella, Ethiopia. 26Department of Gastroenterology,
Hepatology and Infectious Diseases, University Hospital Düsseldorf, Heinrich Heine University Düsseldorf, Düsseldorf, Germany. 27Institut National d’Hygiène, Lomé, Togo.
28Centre Pasteur du Cameroun, Yaounde, Cameroon. 29Centre de Recherche Médicale et Sanitaire (CERMES), Niamey, Niger. 30Macha Research Trust, Choma, Zambia.
31School of Veterinary Medicine, University of Zambia, Lusaka, Zambia. 32Uganda Virus Research Institute, Entebbe, Uganda. 33National Public Health Reference Laboratory,
Ministry of Health, Nairobi, Kenya. 34TiB-Molbiol GmbH, Berlin, Germany. 35Centre Interdisciplinaire de Recherches Médicales de Franceville (CIRMF), Franceville, Gabon.
36School of Medicine, University of Namibia, Windhoek, Namibia. 37Faculty of Medicine, University of Ottawa, Ottawa, Ontario, Canada. 38School of Biomedical Sciences,
Jomo Kenyatta University of Agriculture and Technology, Nairobi, Kenya. 39Kumasi Centre for Collaborative Research in Tropical Medicine (KCCR), Kwame Nkrumah
University of Science and Technology (KNUST), Kumasi, Ghana. 40National Health Laboratory Service Tygerberg Business Unit, Cape Town, South Africa. 41Ministère de la
Santé, Akpakpa, Benin. 42Department of Theoretical and Applied Biology, KNUST, Kumasi, Ghana.
*Corresponding author. Email: felix.drexler@charite.de
The geographic and evolutionary origins of the SARS-CoV-2 Omicron variant (BA.1), which was first
detected mid-November 2021 in Southern Africa, remain unknown. We tested 13,097 COVID-19 patients
sampled between mid-2021 to early 2022 from 22 African countries for BA.1 by real-time RT-PCR. By
November-December 2021, BA.1 had replaced the Delta variant in all African sub-regions following a South-
North gradient, with a peak Rt of 4.1. Polymerase chain reaction and near-full genome sequencing data
revealed genetically diverse Omicron ancestors already existed across Africa by August 2021. Mutations,
altering viral tropism, replication and immune escape, gradually accumulated in the spike gene. Omicron
ancestors were therefore present in several African countries months before Omicron dominated
transmission. These data also indicate that travel bans are ineffective in the face of undetected and
widespread infection.
First release: 1 December 2022
science.org
(Page numbers not final at time of first release) 1
Retracted 20 December 2022. See Retraction.
By September 2022, over 6.5 million persons had died from
coronavirus disease 2019 (COVID-19) (1). The true number of
infections is probably much higher, particularly in Africa
where the diagnostic capacities are low (2, 3). In Africa, the
World Health Organization (WHO) estimates that only 14%
of all SARS-CoV-2 infections are detected (4) and regional
post-mortem data suggest the true COVID-19 death toll may
be underestimated (5).
SARS-CoV-2 has evolved rapidly during intense transmis-
sion throughout the COVID-19 pandemic (6). The most pro-
nounced viral change was the emergence of the Omicron
variant (BA.1), which was first reported on 11 November 2021
in a patient from South Africa (Fig. 1A). Within a few weeks,
BA.1 was reported in 87 countries (7), prevailing over Delta to
become the predominant SARS-CoV-2 variant globally by the
end of December 2021 (8). Divergent Omicron sublineages
termed BA.2, BA.4 and BA.5 emerged globally months later
than BA.1 (Fig. 1A). BA.1 has more than 50 non-synonymous
mutations compared to ancestral SARS-CoV-2 strains, mostly
located in the gene encoding the viral spike protein (9).
Among the key BA.1 spike amino acid substitutions, the
unique combination of K417N, S477N, E484A, and N501Y in
the receptor-binding domain contributes to strong evasion of
immune responses elicited by vaccination or prior infection,
hinting at serotype properties of BA.1 relative to other vari-
ants (10–13). The BA.1 spike gene also harbors three muta-
tions in the region encoding the furin cleavage site, likely
facilitating proteolytic spike maturation (14). Finally, the BA.1
spike S2 subunit contains six unique amino acid substitutions
which reduce cleavage efficiency by the transmembrane ser-
ine protease TMPRSS2, favoring BA.1 entry via the receptor-
independent endosomal pathway and entailing increased
replication of BA.1 in epithelial cells from the upper respira-
tory tract (15). Efficient immune evasion and infection of the
upper respiratory tract are likely key to the explosive global
spread of BA.1 (16).
In response to the emergence of BA.1, the United States of
America, the European Union, and several other countries re-
stricted travel for four to six weeks from Southern and East-
ern African countries, including Botswana, Swaziland,
Lesotho, Mozambique, Namibia, South Africa, and Zimba-
bwe, by the end of November 2021 (17). The direct economic
loss in South Africa alone produced by these travel re-
strictions was roughly 600 million US dollars (18). However,
where and when BA.1 originated remains unknown. Evolu-
tionary reconstructions have projected an ancestor of BA.1
back to mid-2020 (19), which is consistent with preliminary
data on eight partial BA.1-like sequences in samples from pa-
tients in Nigeria collected in August and September 2021 (20)
and with 28 partial Omicron sequences available in the
Global Initiative on Sharing Avian Influenza Data (GISAID)
from samples collected between August and November 2021
in five different Western, Central and Eastern African coun-
tries. Lack of in-depth evolutionary analysis and epidemio-
logical context for putative early Omicron sequences and
regionally heterogeneous testing and reporting of Omicron
infections have prevented definitive assessments of the
spread of BA.1 in Africa. Here, we present the results of a di-
agnostic and evolutionary study to elucidate the emergence
of BA.1 across Africa.
A real-time RT-PCR test was designed to be highly specific
by targeting an BA.1-specific marker (spike 214 EPE inser-
tion) which is near-absent in other Omicron lineages and a
Delta-specific marker (spike deletion 157/158), achieving a di-
agnostic specificity of 98.7% for BA.1 and 99.8% for Delta ac-
cording to GISAID data (table S1). The specificity of BA.1
detection was confirmed by the absence of the BA.1 marker
in 545 SARS-CoV-2-positive respiratory samples from Benin,
Western Africa, collected between January and April 2021
(21). In total, 13,097 samples from laboratory-confirmed
COVID-19 patients from 22 African countries and 200 munic-
ipalities sampled during mid-2021 to early 2022 were in-
cluded in this study (Fig. 1B and fig. S2).
South-North gradient of BA.1 spread in continental
Africa
Across African countries, BA.1 replaced Delta as the predom-
inant SARS-CoV-2 variant already by December 2021 (Fig. 2A,
Delta fraction, and Fig. 2B, BA.1 fraction). Comparing conti-
nent-wide PCR data, BA.1 became the dominant variant
(>50% detection) on 11 November (95% CI, −5/+3 days) in
Southern Africa, 29 November (95% CI, −2/+1 days) in West-
ern Africa, 1 December (95% CI, −2/+3 days) in continental
Eastern Africa, 6 December (95% CI, −3/+3 days) in Central
Africa, and 25 December (95% CI, −1/+1 days) in Northern
Africa (Fig. 2C). The South-North gradient suggested by those
data is consistent with the earliest known BA.1 spread in
Southern Africa and recent genome-based analyses of SARS-
CoV-2 in Africa (22). Delayed BA.1 introduction into Northern
Africa is likely associated with reduced land connectivity im-
posed by the Sahara Desert (23). Similarly, border closure in
Madagascar until late 2021 delayed BA.1 introduction until
January 2022 (table S2). Across African countries, the median
time between the first BA.1 detection and BA.1 predominance
was 28 days (95% CI, 11–66.0) (Fig. 2D), which is comparable
to high-income countries (9, 24). Combining all country-level
data, the BA.1 effective reproduction number Rt peaked at 4.1
only five days after its first detection, which is consistent with
an overall R0 of 3.7 (95% CI, 3.3–4.1) reconstructed for SARS-
CoV-2 variants in Africa (25) and an average Rt of 3.4 among
countries in Africa, the Americas, Asia and Europe (26). In
the combined country-level data, Rt dropped below 1 within
32 days after the detection of the first BA.1 case (Fig. 2E),
likely due to widespread immunity following explosive BA.1
First release: 1 December 2022
science.org
(Page numbers not final at time of first release)
2
Retracted 20 December 2022. See Retraction.
spread (27). This interpretation is in line with the steep in-
crease of reported cases likely corresponding to and the short
duration of the BA.1 wave in Africa (Fig. 2F) (28, 29).
Occurrence of BA.1 ancestors across Africa during
August 2021
Although BA.1 cases were first detected in South Africa and
Botswana and knowing that our PCR data confirmed the ear-
liest predominance of BA.1 in Southern Africa, the variant’s
geographic and evolutionary origin remains unclear. A sys-
tematic search among >6 million GISAID entries based on
the BA.1-specific marker used in this study did not support
existence of potential BA.1 ancestors outside of Africa (table
S1 and fig. S2). In contrast, the earliest detection of BA.1-
specific PCR signals in this study was among 25 patients sam-
pled between August-September 2021 from six different
Western and Eastern African countries (Mali, Benin, Kenya,
Uganda, Ghana, and Niger) (table S2). For comparison, the
first detection of BA.1-specific PCR signal in patients from
Southern Africa only occurred two months later during No-
vember.
To confirm the early and widespread occurrence of poten-
tial Omicron ancestors outside of Southern Africa, near-full
SARS-CoV-2 genome characterization was done in 247 BA.1-
positive and -negative patient samples from Benin, Western
Africa, and 424 patient samples from South Africa using a
combination of deep sequencing- and PCR-based workflows.
The genomic data generated de novo from both countries
confirmed high specificity of around 98% of the BA.1 PCR test
(table S3). Despite low virus concentrations, partial or near-
complete SARS-CoV-2 genomes at a total genome coverage of
71.9% to 98.5% were recovered from five BA.1 PCR-positive
samples from Benin sampled between 22 August and 27 Oc-
tober 2021 (termed Ben-1 to Ben-5) (table S4). Phylogenetic
analyses in Maximum Likelihood and Bayesian frameworks
(Fig. 3, A and B, and fig. S3) robustly identified them as prox-
imal and more distant Omicron ancestors. The proximal Omi-
cron ancestors (Ben-4 and Ben-5) contained 67.7% (42/62)
and 75.0% (39/52) of the Omicron-defining mutations (9) that
were covered by sequencing, sufficient to be classified as BA.1
using the widely used Nextclade and UShER algorithms (Fig.
3C). The distant Omicron ancestors (Ben-1–Ben-3) shared be-
tween 38.3% (23/60) and 66.7% (32/48) of BA.1-defining mu-
tations and were classified as the parental lineage B.1.1 by
Nextclade. Together with other Africa-derived BA.1-like se-
quence entries, the Omicron ancestors clustered in basal sis-
ter relationship between the Omicron clade sensu strictu, the
BA.1 sublineage and Omicron ancestors belonging to the
globally circulating B.1 lineage (Fig. 3, A and B). Phylogeo-
graphic analyses supported BA.1 origins in Western Africa
preceding spread in Southern Africa (fig. S4), which was con-
sistent with the PCR-based data (table S2).
Gradual evolution of Omicron across Africa
BA.1-defining mutations varied among distant and proximal
Omicron ancestors that co-existed temporally, exemplified by
Ben-1, -3 and -4 sampled at nearly identical time points dur-
ing August 2021 in Cotonou, Benin (Fig. 3A), and Ben-5 and
three GISAID sequence entries from neighboring Nigeria (20)
(Fig. 3, A and B, and figs. S3 and S5). Together with evidence
for recombination events in the receptor-binding domain of
Benin-derived Omicron ancestors (fig. S6 and table S5), those
data suggest a non-linear micro-evolutionary pattern of Omi-
cron involving multiple ancestors existing across Western Af-
rican regions over several months (fig. S5). Sequence
comparisons of the distant and proximal Omicron ancestors
from Benin suggested that Omicron accumulated immune es-
cape mutations in the spike gene (11, 12, 30, 31), consistent
with antigenic drift driven by high levels of population im-
munity against prior SARS-CoV-2 variants in Africa (32).
While the spike amino acid exchanges E484A and N501Y
were present among all Omicron ancestors from Benin and
most from GISAID, the occurrence of K417N and S477N var-
ied among proximal Omicron ancestors from Benin and those
available in GISAID. None of those individual mutations are
specific for Omicron and they occur in diverse SARS-CoV-2
lineages (22) (Fig. 3C and fig. S5). Substitutions associated
with immune escape may thus not fully explain Omicron
emergence. Among the Omicron ancestors and the earliest
known African Omicron strains (yellow nodes in Fig. 3, A and
B), only the amino acid exchanges H655Y in the furin cleav-
age site and N969K in the S2 subunit (marked with asterisks
in Fig. 3C) are shared by Omicron sensu strictu sequences in
GISAID, the proximal Omicron ancestors Ben-4/Ben-5 and
available GISAID sequence entries from potential Omicron
ancestors (fig. S5). In contrast, these amino acid exchanges
are absent in other SARS-CoV-2 lineages including the dis-
tant Omicron ancestors Ben-1/-2/-3 from Benin. Both amino
acid substitutions are significantly associated with increased
SARS-CoV-2 fitness in mathematical modeling (33) and are
both present in >99% of all Omicron sequences and highly
specific for Omicron in public databases (table S6). Those
furin cleavage site and S2 mutations may thus represent key
events during the evolution of Omicron beyond immune es-
cape that deserve experimental investigation.
Finally, conjectures addressing the Omicron genealogy in-
clude its evolution in a non-human host (34) or in an immun-
ocompromised individual (35, 36), which would be consistent
with the initial detection of Omicron in South Africa, which
has a high HIV prevalence (37). In stark contrast, the muta-
tion pattern of Omicron ancestors and Omicron strains de-
posited in public databases differed substantially from the
SARS-CoV-2 mutation pattern in immunocompromised indi-
viduals (38) (fig. S7). Our data suggest prolonged and geo-
graphically widespread evolution of Omicron ancestors in
First release: 1 December 2022
science.org
(Page numbers not final at time of first release)
3
Retracted 20 December 2022. See Retraction.
patients across Africa. Although partial evolution of Omicron
ancestors in immunocompromised individuals or non-hu-
man animals cannot be excluded, Omicron did not evolve in
a single infection event according to our continent-wide data.
Eventually, highly transmissible BA.1 sensu strictu emerged,
combining both efficient immune escape and tropism for the
upper respiratory tract. Albeit the evolutionary origins of
other Omicron sublineages than BA.1 remain uncertain, phy-
logenetically ancestral strains existing in Africa during 2021
suggest that diverse Omicron sublineages may share similar
genealogy.
This study is limited by heterogeneous sampling in time
and space and by lack of SARS-CoV-2 genomic data from all
BA.1 PCR-positive patients. However, the PCR test was ex-
haustively validated in two African sub-regions and geo-
graphically widespread testing substantiates robustness of
our findings.
In conclusion, our study clearly shows the need to
strengthen and harmonize surveillance systems on a supra-
national level, establish strategic sampling frameworks, and
the importance of sharing surveillance data (39) to allow ef-
ficient interventions. Without international surveillance and
early detection travel bans have no epidemic containment
value and can cause social and economic harm.
REFERENCES AND NOTES
1. World Health Organization, WHO Coronavirus
(COVID-19) Dashboard;
https://covid19.who.int.
2. A.-L. Sander, A. Yadouleton, A. Moreira-Soto, C. Tchibozo, G. Hounkanrin, Y. Badou,
C. Fischer, N. Krause, P. Akogbeto, E. F. de Oliveira Filho, A. Dossou, S. Brünink, C.
Drosten, M. A. J. Aïssi, M. Harouna Djingarey, B. Hounkpatin, M. Nagel, J. F.
Drexler, An observational
laboratory-based assessment of SARS-CoV-2
molecular diagnostics in Benin, Western Africa. MSphere 6, e00979-20 (2021).
doi:10.1128/mSphere.00979-20 Medline
3. E. Wilkinson, M. Giovanetti, H. Tegally, J. E. San, R. Lessells, D. Cuadros, D. P. Martin,
D. A. Rasmussen, A. N. Zekri, A. K. Sangare, A.-S. Ouedraogo, A. K. Sesay, A.
Priscilla, A.-S. Kemi, A. M. Olubusuyi, A. O. O. Oluwapelumi, A. Hammami, A. A.
Amuri, A. Sayed, A. E. O. Ouma, A. Elargoubi, N. A. Ajayi, A. F. Victoria, A. Kazeem,
A. George, A. J. Trotter, A. A. Yahaya, A. K. Keita, A. Diallo, A. Kone, A. Souissi, A.
Chtourou, A. V. Gutierrez, A. J. Page, A. Vinze, A. Iranzadeh, A. Lambisia, A. Ismail,
A. Rosemary, A. Sylverken, A. Femi, A. Ibrahimi, B. Marycelin, B. S. Oderinde, B.
Bolajoko, B. Dhaala, B. L. Herring, B.-M. Njanpop-Lafourcade, B. Kleinhans, B.
McInnis, B. Tegomoh, C. Brook, C. B. Pratt, C. Scheepers, C. G. Akoua-Koffi, C. N.
Agoti, C. Peyrefitte, C. Daubenberger, C. M. Morang’a, D. J. Nokes, D. G. Amoako,
D. L. Bugembe, D. Park, D. Baker, D. Doolabh, D. Ssemwanga, D. Tshiabuila, D.
Bassirou, D. S. Y. Amuzu, D. Goedhals, D. O. Omuoyo, D. Maruapula, E. Foster-
Nyarko, E. K. Lusamaki, E. Simulundu, E. M. Ong’era, E. N. Ngabana, E. Shumba,
E. El Fahime, E. Lokilo, E. Mukantwari, E. Philomena, E. Belarbi, E. Simon-Loriere,
E. A. Anoh, F. Leendertz, F. Ajili, F. O. Enoch, F. Wasfi, F. Abdelmoula, F. S. Mosha,
F. T. Takawira, F. Derrar, F. Bouzid, F. Onikepe, F. Adeola, F. M. Muyembe, F.
Tanser, F. A. Dratibi, G. K. Mbunsu, G. Thilliez, G. L. Kay, G. Githinji, G. van Zyl, G.
A. Awandare, G. Schubert, G. P. Maphalala, H. C. Ranaivoson, H. Lemriss, H. Anise,
H. Abe, H. H. Karray, H. Nansumba, H. A. Elgahzaly, H. Gumbo, I. Smeti, I. B. Ayed,
I. Odia, I. B. Ben Boubaker, I. Gaaloul, I. Gazy, I. Mudau, I. Ssewanyana, I.
Konstantinus, J. B. Lekana-Douk, J.-C. C. Makangara, J. M. Tamfum, J.-M. Heraud,
J. G. Shaffer, J. Giandhari, J. Li, J. Yasuda, J. Q. Mends, J. Kiconco, J. M. Morobe,
J. O. Gyapong, J. C. Okolie, J. T. Kayiwa, J. A. Edwards, J. Gyamfi, J. Farah, J.
Nakaseegu, J. M. Ngoi, J. Namulondo, J. C. Andeko, J. J. Lutwama, J. O’Grady, K.
Siddle, K. T. Adeyemi, K. A. Tumedi, K. M. Said, K. Hae-Young, K. O. Duedu, L.
Belyamani, L. Fki-Berrajah, L. Singh, L. O. Martins, L. Tyers, M. Ramuth, M.
Mastouri, M. Aouni, M. El Hefnawi, M. I. Matsheka, M. Kebabonye, M. Diop, M.
Turki, M. Paye, M. M. Nyaga, M. Mareka, M.-M. Damaris, M. W. Mburu, M. Mpina,
M. Nwando, M. Owusu, M. R. Wiley, M. T. Youtchou, M. O. Ayekaba, M. Abouelhoda,
M. G. Seadawy, M. K. Khalifa, M. Sekhele, M. Ouadghiri, M. M. Diagne, M. Mwenda,
M. Allam, M. V. T. Phan, N. Abid, N. Touil, N. Rujeni, N. Kharrat, N. Ismael, N. Dia,
N. Mabunda, N. Y. Hsiao, N. B. Silochi, N. Nsenga, N. Gumede, N. Mulder, N. Ndodo,
N. H. Razanajatovo, N. Iguosadolo, O. Judith, O. C. Kingsley, O. Sylvanus, O. Peter,
O. Femi, O. Idowu, O. Testimony, O. E. Chukwuma, O. E. Ogah, C. K. Onwuamah,
O. Cyril, O. Faye, O. Tomori, P. Ondoa, P. Combe, P. Semanda, P. E. Oluniyi, P.
Arnaldo, P. K. Quashie, P. Dussart, P. A. Bester, P. K. Mbala, R. Ayivor-Djanie, R.
Njouom, R. O. Phillips, R. Gorman, R. A. Kingsley, R. A. A. Carr, S. El Kabbaj, S.
Gargouri, S. Masmoudi, S. Sankhe, S. B. Lawal, S. Kassim, S. Trabelsi, S. Metha, S.
Kammoun, S. Lemriss, S. H. A. Agwa, S. Calvignac-Spencer, S. F. Schaffner, S.
Doumbia, S. M. Mandanda, S. Aryeetey, S. S. Ahmed, S. Elhamoumi, S.
Andriamandimby, S. Tope, S. Lekana-Douki, S. Prosolek, S. Ouangraoua, S. A.
Mundeke, S. Rudder, S. Panji, S. Pillay, S. Engelbrecht, S. Nabadda, S. Behillil, S.
L. Budiaki, S. van der Werf, T. Mashe, T. Aanniz, T. Mohale, T. Le-Viet, T. Schindler,
U. J. Anyaneji, U. Chinedu, U. Ramphal, U. Jessica, U. George, V. Fonseca, V. Enouf,
V. Gorova, W. H. Roshdy, W. K. Ampofo, W. Preiser, W. T. Choga, Y. Bediako, Y.
Naidoo, Y. Butera, Z. R. de Laurent, A. A. Sall, A. Rebai, A. von Gottberg, B. Kouriba,
C. Williamson, D. J. Bridges, I. Chikwe, J. N. Bhiman, M. Mine, M. Cotten, S. Moyo,
S. Gaseitsiwe, N. Saasa, P. C. Sabeti, P. Kaleebu, Y. K. Tebeje, S. K. Tessema, C.
Happi, J. Nkengasong, T. de Oliveira, A year of genomic surveillance reveals how
the SARS-CoV-2 pandemic unfolded in Africa. Science 374, 423–431 (2021).
doi:10.1126/science.abj4336 Medline
4. WHO (World Health Organization, 2021), vol. 2022.
5. C. J. Gill, L. Mwananyanda, W. MacLeod, G. Kwenda, R. Pieciak, L. Etter, D. J.
Bridges, C. Chikoti, S. Chirwa, C. Chimoga, L. Forman, B. Katowa, R. Lapidot, J.
Lungu, J. Matoba, G. Mwinga, B. Mubemba, Z. Mupila, W. Muleya, M. Mwenda, B.
Ngoma, R. Nkazwe, D. Nzara, N. Pawlak, L. Pemba, N. Saasa, E. Simulundu, B.
Yankonde, D. M. Thea, Sustained high prevalence of COVID-19 deaths from a
systematic post-mortem study in Lusaka, Zambia: one year later. medRxiv
2022.03.08.22272087
(2022).
https://doi.org/10.1101/2022.03.08.22272087.
[Preprint]
6. W. K. Jo, C. Drosten, J. F. Drexler, The evolutionary dynamics of endemic human
coronaviruses. Virus Evol. 7, veab020 (2021). doi:10.1093/ve/veab020 Medline
7. R. Viana, S. Moyo, D. G. Amoako, H. Tegally, C. Scheepers, C. L. Althaus, U. J.
Anyaneji, P. A. Bester, M. F. Boni, M. Chand, W. T. Choga, R. Colquhoun, M. Davids,
K. Deforche, D. Doolabh, L. du Plessis, S. Engelbrecht, J. Everatt, J. Giandhari, M.
Giovanetti, D. Hardie, V. Hill, N.-Y. Hsiao, A. Iranzadeh, A. Ismail, C. Joseph, R.
Joseph, L. Koopile, S. L. Kosakovsky Pond, M. U. G. Kraemer, L. Kuate-Lere, O.
Laguda-Akingba, O. Lesetedi-Mafoko, R. J. Lessells, S. Lockman, A. G. Lucaci, A.
Maharaj, B. Mahlangu, T. Maponga, K. Mahlakwane, Z. Makatini, G. Marais, D.
Maruapula, K. Masupu, M. Matshaba, S. Mayaphi, N. Mbhele, M. B. Mbulawa, A.
Mendes, K. Mlisana, A. Mnguni, T. Mohale, M. Moir, K. Moruisi, M. Mosepele, G.
Motsatsi, M. S. Motswaledi, T. Mphoyakgosi, N. Msomi, P. N. Mwangi, Y. Naidoo,
N. Ntuli, M. Nyaga, L. Olubayo, S. Pillay, B. Radibe, Y. Ramphal, U. Ramphal, J. E.
San, L. Scott, R. Shapiro, L. Singh, P. Smith-Lawrence, W. Stevens, A. Strydom, K.
Subramoney, N. Tebeila, D. Tshiabuila, J. Tsui, S. van Wyk, S. Weaver, C. K.
Wibmer, E. Wilkinson, N. Wolter, A. E. Zarebski, B. Zuze, D. Goedhals, W. Preiser,
F. Treurnicht, M. Venter, C. Williamson, O. G. Pybus, J. Bhiman, A. Glass, D. P.
Martin, A. Rambaut, S. Gaseitsiwe, A. von Gottberg, T. de Oliveira, Rapid epidemic
expansion of the SARS-CoV-2 Omicron variant in southern Africa. Nature 603,
679–686 (2022). doi:10.1038/s41586-022-04411-y Medline
8. J. Hadfield, C. Megill, S. M. Bell, J. Huddleston, B. Potter, C. Callender, P. Sagulenko,
T. Bedford, R. A. Neher, Nextstrain: Real-time tracking of pathogen evolution.
Bioinformatics 34, 4121–4123 (2018). doi:10.1093/bioinformatics/bty407
Medline
9. E. B. Hodcroft, CoVariants: SARS-CoV-2 Mutations and Variants of Interest (2021);
https://covariants.org/.
10. S. H. Wilks, B. Mühlemann, X. Shen, S. Türeli, E. B. LeGresley, A. Netzl, M. A. Caniza,
J. N. Chacaltana-Huarcaya, V. M. Corman, X. Daniell, M. B. Datto, F. S. Dawood, T.
N. Denny, C. Drosten, R. A. M. Fouchier, P. J. Garcia, P. J. Halfmann, A. Jassem, L.
M. Jeworowski, T. C. Jones, Y. Kawaoka, F. Krammer, C. McDanal, R. Pajon, V.
Simon, M. S. Stockwell, H. Tang, H. van Bakel, V. Veguilla, R. Webby, D. C.
First release: 1 December 2022
science.org
(Page numbers not final at time of first release)
4
Retracted 20 December 2022. See Retraction.
Montefiori, D. J. Smith, Mapping SARS-CoV-2 antigenic relationships and
serological
(2022).
https://doi.org/10.1101/2022.01.28.477987.
responses. bioRxiv 2022.01.28.477987
[Preprint]
11. L. Liu, S. Iketani, Y. Guo, J. F.-W. Chan, M. Wang, L. Liu, Y. Luo, H. Chu, Y. Huang, M.
S. Nair, J. Yu, K. K.-H. Chik, T. T.-T. Yuen, C. Yoon, K. K.-W. To, H. Chen, M. T. Yin,
M. E. Sobieszczyk, Y. Huang, H. H. Wang, Z. Sheng, K.-Y. Yuen, D. D. Ho, Striking
antibody evasion manifested by the Omicron variant of SARS-CoV-2. Nature 602,
676–681 (2022). doi:10.1038/s41586-021-04388-0 Medline
12. Q. Li, J. Nie, J. Wu, L. Zhang, R. Ding, H. Wang, Y. Zhang, T. Li, S. Liu, M. Zhang, C.
Zhao, H. Liu, L. Nie, H. Qin, M. Wang, Q. Lu, X. Li, J. Liu, H. Liang, Y. Shi, Y. Shen, L.
Xie, L. Zhang, X. Qu, W. Xu, W. Huang, Y. Wang, SARS-CoV-2 501Y.V2 variants lack
higher infectivity but do have immune escape. Cell 184, 2362–2371.e9 (2021).
doi:10.1016/j.cell.2021.02.042 Medline
13. W. T. Harvey, A. M. Carabelli, B. Jackson, R. K. Gupta, E. C. Thomson, E. M.
Harrison, C. Ludden, R. Reeve, A. Rambaut, S. J. Peacock, D. L. Robertson; COVID-
19 Genomics UK (COG-UK) Consortium, SARS-CoV-2 variants, spike mutations
and
(2021).
doi:10.1038/s41579-021-00573-0 Medline
immune escape. Nat. Rev. Microbiol.
19, 409–424
14. G. Ye, B. Liu, F. Li, Cryo-EM structure of a SARS-CoV-2 omicron spike protein
ectodomain. Nat. Commun. 13, 1214 (2022). doi:10.1038/s41467-022-28882-9
Medline
15. B. Meng, A. Abdullahi, I. A. T. M. Ferreira, N. Goonawardane, A. Saito, I. Kimura, D.
Yamasoba, P. P. Gerber, S. Fatihi, S. Rathore, S. K. Zepeda, G. Papa, S. A. Kemp,
T. Ikeda, M. Toyoda, T. S. Tan, J. Kuramochi, S. Mitsunaga, T. Ueno, K. Shirakawa,
A. Takaori-Kondo, T. Brevini, D. L. Mallery, O. J. Charles, J. E. Bowen, A. Joshi, A.
C. Walls, L. Jackson, D. Martin, K. G. C. Smith, J. Bradley, J. A. G. Briggs, J. Choi, E.
Madissoon, K. B. Meyer, P. Mlcochova, L. Ceron-Gutierrez, R. Doffinger, S. A.
Teichmann, A. J. Fisher, M. S. Pizzuto, A. de Marco, D. Corti, M. Hosmillo, J. H. Lee,
L. C. James, L. Thukral, D. Veesler, A. Sigal, F. Sampaziotis, I. G. Goodfellow, N. J.
Matheson, K. Sato, R. K. Gupta; CITIID-NIHR BioResource COVID-19
Collaboration; Genotype to Phenotype Japan (G2P-Japan) Consortium; Ecuador-
COVID19 Consortium, Altered TMPRSS2 usage by SARS-CoV-2 Omicron impacts
infectivity and fusogenicity. Nature 603, 706–714 (2022). doi:10.1038/s41586-
022-04474-x Medline
16. J. A. Backer, D. Eggink, S. P. Andeweg, I. K. Veldhuijzen, N. van Maarseveen, K.
Vermaas, B. Vlaemynck, R. Schepers, S. van den Hof, C. B. E. M. Reusken, J.
Wallinga, Shorter serial intervals in SARS-CoV-2 cases with Omicron BA.1 variant
compared with Delta variant, the Netherlands, 13 to 26 December 2021. Euro
Surveill. 27, 2200042 (2022). doi:10.2807/1560-7917.ES.2022.27.6.2200042
Medline
17. M. Mendelson, F. Venter, M. Moshabela, G. Gray, L. Blumberg, T. de Oliveira, S. A.
Madhi, The political theatre of the UK’s travel ban on South Africa. Lancet 398,
2211–2213 (2021). doi:10.1016/S0140-6736(21)02752-5 Medline
18. D. Luke, in Business Insider SA (Business Insider SA, 2022).
19. S. Mallapaty, Where did Omicron come from? Three key theories. Nature 602, 26–
28 (2022). doi:10.1038/d41586-022-00215-2 Medline
20. C. Onwuamah, B. Iwalokun, R. Ahmed, E. Toye, O. Amoo, N. Ndodo, E. Momoh, A.
Okwuraiwe, A. James, C. Okoli, C. Chukwu, J. Shaibu, M. Fowora, J. Fayemi, T.
Onuigbo, M. Sunday, V. Ogbolu, M. Otuonye, N. Mba, R. Audu, O. Ezechi, A.
Abayomi, I. Adetifa, B. Salako The trend of SARS-CoV-2 variants from July 2021 to
Jan 2022 in metropolitan Lagos, the epicentre of the COVID-19 pandemic in
Nigeria.
(2022).
Square
https://www.researchsquare.com/article/rs-1602118/v1.
[Preprint]
Research
21. A. L. Sander, A. Yadouleton, E. F. de Oliveira Filho, C. Tchibozo, G. Hounkanrin, Y.
Badou, P. Adewumi, K. K. René, D. Ange, S. Sourakatou, E. Sedjro, M. A. J. Aïssi, H.
Fidelia, M. H. Djingarey, M. Nagel, W. K. Jo, A. Moreira-Soto, C. Drosten, O. Landt,
V. M. Corman, B. Hounkpatin, J. F. Drexler, Mutations associated with SARS-CoV-
2 variants of concern, Benin, Early 2021. Emerg. Infect. Dis. 27, 2889–2903
(2021). doi:10.3201/eid2711.211353 Medline
22. H. Tegally, J. E. San, M. Cotten, M. Moir, B. Tegomoh, G. Mboowa, D. P. Martin, C.
Baxter, A. W. Lambisia, A. Diallo, D. G. Amoako, M. M. Diagne, A. Sisay, A. N. Zekri,
A. S. Gueye, A. K. Sangare, A.-S. Ouedraogo, A. Sow, A. O. Musa, A. K. Sesay, A. G.
Abias, A. I. Elzagheid, A. Lagare, A.-S. Kemi, A. E. Abar, A. A. Johnson, A. Fowotade,
A. O. Oluwapelumi, A. A. Amuri, A. Juru, A. Kandeil, A. Mostafa, A. Rebai, A. Sayed,
A. Kazeem, A. Balde, A. Christoffels, A. J. Trotter, A. Campbell, A. K. Keita, A. Kone,
A. Bouzid, A. Souissi, A. Agweyu, A. Naguib, A. V. Gutierrez, A. Nkeshimana, A. J.
Page, A. Yadouleton, A. Vinze, A. N. Happi, A. Chouikha, A. Iranzadeh, A. Maharaj,
A. L. Batchi-Bouyou, A. Ismail, A. A. Sylverken, A. Goba, A. Femi, A. E. Sijuwola, B.
Marycelin, B. L. Salako, B. S. Oderinde, B. Bolajoko, B. Diarra, B. L. Herring, B.
Tsofa, B. Lekana-Douki, B. Mvula, B.-M. Njanpop-Lafourcade, B. T. Marondera, B.
A. Khaireh, B. Kouriba, B. Adu, B. Pool, B. McInnis, C. Brook, C. Williamson, C.
Nduwimana, C. Anscombe, C. B. Pratt, C. Scheepers, C. G. Akoua-Koffi, C. N.
Agoti, C. M. Mapanguy, C. Loucoubar, C. K. Onwuamah, C. Ihekweazu, C. N.
Malaka, C. Peyrefitte, C. Grace, C. E. Omoruyi, C. D. Rafaï, C. M. Morang’a, C.
Erameh, D. B. Lule, D. J. Bridges, D. Mukadi-Bamuleka, D. Park, D. A. Rasmussen,
D. Baker, D. J. Nokes, D. Ssemwanga, D. Tshiabuila, D. S. Y. Amuzu, D. Goedhals,
D. S. Grant, D. O. Omuoyo, D. Maruapula, D. W. Wanjohi, E. Foster-Nyarko, E. K.
Lusamaki, E. Simulundu, E. M. Ong’era, E. N. Ngabana, E. O. Abworo, E. Otieno, E.
Shumba, E. Barasa, E. B. Ahmed, E. A. Ahmed, E. Lokilo, E. Mukantwari, E.
Philomena, E. Belarbi, E. Simon-Loriere, E. A. Anoh, E. Manuel, F. Leendertz, F. M.
Taweh, F. Wasfi, F. Abdelmoula, F. T. Takawira, F. Derrar, F. V. Ajogbasile, F.
Treurnicht, F. Onikepe, F. Ntoumi, F. M. Muyembe, F. E. Z. Ragomzingba, F. A.
Dratibi, F.-A. Iyanu, G. K. Mbunsu, G. Thilliez, G. L. Kay, G. O. Akpede, G. U. van Zyl,
G. A. Awandare, G. S. Kpeli, G. Schubert, G. P. Maphalala, H. C. Ranaivoson, H. E.
Omunakwe, H. Onywera, H. Abe, H. Karray, H. Nansumba, H. Triki, H. A. A. Kadjo,
H. Elgahzaly, H. Gumbo, H. Mathieu, H. Kavunga-Membo, I. Smeti, I. B. Olawoye, I.
M. O. Adetifa, I. Odia, I. B. Ben Boubaker, I. A. Muhammad, I. Ssewanyana, I. Wurie,
I. S. Konstantinus, J. W. A. Halatoko, J. Ayei, J. Sonoo, J.-C. C. Makangara, J. M.
Tamfum, J.-M. Heraud, J. G. Shaffer, J. Giandhari, J. Musyoki, J. Nkurunziza, J. N.
Uwanibe, J. N. Bhiman, J. Yasuda, J. Morais, J. Kiconco, J. D. Sandi, J. Huddleston,
J. K. Odoom, J. M. Morobe, J. O. Gyapong, J. T. Kayiwa, J. C. Okolie, J. S. Xavier, J.
Gyamfi, J. F. Wamala, J. H. K. Bonney, J. Nyandwi, J. Everatt, J. Nakaseegu, J. M.
Ngoi, J. Namulondo, J. U. Oguzie, J. C. Andeko, J. J. Lutwama, J. J. H. Mogga, J.
O’Grady, K. J. Siddle, K. Victoir, K. T. Adeyemi, K. A. Tumedi, K. S. Carvalho, K. S.
Mohammed, K. Dellagi, K. G. Musonda, K. O. Duedu, L. Fki-Berrajah, L. Singh, L. M.
Kepler, L. Biscornet, L. de Oliveira Martins, L. Chabuka, L. Olubayo, L. D. Ojok, L.
L. Deng, L. I. Ochola-Oyier, L. Tyers, M. Mine, M. Ramuth, M. Mastouri, M.
ElHefnawi, M. Mbanne, M. I. Matsheka, M. Kebabonye, M. Diop, M. Momoh, M. D.
L. Lima Mendonça, M. Venter, M. F. Paye, M. Faye, M. M. Nyaga, M. Mareka, M.-M.
Damaris, M. W. Mburu, M. G. Mpina, M. Owusu, M. R. Wiley, M. Y. Tatfeng, M. O.
Ayekaba, M. Abouelhoda, M. A. Beloufa, M. G. Seadawy, M. K. Khalifa, M. M.
Matobo, M. Kane, M. Salou, M. B. Mbulawa, M. Mwenda, M. Allam, M. V. T. Phan,
N. Abid, N. Rujeni, N. Abuzaid, N. Ismael, N. Elguindy, N. M. Top, N. Dia, N.
Mabunda, N. Y. Hsiao, N. B. Silochi, N. M. Francisco, N. Saasa, N. Bbosa, N.
Murunga, N. Gumede, N. Wolter, N. Sitharam, N. Ndodo, N. A. Ajayi, N. Tordo, N.
Mbhele, N. H. Razanajatovo, N. Iguosadolo, N. Mba, O. C. Kingsley, O. Sylvanus, O.
Femi, O. M. Adewumi, O. Testimony, O. A. Ogunsanya, O. Fakayode, O. E. Ogah,
O.-E. Oludayo, O. Faye, P. Smith-Lawrence, P. Ondoa, P. Combe, P. Nabisubi, P.
Semanda, P. E. Oluniyi, P. Arnaldo, P. K. Quashie, P. O. Okokhere, P. Bejon, P.
Dussart, P. A. Bester, P. K. Mbala, P. Kaleebu, P. Abechi, R. El-Shesheny, R.
Joseph, R. K. Aziz, R. G. Essomba, R. Ayivor-Djanie, R. Njouom, R. O. Phillips, R.
Gorman, R. A. Kingsley, R. M. D. E. S. A. Neto Rodrigues, R. A. Audu, R. A. A. Carr,
S. Gargouri, S. Masmoudi, S. Bootsma, S. Sankhe, S. I. Mohamed, S. Femi, S.
Mhalla, S. Hosch, S. K. Kassim, S. Metha, S. Trabelsi, S. H. Agwa, S. W. Mwangi, S.
Doumbia, S. Makiala-Mandanda, S. Aryeetey, S. S. Ahmed, S. M. Ahmed, S.
Elhamoumi, S. Moyo, S. Lutucuta, S. Gaseitsiwe, S. Jalloh, S. F. Andriamandimby,
S. Oguntope, S. Grayo, S. Lekana-Douki, S. Prosolek, S. Ouangraoua, S. van Wyk,
S. F. Schaffner, S. Kanyerezi, S. Ahuka-Mundeke, S. Rudder, S. Pillay, S. Nabadda,
S. Behillil, S. L. Budiaki, S. van der Werf, T. Mashe, T. Mohale, T. Le-Viet, T. P.
Velavan, T. Schindler, T. G. Maponga, T. Bedford, U. J. Anyaneji, U. Chinedu, U.
Ramphal, U. E. George, V. Enouf, V. Nene, V. Gorova, W. H. Roshdy, W. A. Karim,
W. K. Ampofo, W. Preiser, W. T. Choga, Y. A. Ahmed, Y. Ramphal, Y. Bediako, Y.
Naidoo, Y. Butera, Z. R. de Laurent, A. E. O. Ouma, A. von Gottberg, G. Githinji, M.
Moeti, O. Tomori, P. C. Sabeti, A. A. Sall, S. O. Oyola, Y. K. Tebeje, S. K. Tessema,
T. de Oliveira, C. Happi, R. Lessells, J. Nkengasong, E. Wilkinson; Africa Pathogen
Genomics Initiative (Africa PGI), The evolving SARS-CoV-2 epidemic in Africa:
Insights from rapidly expanding genomic surveillance. Science 378, eabq5358
(2022). doi:10.1126/science.abq5358 Medline
23. E. Strano, M. P. Viana, A. Sorichetta, A. J. Tatem, Mapping road network
communities for guiding disease surveillance and control strategies. Sci. Rep. 8,
First release: 1 December 2022
science.org
(Page numbers not final at time of first release)
5
Retracted 20 December 2022. See Retraction.
4744 (2018). doi:10.1038/s41598-018-22969-4 Medline
24. K. Sun, S. Tempia, J. Kleynhans, A. von Gottberg, M. L. McMorrow, N. Wolter, J. N.
Bhiman, J. Moyes, M. du Plessis, M. Carrim, A. Buys, N. A. Martinson, K. Kahn, S.
Tollman, L. Lebina, F. Wafawanaka, J. D. du Toit, F. X. Gómez-Olivé, T. Mkhencele,
C. Viboud, C. Cohen, SARS-CoV-2 transmission, persistence of immunity, and
estimates of Omicron’s impact in South African population cohorts. Sci. Transl.
Med. 14, eabo7081 (2022). doi:10.1126/scitranslmed.abo7081 Medline
25. S. A. Iyaniwura, M. Rabiu, J. F. David, J. D. Kong, The basic reproduction number
(2022).
across Africa. PLOS ONE
e0264455
17,
of COVID-19
doi:10.1371/journal.pone.0264455 Medline
26. Y. Liu, J. Rocklöv, The effective reproductive number of the Omicron variant of
SARS-CoV-2 is several times relative to Delta. J. Travel Med. 29, taac037 (2022).
doi:10.1093/jtm/taac037 Medline
27. H. Tegally, M. Moir, J. Everatt, M. Giovanetti, C. Scheepers, E. Wilkinson, K.
Subramoney, Z. Makatini, S. Moyo, D. G. Amoako, C. Baxter, C. L. Althaus, U. J.
Anyaneji, D. Kekana, R. Viana, J. Giandhari, R. J. Lessells, T. Maponga, D.
Maruapula, W. Choga, M. Matshaba, M. B. Mbulawa, N. Msomi, Y. Naidoo, S. Pillay,
T. J. Sanko, J. E. San, L. Scott, L. Singh, N. A. Magini, P. Smith-Lawrence, W.
Stevens, G. Dor, D. Tshiabuila, N. Wolter, W. Preiser, F. K. Treurnicht, M. Venter,
G. Chiloane, C. McIntyre, A. O’Toole, C. Ruis, T. P. Peacock, C. Roemer, S. L.
Kosakovsky Pond, C. Williamson, O. G. Pybus, J. N. Bhiman, A. Glass, D. P. Martin,
B. Jackson, A. Rambaut, O. Laguda-Akingba, S. Gaseitsiwe, A. von Gottberg, T. de
Oliveira; NGS-SA consortium, Emergence of SARS-CoV-2 Omicron lineages BA.4
and BA.5 in South Africa. Nat. Med. 28, 1785–1790 (2022). doi:10.1038/s41591-
022-01911-2 Medline
28. K. Y. Oróstica, S. Contreras, A. Sanchez-Daza, J. Fernandez, V. Priesemann, Á.
Olivera-Nappa, New year, new SARS-CoV-2 variant: Resolutions on genomic
surveillance protocols to face Omicron. Lancet Reg Health Am 7, 100203 (2022).
doi:10.1016/j.lana.2022.100203 Medline
29. P. Elliott, O. Eales, N. Steyn, D. Tang, B. Bodinier, H. Wang, J. Elliott, M. Whitaker,
C. Atchison, P. J. Diggle, A. J. Page, A. J. Trotter, D. Ashby, W. Barclay, G. Taylor,
H. Ward, A. Darzi, G. S. Cooke, C. A. Donnelly, M. Chadeau-Hyam, Twin peaks: The
Omicron SARS-CoV-2 BA.1 and BA.2 epidemics in England. Science 376, eabq4411
(2022). doi:10.1126/science.abq4411 Medline
30. J. Wise, Covid-19: The E484K mutation and the risks it poses. BMJ 372, n359
(2021). doi:10.1136/bmj.n359 Medline
31. Y. Cao, A. Yisimayi, Y. Bai, W. Huang, X. Li, Z. Zhang, T. Yuan, R. An, J. Wang, T.
Xiao, S. Du, W. Ma, L. Song, Y. Li, X. Li, W. Song, J. Wu, S. Liu, X. Li, Y. Zhang, B. Su,
X. Guo, Y. Wei, C. Gao, N. Zhang, Y. Zhang, Y. Dou, X. Xu, R. Shi, B. Lu, R. Jin, Y. Ma,
C. Qin, Y. Wang, Y. Feng, J. Xiao, X. S. Xie, Humoral immune response to circulating
SARS-CoV-2 variants elicited by inactivated and RBD-subunit vaccines. Cell Res.
31, 732–741 (2021). doi:10.1038/s41422-021-00514-9 Medline
32. S. A. Madhi, G. Kwatra, J. E. Myers, W. Jassat, N. Dhar, C. K. Mukendi, A. J. Nana,
L. Blumberg, R. Welch, N. Ngorima-Mabhena, P. C. Mutevedzi, Population
immunity and Covid-19 severity with Omicron variant in South Africa. N. Engl. J.
Med. 386, 1314–1326 (2022). doi:10.1056/NEJMoa2119658 Medline
33. F. Obermeyer, M. Jankowiak, N. Barkas, S. F. Schaffner, J. D. Pyle, L. Yurkovetskiy,
M. Bosso, D. J. Park, M. Babadi, B. L. MacInnis, J. Luban, P. C. Sabeti, J. E. Lemieux,
Analysis of 6.4 million SARS-CoV-2 genomes identifies mutations associated with
fitness. Science 376, 1327–1332 (2022). doi:10.1126/science.abm1208 Medline
34. V. L. Hale, P. M. Dennis, D. S. McBride, J. M. Nolting, C. Madden, D. Huey, M.
Ehrlich, J. Grieser, J. Winston, D. Lombardi, S. Gibson, L. Saif, M. L. Killian, K. Lantz,
R. M. Tell, M. Torchetti, S. Robbe-Austerman, M. I. Nelson, S. A. Faith, A. S.
Bowman, SARS-CoV-2 infection in free-ranging white-tailed deer. Nature 602,
481–486 (2022). doi:10.1038/s41586-021-04353-x Medline
35. S. T. Sonnleitner, M. Prelog, S. Sonnleitner, E. Hinterbichler, H. Halbfurter, D. B. C.
Kopecky, G. Almanzar, S. Koblmüller, C. Sturmbauer, L. Feist, R. Horres, W. Posch,
G. Walder, Cumulative SARS-CoV-2 mutations and corresponding changes in
immunity in an immunocompromised patient indicate viral evolution within the
host. Nat. Commun. 13, 2560 (2022). doi:10.1038/s41467-022-30163-4 Medline
36. S. A. Clark, L. E. Clark, J. Pan, A. Coscia, L. G. A. McKay, S. Shankar, R. I. Johnson,
V. Brusic, M. C. Choudhary, J. Regan, J. Z. Li, A. Griffiths, J. Abraham, SARS-CoV-
2 evolution in an immunocompromised host reveals shared neutralization escape
mechanisms. Cell 184, 2605–2617.e18 (2021). doi:10.1016/j.cell.2021.03.027
Medline
37. The World Bank Group, World Bank Open Data, Prevalence of HIV, total (% of
(2022);
population
15-49)
https://data.worldbank.org/indicator/SH.DYN.AIDS.ZS.
ages
38. S. Harari, M. Tahor, N. Rutsinsky, S. Meijer, D. Miller, O. Henig, O. Halutz, K.
Levytskyi, R. Ben-Ami, A. Adler, Y. Paran, A. Stern, Drivers of adaptive evolution
during chronic SARS-CoV-2 infections. Nat. Med. 28, 1501–1508 (2022).
doi:10.1038/s41591-022-01882-4 Medline
39. W. Preiser, S. Engelbrecht, T. Maponga, No point in travel bans if countries with
poor surveillance are ignored. Lancet 399, 1224 (2022). doi:10.1016/S0140-
6736(22)00363-4 Medline
40. C. Fischer, Gradual emergence followed by exponential spread of the SARS-CoV-
(2022);
Zenodo
version
1.4,
2 Omicron
variant
https://doi.org/10.5281/zenodo.7342478.
Africa,
in
41. S. Engelbrecht, K. Delaney, B. Kleinhans, E. Wilkinson, H. Tegally, T. Stander, G.
van Zyl, W. Preiser, T. de Oliveira, Multiple early introductions of SARS-CoV-2 to
Cape Town, South Africa. Viruses 13, 526 (2021). doi:10.3390/v13030526
Medline
42. F. Grabowski, M. Kochańczyk, T. Lipniacki, The spread of SARS-CoV-2 variant
Omicron with a doubling time of 2.0-3.3 days can be explained by immune
evasion. Viruses 14, 294 (2022). doi:10.3390/v14020294 Medline
43. A. J. Drummond, M. A. Suchard, D. Xie, A. Rambaut, Bayesian phylogenetics with
BEAUti and the BEAST 1.7. Mol. Biol. Evol. 29, 1969–1973 (2012).
doi:10.1093/molbev/mss075 Medline
44. P. Lemey, A. Rambaut, A. J. Drummond, M. A. Suchard, Bayesian phylogeography
finds
Comput.
(2009).
doi:10.1371/journal.pcbi.1000520 Medline
e1000520
roots.
PLOS
Biol.
its
5,
45. F. Bielejec, G. Baele, B. Vrancken, M. A. Suchard, A. Rambaut, P. Lemey, SpreaD3:
Interactive visualization of spatiotemporal history and trait evolutionary
processes. Mol. Biol. Evol. 33, 2167–2169 (2016). doi:10.1093/molbev/msw082
Medline
46. V. M. Corman, O. Landt, M. Kaiser, R. Molenkamp, A. Meijer, D. K. W. Chu, T.
Bleicker, S. Brünink, J. Schneider, M. L. Schmidt, D. G. J. C. Mulders, B. L.
Haagmans, B. van der Veer, S. van den Brink, L. Wijsman, G. Goderski, J.-L.
Romette, J. Ellis, M. Zambon, M. Peiris, H. Goossens, C. Reusken, M. P. G.
Koopmans, C. Drosten, Detection of 2019 novel coronavirus (2019-nCoV) by real-
time RT-PCR. Euro Surveill. 25, 2000045 (2020). doi:10.2807/1560-
7917.ES.2020.25.3.2000045 Medline
ACKNOWLEDGMENTS
We thank our colleagues who provided insight and expertise that greatly assisted the
research: Sarah Belkalem, National Influenza Centre, Viral Respiratory
Laboratory, Department of Virology, Institut Pasteur of Algeria, Algeria; El Alia
Gradi, National Influenza Centre, Viral Respiratory Laboratory, Department of
Virology, Institut Pasteur of Algeri, Algeria; Kahina Izri, National Influenza Centre,
Viral Respiratory Laboratory, Department of Virology, Institut Pasteur of Algeri,
Algeria; Sonia Carvalho, Laboratório de Biologia Molecular, Instituto Nacional de
Investigação em Saúde (INIS), Luanda, Angola; Joana P. da Paixão, Laboratório
de Biologia Molecular, Instituto Nacional de Investigação em Saúde (INIS),
Luanda, Angola; Susana Daniel, Laboratório de Biologia Molecular, Instituto
Nacional de Investigação em Saúde (INIS), Luanda, Angola; Kumbelembe David,
Laboratório de Biologia Molecular, Instituto Nacional de Investigação em Saúde
(INIS), Luanda, Angola; Moisés Dembo, Laboratório de Biologia Molecular,
Instituto Nacional de Investigação em Saúde (INIS), Luanda, Angola; Julio
Estobre, Laboratório de Biologia Molecular, Instituto Nacional de Investigação
em Saúde (INIS), Luanda, Angola; Luzia Inglês, Laboratório de Biologia
Molecular, Instituto Nacional de Investigação em Saúde (INIS), Luanda, Angola;
Domingos Jandondo, Laboratório de Biologia Molecular, Instituto Nacional de
Investigação em Saúde (INIS), Luanda, Angola; Agostinho Paulo, Laboratório de
Biologia Molecular, Instituto Nacional de Investigação em Saúde (INIS), Luanda,
Angola; Amilton Pereira, Laboratório de Biologia Molecular, Instituto Nacional de
Investigação em Saúde (INIS), Luanda, Angola; Ramaliatou Chabi Nari,
Laboratoire des fievres hemorragiques virales de Cotonou; Akpakpa, Cotonou,
Benin; Rejeunie P.J. Mindzie Ngomo, Laboratoire des fievres hemorragiques
virales de Cotonou; Akpakpa, Cotonou, Benin; Stéphane Sohou, Laboratoire des
fievres hemorragiques virales de Cotonou; Akpakpa, Cotonou, Benin; Ragive
First release: 1 December 2022
science.org
(Page numbers not final at time of first release)
6
Retracted 20 December 2022. See Retraction.
Parode Takale, Molecular diagnostic Laboratory HDL, Pointe-Noire, Congo;
Negeri Debela, Hirsch Institute of Tropical Medicine, Asella, Ethiopia; Friederike
Hunstig, Department of Gastroenterology, Hepatology and Infectious Diseases,
University Hospital Düsseldorf, Heinrich Heine University Düsseldorf, Germany &
Hirsch Institute of Tropical Medicine, Asella, Ethiopia; Tom Lüdde, Department of
Gastroenterology, Hepatology and Infectious Diseases, University Hospital
Düsseldorf, Heinrich Heine University Düsseldorf, Germany & Hirsch Institute of
Tropical Medicine, Asella, Ethiopia; Julia Cyrielle Andeko, Centre
Interdesciplinaire de Recherches M´dicales de Franceville (CIRMF), Gabon; Lucie
Marquet, Centre Interdesciplinaire de Recherches M´dicales de Franceville
(CIRMF), Gabon; Mary Basiru Njai, Medical Research Council Unit at London
School of Hygiene and Tropical Medicine, Gambia; Ebrima Ceesay, Medical
Research Council Unit at London School of Hygiene and Tropical Medicine,
Gambia; Fatoumatta Cham, Medical Research Council Unit at London School of
Hygiene and Tropical Medicine, Gambia; Hoja Gaye, Medical Research Council
Unit at London School of Hygiene and Tropical Medicine, Gambia; Yusupha
Jallow, Medical Research Council Unit at London School of Hygiene and Tropical
Medicine, Gambia; Mamlie Touray, Medical Research Council Unit at London
School of Hygiene and Tropical Medicine, Gambia; Mariama Touray, Medical
Research Council Unit at London School of Hygiene and Tropical Medicine,
Gambia; Wendy Karen Jó Lei, Charité – Universitätsmedizin Berlin, corporate
member of Freie Universität Berlin and Humboldt Universität zu Berlin, Institute
of Virology, Germany; Anna-Lena Sander, Charité – Universitätsmedizin Berlin,
corporate member of Freie Universität Berlin and Humboldt Universität zu
Berlin, Institute of Virology, Germany; Henry Acheampong, KCCR, UPO, PMB,
KNUST, Kumasi, Ghana; Millicent Afatodzie, Noguchi Memorial Institute for
Medical Research, Ghana; Sherihane Aryeetey, KCCR, UPO, PMB, KNUST,
Kumasi, Ghana; Christopher Dorcoo, Noguchi Memorial Institute for Medical
Research, University of Ghana, Ghana; Elvis Lomotey, Noguchi Memorial
Institute for Medical Research, University of Ghana, Ghana; Daniel Odumang,
Noguchi Memorial Institute for Medical Research, University of Ghana, Ghana;
Millicent Opoku, Noguchi Memorial Institute for Medical Research, University of
Ghana, Ghana; Grace Opoku-Gyamfi, Noguchi Memorial Institute for Medical
Research, University of Ghana, Ghana; Millicent Oye Kyei, Noguchi Memorial
Institute for Medical Research, University of Ghana, Ghana; Shalyn Akasa,
National Public Health Reference Laboratory, Ministry of Health, Kenya; Claudio
Raharinandrasana, Insitut Pasteur de Madagascar - Virology Unit, Madagaskar;
Anne-Marie Ratsimbazafy, Insitut Pasteur de Madagascar - Virology Unit,
Madagaskar; Idrissa Coulibaly, Health Clinic, Sanso, Morila SA, Mali; Mamoudou
Kanoute, Health Clinic, SOMISY, Mali; Mama Kanta, Health Clinic, SOMILO, Mali;
Aminata Maiga, Laboratoire d’analyses médicales, CHU Point G, Bamako, Mali,
Mali; Didier Ndane, Health Clinic, SMK, Mali, Mali; Mamadou Sanghata, Health
Clinic, Sanso, Morila SA, Mali; Abdoul Aziz Sow, Health Clinic, SOMILO, Mali;
Abdelmajid Eloualid, Institut Pasteur du Maroc, Casablanca, Marocco; Abdellah
Faouzi, Institut Pasteur du Maroc, Casablanca, Marocco; Saloua Nadifiyine,
Institut Pasteur du Maroc, Casablanca, Marocco; Deborah Goudiaby, Institut
Pasteur de Dakar (IPD), Senegal; Davy Evrard Kiori, Institut Pasteur de Dakar
(IPD), Senegal; Marie Pedapa Mendy, Institut Pasteur de Dakar (IPD), Senegal;
Mathilda Claassen, Stellenbosch University, National Health Laboratory Service
Tygerberg, South Africa; Bronwyn Roberts, Stellenbosch University, South
Africa; Shannon Wilson, Stellenbosch University, South Africa; Fortune Salah,
Institut National d’Hygiène, Lomé, Togo; Micheline Tettekpoe, Institut National
d’Hygiène, Lomé, Togo; Fridah Aryemo, Gulu University Multifunctional Research
Laboratories, Gulu, Uganda; Doreen Ato, Gulu University Multifunctional
Research Laboratories, Gulu, Uganda; Ian Goodhead, Gulu University
Multifunctional Research Laboratories, Gulu, Uganda and School of Science,
Engineering and Environment, University of Salford, United Kingdom; Pamela
Lamwaka, Gulu University Multifunctional Research Laboratories, Gulu, Uganda.
Funding: African Academy of Sciences grant SARSCov2-4-20-004 (IOD). Bill &
Melinda Gates Foundation grant INV-005971 (JFD, CD); grant INV-024130 (IOD);
The findings and conclusions contained within are those of the authors and do
not necessarily reflect positions or policies of the Bill & Melinda Gates
Foundation. CIRMF is a member of CANTAM funding by EDCTP CSA2020NoE-
3100 – CANTAM 3 and supported by the Gabonese Government and Total
Gabon. Horizon 2020, European and Developing Countries Clinical Trials
Partnership (EDCTP2) program, PANDORA-ID-NET Grant RIA2016E-1609
(AASy, ROP). Poliomyelitis Research Foundation grant 21/40 (KKP). Programa
de Desenvolvimento de Ciência e Tecnologia - BONGOLA Project grant
Nº11/MESCTI/PDCT/2020 (JFMdM). South African Department of Science and
Innovation (sub-award via University of KwaZulu-Natal) grant S006872 (TGM).
Stellenbosch University Postgraduate Scholarship Programme grant 25095676
(KKP). UKRI Global Challenges Research Fund, grant number NF118 (RB, RE,
JMw). World Health Organization grant 2021/1113013-0 (IOD). Author
contributions: Resources: TGM, AY, NA, EA, PA, PAf, JA, SFA, LA, YA, MAB, AB,
RB, ALMB, FB, MC, PC, RMC, JC, GC, AC, UDA, XNdL, JFMdM, FD, ND, YD, LD,
PD, RE, YE, AE, OF, TF, AH, PVI, NI, RJ, SJ, BK, JK, LK, OK, VL, AL, OL, SELD,
JBLD, EL, HL, JL, SM, IM, BM, PAM, JM, LM, JMw, NN, CAN, MON, EM, RN, JN,
SGN, EOO, AO, JBO, MO IOD, KKP, ROP, WP, VR, FS, SS, AAS, AASy, PATN, ZT,
FOT, TBT. Supervision: JFD. Validation: CF, TGM, AY, JFD. Visualization: CF.
Writing – original draft: CF, JFD. Writing – review & editing: CF, TGM, AY, JFD.
Conceptualization: CF, JFD. Data curation: CF, AF. Formal Analysis: CF, AF, JFD.
Funding acquisition: TGM, RB, JFMdM, CD, IOD, KKP, ROP, AASy, JFD.
Investigation CF, TGM, AY, NA, EA, PA, PAf, JA, SFA, LA, YA, MAB, AB, RB, ALMB,
FB, MC, PC, RMC, JC, GC, AC, UDA, XNdL, JFMdM, FD, ND, YD, LD, PD, RE, YE,
AE, OF, TF, AH, PVI, NI, RJ, SJ, BK, JK, LK, OK, VL, AL, OL, SELD, JBLD, EL, HL,
JL, SM, IM, BM, PAM, JM, LM, AMS, JMw, NN, CAN, MON, EM, RN, JN, SGN, EOO,
AO, JBO, MO IOD, KKP, ROP, WP, VR, FS, SS, AAS, AASy, PATN, ZT, FOT, TBT.
Methodology: CF, JFD. Project Administration: AK, JFD. Competing interests:
Olfert Landt is the former owner of TIB Molbiol, the company that produced the
kits provided to African partner laboratories within the framework of this study.
The kits are not commercially available. All authors declare that they have no
competing interests. Data and materials availability: Respiratory samples sent
to laboratories for COVID-19 testing by attending physicians were re-tested in an
anonymized fashion using the COVID-19 test developed in this study. Ethical
approval for re-testing and scientific usage was provided by the institutional
research ethics board (IRB) of Charité-Universitätsmedizin Berlin (EA2/028/22)
and by IRBs from Burkina Faso, Laboratoire National de Référence-Grippes
(2020-7-126), Cameroon, Centre Pasteur du Cameroun
(2020/05/1224/CE/CNERSH/SP), Ghana, Kumasi Centre for Collaborative
Research in Tropical Medicine (KCCR), KNUST (CHRPE/AP/566/21), Kenya,
Jomo Kenyatta University of Agriculture and Technology, Department of
Biochemistry (JKU/2/4/896B), Uganda, Gulu University Multifunctional
Laboratories (GUREC-093-20),, Makerere University, College of Health Science,
Kamala, Uganda (SBS-2022-130) and ZALMBia, Tropical Diseases Research
Centre, Ndola Teaching Hospital (00003729). In all other countries, IRB
approval for re-testing anonymized specimens was not required. GISAID
sequences used for analyses and R scripts can be found on GitHub (40). Genome
sequences generated in this study from samples from Benin are available at
GenBank (submission numbers OP537261 - OP537513), sequences from South
Africa are available in GISAID (40). License information: This work is licensed
under a Creative Commons Attribution 4.0 International (CC BY 4.0) license,
which permits unrestricted use, distribution, and reproduction in any medium,
provided the original work is properly cited. To view a copy of this license, visit
https://creativecommons.org/licenses/by/4.0/. This license does not apply to
figures/photos/artwork or other content included in the article that is credited
to a third party; obtain authorization from the rights holder before using such
material.
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add8737
Materials and Methods
Figs. S1 to S9
Tables S1 to S7
References (41–46)
MDAR Reproducibility Checklist
Submitted 11 July 2022; accepted 23 November 2022
Published online 1 December 2022
10.1126/science.add8737
First release: 1 December 2022
science.org
(Page numbers not final at time of first release)
7
Retracted 20 December 2022. See Retraction.
Fig. 1. Study setup. (A) Timeline of known events in the evolution and emergence of Omicron. (B) Geographic
distribution of sampling sites and countries (alpha-3 country codes) represented in this study.
First release: 1 December 2022
science.org
(Page numbers not final at time of first release)
8
Retracted 20 December 2022. See Retraction.
Fig. 2. Epidemiology of Omicron/BA.1 in Africa. (A) Fraction of samples positive for the Delta marker.
(B) Fraction positive for the BA.1 marker. (C) Modeled increase in BA.1 fraction of all SARS-CoV-2 infections per
African region based on PCR testing. (D) Days until BA.1 became the dominant SARS-CoV-2 variant after its first
detection by PCR. (E) Rt and the incidence among countries represented in this study. (F) Daily reported SARS-
CoV-2 cases in Africa (28).
First release: 1 December 2022
science.org
(Page numbers not final at time of first release)
9
Retracted 20 December 2022. See Retraction.
Fig. 3. Evolution of Omicron/BA.1 in Africa. (A) Time-resolved Nextstrain phylogeny of BA.1. (B) Approximate
Maximum Likelihood phylogeny of African Omicron ancestors. (C) BA.1-defining mutations in samples from
Benin (9). Asterisks, spike amino acid substitutions H655Y and N969K. HTS, high-throughput sequencing.
Lineage assignment was conducted using the Nextclade (https://clades.nextstrain.org/) and UShER
(https://genome.ucsc.edu/cgi-bin/hgPhyloPlace) online tools. Benin-derived sequences are available in
GenBank (accession numbers OP537480-OP537485).
First release: 1 December 2022
science.org
(Page numbers not final at time of first release) 10
Retracted 20 December 2022. See Retraction.
|
10.1126_science.ade1702
|
RES EARCH
SOCIAL LEARNING
Social signal learning of the waggle dance
in honey bees
Shihao Dong1†, Tao Lin1†, James C. Nieh2*, Ken Tan1*
Honey bees use a complex form of spatial referential communication. Their “waggle dance”
communicates the direction, distance, and quality of a resource to nestmates by encoding celestial
cues, retinal optic flow, and relative food value into motion and sound within the nest. We show that
correct waggle dancing requires social learning. Bees without the opportunity to follow any dances
before they first danced produced significantly more disordered dances with larger waggle angle
divergence errors and encoded distance incorrectly. The former deficit improved with experience,
but distance encoding was set for life. The first dances of bees that could follow other dancers
showed neither impairment. Social learning, therefore, shapes honey bee signaling, as it does
communication in human infants, birds, and multiple other vertebrate species.
learn resource location and quality. However,
it has not been previously determined wheth-
er dance following can improve the dance
performances of young waggle dancers or
whether the dance is completely genetically
preprogrammed (innate).
The waggle dance is a sophisticated form
of spatial referential communication (9). The
dancer repeatedly circles in a figure-eight pat-
tern centered around a waggle run in which
the bee waggles its abdomen as it moves for-
ward (Fig. 1). Referential communication codes
information, and the dancer encodes the po-
lar coordinates of a resource relative to the
nest. Longer waggle runs communicate greater
distances (more retinal optical flow), and the
waggle direction angle communicates resource
direction. When a bee dances on a vertical comb
in the dark, the bee points in the direction of
the resource relative to the sun, as transposed
to the vertical in relation to gravity. The qual-
ity of the food relative to colony need and the
dancer’s prior experiences (10) are encoded in
the number of waggle run repetitions and the
speed with which the dancer returns to re-
peat each successive waggle run (11).
There is a strong genetic component to the
dance: Different honey bee species have dis-
tinctive distance encodings (calibrations) that
persist even when they are cross-fostered
(12, 13). An encoding is a curve that describes
the relationship between physical distance
and the duration of waggle runs for resources
at those distances (14). Theoretically, novice
dancers could benefit by learning from ex-
perienced dancers because waggle dancing
requires retrieving navigational memory and
using detailed motor programs and real-
time feedback to translate resource location
(15). Dances occur on the dance floor, which
often consists of colony-specific, uneven, and
convoluted comb surfaces (Fig. 1 and fig. S1) (16)
that dancers must negotiate at relatively high
velocities. On average, they cover more than
their body length in 1 s (waggle running at
S ocial learning occurs when one individ-
ual learns by observing or interacting
with another (1) and is particularly use-
ful when complex behaviors must be
tuned to specific environmental circum-
stances or honed by practice or social shap-
ing. For example, human infant babbling and
young songbird subsongs are shaped by social
feedback into more mature vocal behavior (2),
and young naked mole rats learn distinctive
colony dialects from older rats (3). Longer
periods of interaction, such as those occur-
ring between parents and offspring, can favor
the evolution of such open programs (4),
which allow novices to acquire skills more
rapidly from experienced individuals than
they could on their own (5). Proficient individ-
uals have had more opportunities to fine-tune
their brains and motor outputs to environ-
mental circumstances (5); thus, learning from
them can be beneficial.
Eusocial insects use social learning, but it is
unclear whether this learning shapes their
communication, which can be remarkably so-
phisticated and cognitively complex. Polistes
fuscatus wasps use social eavesdropping, a
form of social learning, to observe conflicts
and to assess and remember rivals through
facial recognition (6). Bumble bees can learn
by observation to copy or avoid the foraging
choices of other bumble bees through their
previous experiences of reward or punish-
ment (7). These bees can also learn to obtain
a nectar reward by watching their nestmates
perform a new behavior and can then in-
novate and solve the problem more effi-
ciently (8). Honey bee workers use social
learning when following the waggle dance to
1CAS Key Laboratory of Tropical Forest Ecology, Xishuangbanna
Tropical Botanical Garden, Chinese Academy of Sciences,
Kunming 650000, Yunnan, China. 2School of Biological
Sciences, Department of Ecology, Behavior, and Evolution,
University of California San Diego, La Jolla, CA 92093, USA.
*Corresponding author. Email: kentan@xtbg.ac.cn (K.T.);
jnieh@ucsd.edu (J.C.N.)
†These authors contributed equally to this work.
Fig. 1. Waggle dance directional error was highest in the first dances of naive bees that could
not follow waggle dances. (A) The dancer (w) shakes its abdomen (i-ii-iii, creating one cycle) during
the waggle run (1-2-3), whose angle (a) communicates direction, and then makes a semicircular
return while being tracked by dance followers (f). (B) Divergence error angles decreased with experience
in experimental colonies but not in control colonies, in which errors were consistently low (different
letters indicate significant differences, Tukey HSD test, P < 0.05). (Inset) Dancers typically perform
on irregular surfaces that vary between colonies. Data (black circles), notched box plots, and violin plots
are shown in all figures.
Dong et al., Science 379, 1015–1018 (2023)
10 March 2023
1 of 4
RES EARCH | R E S E A R C H A R T I C L E
15 mm/s and returning back at 20 mm/s) while
sensing the positions of their bodies relative
to gravity and producing the correct waggle
frequency and angle (17). Thus, errors occur.
A dancer’s successive waggle runs can point
to different angles, resulting in directional er-
rors (18). Similarly, waggle runs within the
same dance can vary in duration, conveying
distance errors (19).
Foragers have the opportunity to learn from
more experienced dancers. Workers become
foragers as they age: They begin following wag-
gle dancers when they are 8 days old and sub-
sequently perform their first waggle dances
when they are 12 days old (20). All workers
follow waggle dances before they waggle dance
(20), and most follow dances performed by
older bees that have previously danced (11).
We therefore predicted that the first waggle
dances of foragers will exhibit more errors if
they are reared in an environment in which
they cannot follow other waggle dancers be-
fore they begin to dance.
We therefore created colonies in which we
observed the first waggle dances produced by
foragers (all individually marked and trained
to 55% w/v sucrose feeders located 150 m from
their colonies) that either could or could not fol-
low other waggle dancers (table S1). Each of our
five experimental colonies was established with
a single cohort of 1-day-old bees. As these bees
aged, we monitored the colonies until we ob-
served the first waggle dances and then ob-
served the same dancers 20 days later when
they had more foraging and dancing experi-
ence. Naive dancers could not follow any other
dancers before their first dances because all
bees in the colony were the same age, but as
these dancers grew older, they followed other
waggle dancers and had more experience danc-
ing. In five control colonies that we established
at the same time with adult bees of all ages and
in which we observed waggle dancing within
1 to 2 days of colony creation, we measured
the waggle dances of control bees at two com-
parable stages: the first waggle dances in the
control colonies (C1First Dances) and the waggle
dances of the same dancers 20 days later when
they had more foraging and dancing experi-
ence (C2Older Dancers).
We observed no waggle dancing in all exper-
imental colonies before the first group of bees
aged into foraging and dancing (E1First Dances naive;
9.0 ± 2.0 days old). Although we did not track
all behaviors of these same bees until 20 days
later, when they were older and had experi-
ence dancing and following other dancers
(E2Older Dancers), on each observation day we
saw multiple E1First Dances naive bees following
waggle dancers for natural food sources. In
all control colonies, we had a marked cohort
of bees of known age and likewise observed
that they followed waggle dances before they
performed their first dances (C1First Dances;
Table 1. Summary of statistical results for all experiments. Colony type is either experimental
(E) or control (C), and time point refers to (1) the first dances of bees or (2) subsequent dances of
the same bees observed 20 days later.
Measure
Model R2
adj Colony type
Time point
Interaction, colony type
by time point
0.12
0.57
0.79
0.34
0.02
waggle run
waggle run CV
F1,34 = 1.52,
P = 0.23
F1,34 = 12.93,
P = 0.001
F1,34 = 1.32,
P = 0.26
F1,34 = 0.09,
P = 0.77
F1,34 = 0.28,
P = 0.60
F1,34 = 1.01,
P = 0.32
F1,34 = 0.94,
P = 0.34
F1,34 = 12.48,
P = 0.0012
F1,34 = 0.67,
P = 0.42
F1,34 = 14.99,
P = 0.0005
F1,34 = 10.18,
P = 0.003
F1,34 = 1.46,
P = 0.24
F1,34 = 3.83,
P = 0.06
F1,34 = 22.80,
P < 0.0001
Food direction
.....................................................................................................................................................................................................................
F1,30 = 5.85,
P = 0.02
Divergence angle
.....................................................................................................................................................................................................................
Food distance
.....................................................................................................................................................................................................................
F1,32 = 157.20,
P < 0.0001
Waggle duration
.....................................................................................................................................................................................................................
F1,30 = 20.08,
P = 0.0001
Waggle duration range error
.....................................................................................................................................................................................................................
F1,30 = 0.03,
P = 0.86
Waggle duration CV
.....................................................................................................................................................................................................................
F1,27 = 88.26,
Number of waggles per
P < 0.0001
.....................................................................................................................................................................................................................
F1,28 = 0.87,
Number of waggles per
P = 0.36
.....................................................................................................................................................................................................................
F1,29 = 8.56,
P = 0.007
Return flight time
.....................................................................................................................................................................................................................
Food quality
.....................................................................................................................................................................................................................
F1,30 = 4.99,
P = 0.03
Number of waggle runs
.....................................................................................................................................................................................................................
F1,30 = 11.68,
P = 0.002
Return-phase duration
.....................................................................................................................................................................................................................
F1,31 = 0.13,
P = 0.72
Return-phase CV
.....................................................................................................................................................................................................................
Dance quality
.....................................................................................................................................................................................................................
F1,30 = 7.35,
P = 0.011
Disorder proportion
.....................................................................................................................................................................................................................
F1,31 = 180.07,
P < 0.0001
Number of followers
.....................................................................................................................................................................................................................
F1,34 = 21.58,
P < 0.0001
F1,34 = 15.15,
P = 0.0004
F1,34 = 0.60,
P = 0.45
F1,34 = 1.35,
P = 0.25
F1,34 = 2.54,
P = 0.12
F1,34 = 0.05,
P = 0.82
F1,34 = 20.43,
P < 0.0001
F1,34 = 49.60,
P < 0.0001
F1,34 = 6.46,
P = 0.02
F1,34 = 17.46,
P = 0.0002
<0.001
0.08
0.03
0.56
0.52
0.76
0.12
9.9 ± 1.0 days old) and continued to follow
waggle dances over the next 20 days. All sta-
tistical results are reported in Table 1.
Food direction and distance
E1First Dances naive bees had significantly greater
divergence angles (higher directional error)
that decreased when they became E2Older Dancers
bees [Tukey honestly significant difference
(HSD) test, P < 0.05, Fig. 1B]. The dances of
C1First Dances and C2Older Dancers bees did not
have significantly different divergence errors.
The dances of E1First Dances naive and E2Older Dancers
bees had longer waggle run durations than
those of C1 First Dances or C2 Older Dancers bees
(Tukey HSD test, P < 0.05, Fig. 2), suggesting
that distance encoding was disrupted when
bees could not follow experienced dancers
and that disruption persisted even after they
had more practice dancing and following
other dancers. The reasons for this disruption
are unclear, but E1First Dances naive foragers
had longer return flight times than those of all
other bee types (Tukey HSD test, P < 0.05). If
E1First Dances naive bees thereby experienced
greater retinal optic flow, this should translate
into longer waggle run durations (21). However,
when the same bees were 20 days older, they had
shorter flight durations and yet persisted in
making the same distance-encoding errors.
The waggle duration range error was signif-
icantly higher in the dances of E1First Dances naive
bees than in those of C1First Dances or C2Older Dancers
bees (Tukey HSD test, P < 0.05), although it
was not different between E1First Dances naive
and E2Older Dancers bees, again suggesting a
lifetime disruption of distance communica-
tion as a result of our treatment. In accordance
with the waggle duration trends, the dances of
E1First Dances naive and E2Older Dancers bees had more
waggles per waggle run than those of C1First Dances
or C2Older Dancers bees (Tukey HSD test, P < 0.05).
There were no significant differences between
coefficients of variation (CV) for waggle run
duration or the number of waggles per wag-
gle run (Tukey HSD test, P > 0.05).
Food quality
Bees signal higher-quality food relative to colony
needs by increasing the number of waggle runs
Dong et al., Science 379, 1015–1018 (2023)
10 March 2023
2 of 4
RES EARCH | R E S E A R C H A R T I C L E
per dance and performing shorter return phases
(11). In general, our dancers tended to signal a
higher value for identical sucrose solutions
when their colonies were older and larger
(E2Older Dancers and C2Older Dancers phases) than
when their colonies were smaller and younger
(Fig. 3A, E1First Dances naive and C1First Dances
phases), perhaps reflecting greater colony
need. The dances of E2Older Dancers bees had
significantly more waggle runs than those
of E1First Dances naive bees, but in control
colonies there were no significant differences
between the first waggle dances of bees and
their waggle dances 20 days later (Tukey HSD
tests, P > 0.05). Return-phase durations were
only shorter for C2Older Dancers bees as com-
pared with all other forager types (Tukey HSD
tests, P < 0.05). There were no significant dif-
ferences in return-run duration CV (Tukey HSD
test, P > 0.05).
Dance quality
The dances of E1First Dances naive bees were sig-
nificantly more disordered than the dances
of E2Older Dancers, C1First Dances, or C2Older Dancers
bees (Tukey HSD test, P < 0.05, Fig. 3B). The
number of dance followers per dance was sig-
nificantly lower for experimental colonies than
for control colonies (Tukey HSD test, P < 0.05),
but was not different between E1First Dances naive
and E2Older Dancers bees. Increasing dance dis-
order was positively correlated with higher di-
vergence angle errors for E1First Dances naive and
C2Older Dancers bees (F1,16 ≥ 4.72, P ≤ 0.045)
but not for E2First Dances or C1First Dances bees
(F1,16 ≤ 0.43, P ≥ 0.52, Fig. 3C).
Our results suggest that social signal learn-
ing can improve waggle dancing. The dances
of E1First Dances naive bees who could not follow
dances before they first danced had greater
divergence angle errors, signaled greater dis-
tances, and were significantly more disordered
than those of C1First Dances bees that were exposed
to waggle dancing. Once the same bees were
older and had experience with dance follow-
ing and dancing (E2Older Dancers), they signifi-
cantly decreased divergence angle errors and
performed more orderly dances. However, they
were never able to produce normal distance
encoding. Greater age, more experience follow-
ing dances, additional practice with flying and
foraging, or a combination of these factors
could account for the improvements between
E2Older Dancers and E1First Dances naive dances.
Control bees improved by reducing distance
range errors only when they were 20 days older
(C2Older Dancers versus C1First Dances). Following
experienced dancers before they first danced was
sufficient for C1First Dances bees to correctly order
their dances with the lower number of direc-
tional errors typical of older, experienced bees.
Why should honey bees use social learning
to improve their waggle dancing? Learning is
a useful way to refine behaviors for local con-
Food distance
A
)
s
m
(
s
n
o
i
t
a
r
u
d
n
u
r
e
l
g
g
a
W
r
o
r
r
e
e
g
n
a
r
n
o
i
t
a
r
u
d
n
u
r
e
l
g
g
a
W
n
u
r
e
l
g
g
a
w
r
e
p
s
e
l
g
g
a
w
f
o
.
o
N
)
s
(
s
e
m
i
t
t
h
g
i
l
f
n
r
u
t
e
R
500
400
300
200
100
0
300
200
100
0
6
5
4
3
2
1
0
200
150
100
50
0
A
A
B
B
AB
B
C
A
A
B
B
A
B
B
B
E1 First Dances naive
E2 Older Dancers C1 First Dances C2 Older Dancers
Bee type
Fig. 2. Naive dancers that could not follow
other dancers had disrupted distance encoding
(waggle run durations and the number of
waggles per waggle run) that persisted
throughout their lifetimes. However, return
flight times in experimental colonies significantly
declined with experience. Different letters indicate
significant differences.
ditions. We suggest that the distinct topologies
of each colony’s dance floor make it advanta-
geous for novice dancers to learn from more
experienced ones. Another possibility is that
experienced dancers could transmit to nest-
mates distance encodings that are based on
local optic flow. Theoretically, distance encod-
ings should be optimized according to the en-
vironment: the locations of food and the
amount of optic flow that foragers experi-
ence when flying to this food. Because honey
Fig. 3. Dance disorder was highest in naive
first dancers and was positively correlated
with angular error. Between groups, there were
changes in (A) the communication of food quality
and (B) dance quality and the number of dance
followers (different letters indicate significant
differences). (C) Directional error was positively
correlated with dance disorder in E1First Dances naive
and C2Older Dancers bees.
Dong et al., Science 379, 1015–1018 (2023)
10 March 2023
3 of 4
5. C. P. van Schaik, in Animal Behaviour: Evolution
25. J. B. Free, Y. Spencer-Booth, Proc. R. Entomol. Soc. Lond., Ser.
RES EARCH | R E S E A R C H A R T I C L E
bee species can inhabit very different envi-
ronments, distance encodings can be signif-
icantly different between species (14) and
within species for Apis florea (22) and Apis
mellifera (23). Given the imprecision inher-
ent in waggle dances, the importance of these
differences is not clear. Schürch et al. (24) com-
pared the distance encodings of A. mellifera
dancers in environments with different optic
flow levels and found significant differences
in the encoding line intercepts but not in the
slopes. Our results indicate that we perma-
nently altered distance encoding in our experi-
mental colonies: After our treatment, novice
dancers continued to make the same distance-
encoding errors even near the end of their
adult lives (25) despite decreasing their di-
rectional errors and dance disorder. Some
aspects of the waggle dance can evidently be
altered in young bees and are irreversible.
Thus, we argue that the cultural modification
and transmission of signals may be possible in
social insects.
RE FE RENCES AND N OT ES
and Mechanisms, P. Kappeler, Ed. (Springer, 2010),
pp. 623–653.
6. E. A. Tibbetts, E. Wong, S. Bonello, Curr. Biol. 30, 3007–3010.
e2 (2020).
7. E. H. Dawson, A. Avarguès-Weber, L. Chittka, E. Leadbeater,
Curr. Biol. 23, 727–730 (2013).
8. O. J. Loukola, C. Solvi, L. Coscos, L. Chittka, Science 355,
833–836 (2017).
9. K. von Frisch, The Dance Language and Orientation of Bees
(Belknap Press, 1967).
10. C. Grüter, T. J. Czaczkes, Anim. Behav. 151, 207–215
(2019).
11. T. D. Seeley, The Wisdom of the Hive: The Social Physiology of
Honey Bee Colonies (Harvard Univ. Press, 1995).
12. S. Su et al., PLOS ONE 3, e2365 (2008).
13. K. Tan et al., Naturwissenschaften 95, 1165–1168 (2008).
14. P. L. Kohl et al., Proc. Biol. Sci. 287, 20200190
(2020).
15. A. B. Barron, J. A. Plath, J. Exp. Biol. 220, 4339–4346
(2017).
16. M. L. Smith, N. Napp, K. H. Petersen, Proc. Natl. Acad. Sci. U.S.A.
118, e2103605118 (2021).
17. T. Landgraf, R. Rojas, H. Nguyen, F. Kriegel, K. Stettin,
PLOS ONE 6, e21354 (2011).
18. K. Preece, M. Beekman, Anim. Behav. 94, 19–26
(2014).
19. R. J. De Marco, J. M. Gurevitz, R. Menzel, J. Exp. Biol. 211,
1635–1644 (2008).
20. H. Ai et al., bioRxiv 179408 [Preprint] (2017). https://doi.org/
10.1101/179408
21. H. E. Esch, S. Zhang, M. V. Srinivasan, J. Tautz, Nature 411,
1. E. Leadbeater, L. Chittka, Curr. Biol. 17, R703–R713
581–583 (2001).
(2007).
22. E. A. George et al., J. Exp. Biol. 224, jeb242404
2. M. H. Goldstein, A. P. King, Proc. Natl. Acad. Sci. U.S.A. 100,
(2021).
8030–8035 (2003).
3. A. J. Barker et al., Science 371, 503–507 (2021).
4. E. Mayr, Am. Sci. 62, 650–659 (1974).
23. J. Tautz et al., PLOS Biol. 2, E211 (2004).
24. R. Schürch et al., Anim. Behav. 150, 139–145
(2019).
A Gen. Entomol. 34, 141–150 (1959).
26. S. Dong, T. Lin, J. C. Nieh, K. Tan, Social signal learning of the
waggle dance in honey bees, Zenodo (2023); https://doi.org/10.
5281/zenodo.7301648.
AC KNOWLED GME NTS
Funding: This work was supported by the CAS Key Laboratory of
Tropical Forest Ecology, Xishuangbanna Tropical Botanical Garden,
Chinese Academy of Sciences. Additional funding was provided
by the CAS 135 program (2017XTBG-T01) and the National Natural
Science Foundation of China (31770420) to K.T. Author
contributions: All authors contributed to the conceptualization
and design of this research. S.D. and T.L. conducted the
experiment, J.C.N analyzed the data, and S.D., K.T., and J.C.N.
contributed to the writing of the manuscript. Competing
interests: The authors declare that they have no competing
interests. Data and materials availability: Data are available at
Zenodo (26). License information: Copyright © 2023 the authors,
some rights reserved; exclusive licensee American Association
for the Advancement of Science. No claim to original US
government works. https://www.science.org/about/science-
licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade1702
Materials and Methods
Fig. S1
Table S1
Movies S1 to S2
References (27–30)
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol
Submitted 28 July 2022; accepted 19 January 2023
10.1126/science.ade1702
Dong et al., Science 379, 1015–1018 (2023)
10 March 2023
4 of 4
|
10.1126_science.ade2038
|
RES EARCH
BIOMATERIALS
Deformable hard tissue with high fatigue resistance
in the hinge of bivalve Cristaria plicata
Xiang-Sen Meng1†, Li-Chuan Zhou2,3†, Lei Liu1†, Yin-Bo Zhu2, Yu-Feng Meng1, Dong-Chang Zheng2,
Bo Yang1, Qi-Zhi Rao4, Li-Bo Mao1*, Heng-An Wu2*, Shu-Hong Yu1,5*
The hinge of bivalve shells can sustain hundreds of thousands of repeating opening-and-closing valve
motions throughout their lifetime. We studied the hierarchical design of the mineralized tissue in the
hinge of the bivalve Cristaria plicata, which endows the tissue with deformability and fatigue resistance
and consequently underlies the repeating motion capability. This folding fan–shaped tissue consists of
radially aligned, brittle aragonite nanowires embedded in a resilient matrix and can translate external
radial loads to circumferential deformation. The hard-soft complex microstructure can suppress stress
concentration within the tissue. Coherent nanotwin boundaries along the longitudinal direction of the
nanowires increase their resistance to bending fracture. The unusual biomineral, which exploits the
inherent properties of each component through multiscale structural design, provides insights into
the evolution of antifatigue structural materials.
B rittle materials are extensively used as
structural or functional components in
various fields such as aerospace, tissue
engineering, and electronics (1–3). How-
ever, artificial brittle materials are sen-
sitive to microcracks and imperceptible defects
and thus suffer from the risk of cumulative
fatigue damage caused by the prolonged cyclic
loading, which can eventually cause catastrophic
failure (4–7). In this respect, living organisms
produce rigid biominerals such as nacre (8),
bone (9), and tooth (10) in which the fatigue
tolerance is greatly improved but the native
high strength and rigidity of the minerals are
retained. The fatigue damage in these biomin-
erals, such as crack propagation, can be shielded
by extrinsic mechanisms, including crack bridg-
ing, deflection, and branching (8–11).
Although these biominerals have inspired
fatigue-tolerant artificial materials such as
tough biomimetic ceramics (12, 13), more and
diverse design principles have to be uncovered
to extend the scope of artificial materials. For
example, materials with high flexibility have
received intense focus owing to the develop-
ment of foldable and wearable devices (14, 15).
Existing antifatigue models derived from very
1Department of Chemistry, Institute of Biomimetic Materials
and Chemistry, Anhui Engineering Laboratory of Biomimetic
Materials, New Cornerstone Science Laboratory, Division of
Nanomaterials and Chemistry, Hefei National Research
Center for Physical Sciences at the Microscale, University of
Science and Technology of China, Hefei 230026, China.
2CAS Key Laboratory of Mechanical Behavior and Design of
Materials, Department of Modern Mechanics, CAS Center
for Excellence in Complex System Mechanics, University of
Science and Technology of China, Hefei 230027, China.
3School of Mechanical Engineering, Hefei University of
Technology, Hefei 230009, China. 4Anhui Shuyan Intelligent
Technologies Co., Wuhu 241200, China. 5Institute of
Innovative Materials, Department of Materials Science and
Engineering, Department of Chemistry, Southern University
of Science and Technology, Shenzhen 518055, China.
*Corresponding author. Email: maolb@ustc.edu.cn (L.-B.M.);
wuha@ustc.edu.cn (H.-A.W); shyu@ustc.edu.cn (S.-H.Y.)
†These authors contributed equally to this work.
rigid biominerals such as nacre are not effective
for the design of such flexible materials. Besides
the lack of flexibility, the fatigue tolerance of
these models strongly relies on the rising R-curve
behavior during crack propagation (11), yet the
crack extension can cause irreversible impacts
on the device performances (16, 17). This indi-
cates that the protection of the brittle compo-
nents in such devices against fatigue should
not be overdependent on the mechanisms that
only come into effect in the crack wake (2, 11).
Macrostructures and mechanical performance
We report the antifatigue design of the de-
formable calcareous tissue in the hinge of
the bivalve shell Cristaria plicata (Fig. 1A, i,
and fig. S1A) (18). This tissue exhibits both
high deformability and exceptional fatigue
resistance during the hundreds of thousands
of repeating opening-and-closing (ROAC) mo-
tions of the shell valves. The hinge is located
at the dorsal edge of the bivalve shell, by which
the two valves are joined together, and func-
tions as the axis of the ROAC motions (Fig. 1,
A, ii, and B, i; fig. S1; and movie S1) (19). It un-
dergoes large deformation during the clos-
ing process of two rigid valves driven by the
adductor muscles and provides the driving
force for the spontaneous valve opening by
releasing the stored elastic potential energy
(movie S2) (20).
To verify the mechanical and antifatigue
performance of the hinge, we performed cyclic
loading tests that simulated the ROAC mo-
tions on freshly prepared C. plicata samples
(fig. S2A and movie S3) (21). No obvious fatigue
failure was observed even after 1,500,000 cycles
under natural working conditions (Fig. 1C and
figs. S3 and S4A). We also imposed a larger
load (about 20% overload) on the edge-trimmed
samples (fig. S2), and the hinge exhibited
similar behavior in this case (Fig. 1C and figs. S4B
and S5). To investigate the antifatigue per-
formance of the hinge under much larger loads,
we further increased the load stepwise to about
100% overload in the cyclic tests (fig. S6). Its
function is still largely preserved, although the
resilience of the hinge decreases under such
conditions.
Reconstructed x-ray microcomputed tomo-
graphic (mCT) images reveal the inhomoge-
neous electron density distribution in the hinge,
which is surrounded by a porous tissue and two
slab areas (Fig. 1B, ii, and fig. S1F). Given the
differences in both optical and mCT images
(Fig. 1B), the hinge area can be divided into two
distinct regions: a folding fan–shaped region
(FFR) and an outer ligament (OL). Elemental
mapping images reveal a sharp increase of cal-
cium concentration from the OL to the FFR
(Fig. 1D), which is consistent with the elastic
modulus and hardness variations of the re-
gions in the dried hinge and the density dif-
ference in the mCT image (Fig. 1E and fig. S7)
(22). This correlation can be explained by cal-
cium carbonate in the form of aragonite being
the only mineral phase in the two hinge re-
gions that can be observed with powder x-ray
diffraction (fig. S8) (23).
We monitored the two hinge regions to un-
ravel their roles during the ROAC motions
(Fig. 2A and movie S4). During closing, the
FFR region is pushed by the rotating slab areas
that are bound to the valves. The two sides of
the FFR then rotate, and the whole FFR re-
gion simultaneously bends and deforms in the
circumferential direction, which is accompa-
nied with the stretch of the OL (Fig. 2A). The
FFR undertakes a minor radial deformation
and provides robust radial support to fix the
OL to ensure effective OL circumferential stretch.
We validated this behavior by means of finite
element analysis (FEA) during the closing
process (Fig. 2B, fig. S9, and movie S5), and
the behavior reverses during opening. There-
fore, the structure of the FFR, which is sim-
ilar to that of traditional arch structures, can
effectively translate the external radial load
to the circumferential deformation. We also
calculated the stress distributions in the FEA
model along both circumferential and radial
directions at the open and closed states, re-
spectively (Fig. 2, C to E, and supplementary
text 1). The circumferential tensile stress in
the hinge is mainly borne by the OL (Fig. 2D
and fig. S10), which agrees with the mea-
sured result that the stress along the circum-
ferential direction is much higher in the OL
than that in the FFR (Fig. 2, F and G, and
fig. S10).
In addition, given that the OL and the out-
most edge of the FFR are bound together,
the deformations are supposed to be similar
for both when stretched. The elastic strain
energy released by the FFR and the OL was
derived by integrating the products of the
energy density of each element by its volume
Meng et al., Science 380, 1252–1257 (2023)
23 June 2023
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
A z
x
B
z
z
y
y
i
ii
i
ii
C
150
)
%
(
d
a
o
l
e
v
i
t
a
e
R
l
120
90
60
30
MaxNWC
MinNWC
Max120%
Min120%
Hinge
FFR
OL
SA
PT
Valve
1
10,000
100,000
Number of cycles
1,000,000
D
Ca
E
O
y
t
i
s
n
e
t
n
i
e
v
i
t
a
e
R
l
O
OL
FFR
0
600
1200
Distance (µm)
1800
4
3
2
1
0
)
a
P
G
(
s
s
e
n
d
r
a
H
Ca
SA
2400
)
a
P
G
(
s
s
e
n
d
r
a
H
3.0
2.0
1.0
0.0
OL
FFR
SA
60
40
20
0
64
48
32
16
0
)
a
P
G
(
l
s
u
u
d
o
M
)
a
P
G
(
l
s
u
u
d
o
M
0
150
300
Distance (µm)
450
600
Fig. 1. Structural features and mechanical properties of the hinge. (A) (i)
Optical image of C. plicata and (ii) sectional photograph of the shell at the cutting
plane (dotted rectangle) shown in (i). The yellow dash-dotted line in (i) represents
the hinge axis as well as the axis of the ROAC motions. This axis is indicated
with the circled dot in (ii). Scale bars, 2 cm. (B) (i) Optical image and (ii)
reconstructed three-dimensional mCT image of a polished hinge sample viewed in the
longitudinal direction [(A), ii, dashed box]. The different colors in (ii) indicate the
distinct regions in the hinge: FFR, blue; OL, saddle brown; slab areas (SAs), sienna;
porous tissue (PT), light cyan; and valve, pale yellow. The SAs exhibit a slightly
lighter gray color than that of the nacreous part of the valves, indicating a higher
electron density. Scale bars, 1 cm. (C) Relative maximum and minimum loads in
the fatigue tests under the natural working condition (NWC) and overloading
condition (about 20% overload) (supplementary materials, materials and methods).
(D) Elemental maps and line scans showing the calcium (Ca) and the oxygen (O)
distributions in the yellow square in (B), in which the FFR, the OL, and the SAs
are included. Scale bars, 500 mm. (E) Maps of the elastic modulus and hardness of
the yellow square in (B) and the corresponding line scans from the OL to the FFR
and then the SA of a dried sample. Scale bars, 200 mm.
in the numerical model. The results suggest
that the OL stores most of the elastic strain
energy during closing, which is then released
to sustain valve opening (supplementary text
1 and fig. S11). Compressive tests show that
the FFR is highly anisotropic (Fig. 2, G and
H). In the radial direction, the whole FFR is
under compressive stress (Fig. 2E), which is
associated with the radial support of the FFR
to the OL. Nevertheless, the deformation of
the FFR in this direction is limited by the space
of the entire hinge area (Fig. 2, A and B). Com-
pensatorily, the FFR has a much larger tangent
modulus along the radial direction compared
with that along the circumferential direction
(Fig. 2, G and H). Therefore, it can translate
sufficient radial load to support the OL within
a very small radial deformation. By contrast,
the FFR can deform much easier circumfer-
entially, which allows the FFR to adapt itself
to the hinge deformation. These observations
reveal that as a dense and relatively rigid cal-
careous tissue (figs. S8 and S12 and tables S1
and S2), the FFR endures a large radial load to
support the OL while undertaking a large cir-
cumferential deformation during the ROAC mo-
tions of the valves without fatigue failure (more
than 1,500,000 cycles). This distinguishes the
FFR from other biominerals such as nacre, cor-
tical bone, and human enamel (table S3). Once
produced, the acellular FFR begins to func-
tion during the ROAC motions. Yet it neither
comes in contact with cells, like the nacreous
inner layer of the valves does, nor contains any
cells inside and thus cannot be repaired after
being damaged, suggesting that the FFR has
to be inherently robust (fig. S13) (24, 25).
Microstructures and crystallographic features
To understand the mechanisms of the mechan-
ical function and antifatigue performance of
the FFR, we analyzed its microstructures and
crystallographic features. The fracture surface
of the FFR exhibits a concentrically laminated
structure (Fig. 1B, i, and fig. S14). Each layer
consists of tightly aligned and stacked, long
nanowires with diameters of about 100 to
200 nm (Fig. 3A). Detailed measurements re-
veal that the diameters increase gradually
from the inner part of one nanowire layer to
the outer part (fig. S15). Such laminated structure
Meng et al., Science 380, 1252–1257 (2023)
23 June 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Transmission relationship of the ROAC motions. (A) Optical images
and (B) FEA models of the hinge in an open state and a closed state. The contours of
the FFR and the OL at the two states are compared. Yellow, open state; red, closed
state. Scale bar, 500 mm. (C to E) FEA of the stress distributions in (C) the hinge
along (D) the circumferential and (E) radial directions. The red arrows indicate the
directions of the stresses. (F) Tensile tests of the OL and (G and H) compressive tests
of the FFR (freshly prepared wet sample) along the (G) circumferential and (H) radial
directions, showing the different mechanical performances of these regions.
and varying nanowire diameters can help the
FFR accommodate the nanowire space-filling
pattern.
Additionally, the cross section of each nano-
wire exhibits a pseudo-hexagonal shape (Fig.
3B) (26). We decalcified an FFR sample with
an ethylenediaminetetraacetic acid disodium
salt dihydrate solution to remove the aragonite
minerals (27). The remnant shows a typical
honeycomb-like structure, indicating that the
aragonite nanowires are embedded in a con-
tinuous organic matrix (Fig. 3, C and D; fig.
S12; and table S2). The nanowire orientations
at several sampling positions in the fracture
surface of the FFR suggest that they are ra-
dially oriented in the whole region, which are
similar to the ribs in a folding fan (Fig. 3I and
fig. S16). Micro x-ray diffraction mapping of
the same surface reveals that the aragonite
i
nanowires uniformly grow along the 002
crystallographic direction (Fig. 3J) (28), which
agrees with the high-resolution transmission
electron microscopy (HRTEM) analysis (Fig. 3,
E to H). These results indicate that the mor-
h
h
h
i crystallographic ori-
phological and the 002
entations of the aragonite nanowires match
i growth orien-
each other. This preferred 002
tation of the nanowires is also the fastest growth
direction during abiotic aragonite crystallization
(29, 30). Although biological control can substan-
tially change the mineralization process, the
crystallographic anisotropy in growth rate can
contribute to the biominerals formation with
highly anisotropic morphologies, leading to a
more cost-effective biomineralization process
(31, 32).
Origins of deformability and fatigue resistance
To study the role of the morphological and
crystallographic orientations of the aragonite
nanowires in functionality, we extracted a rep-
resentative volume element (RVE) from the
FFR and simulated its response to loads along
different directions (fig. S17). In the direction
that is normal to the nanowires, which is the
circumferential direction in the FFR, the RVE
can easily deform because of its relatively low
modulus (fig. S17, A and C). In the direction
h
i= 010h
that is parallel to the nanowires, which is the
radial direction in the FFR, the RVE can trans-
fer a large load with a limited deformation (fig.
S17, B and D). We also found that the RVE with
aragonite nanowires uniformly oriented in
i direction exhibits a higher stress level
h
the 002
than that oriented in the 100h
i at the
same strain (fig. S17D), indicating that the
former can transfer larger radial loads. Given
i crystallographic direction has
that the 002
the fastest growth rate during the aragonite
crystallization (30), the above analyses suggest
an elegant consistency of the mechanically
favored orientation, the thermodynamically
favored orientation, and the actual nanowire
crystal growth direction. Consequently, although
the FFR is fabricated in a cost-effective way con-
cerning the nanowire growth direction, the
microstructures and crystallographic features
of the FFR are also optimized for the sake of
good circumferential deformability and radial
load translation capability, both of which are
the cornerstones of the long-term functioning
performance of the FFR (Fig. 2).
Meng et al., Science 380, 1252–1257 (2023)
23 June 2023
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Microstructural and crystallographic features of the aragonite nano-
wires in the FFR. (A) Longitudinal and (B) cross-sectional fracture surfaces of the
FFR showing the oriented nanowires with a pseudo-hexagonal shape. Scale bars,
(A) 1 mm and (B) 400 nm. (C) Longitudinal and (D) cross-sectional fracture surfaces
of the decalcified FFR showing the honeycomb structure. Scale bars, 400 nm.
(E) TEM image of a transverse ultrathin section of the FFR. The yellow arrows indicate
the twin boundaries in the nanowires. Not all the boundaries are visible because
of the tilting angles. Scale bar, 200 nm. (F) HRTEM image of an aragonite nanowire
showing the twin boundary. (Inset) Selected area electron diffraction image of the
twin structure revealing that the twin boundary is the (110) plane. Scale bars,
h
5 nm and (inset) 2 nm−1. (G) TEM and (H) HRTEM images of an ultrathin longitudinal
i direction. Scale bars,
section of the FFR. The aragonite crystals grow along the 002
(G) 200 nm and (H) 3 nm. (I) Morphological orientation of nanowires in the FFR
(indicated with the white fan shape). The images in the region were obtained by
means of fast Fourier transform (FFT) of the scanning electron microscopy (SEM)
images (for example, the FFT image in the box is transformed from the SEM
image in the dashed box). Scale bar, 1 mm. (J) Crystallographic orientation 002
(black lines) of the aragonite nanowires in the FFR, which is acquired from
the crystal diffractogram. The color of each square indicates the deflection angle of
h
002
i orientation against the mirror plane of the FFR.
h
i
Because the morphological and crystallo-
graphic orientations of these ordered nano-
wires are the same, the internal stress in the
FFR can be deduced from the stress-induced
lattice distortion of the aragonite crystals (33).
We therefore investigated the stress states of
some microdomains in the FFR by means of
high-resolution synchrotron x-ray diffraction.
The cell parameter c of the aragonite nano-
wires at the closed state decreases compared
with that at the open state (Fig. 4A and fig. S18),
whereas the changes of parameters a and b
are much smaller. Because the nanowires grow
i direction, the result indicates
along the 002
that each aragonite nanowire experiences a
h
large radial load along the longitudinal direc-
tion. However, although the FFR endures a
larger deformation in the circumferential di-
rection, its elastic modulus along this direction
is smaller, and it is the organic matrix that
bears most of the deformation (fig. S17). This
explains the much smaller changes of a and b
as well as the small circumferential loads on
the nanowires. The result is thus in good agree-
ment with the properties of the RVE.
The ROAC motions can be implemented hun-
dreds of thousands of times during the course
of the shell’s lifetime, and thus the FFR that
exerts distinct functionalities in different direc-
tions also has to be fatigue resistant (Fig. 1C and
fig. S4). Because the aragonite nanowires are
extremely long and brittle (34), they should
have broken off easily as the FFR undergoes
the radial load from the OL. In this respect, the
organic matrix, which is resistant to fracture
and wraps around each nanowire, can prevent
the fragile nanowire from bending and break-
ing (fig. S19). We evaluated the stress states of
the nanowires during the ROAC motions to see
how the nanowires can sustain the motions. Be-
cause the FFR bears a circumferential deforma-
tion, lateral compressive stress or tensile stress
should thus be applied to the nanowires.
Atomic-force microscopy (AFM) observation
of the FFR fracture surface in liquid reveals,
Meng et al., Science 380, 1252–1257 (2023)
23 June 2023
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. In situ stress state analyses of the nano-
wires. (A) High-resolution synchrotron x-ray diffraction
analysis of the lattice constant variation (from open to
closed) of the aragonite crystals. (B) AFM images
obtained at “up,” “middle,” and “down” positions in the
FFR (fig. S22B) at the open and closed states of the
valves. Scale bars, 500 nm. (C) Shift of relative
position (Dd) of two adjacent nanowires.
A
Up
Middle
Down
Δa (%)
Δb (%)
Δc (%)
-
0
-
0
-
0
0
0123
0
3
1
2
.
.
.
.
-
0
-
0
-
0
0
0123
0
3
1
2
.
.
.
.
-
0
-
0
-
0
0
0123
0
3
1
2
.
.
.
.
B
p
U
d
c
d
o
120 nm
Up
Middle
C
n
o
i
t
i
s
o
p
g
n
i
l
p
m
a
S
-110 nm
340 nm
-370 nm
230 nm
Down
-300 nm
40
60 70
50
Δd (nm)
d
o
d
o
l
e
d
d
M
i
d
c
Deformability
n
w
o
D
d
c
II
Folding fan shape
Radially aligned
nanowires
I
II
III
III
Weak interfaces
Hard fibers and soft matrix
Bicontinuous phase
networks
Cross-linked networks
and crystalline domains
Nanorod
remineralization
Microcracks
and ligament bridges
Grain refinement
and Inclusions
IV
Protein
Hard-soft combination
IV
Nanorod
TB
GB
Chemical and
structural gradients
Sliding and
sacrificial bonding
Grain boundary
and twin boundary
Twin boundary
Enamel
Cortical bone
Metal
FFR
Elastomer
Fig. 5. Antifatigue design of typical biological and artificial structural
materials. The antifatigue performance of the structural materials with diverse
stiffness relies on their specific structures. Enamel: Chemical and structural
gradients inhibit the fatigue crack initiation and propagation; damages can be
repaired by the hydroxyapatite nanorod remineralization (10, 39). Cortical bone:
Sacrificial bond rupture and collagen fibril sliding impede the crack initiation;
microcrack, ligament bridges, and weak interface induce the crack deflection and
slows the crack propagation; remodeling substitutes old or damaged bone with
new bone tissue (9, 40). Metal: Twin boundaries (TB) and grain boundaries (GB)
suppress dislocation activities; nano inclusions retard microcrack propagation (41).
FFR: External load is translated by the folding fan–like structure (I) to
circumferential deformation without inducing stress concentration; the hard-soft
complex structure consisting of radially aligned nanowires (II) deforms
compatibly to sustain the circumferential deformation, while the soft matrix (III)
undertakes most of the volume compression; twin boundaries (IV) along the
length direction of the nanowires provide a bottom-level guarantee to the
aragonite nanowires against bending fracture. Elastomer: Cross-linked networks
and crystalline domains raise the threshold of energy release rate; bicontinuous
phase networks and a hard fiber and soft matrix complex enhance the crack
deceleration and blunting (42).
Meng et al., Science 380, 1252–1257 (2023)
23 June 2023
5 of 6
m
m
0
1
-
m
µ
0
1
m
µ
0
1
-
m
n
0
0
1
m
n
0
0
1
-
0
1
m
n
0
1
<
RES EARCH | R E S E A R C H A R T I C L E
through comparison of the relative positions of
adjacent nanowires, the nanowires’ axial sliding
indicating a shear force around the nanowires as
well as the lateral stresses (Fig. 4, B and C, and
fig. S20) (35). We thus simulated the three stress
states—compression, tension, and shear—in a
two-dimensional numerical model (fig. S21). The
FEA results suggest that it is the matrix that
bears most of the compressive and shear strains
in all the states; no stress concentration is
found in the nanowires or the organic matrix.
Therefore, the deformation compatibility of
this hard-soft complex structure can effectively
reduce the possibility of brittle rupture of the
aragonite nanowires, by which the fatigue
damage of the FFR can be suppressed.
Aragonite belongs to the Pmcn space group
in which (110) twinning planes can easily de-
velop in aragonite crystals. This makes these
orthorhombic crystals exhibit a pseudohex-
agonal appearance (Fig. 3, B and E) (36). As
the aragonite nanowires in the FFR grow along
i direction, abundant twin boundaries
h
the 002
are found in parallel with the axial direction of
the nanowires (Fig. 3, E and F). Such twin
boundaries can improve the nanowire re-
sistance to bending fracture (37, 38). There-
fore, in addition to the higher stress transfer
i orientation of the arag-
capability of the 002
onite nanowires, this orientation is also asso-
ciated with the nanotwin boundary formation
that plays a key role in preventing the fracture
of the nanowires under cyclic stress states.
h
Conclusions and outlook
We have revealed the hierarchical structure
design of the FFR in the hinge of C. plicata,
which spans from the macroscale level down
to the lattice level. This design is not a simple
accumulation of isolated antifatigue mecha-
nisms; rather, each aspect works synergisti-
cally (Fig. 5). The notable deformability and
load translation capability of the FFR orig-
inate from the hierarchical structures, which
cannot be achieved by any specific mechanism
that acts at only a few length scales. The com-
bination of the functionality and fatigue re-
sistance of the FFR exemplifies how the service
life of a material can be prolonged by exploit-
ing the intrinsic properties of each component.
Furthermore, C. plicata shell also exploits the
inherent crystallographic characteristics of
aragonite for the biomineralization of the
FFR; this tactic is rarely used in the fabrication
of artificial structural materials. Although there
is much to learn from this biomineral, we fab-
ricated a proof-of-concept glass-polymer com-
posite with the FFR-like microstructure as a
primitive demonstration (fig. S22). The com-
posite exhibits anisotropic mechanical behav-
iors and good fatigue resistance similar to
those of the FFR (supplementary text 2). This
biomineral provides a multiscale model for de-
signing structural materials that contain brit-
tle components, such as ceramics, yet require
both deformability and fatigue resistance.
RE FERENCES AND NOTES
1. D. Gu et al., Science 372, eabg1487 (2021).
2. S. Bhumiratana et al., Sci. Transl. Med. 8, 343ra83 (2016).
3. S. Park, H. Park, S. Seong, Y. Chung, Sci. Rep. 10, 7660 (2020).
4. K. J. Miller, Mater. Sci. Technol. 9, 453–462 (1993).
5. C. E. Renshaw, E. M. Schulson, Nature 412, 897–900 (2001).
6.
7.
8. Y. Shao, H.-P. Zhao, X.-Q. Feng, H. Gao, J. Mech. Phys. Solids
J. Schijve, Int. J. Fatigue 25, 679–702 (2003).
J. J. Kruzic, Science 325, 156–158 (2009).
60, 1400–1419 (2012).
9. M. E. Launey, M. J. Buehler, R. O. Ritchie, Annu. Rev. Mater.
Res. 40, 25–53 (2010).
10. D. Bajaj, A. Nazari, N. Eidelman, D. D. Arola, Biomaterials 29,
4847–4854 (2008).
11. R. O. Ritchie, Nat. Mater. 10, 817–822 (2011).
12. Z. Ma et al., Sci. Rep. 6, 22156 (2016).
13. T. Magrini et al., Nat. Commun. 10, 2794 (2019).
14. M. J. Kim et al., Sci. Adv. 7, eabg8169 (2021).
15. C. Xu, Y. Song, M. Han, H. Zhang, Microsyst. Nanoeng. 7, 25
(2021).
16. E. H. Cho et al., Nanoscale 10, 628–638 (2018).
17. G. Zan et al., Matter 4, 3232–3247 (2021).
18. E. Savazzi, Y. A. O. Peiyi, Lethaia 25, 195–209 (1992).
19. S. A. Wainwright, Nature 224, 777–779 (1969).
20. K. Ono, Y. Kikuch, K. Higashi, N. Tamiya, N. Yasuoka, J. Biomech.
23, 307–312 (1990).
21. A. Fernández-Canteli, S. Blasón, B. Pyttel, M. Muniz-Calvente,
E. Castillo, Int. J. Fract. 223, 189–196 (2020).
22. J. C. Weaver et al., Science 336, 1275–1280 (2012).
23. D. Labonte, A.-K. Lenz, M. L. Oyen, Acta Biomater. 57, 373–383
(2017).
24. G. Bevelander, H. Nakahara, Calcif. Tissue Res. 4, 101–102 (1969).
25. C. Ituarte, D. G. Zelaya, J. Molluscan Stud. 81, 455–465 (2015).
26. M. E. Marsh, R. L. Sass, Science 208, 1262–1263 (1980).
27. C. Picart, D. E. Discher, Nature 448, 879–880 (2007).
28. O. Paris, Biointerphases 3, FB16 (2008).
29. A. G. Checa, T. Okamoto, J. Ramírez, Proc. Biol. Sci. 273, 1329–1337 (2006).
30. V. Schoeppler et al., Proc. Natl. Acad. Sci. U.S.A. 116,
20388–20397 (2019).
31. P. U. P. A. Gilbert et al., J. Am. Chem. Soc. 130, 17519–17527 (2008).
32. F. Nudelman, Semin. Cell Dev. Biol. 46, 2–10 (2015).
33. B. Pokroy, J. P. Quintana, E. N. Caspi, A. Berner, E. Zolotoyabko,
Nat. Mater. 3, 900–902 (2004).
34. Z. Wu, Y.-W. Zhang, M. H. Jhon, H. Gao, D. J. Srolovitz, Nano Lett.
12, 910–914 (2012).
35. M. Frey et al., Adv. Sci. (Weinh.) 6, 1802190 (2019).
36. W. L. Bragg, Proc. R. Soc. A Contain. Pap. Math. Phys. Character
105, 16–39 (1924).
37. A. J. Cao, Y. G. Wei, S. X. Mao, JOM 60, 85–88 (2008).
38. Y. Tian et al., Nature 493, 385–388 (2013).
39. K. A. DeRocher et al., Nature 583, 66–71 (2020).
40. J. Li, H. Gong, Acta Mech. Sin. 37, 516–526 (2020).
41. D. Rozumek, Metals (Basel) 11, 1957 (2021).
42. Y. Huang et al., Soft Matter 18, 5153–5165 (2022).
43. S.-H. Yu et al., Deformable hard tissue with high fatigue
resistance in the hinge of bivalve Cristaria plicata. Dryad (2023);
https://doi.org/10.5061/dryad.69p8cz95j.
AC KNOWLED GME NTS
The authors thank P. Fratzl from Max Planck Institute for helpful
comments and discussions on this work and the manuscript. The
authors also thank Z.-L. Zhu for providing micro x-ray diffraction
mapping; Y.-G. Gu for fatigue tests; J. Tian and S.-Q. Fu for
providing scanning electron microscopy tests and element
mapping; T.-W. Li and Y.-R. Wang for providing high-resolution
transmission electron microscopy analyses; Z. He, S.-C. Zhang,
and J. Pang for AFM tests; Y.-T. Wei (Bruker) and J.-P. Wang
(Bruker) for supporting nanoindentation tests; G.-Y. Gao for
powder x-ray diffraction tests; W. Wen for high-resolution x-ray
diffraction tests; and B.-H. Zhan and R.-D. Wang (Beijing Institute
of Fashion Technology) and C.-X. Yu (Anhui University) for their
assistance in visualizing this work. This research used Beamline
BL14W of the Shanghai Synchrotron Radiation Facility (SSRF)
for high-resolution synchrotron x-ray diffraction tests. The authors
also thank Shiyanjia Lab (www.shiyanjia.com) for hematoxylin and
eosin staining. This work was partially carried out at the USTC
Center for Micro and Nanoscale Research and Fabrication.
Funding: This work was supported by financial support from the
National Key Research and Development Program of China
(grants 2021YFA0715700 and 2018YFE0202201) and the National
Natural Science Foundation of China (grants 21701161, 22293044,
12232016, and 12172346). Y.-B.Z. acknowledges funding support
from the Youth Innovation Promotion Association CAS (2022465).
This work has been supported by New Cornerstone Science
Foundation. Author contributions: S.-H.Y. and L.-B.M. conceived
the idea and designed the experiments. S.-H.Y., L.-B.M., X.-S.M.,
L.-C.Z., and L.L. wrote and edited the paper. X.-S.M., Y.-F.M., B.Y.,
and Q.-Z.R. performed the experiments and analyzed the data.
L.-C.Z., Y.-B.Z., and H.-A.W. performed theoretical analyses. Y.-B.Z.
and D.-C.Z. provided valuable advice for mechanical analyses.
Y.-B.Z. and L.-B.M. contributed to antifatigue mechanisms. All
authors discussed the results. Competing interests: All authors
declare that they have no competing interests. Data and
materials availability: Data are available in the manuscript or the
supplementary materials or are deposited in Dryad (43). License
information: Copyright © 2023 the authors, some rights reserved;
exclusive licensee American Association for the Advancement of
Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade2038
Materials and Methods
Supplementary Text
Figs. S1 to S23
Tables S1 to S3
References (44–53)
MDAR Reproducibility Checklist
Movies S1 to S5
Submitted 1 August 2022; accepted 25 April 2023
10.1126/science.ade2038
Meng et al., Science 380, 1252–1257 (2023)
23 June 2023
6 of 6
|
10.1126_science.add9884
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
GREEN NUDGES
Reducing single-use cutlery with green nudges:
Evidence from China’s food-delivery industry
Guojun He*, Yuhang Pan, Albert Park, Yasuyuki Sawada, Elaine S. Tan
INTRODUCTION: Plastic waste is a global envi-
ronmental threat that endangers marine and
freshwater ecosystems worldwide. More recent-
ly, as food-delivery services became increasing-
ly popular during the COVID-19 pandemic, the
surge in plastic waste generated by single-use
cutlery (SUC) has become a key environmental
concern. Yet effective policies that control SUC
waste are largely nonexistent, and it is important
to find ways to encourage individuals to reduce
their SUC consumption. Using data from China,
the world’s largest producer and consumer of
SUC, we investigated how green nudges can af-
fect individuals’ cutlery decisions when placing
food-delivery orders.
RATIONALE: From 2019 to 2020, three Chinese
cities (Beijing, Shanghai, and Tianjin) introduced
regulations that prohibited online food-delivery
companies from including SUC unless it was
explicitly requested. To comply with the regula-
tions, Alibaba’s food-delivery company, Eleme,
changed its app in the following ways: (i) by add-
ing a pop-up window that required customers to
explicitly choose the number of SUC sets to be
included with their orders, (ii) by setting the
default for this pop-up window to be “no cut-
lery,” and (iii) by providing a small nonpecuniary
incentive––several Ant Forest green points––to
those who chose the “no cutlery” option. The
green points do not have a monetary value,
but if one accumulates enough points (roughly
by placing more than 1000 online food orders),
they can be redeemed in exchange for planting
a real tree (under the customer’s name) in a
desert area in China. The changes in the app’s
25
20
15
10
5
O
C
N
S
0
–12
SNCO under new
checkout interface (with green nudges)
Cutlery choice
Cutlery
(based on food amount)
One set of cutlery
Two sets of cutlery
.
.
.
.
.
Confirm
SNCO under old
checkout interface
Cutlery choice
No cutlery
16 green points reward to plant trees
Cutlery (based on food amount)
One set of cutlery
.
.
.
No cutlery as default choice
Confirm
–8
–4
0
4
8
12
Months before and after changing Eleme’s checkout interface
user interface embody the concept of “nudging”
from behavioral economics and social psychol-
ogy, which describes approaches that change the
choice environment (or choice architecture) or
provide indirective information to influence
the behaviors and decision-making processes
of individuals. Using customer-level data from
Alibaba in 10 cities from 2019 to 2020, we
compared behavioral differences between the
customers in the “nudged” cities and those in
the control cities before and after the introduc-
tion of green nudges.
RESULTS: The green nudges, on average, in-
creased an individual’s share of no-cutlery or-
ders by 20.1 percentage points, which was a
648% increase relative to the baseline group.
Meanwhile, the green nudges incentivized a
large portion of individuals to somewhat
change their behaviors rather than encourag-
ing only a small portion to change their be-
haviors substantially. Women, older individuals,
frequent food-delivery-service users, and
wealthy individuals were more responsive to
the green nudges. Importantly, Alibaba’s busi-
ness performance was not affected by the
green nudges, suggesting that this could be a
highly cost-effective way to reduce SUC waste.
Additional mechanism analyses revealed that
the default change and increased salience
of the no-cutlery option were the main drivers
of the observed behavioral changes, whereas
the incentive to accumulate green points to
plant trees played a relatively muted role. We
estimate that if green nudges were applied to
all of China, more than 21.75 billion sets of
SUC could be saved annually, which is equiv-
alent to preventing the generation of 3.26 mil-
lion metric tons of plastic waste and saving
5.44 million trees.
CONCLUSION: Our study provides compelling
evidence that nudges can be a powerful tool
for changing behaviors. It also suggests that
private sector and platform companies can
provide highly cost-effective solutions to pro-
mote prosocial behaviors among their custom-
ers. In this study, the costs of implementing
the green nudges were almost negligible (i.e.,
several hours of work to redesign the user
interface), yet the aggregated environmental
benefits were tremendous. We thus recom-
mend that other online food-delivery plat-
forms, such as DoorDash and Uber Eats, try
similar green nudges to reduce global plas-
.
.
.
.
.
tic waste.▪
Green nudges reduce SUC. The graph illustrates the trends in the share of no-cutlery orders (SNCO) among
Alibaba’s Eleme customers in three cities (Beijing, Shanghai, and Tianjin) before and after changing the
app’s checkout interface, with dashed orange and teal lines indicating the average SNCO. Compared with
the old interface, the new interface has three nudging components: (i) a pop-up window that requires
customers to explicitly choose the number of SUC sets to be included with their orders, (ii) the default for
this pop-up window set to “no cutlery,” and (iii) a small nonpecuniary incentive—some Ant Forest green
points—that is given to those who choose the no-cutlery option.
The list of author affiliations is available in the full article online.
*Corresponding author. Email: gjhe@hku.hk
Cite this article as G. He et al., Science 381, eadd9884 (2023).
DOI: 10.1126/science.add9884
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.add9884
He et al., Science 381, 1064 (2023)
8 September 2023
1 of 1
RES EARCH
R E S E A R C H A R T I C L E
◥
GREEN NUDGES
Reducing single-use cutlery with green nudges:
Evidence from China’s food-delivery industry
Guojun He1,2*, Yuhang Pan3, Albert Park4,5,6,7, Yasuyuki Sawada8,9, Elaine S. Tan4
Rising consumer demand for online food delivery has increased the consumption of disposable cutlery,
leading to plastic pollution worldwide. In this work, we investigate the impact of green nudges on
single-use cutlery consumption in China. In collaboration with Alibaba’s food-delivery platform, Eleme
(which is similar to Uber Eats and DoorDash), we analyzed detailed customer-level data and found that
the green nudges—changing the default to “no cutlery” and rewarding consumers with “green points”—
increased the share of no-cutlery orders by 648%. The environmental benefits are sizable: If green
nudges were applied to all of China, more than 21.75 billion sets of single-use cutlery could be saved
annually, equivalent to preventing the generation of 3.26 million metric tons of plastic waste and saving
5.44 million trees.
P lastic waste is a global threat that en-
dangers marine and freshwater ecosys-
tems worldwide (1–3). In 2021, more than
400 million metric tons of plastic waste
were produced worldwide, and it is pre-
dicted that the world’s plastic waste growth will
continue to outpace the efforts to reduce plas-
tic pollution in the coming decades (3). More
recently, as food-delivery services became in-
creasingly popular during the COVID-19 pan-
demic, the surge in plastic waste generated
by single-use cutlery (SUC) has become a key
environmental challenge faced by many coun-
tries (4).
Reducing SUC waste in the food-delivery in-
dustry is particularly important in China, the
world’s largest producer and consumer of SUC.
As of 2019, more than 540 million Chinese were
active users of food-delivery services and each
day consumed more than 50 million sets of
SUC that were not adequately recycled or dis-
posed of (5). SUC in China typically includes
a plastic fork, a plastic spoon, a pair of wood-
en chopsticks, and a napkin. Consequently,
SUC usage in China not only generates large
amounts of plastic waste (i.e., used forks and
spoons) but also consumes a large number
of trees (discarded chopsticks and napkins),
1Faculty of Business and Economics, University of Hong Kong,
Hong Kong SAR, China. 2Institute for Climate and Carbon
Neutrality, University of Hong Kong, Hong Kong SAR, China.
3Institute for Global Health and Development, Peking University,
Beijing, China. 4Asian Development Bank, Metro Manila,
Philippines. 5Department of Economics, Hong Kong University
of Science and Technology, Clear Water Bay, Kowloon, Hong
Kong SAR, China. 6Division of Social Science, Hong Kong
University of Science and Technology, Clear Water Bay,
Kowloon, Hong Kong SAR, China. 7Division of Public Policy,
Hong Kong University of Science and Technology, Clear Water
Bay, Kowloon, Hong Kong SAR, China. 8Faculty of Economics,
University of Tokyo, Tokyo, Japan. 9Asian Development Bank
Institute, Tokyo, Japan.
*Corresponding author. Email: gjhe@hku.hk
which reduces forest area, threatens ecosys-
tems, and ultimately poses substantial health
risks to humans (6).
To reduce SUC consumption, policy-makers
in China set a target of reducing SUC usage in
food deliveries by 30% by 2025. Online plat-
forms were required to establish plans to achieve
that target (7). From 2019 to 2020, Shanghai,
Beijing, and Tianjin further introduced pilot
city-wide regulations that prohibited online
food-delivery companies from including SUC
in their orders unless it was explicitly requested
by customers. To comply with the new rules,
online food-delivery platforms redesigned their
app user interfaces and required that cus-
tomers explicitly request SUC when placing
an order.
In this study, we collaborated with Alibaba’s
online food-ordering platform Eleme to eval-
uate the effectiveness of their efforts to reduce
SUC consumption. Eleme, which is similar to
Uber Eats and DoorDash, is China’s second-
largest food-delivery company, with more than
753 million users in 2022. Hereafter, we refer
to the platform simply as “Alibaba.” To comply
with city regulations, Alibaba changed its app
in the following ways: (i) by adding a pop-up
window that required customers to explicitly
choose the number of SUC sets with their or-
ders, (ii) by setting the default for this pop-up
window to be “no cutlery,” and (iii) by provid-
ing a small nonpecuniary incentive––some “Ant
Forest green points” (hereafter “green points”)—
to those choosing the no-cutlery option (see
supplementary note 1). The green points do not
have a monetary value, but if one accumulates
enough points (roughly, by placing more than
1000 online food orders), they can be redeemed
in exchange for planting a real tree (under the
customer’s name) in a desert area in China.
The changes in the Alibaba app’s user inter-
face embody the concept of “nudging” from
behavioral economics and social psychology,
which describes approaches that change the
choice environment (or choice architecture)
or provide indirective information to influ-
ence the behaviors and decision-making of
individuals (8, 9). The key feature of nudges
is that they do not limit the choices available
to individuals, nor do they change the mone-
tary incentives (8, 9). In the past two decades,
nudges have been used in many social do-
mains (10), including green nudges that pro-
mote pro-environmental behaviors (11, 12).
However, to what extent green nudges can
be effective remains controversial (11, 13–16),
and there are few opportunities for research-
ers to examine the impacts of green nudges at
scale and for an extended period of time.
Using detailed customer-level data from
Alibaba (17), we investigated how Alibaba’s
green nudges affected individuals’ cutlery de-
cisions. Specifically, we applied a difference-in-
differences model to the data and compared
the food-ordering behaviors of individuals in
the pilot (hereafter, “treated”) cities with green
nudges with those of individuals in the “con-
trol” cities without green nudges before and
after Alibaba changed the app interface (Mate-
rials and methods). Our research findings not
only provide compelling evidence for how green
nudges can affect pro-environmental behaviors
but also generate important implications for the
understanding of nudges in general.
Specifically, this study has several distinc-
tive features that distinguish it from the vast
literature on nudges. First, there is an ongoing
debate about why nudges work and under what
conditions they work (18–21), and the rich data
in this study allowed us to explore the under-
lying mechanisms through which the green
nudges affect individuals’ behaviors. Second,
with a few exceptions (22, 23), most studies in
the nudge literature focus on the short-term
impacts; by contrast, our study can follow indi-
viduals’ repeated decisions, explore the per-
sistence of the nudging effect, and examine
whether nudges work in situations where in-
dividuals can easily switch to other options and
set other options as the default. Examining the
medium- and long-term effects is particularly
important in our setting because short-term
analysis might exaggerate the treatment effects.
Third, nudges have been criticized for being
manipulative because they are not always trans-
parent and can take advantage of individuals’
ignorance or lack of awareness (24–29). How-
ever, the green nudges we studied here are easy
to understand and completely transparent to
users (i.e., the pop-up window that requires in-
dividuals to make explicit choices). Fourth, the
panel data of individuals allow us to estimate
individual-specific effects, which can inform us
of how many people are “nudgeable” and the
entire distribution of the nudge effects. Finally,
few policies target plastic-waste production at
He et al., Science 381, eadd9884 (2023)
8 September 2023
1 of 8
RES EARCH | R E S E A R C H A R T I C L E
the consumer level, except charges on plastic
bags (30–33). Thus, our results can help policy-
makers better understand whether it is feasi-
ble to reduce consumer plastic consumption
by regulating platform companies.
Methods
Changes in Alibaba’s app user interface
Alibaba redesigned its app’s user interface to
comply with the new environmental regula-
tions in Shanghai (from 1 July 2019), Beijing
(from 1 May 2020), and Tianjin (from 1 Decem-
ber 2020). Figure 1 depicts the changes. Fig. 1A
shows the old user interface: The options for
SUC were placed at the end of the food-ordering
page, with the default being “with cutlery (num-
ber of sets based on the amount of food).” Res-
taurants typically provided SUCs freely and
sometimes even gave more SUCs than needed
to avoid negative reviews (34).
In the new user interface, as shown in Fig.
1B, when a customer in a treated city placed an
order and clicked the check-out button, a win-
dow emerged that required the customer to
explicitly choose the number of sets of SUC. Im-
portantly, on this pop-up window, the default
became “no cutlery,” and customers had to scroll
down the screen to choose a different option.
Furthermore, to prevent the window from ap-
pearing in their future orders, a customer could
set any of the options as their new device-level
default by clicking the “set as default” button.
The no-cutlery option came with a small non-
pecuniary incentive, that is, the green points.
When a customer placed an order without SUC,
16 green points would be awarded to the per-
son, which could be stored and accumulated
in Alibaba’s payment app, Alipay. When con-
sumers accumulated 16,000 green points, they
could spend them to request that Alibaba plant
a real tree named after the consumer in a des-
ert region of China. There are also other ways to
obtain green points under Alibaba’s platform,
such as walking more, taking more public
transportation, selling used items, and so on.
The green points have helped plant more than
223 million trees in China from 2016 to 2020
(35). Getting a tree planted only by placing no-
cutlery orders is challenging. According to
our data, the average number of monthly no-
cutlery orders for a consumer was only 0.77
from 2019 to 2020.
We could identify three nudging compo-
nents of the changes to the user interface: (i)
a pop-up window that increased the salience
of available cutlery options to the customers,
(ii) the change of the default selection to “no
cutlery,” and (iii) the inclusion of a small non-
pecuniary incentive (green points) as part of
the new default option. The effect we estimated
in this study thus came from a mixture of dif-
ferent nudging incentives rather than just one.
According to the nudge taxonomy suggested by
Münscher et al. (36), nudges can be classified
into three types: those that provide decision
information (like reframing, increasing vis-
ibility, and providing social reference points),
those that provide decision assistance (like re-
minders and commitment devices), and those
that change the decision structure (like chang-
ing the defaults, the option-related efforts, the
range of composition of options, and the op-
tion consequences). In our setting, the salience
intervention provided decision information,
that is, it made the cutlery options more ex-
plicit, transparent, and visible. The other two
nudging elements altered the decision struc-
ture, including a standard change in the de-
fault (i.e., “no cutlery”) and a change in the option
consequences (i.e., green points for planting
trees). The green-points incentive could also
be viewed as a symbolic award, which pro-
vided individuals with an intrinsic value that
went beyond and exceeds that of an expend-
able or disposable item (37).
Our data included each user’s monthly
food-ordering history from 1 January 2019 to
31 December 2020 in 10 major Chinese cities
(17). These include the three treated cities
with green nudges (i.e., Beijing, Shanghai, and
Tianjin) and the seven control cities without
the nudges (Qingdao, Xi’an, Guangzhou, Nanjing,
Hangzhou, Wuhan, and Chengdu). Among these
cities, we randomly sampled about 200,000
“active” users (i.e., those who placed at least
one order between 2019 and 2020). We then
A Old Check-Out Interface
B
New Check-Out Interface with Green Nudges in Treated Cities
Fig. 1. Green-nudge interventions at the Eleme application interface.
(A) The old interface, in which the default cutlery option at the check-out page
was preset as “with cutlery.” Users had to click on the pop-up window and scroll
to the bottom to choose the no-cutlery option. (B) The new interface in the
treated cities. A window about cutlery automatically popped up during checkout
and required the consumer to choose the number of sets of SUC. The default
cutlery option was “no cutlery.” Users could make any option their default by clicking
the “set as default” button (so that this window would not pop up in their future
orders). The no-cutlery option came with a small nonpecuniary incentive, that is,
16 Ant Forest green points. Contents are translated by the authors.
He et al., Science 381, eadd9884 (2023)
8 September 2023
2 of 8
RES EARCH | R E S E A R C H A R T I C L E
compared the behavioral differences between
the users in the treated cities and the users in
the control cities before and after the intro-
duction of green nudges using a difference-
in-differences model (Eq. 1 in Materials and
methods). We asked the following four re-
search questions (RQs):
RQ 1. Did the green nudges affect individuals’
no-cutlery orders and did the effects persist?
We hypothesized that the green nudges should
increase the no-cutlery orders because both
the default change and the green points incen-
tivized them to do so.
RQ 2. Did the green nudges negatively affect
the platform’s business performance? Because
the new interface did not reduce the options
available to customers and offered additional
green-point incentives, we hypothesized that
the user experience of the platform would not
deteriorate, and, therefore, its business per-
formance would not be negatively affected.
RQ 3. Which types of customers, if any, were
more likely to respond to the green nudges?
The answer depended on the costs and benefits
of choosing the no-cutlery option for different
individuals, as well as their general “nudgeabil-
ity.” For example, those who were more envi-
ronmentally conscious and ate at home could
be more nudgeable than others.
RQ 4. Were the aggregate environmental
benefits large? The answer to this question de-
pended on the extent to which the green nudges
affected SUC consumption and how much waste
reduction this implies.
Results
Green nudges substantially increased the
no-cutlery orders
Figure 2A depicts the changes in the raw data.
We observed that the share of no-cutlery orders
significantly increased in the treated cities after
the intervention, whereas the share remained
relatively unchanged throughout the study pe-
riod in the control cities.
We then estimated the difference-in-differences
model (Eq. 1 in Materials and methods) and
summarized the results in Fig. 2B. We observed
that the green nudges increased the share of
no-cutlery orders by 19.3 percentage points
in Shanghai, 21.2 percentage points in Beijing,
and 20.4 percentage points in Tianjin (regres-
sion results in table S1). On average, the green
nudges increased an individual’s share of no-
cutlery orders by 20.1 percentage points. Be-
cause the share of no-cutlery orders before the
treatment was around 3.1 percentage points,
the frequency of no-cutlery orders increased by
648% (20.1 divided by 3.1 and multiplied by 100).
The impacts of green nudges observed in this
study are significantly larger than those found
in the literature, as summarized in table S2.
We conducted an event-study analysis to
examine the plausibility of the parallel trend
assumption in the difference-in-differences
model (Eq. 2 in Materials and methods). We
observed that before the introduction of the
green nudges, the control and treated cities
followed essentially the same trends (Fig. 2C
and table S3). Immediately after the intro-
duction of green nudges, however, the share of
no-cutlery orders in the treated cities signif-
icantly increased. These results suggest that
A
C
B
D
Fig. 2. Investigation of the impacts of green nudges. (A) Trends in the share
of no-cutlery orders (SNCO). The vertical green and blue dashed lines indicate
introductions of the green nudges in Shanghai and Beijing, respectively. The green
nudges were implemented in December 2020 in Tianjin. White circles refer to
SNCO in the other seven cities in the control group. (B) The estimated impacts of
green nudges on SNCO using individual-level data. The first three bars present the
estimated effects of green nudges in each treated city separately. The fourth bar
reports the average treatment effect across all cities. (C) The event-study results,
where the reference group in the regression is 1 month before the introduction
of the green nudges. CI, confidence interval. (D) The distribution of individual
treatment-effect estimates for all the consumers in the treated cities. The x and
y axes indicate the effect size and density of the estimates, respectively.
He et al., Science 381, eadd9884 (2023)
8 September 2023
3 of 8
RES EARCH | R E S E A R C H A R T I C L E
in the absence of green nudges, customers in
the two groups of cities would follow similar
SUC consumption patterns. In supplementary
note 2, we further show that the media and
public attention to plastic waste and pollution
did not change significantly when the city-wide
regulations were introduced, which ensures
that our findings are driven by the app changes,
rather than by the regulations directly.
We further estimated the individual treat-
ment effects and summarized the estimates in
Fig. 2D (Eq. 1 in Materials and methods). We
made several observations. First, nearly 83% of
the individuals in the treated cities responded
positively to the green nudges (positive and
statistically significant at the 10% level). Sec-
ond, cutlery choices did not change much for
about 6% of the individuals in the treated cities
(close to zero and statistically nonsignificant
at the 10% level). Third, the remaining 11% of
the customers were nudge defiers and behaved
in the opposite direction to what was encour-
aged. Figure S1 further reveals that most nudge
defiers were customers who had previously
placed no-cutlery orders, suggesting that a
small share of environmentally conscious peo-
ple actually disliked being nudged. Fourth, al-
though most of the population responded
positively to the green nudges, the density of
the positive estimates was negatively corre-
lated with the effect magnitude. Finally, most
of the consumers did not set “no cutlery” as
their default option (otherwise, most estimates
would be close to or equal to one) nor did most
of them set “with cutlery” as their default op-
tion (otherwise most estimates would be close
to or equal to zero).
Green nudges did not harm business performance
One concern regarding the green nudges was
that some customers might dislike the new
choice architecture and became less likely to
use the Alibaba platform for food deliveries. If
this were to happen, such nudges could become
unsustainable because the platform would lose
revenue in the long run, with less environ-
mentally friendly rivals gaining a competitive
advantage.
Figure 3 depicts the changes in customers’
total spending on the platform each month
and the total number of orders they placed.
Both the total order amount and the total num-
ber of orders followed the same trend in the
treated and control cities (results in table S4).
The total spending and total number of or-
ders substantially dropped in early 2020 in all
the cities, which was caused by the Chinese
government’s stringent anticontagion policies
during the COVID-19 pandemic. Therefore, we
conclude that the impact of green nudges on
the platform’s business performance is negli-
gible and statistically nonsignificant.
Certain subpopulations responded more
to the green nudges
We examined the heterogeneous impacts of
nudges for different subpopulations. The re-
sults are summarized in Fig. 4 (regression re-
sults in tables S5 to S11).
Several patterns emerged. First, the effects
were larger for women. Specifically, the no-
cutlery orders of women increased by 21.4
percentage points after the nudges were in-
troduced compared with an 18.4 percentage
point increase for men. Second, the effects also
differed across age groups. For customers be-
tween the ages of 18 and 24, green nudges only
increased the share of no-cutlery orders by 11.9
percentage points, whereas for the middle-aged
and elderly, the increase was as much as 30
to 34 percentage points. Each difference was
statistically significant. Third, frequent users
(i.e., those that rarely cooked) responded less
to the green nudges than infrequent users.
Fourth, there were significant differences across
wealth groups, as defined by the value of in-
dividuals’ cell phones. For those using high-
value cell phones (>CNY 8000 or USD 1151 as
of January 2020), green nudges increased the
share of no-cutlery orders by 22.2 percentage
points, which was 3.92 percentage points higher
than for those using low-value cell phones
(≤CNY 1900 or USD 273). Fifth, and relatedly,
those who ordered more-expensive meals (based
on per-order expenditure) responded more to
the green nudges. Thus, whether measured by
cell phone value or meal price, more-affluent
individuals were more responsive to the green
nudges. Finally, the green nudges increased
the share of no-cutlery orders by 24 percent-
age points for individuals who had previously
placed no-cutlery orders before the interven-
tion, which was 4 to 5 percentage points higher
than for those who had never placed a no-
cutlery order before. It is possible that this dif-
ference stems from a difference in inherent
environmental consciousness.
Aggregate environmental benefits were nontrivial
Based on the findings reported in the previous
sections, we estimated that the green nudges,
in total, reduced the number of SUCs in the
A Customers’ Total Spending
B Customers’ Total Number of Orders
200
150
COVID−19 & Lockdown
200
150
COVID−19 & Lockdown
01/2019=100
01/2019=100
50
50
Shanghai
Beijing
Tianjin
Control Group Cities
0
01/2019
07/2019
01/2020
07/2020
01/2021
0
01/2019
07/2019
01/2020
07/2020
01/2021
Fig. 3. Changes in food-delivery business. Changes in (A) customers’ total spending and (B) customers’ total number of orders. The vertical green and blue dashed
lines indicate the introductions of the green nudges in Shanghai and Beijing, respectively. The green nudges were implemented in December 2020 in Tianjin.
He et al., Science 381, eadd9884 (2023)
8 September 2023
4 of 8
RES EARCH | R E S E A R C H A R T I C L E
A
D
B
E
C
F
Fig. 4. Heterogeneous impacts of green nudges on different subpopulations. Subsample analysis based on consumer characteristics (A) gender, (B) age,
(C) order frequency, (D) average expenditure per order, (E) cellphone value, and (F) previous no-cutlery orders before the introduction of green nudges. Each bar
represents a separate regression. Values on the y axes are percentage points (pp). Corresponding results are reported in tables S5 to S10.
treated cities by more than 225.33 million sets
during the 27 months of our study (18 months
in Shanghai, 8 months in Beijing, and 1 month in
Tianjin) (Fig. 5). Based on the weight and com-
position of a typical set of SUC, this reduction
in SUC consumption may have prevented the
generation of 4506.52 metric tons of waste and
saved 56,333 trees (supplementary note 3). Fur-
ther, if all the green points rewarded through
the green nudges were used to plant trees,
112,665 additional trees could have been
planted by Alibaba. If Alibaba introduced the
green nudges to the entire country, more than
8.7 billion sets of SUC would be saved annually
in China (based on 2020 data; supplementary
note 3). Additionally, if all food-delivery ser-
vices in the country adopted green nudges, the
total SUC consumption would decrease by
21.75 billion sets (Fig. 5), which is equivalent
to eliminating 3.26 million metric tons of plas-
tic waste and saving 5.44 million trees.
These numbers should be interpreted as
the upper-bound environmental benefits for
two reasons. First, in reality, despite custo-
mers choosing the no-cutlery option when
placing orders, some restaurants neverthe-
less provided SUC. This would occur, for ex-
ample, when the restaurant was too busy to
customize its orders or was concerned that
the consumers might have mistakenly cho-
sen the no-cutlery option. Second, people often
wasted their green points or did not accumu-
late enough points to plant trees (supplemen-
tary note 1).
Mechanisms behind the behavioral changes
Alibaba’s app changes included three nudging
components: default change, green-points in-
centive, and increased salience of the SUC
options. We conducted additional analyses to
understand the contributions of different nudg-
ing components to the observed nudge effects.
First, our analyses described in supplemen-
tary note 4 show that the incentive to accumu-
late green points was unlikely to be the main
driving force for individuals’ behavioral changes.
This is because (i) the effect of the green nudges
did not depend on an individual’s pretreat-
ment green points (table S12), (ii) the effect
was similar for those who had previously planted
a tree and for those who had never done so
(table S13), (iii) the effect was similar for users
whose green points were closer to the thresh-
old of being able to plant a tree (table S14), and
(iv) the green nudges did not significantly in-
crease individuals’ total green points and most
customers were not actively accumulating green
points by forgoing SUC (table S15).
We then examined the underlying channels
through which the default change could change
individuals’ SUC choices. The previous litera-
ture suggested several theories that might
explain the default effect, including switching
cost, present-biased preferences and procras-
tination (38–40), reference dependence (41–44),
inattention to default change (29, 45–49), and
endorsement effect (42, 50, 51). We carefully
examined these theories in supplementary
note 5 and reached the following conclusions.
First, because the switching cost was negli-
gible in our setting, it was unlikely to be the
driving force of the default effect. Second, the
consumers’ decisions did not involve compli-
cated monetary or discounting calculations
and they needed to make decisions repeatedly,
so present-biased preferences and procrasti-
nation also fell short in explaining the ob-
served behavioral changes. Third, reference
dependence and loss aversion could not ex-
plain the default effect in our setting either.
Under the no-cutlery default, consumers were
not choosing to forgo something of value; there-
fore, loss aversion vis-à-vis the default choice
did not apply. Fourth, given the repeated na-
ture of online food orders and the new default
popping up for each order, inattention to the
default change also failed to explain the ob-
served effects.
Therefore, we conclude that much of the ob-
served effects should be driven by the endorse-
ment effect brought by the default change and
increased salience. Seeing the new user in-
terface, individuals inferred that the default
He et al., Science 381, eadd9884 (2023)
8 September 2023
5 of 8
RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. Estimated reduc-
tion in SUC consumption
from green nudges.
(A) Estimated SUC reduc-
tions in pilot cities. The
green bars illustrate the
counterfactual reductions in
SUC in Beijing and Shanghai
after the introduction of
green nudges. (B) Estimated
potential reduction in SUC
for all of China if the green
nudges had been introduced
nationally in July 2019.
A Estimated SUC Reductions in Three Pilot Cities
Nudge Started in Beijing
Observed SUC
SUCs Without Green Nudges
Reductions in SUC
Nudge Started in Shanghai
300
200
100
s
t
e
s
d
n
a
s
n
u
o
h
t
,
y
r
e
l
t
l
u
C
e
s
U
−
e
g
n
S
d
e
c
u
d
e
R
i
0
01/2019
07/2019
01/2020
07/2020
01/2021
B Simulated Reductions in the Entire Country
s
t
e
s
n
o
i
l
l
i
b
,
y
r
e
l
t
u
C
e
s
U
−
e
g
n
S
n
i
l
i
s
n
o
i
t
c
u
d
e
R
Monthly Reductions
Cumulative Reductions
Nudge Started in the Entire Country
4
3
2
1
40
30
20
10
0
01/2019
07/2019
01/2020
07/2020
0
01/2021
s
t
e
s
n
o
i
l
l
i
b
,
y
r
e
l
t
u
C
e
s
U
−
e
g
n
S
n
i
l
i
s
n
o
i
t
c
u
d
e
R
e
v
i
t
a
u
m
u
C
l
option was encouraged and endorsed by the
platform and thus became more likely to choose
it. Meanwhile, the pop-up window made the
decision about SUC more salient and reminded
consumers that the no-cutlery option was avail-
able, so the increased salience of the new de-
fault could also have played a role.
Discussion
Using data from China, we show in this work
that green nudges can substantially reduce
plastic waste generated by the food-delivery
industry. Changing the default to “no cutlery”
and rewarding consumers with green points
on Alibaba’s food-delivery platform increased
the share of no-cutlery orders by 648%, and
the effect persisted. If green nudges were ap-
plied to all of China, more than 21.75 billion sets
of SUC would be saved annually—equivalent to
20.4% of plastic waste reduction in the food
delivery industry—which would eliminate 3.26
million metric tons of plastic waste and save
5.44 million trees.
There are several implications of this study.
First, it offers compelling evidence that nudges
can be a powerful tool to change behaviors.
This finding contrasts with some recent reviews
that show that nudges are largely ineffective,
especially after accounting for publication
bias (52, 53). Two factors may help explain the
large differences between our estimates and
those in the literature. First, Alibaba’s green
nudges were introduced into a new domain
where most individuals had not been affected
by other policies or other types of encourage-
ment, so the baseline take-up rate was low.
Second, the endorsement effect might be par-
ticularly strong in our setting because adher-
ing to the default does not incur a notable cost
for most individuals. We thus conclude that
the effectiveness of nudges is highly context-
specific, and more nudges should be explored
in new domains.
The results also help us better understand
how individuals adjusted their behaviors. For
example, the impacts of nudging were greatest
in the first few months after the introduction
and decreased slightly over time, indicating
that some customers later decided to override
the default no-cutlery option. This is consistent
with previous literature on the decaying effect
of default (23). We also found that women,
middle-aged or older individuals, frequent
users, affluent consumers, and environmen-
tally conscious customers were more respon-
sive to the green nudges. This information can
be used to improve the targeting of future pro-
environmental interventions. Further, the in-
dividual treatment effects revealed that the
green nudges incentivized a large portion of
individuals to somewhat change their behav-
iors rather than encouraging only a small por-
tion to change their behaviors substantially
and that people seemed to get prompted every
time and occasionally opted not to have SUCs,
rather than setting their own defaults. We also
documented a somewhat puzzling phenome-
non: A small portion of individuals were nudge
defiers and responded in the opposite di-
rection to what was encouraged. The reasons
for such antinudging behaviors remain largely
unknown.
Another important finding is that green
nudges did not negatively affect Alibaba’s
business, suggesting that they can be a highly
cost-effective tool to promote individuals’ pro-
environmental behaviors. For Alibaba, the cost
of implementing the green nudges was trivial
because it only required several software en-
gineers to redesign the user interface. Yet the
environmental benefits were tremendous. We
thus recommend that other online food-delivery
platforms, such as DoorDash and Uber Eats,
try similar green nudges to reduce global plas-
tic waste. More generally, we think that the
private sector and platform companies can
play a powerful role in promoting prosocial
behaviors among their customers. Better align-
ment between their corporate social responsi-
bilities and ecofriendly initiatives could bring
about far-reaching impacts to our planet.
We conclude by noting the caveats of this
study and future directions. First, because of
the lack of data from other food-delivery plat-
forms, we cannot examine whether Alibaba’s
green nudges had positive spillovers for tran-
sactions on their platforms. If positive spill-
overs exist, the benefits of green nudges would
be even greater. Second, through field research,
we learned that some restaurants provided
cutlery to customers even when the no-cutlery
option was chosen. This could undermine the
power of the green nudges and reduce the
environmental benefits associated with cus-
tomers’ pro-environmental behaviors. Future
research is warranted to find ways to encour-
age restaurants to better comply with custom-
ers’ requests. Finally, although the aggregated
environmental benefits of the green nudges
were nontrivial, a substantial portion of the
solid and plastic waste in the food-delivery in-
dustry comes from packaging (i.e., using dis-
posable food containers and plastic bags for
delivery services). In typical Chinese food-delivery
orders, packaging waste accounts for more
than 80% of the food-delivery waste (54). Ad-
ditional policies should be introduced to better
control the waste generated in the packaging
process.
Materials and methods
Data
We obtained data from Eleme, Alibaba’s food-
ordering and -delivery platform, which is similar
He et al., Science 381, eadd9884 (2023)
8 September 2023
6 of 8
RES EARCH | R E S E A R C H A R T I C L E
to DoorDash and Uber Eats. From June 2019
to June 2022, the total number of monthly active
users on Eleme increased from 709 million to
753 million, making it the second-largest plat-
form in China (next to Meituan) (55).
The data included 197,062 randomly selected
users’ monthly food-ordering history, their
green-points history, their tree-planting records,
and their personal characteristics from 1 January
2019 to 31 December 2021 in 10 major Chinese
cities. All of the consumers in the sample set
placed at least one food delivery order during
the research period.
Among the 10 cities, there were three treated
cities, including Beijing, Shanghai, and Tianjin.
Seven cities were the control cities, including
Qingdao, Xi’an, Guangzhou, Nanjing, Hangzhou,
Wuhan, and Chengdu. According to China’s
provincial statistical yearbook, these cities had
a combined total population of 157.36 million
as of 2020.
Eleme provided monthly data on each cus-
tomer’s food-ordering history, including the
number of orders, the number of no-cutlery
orders, and total expenditures. Based on this
information, we calculated the share of no-
cutlery orders (SNCO) in percentage terms,
that is, we divided the number of no-cutlery
orders by the total number of orders and mul-
tiplied by 100. Data on green points included
total rewarded green points, total rewarded
green points from no-cutlery orders, and total
harvested and/or accumulated green points.
More details about the green points are dis-
cussed in supplementary note 1. Data on tree-
planting records included each customer’s
cumulative and current-month tree-planting
activities. Customers’ personal characteristics
included gender, age group, and approximate
value of their cellphone, which was computed
by Eleme’s algorithm. The correlations between
pretreatment SNCO behaviors and consumers’
characteristics are summarized in table S16.
Model
We used a difference-in-differences approach
to quantify the impact of the green nudges on
individuals’ SUC-ordering behaviors. Specifi-
cally, the following model was estimated:
SNCOict = a + b*Nudgeict + pi + rt + eict (1)
Total number of orders
where the dependent variable is the SNCO for in-
dividual i in city c in year-month t, which is cal-
culated by SNCOict ¼ Number of orders without SUC
(cid:2) 100.
If a customer did not make any food-delivery
order in a specific month, we dropped this ob-
servation (rather the entire consumer’s record)
from the regression. Nudgeict is the treatment
variable that equals one if individual i in city
c received the green nudges in a specific year
and month and is zero otherwise. We also
controlled for individual fixed effects, pi, and
time (year-by-month) fixed effects, rt. a is the
intercept, b is the coefficient of interest, and eict
is the error term. Standard errors were two-way
clustered at the individual level and at the city-
year-month level (56).
The difference-in-differences model estimates
the treatment effects of nudges, b, by compar-
ing the changes in the share of no-cutlery or-
ders in the treated cities (Beijing, Shanghai,
and Tianjin) with changes in the control cities
(seven cities in the control group) before and
after the nudges were implemented. The iden-
tifying assumption was that, in the absence of
the green nudges, customers in the treated
cities would display similar SUC usage trends
as customers in the control cities. We tested
this assumption by examining the trends in
SNCO for both treated and control cities be-
fore and after the introduction of the green
nudges. Specifically, we estimated an event
study specification as follows:
SNCOict ¼ a þ
X
k¼9
k≥(cid:3)9;k≠(cid:3)1
bk(cid:2)D Nudgeict;k þ
pi þ rt
þ eict
ð2Þ
where D_Nudgeict,k is a set of dummy variables
indicating the treatment status in different
periods: For individual i in city c who was never
nudged, D_Nudgeict,k was always set equal to
zero; but for individual i in city c who received
the green nudges, we first defined Sct as the
period during which the nudges were intro-
duced and then we defined D_Nudgeict,k = 1 if t –
Sct = k and equal to zero otherwise, where
k ∈ (cid:3)9; 9
(cid:4). The dummy for k = –1 was omitted
½
in the regression as the reference group. There-
fore, bk dynamically captured the impacts of
green nudges from when they were first in-
troduced until 9 months later. By including
leads of treatment timing, we could test whether
pretrends differed for treatment and control
cities by examining whether green nudges had
any impact on outcomes up to 9 months be-
fore the actual implementation.
To understand how each individual was
affected by the green nudges relative to the
control group, we estimated the individual treat-
ment effects using Eq. 1 and only a subsample
for the estimation: individuals in the control
cities plus one individual in the treated city.
The standard errors were then clustered at
the individual level. We then repeated the re-
gression for all the treated individuals, ob-
taining a large number of individual-specific
treatment-effect estimates. Note that the in-
dividual effect estimates should be interpreted
with caution because the statistical power was
limited given the small sample size for each
individual.
To investigate whether the treatment effect of
the green nudges was stronger for users whose
points were closer to the threshold of being able
to plant a tree (and thus revealed a strong mo-
tivation to respond to the green-points incentive),
we followed Grembi et al. (57) and Giambona
and Ribas (58) and applied a difference-in-
discontinuities specification using the follow-
ing model:
SNCOict = b1Belowi + b2Diffi + b3Belowi · Diffi
+ g0Nudgeict + g1Belowi · Nudgeict + g2Diffi ·
Nudgeict + g3Belowi · Diffi · Nudgeict + eict
subject to –h ≤ Diffi ≤ h
(3)
where SNCOict is the share of no-cutlery orders
for individual i in city c in year-month t. Dum-
my variable Belowi was equal to one when indi-
vidual i’s points were below 16,000 points—
because the minimum requirement for planting
a tree was 16,000 points—before the treatment
and equal to zero if individual i’s points were
above or equal to the threshold. Diffi is the
running variable calculated by the difference
between individual i’s points and the thresh-
old of 16,000 points before the treatment. A
negative value for Diffi suggests that individ-
ual i’s points are below 16,000. Nudgeict is the
treatment variable that equals one if individual
i in city c received the green nudges in a specific
year and month and that equals zero otherwise.
h is the bandwidth chosen by researchers to esti-
mate the discontinuity. Individual fixed effects
and year-month fixed effects were included in
the regression. Standard errors were two-way
clustered both at the individual level and at
the city-year-month level. The difference-in-
discontinuity estimator is g1, which measures
the treatment effect on individuals who were
below the threshold.
REFERENCES AND NOTES
1.
J. R. Jambeck et al., Plastic waste inputs from land into the
ocean. Science 347, 768–771 (2015). doi: 10.1126/
science.1260352; pmid: 25678662
2. L. C. M. Lebreton et al., River plastic emissions to the world’s
oceans. Nat. Commun. 8, 15611 (2017). doi: 10.1038/
ncomms15611; pmid: 28589961
3. S. B. Borrelle et al., Predicted growth in plastic waste exceeds
efforts to mitigate plastic pollution. Science 369, 1515–1518
(2020). doi: 10.1126/science.aba3656; pmid: 32943526
4. T. M. Adyel, Accumulation of plastic waste during COVID-19.
Science 369, 1314–1315 (2020). doi: 10.1126/science.abd9925;
pmid: 32913095
5. Xinhua Net; http://www.xinhuanet.com/fortune/2019-12/19/
c_1125362989.htm.
6. R. Qi, D. L. Jones, Z. Li, Q. Liu, C. Yan, Behavior of microplastics
and plastic film residues in the soil environment: A critical
review. Sci. Total Environ. 703, 134722 (2020). doi: 10.1016/
j.scitotenv.2019.134722; pmid: 31767311
7. National Development and Reform Commission, Ministry of
Ecology and Environment, “Opinions on further strengthening
the cleanup of plastic pollution” (2020); https://www.ndrc.gov.
cn/xxgk/zcfb/tz/202001/t20200119_1219275.html.
8. R. H. Thaler, From cashews to nudges: The evolution of
behavioral economics. Am. Econ. Rev. 108, 1265–1287 (2018).
doi: 10.1257/aer.108.6.1265
9. R. H. Thaler, Nudge, not sludge. Science 361, 431–431 (2018).
doi: 10.1126/science.aau9241; pmid: 30072515
10. R. H. Thaler, C. R. Sunstein, Nudge: Improving Decisions
About Health, Wealth, and Happiness (Penguin Books,
2009).
11. F. Carlsson, C. Gravert, O. Johansson-Stenman, V. Kurz,
The use of green nudges as an environmental policy
He et al., Science 381, eadd9884 (2023)
8 September 2023
7 of 8
RES EARCH | R E S E A R C H A R T I C L E
instrument. Rev. Environ. Econ. Policy 15, 216–237 (2021).
doi: 10.1086/715524
waste. Front. Psychol. 10, 266 (2019). doi: 10.3389/
fpsyg.2019.00266; pmid: 30863332
12. C. Schubert, Green nudges: Do they work? Are they ethical?
31. H. He, Effects of environmental policy on consumption:
Ecol. Econ. 132, 329–342 (2017). doi: 10.1016/
j.ecolecon.2016.11.009
13. J. Wensing, V. Caputo, L. Carraresi, S. Bröring, The effects of
green nudges on consumer valuation of bio-based plastic
packaging. Ecol. Econ. 178, 106783 (2020). doi: 10.1016/
j.ecolecon.2020.106783
14. R. Croson, N. Treich, Behavioral environmental economics:
Promises and challenges. Environ. Resour. Econ. 58, 335–351
(2014). doi: 10.1007/s10640-014-9783-y
15. M. Akbulut-Yuksel, C. Boulatoff, The effects of a green nudge
on municipal solid waste: Evidence from a clear bag policy.
J. Environ. Econ. Manage. 106, 102404 (2021). doi: 10.1016/
j.jeem.2020.102404
16. D. Miłaszewicz, Survey results on using nudges for choice of
green-energy supplier. Energies 15, 2679 (2022). doi: 10.3390/
en15072679
17. G. He, Y. Pan, A. Park, Y. Sawada, E. Tan, Data and code
info for: Reducing single-use cutlery with green nudges:
Evidence from China’s food delivery industry. Dryad (2023);
https://doi.org/10.5061/dryad.wm37pvmsx.
18. M. Fowlie et al., Default effects and follow-on behaviour:
Evidence from an electricity pricing program. Rev. Econ. Stud.
88, 2886–2934 (2021). doi: 10.1093/restud/rdab018
19. J. Blumenstock, M. Callen, T. Ghani, Why do defaults affect
behavior? Experimental evidence from Afghanistan. Am. Econ.
Rev. 108, 2868–2901 (2018). doi: 10.1257/aer.20171676
20. B. D. Bernheim, A. Fradkin, I. Popov, The welfare economics of
default options in 401(k) plans. Am. Econ. Rev. 105,
2798–2837 (2015). doi: 10.1257/aer.20130907
21. R. Chetty, J. N. Friedman, S. Leth-Petersen, T. H. Nielsen,
T. Olsen, Active vs. passive decisions and crowd-out in retirement
savings accounts: Evidence from Denmark. Q. J. Econ. 129,
1141–1219 (2014). doi: 10.1093/qje/qju013
22. U. Liebe, J. Gewinner, A. Diekmann, Large and persistent
effects of green energy defaults in the household and business
sectors. Nat. Hum. Behav. 5, 576–585 (2021). doi: 10.1038/
s41562-021-01070-3; pmid: 33707659
23. A. Brandon, P. Ferraro, J. List, R. Metcalfe, M. Price,
F. Rundhammer, “Do the effects of nudges persist? Theory and
evidence from 38 natural field experiments,” Working Paper
w23277, National Bureau of Economic Research, Cambridge,
MA, March 2017.
24. L. Bovens, “The ethics of nudge” in Preference Change,
T. Grüne-Yanoff, S. O. Hansson, Eds. (Springer, 2009),
pp. 207–219.
25. N. Dhingra, Z. Gorn, A. Kener, J. Dana, The default pull:
An experimental demonstration of subtle default effects on
preferences. Judgm. Decis. Mak. 7, 69–76 (2012). doi: 10.1017/
S1930297500001844
26. P. G. Hansen, A. M. Jespersen, Nudge and the manipulation of
choice: A framework for the responsible use of the nudge
approach to behaviour change in public policy. Eur. J. Risk
Regul. 4, 3–28 (2013). doi: 10.1017/S1867299X00002762
27. H. Bruns, E. Kantorowicz-Reznichenko, K. Klement,
M. Luistro Jonsson, B. Rahali, Can nudges be transparent and
yet effective? J. Econ. Psychol. 65, 41–59 (2018). doi: 10.1016/
j.joep.2018.02.002
28. F. Ebeling, S. Lotz, Domestic uptake of green energy promoted
by opt-out tariffs. Nat. Clim. Chang. 5, 868–871 (2015).
doi: 10.1038/nclimate2681
29. N. C. Smith, D. G. Goldstein, E. J. Johnson, Choice without
awareness: Ethical and policy implications of defaults.
J. Public Policy Mark. 32, 159–172 (2013). doi: 10.1509/
jppm.10.114
30. G. O. Thomas, E. Sautkina, W. Poortinga, E. Wolstenholme,
Lessons from the Chinese plastic bag regulation. Environ. Dev.
Econ. 17, 407–431 (2012). doi: 10.1017/S1355770X1200006X
32. S. Larcom, L. A. Panzone, T. Swanson, Follow the leader?
Testing for the internalization of law. J. Legal Stud. 48,
217–244 (2019). doi: 10.1086/699817
33. J. M. Cabrera, M. Caffera, A. Cid, Modest and incomplete
incentives may work: Pricing plastic bags in Uruguay.
J. Environ. Econ. Manage. 110, 102525 (2021). doi: 10.1016/
j.jeem.2021.102525
34. C. Lai, “Chongqing green group sues three delivery services
over disposable cutlery,” Hong Kong Free Press, 6 September
2017; https://hongkongfp.com/2017/09/06/chongqing-
green-group-sues-three-delivery-services-disposable-cutlery/.
35. http://dse.rcees.cas.cn/dtxx/dtxxwz/202103/
t20210331_632275.html
36. R. Münscher, M. Vetter, T. Scheuerle, A review and taxonomy
of choice architecture techniques. J. Behav. Decis. Making 29,
511–524 (2016). doi: 10.1002/bdm.1897
37. J. Gallus, Fostering public good contributions with symbolic
awards: A large-scale natural field experiment at Wikipedia.
Manage. Sci. 63, 3999–4015 (2017). doi: 10.1287/mnsc.2016.2540
38. T. O’Donoghue, M. Rabin, Choice and procrastination. Q. J. Econ.
116, 121–160 (2001). doi: 10.1162/003355301556365
39. J. J. Choi, D. Laibson, B. C. Madrian, A. Metrick, Optimal
defaults. Am. Econ. Rev. 93, 180–185 (2003). doi: 10.1257/
000282803321947010
40. G. D. Carroll, J. J. Choi, D. Laibson, B. C. Madrian, A. Metrick,
Optimal defaults and active decisions. Q. J. Econ. 124,
1639–1674 (2009). doi: 10.1162/qjec.2009.124.4.1639;
pmid: 20041043
41. R. H. Thaler, S. Benartzi, Save More TomorrowTM: Using
behavioral economics to increase employee saving. J. Polit.
Econ. 112, S164–S187 (2004). doi: 10.1086/380085
42. B. C. Madrian, D. F. Shea, The power of suggestion: Inertia in
401(k) participation and savings behavior. Q. J. Econ. 116,
1149–1187 (2001). doi: 10.1162/003355301753265543
43. K. M. Marzilli Ericson, A. Fuster, The endowment effect.
Annu. Rev. Econ. 6, 555–579 (2014). doi: 10.1146/
annurev-economics-080213-041320
44. R. H. Thaler, Toward a positive theory of consumer choice.
J. Econ. Behav. Organ. 1, 39–60 (1980). doi: 10.1016/
0167-2681(80)90051-7
45. R. Chetty, A. Looney, K. Kroft, Salience and taxation: Theory
and evidence. Am. Econ. Rev. 99, 1145–1177 (2009).
doi: 10.1257/aer.99.4.1145
46. H. Allcott, Social norms and energy conservation. J. Public Econ.
95, 1082–1095 (2011). doi: 10.1016/j.jpubeco.2011.03.003
47. D. Hirshleifer, S. S. Lim, S. H. Teoh, Driven to distraction:
Extraneous events and underreaction to earnings news.
J. Finance 64, 2289–2325 (2009). doi: 10.1111/
j.1540-6261.2009.01501.x
48. B. Bollinger, P. Leslie, A. Sorensen, Calorie posting in chain
restaurants. Am. Econ. J. Econ. Policy 3, 91–128 (2011).
doi: 10.1257/pol.3.1.91
49. J. R. Kling, S. Mullainathan, E. Shafir, L. C. Vermeulen,
M. V. Wrobel, Comparison friction: Experimental evidence from
Medicare drug plans. Q. J. Econ. 127, 199–235 (2012).
doi: 10.1093/qje/qjr055; pmid: 22454838
50. C. R. M. McKenzie, M. J. Liersch, S. R. Finkelstein, Recommendations
implicit in policy defaults. Psychol. Sci. 17, 414–420 (2006).
doi: 10.1111/j.1467-9280.2006.01721.x; pmid: 16683929
51. J. R. Brown, A. M. Farrell, S. J. Weisbenner, Decision-making
approaches and the propensity to default: Evidence and
implications. J. Financ. Econ. 121, 477–495 (2016).
doi: 10.1016/j.jfineco.2016.05.010
architecture interventions across behavioral domains.
Proc. Natl. Acad. Sci. U.S.A. 119, e2107346118 (2022).
doi: 10.1073/pnas.2107346118; pmid: 34983836
53. M. Maier et al., No evidence for nudging after adjusting for
publication bias. Proc. Natl. Acad. Sci. U.S.A. 119,
e2200300119 (2022). doi: 10.1073/pnas.2200300119;
pmid: 35858389
54. Y. Zhang, Z. Wen, Mapping the environmental impacts and
policy effectiveness of takeaway food industry in China.
Sci. Total Environ. 808, 152023 (2022). doi: 10.1016/
j.scitotenv.2021.152023; pmid: 34861304
55. Data sources are as follows: https://www.chyxx.com/industry/
202009/896228.html#:~:text=2015%2D2019%E5%B9%B4%
E9%A5%BF%E4%BA%86,%E5%A2%9E%E5%8A%A0%E4%
BA%860.08%E4%BA%BF%E4%BA%BA%E3%80%82 and
https://www.statista.com/statistics/1218045/china-leading-
food-delivery-apps-based-on-monthly-active-users/
#statisticContainer.
56. A. C. Cameron, J. B. Gelbach, D. L. Miller, Robust inference with
multiway clustering. J. Bus. Econ. Stat. 29, 238–249 (2011).
doi: 10.1198/jbes.2010.07136
57. V. Grembi, T. Nannicini, U. Troiano, Do fiscal rules matter?
Am. Econ. J. Appl. Econ. 8, 1–30 (2016). doi: 10.1257/
app.20150076
58. E. Giambona, R. P. Ribas, Unveiling the Price of Obscenity:
Evidence from Closing Prostitution Windows in Amsterdam.
J. Policy Anal. Manage. 42, 677–705 (2023). doi: 10.1002/
pam.22459
AC KNOWLED GME NTS
We benefited from L. Goette’s insights and thank seminar
participants at the ADB project workshop for their comments and
suggestions. We also thank two Alibaba project coordinators,
R. Zhang and Y. Wang, for providing data. S. Kimura, D. Guo,
A. Tan, J. Li, and Y. Quan provided excellent coordination and
research assistance. Funding: We acknowledge funding support
from the Research Grant Council of Hong Kong (theme-based
research grant T31-603/21-N) and Peking University (early career
grant 7101303264). The views expressed in this paper are those
of the authors only and do not necessarily reflect the views and
policies of the Asian Development Bank or its board of governors or
the governments they represent. Author contributions: All authors
contributed equally to this work and are listed alphabetically in the
author list. Conceptualization: G.H., A.P., Y.S., E.S.T.; Methodology and
data collection: Y.P., E.S.T.; Analysis: G.H., Y.P., A.P., Y.S., E.S.T.;
Writing – original draft: G.H., Y.P.; Writing – review and editing: G.H.,
Y.P., A.P., Y.S., E.S.T. Competing interests: The authors declare that
they have no competing interests. Data and materials availability:
The data usage guidelines and codes necessary to re-produce and
extend all the tables from this research can be accessed through
Dryad (17). The customer-level data contain private information and
are thus stored by a designated server provided by Alibaba. Those
who want to access the data for replication should file an application
following the instructions provided by (17). License information:
Copyright © 2023 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science. No
claim to original US government works. https://www.science.org/
about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add9884
Supplementary Text
Figs. S1 to S4
Tables S1 to S16
References (59–71)
MDAR Reproducibility Checklist
L. Whitmarsh, The English plastic bag charge changed behavior
and increased support for other charges to reduce plastic
52. S. Mertens, M. Herberz, U. J. J. Hahnel, T. Brosch, The
effectiveness of nudging: A meta-analysis of choice
Submitted 18 July 2022; accepted 26 July 2023
10.1126/science.add9884
He et al., Science 381, eadd9884 (2023)
8 September 2023
8 of 8
|
10.1126_science.ade2650
|
RES EARCH
BIOPHYSICS
MINFLUX dissects the unimpeded walking of kinesin-1
Jan O. Wirth1†, Lukas Scheiderer1†, Tobias Engelhardt1‡, Johann Engelhardt1,
Jessica Matthias1‡, Stefan W. Hell1,2*
We introduce an interferometric MINFLUX microscope that records protein movements with up to
1.7 nanometer per millisecond spatiotemporal precision. Such precision has previously required
attaching disproportionately large beads to the protein, but MINFLUX requires the detection of only
about 20 photons from an approximately 1-nanometer-sized fluorophore. Therefore, we were able to
study the stepping of the motor protein kinesin-1 on microtubules at up to physiological adenosine-5′-
triphosphate (ATP) concentrations. We uncovered rotations of the stalk and the heads of load-free
kinesin during stepping and showed that ATP is taken up with a single head bound to the microtubule
and that ATP hydrolysis occurs when both heads are bound. Our results show that MINFLUX quantifies
(sub)millisecond conformational changes of proteins with minimal disturbance.
E xploring movements and conformational
changes of proteins lies at the heart of
unraveling the inner workings of a cell,
but the tools for accomplishing this task
have so far been limited. Nanometer-sized
protein motions of millisecond duration can
be retrieved by tethering the protein to a bead
held in an infrared optical trap and measuring
the bead’s movement (1–3). However, this prep-
aration subjects the protein to a force and
thus does not allow the direct observation of
its entirely free motion. The 70- to 500-nm-
diameter beads required for optical trapping
are also orders of magnitude larger than the
protein itself, which also causes problems, in-
cluding susceptibility to laser-induced heating
(4). Alternatively, a protein can be observed
with no or minimal restrictions by labeling it
with an ~1-nm-sized organic fluorophore and
recording its motion with the camera of a light
microscope (5, 6). The position of the label is
then inferred from the peak of the fluorescence
diffraction pattern rendered by N camera-
detected photons. Unfortunately, the result-
ffiffiffiffi
N
ing localization precision s scales with 1=
,
meaning that s = 1 to 2 nm typically requires
N > 2500 photons (7). Thus, even the brightest
fluorophores entail localization times of
hundreds of milliseconds. Fluorescence-
based localization therefore cannot live up to
the spatiotemporal resolution (STR) provided
by optical traps. Replacing the tiny fluoro-
phore with a laser-scattering gold bead of
≥30 nm diameter (8–10) can compensate
for this shortfall, but the volume, drag, and
electrostatic interactions of the gold bead
preclude unimpeded protein motion.
p
1Department of Optical Nanoscopy, Max Planck Institute for
Medical Research, Heidelberg, Germany. 2Department of
NanoBiophotonics, Max Planck Institute for Multidisciplinary
Sciences, Göttingen, Germany.
*Corresponding author. Email: shell@gwdg.de
†These authors contributed equally to this work.
‡Present address: Abberior Instruments GmbH, Göttingen,
Germany.
These limits are also reflected in the under-
standing of the arguably best-studied moving
protein, the homodimeric motor protein kinesin-1,
hereafter called kinesin, which is responsible
for the anterograde transport on microtubules
and the malfunction of which is linked to
diseases (11–14). Although the above tools have
greatly advanced our understanding of how
kinesin walks, many details of its mechano-
chemical cycle have remained controversial or
elusive (15, 16).
ffiffi
t
p
We reasoned that MINFLUX (17), a recently
introduced microscopy method for localizing
fluorophores with a minimal rather than a
maximal number of detected photons N,
should greatly improve the study of pro-
tein movements. For a given N, MINFLUX
(17–20) typically renders an STR of s
with
~10-fold improved s, or a 100-fold increased
temporal resolution t compared with popular
camera-based localizations (18). Using a single
fluorophore of ~1 nm in size, an STR is attained
that has so far required the use of bulky beads.
This combination of STR and a small label has
motivated us to revisit the walking of kinesin.
Here, we report on an interferometric MINFLUX
implementation that delivers nanometer/
submillisecond STR in protein tracking. Har-
nessing this STR, we determined the steps
and substeps of the heads and the stalk of
kinesin. The direct observation of unhindered
substeps allowed us to determine in which
state adenosine-5′-triphosphate (ATP) binds
and hydrolyzes and to uncover orientation
changes of functional subunits of kinesin
during stepping. Our study concomitantly
establishes MINFLUX as a tool for exam-
ining fast protein movements and confor-
mational changes at nanometer scale with
minimal or no impediment.
Interferometric MINFLUX maximizes
fluorophore localization precision
A scanning MINFLUX microscope features a
beam for fluorophore excitation with a cen-
tral intensity minimum (“zero”) that is posi-
tioned in the sample with subnanometer
precision. Emitted photons are counted by a
confocal point detector (Fig. 1A). The closer
the central excitation minimum is to the
fluorophore, the lower the fluorescence rate,
meaning that the number of detections readily
discloses the distance between the unknown
position of the fluorophore and the perfectly
known position of the minimum. In fact, the
intensity of the excitation beam around the
minimum increases quadratically with dis-
tance to the minimum (Fig. 1B), with steep-
ness depending on the beam’s focusing angle,
wavelength, and power. Therefore, the fluo-
rescence detection rate displays the same
quadratic dependence on the fluorophore-
to-minimum distance (Fig. 1B). If the rate
is minimal, i.e., down to background level,
then the fluorophore is localized because
the position of the fluorophore coincides
with that of the excitation beam minimum.
However, because of the adverse role of
background, matching the two positions at
the angstrom level is usually not possible.
Fortunately, such perfect matching is not
needed because the mismatch and thus the
fluorophore position can be precisely derived
from a relatively small number of photons N
gained by targeting the minimum to two or
more positions within a small spatial interval
L containing the fluorophore (Fig. 1B).
(cid:4)
MINFLUX localization of a fluorophore
located at an unknown position within the
diffraction limit (~200 nm) is performed
iteratively (19) by continually shifting the
minimum closer to the fluorophore. The
localization usually starts out from an inter-
val L of ~200 nm, which is then reduced on
the basis of the initially derived precision s0.
In theory, an iterative reduction of L in pro-
portion to the precision sk–1 of the previous
step, Lk = ask–1, gives rise to an exponential
increase in precision after k steps: sk ¼
(cid:3)
s0exp (cid:2) 8N
. The parameter a ensures that
ea2
the next Lk is small enough to quickly zoom
in on the fluorophore but large enough to
keep the fluorophore within the next interval.
This exponential increase in precision with N
signifies a most efficient use of the detected
photons and should be contrasted with the
sluggish 1=
dependence in camera-based
localization (see supplementary text section
2.2). The reduction of Lk ends just before
the quadratic dependence disappears amid
background. Therefore, practical MINFLUX
localization precision s is limited by the
steepness-to-background ratio of the excita-
tion beam. In principle, steepness can be
arbitrarily increased by increasing the beam
power, but because this measure also increases
the background, we designed a MINFLUX
system that inherently yields higher steep-
ness compared with reported donut-based
systems (17–20).
ffiffiffiffi
N
p
Wolff et al., Science 379, 1004–1010 (2023)
10 March 2023
1 of 7
RES EARCH | R E S E A R C H A R T I C L E
A
B
x
100 nm
-L/2 0 L/2
Δφ=0 π/2
π
3π/2
Lens
X
Y
Δφ
DM
Lens
PH
APD
C
)
m
n
(
σ
5
4
3
2
Phase/amplitude
modulators
r
e
s
a
L
FPGA
L= 30 nm
L= 16 nm
Δt=631 µs
2.1 nm
1.7 nm
y
t
i
s
n
e
t
n
I
l
a
n
g
S
i
80
60
40
20
0
D
)
m
n
(
n
o
i
t
i
s
o
P
Background
-L/2
xM
0
X
1
10
20
40
60
Photons per dimension
-20
0.0
0.5
1.0
Time (s)
φ-L/2
φ0
φ+L/2
L/2
x
x
y
2.2 nm
2 nm
2
1.5
Fig. 1. Interferometric MINFLUX microscope provides nanometer localization precision of a fluoro-
phore with 20 to 40 detected photons. (A) Simplified setup. A 640-nm laser beam is shaped by a phase
modulator and an amplitude modulator to create a pair of beams with a defined phase difference ϕ in
the entrance pupil of an objective lens. Their interference creates an intensity pattern with a local line-shaped
minimum in the focal plane. The minimum is shifted by changing ϕ with the phase/amplitude modulator.
Two orthogonal pairs of beams are used for covering both the x and y directions. The fluorescence
collected by the lens passes a dichroic mirror (DM) that otherwise deflects the laser light, is focused onto
a confocal pinhole (PH), and is detected by an avalanche photodiode (APD). The (x,y) localization algorithm
is implemented in a field programmable gate array (FPGA) directing the minimum to specified (x,y)
positions, depending on the number of photons detected by the APD. (B) Top: MINFLUX localization in
one dimension, using a change of ϕ to place the minimum at positions –L/2, 0, and L/2 of a linear interval
around the expected molecule position xM. Bottom: The photons counts measured with the minimum at these
three points allow retrieving xM by fitting with a parabola. In iterative MINFLUX localizations in which the
excitation intensity minimum approaches the fluorophore, the laser power is increased accordingly to
keep the fluorescence rate at the same level. Background is caused by nonvanishing excitation intensity
at the minimum, stray light, and detector dark counts. (C) Localization precision s of single surface–
immobilized ATTO 647N fluorophores using L = 30 nm (263 fluorophores) and L = 16 nm (232 fluorophores),
yielding s = 2.1 and 1.7 nm, respectively. (D) Tracking of a single fluorophore moved by a piezoelectric
stage [L = 30 nm, 2 nm residual noise, 0.607 ms temporal resolution, 70 photons per (x,y) localization]
with corresponding step fit in the x direction and constant fit in the y direction.
Specifically, our MINFLUX setup featured
two pairs of oblique beams that interfered de-
structively in the focal plane (Fig. 1A). One
of the pairs was arranged in the x direction,
rendering a y-oriented line-shaped mini-
mum for x localization; the pair for y localization
was arranged accordingly in the y direction.
Line-shaped minima have also been used in
stimulated emission depletion (STED) micros-
copy (21) because they require fewer polariza-
tion and aberration optimizations while pro-
viding higher steepness (fig. S1) and lower
background. Altering the phase difference of
the x-arranged beams moved the y-oriented
line-shaped minimum with angstrom precision
in the x direction and vice versa. By targeting
the minima to coordinates –Lk/2, 0 and Lk/2
around the last estimated fluorophore posi-
tion, the position was iteratively established
for each dimension (x and y) on the basis of
the number of detections (Fig. 1B). The (x,y)
trajectory was obtained by repeatedly switch-
ing between x and y using an electro-optical
modulator (figs. S2 and S3). Once Lk =16 nm
was reached, as few as ~20 detected photons
sufficed to localize single immobilized ATTO
647N fluorophores with an average precision
s = 1.7 nm per dimension (Fig. 1C). For Lk =
30 nm, a precision s = 2.1 nm was obtained
with ~28 photons. Because the average signal-
to-background ratio was more than three
times higher for Lk = 30 nm, we performed
all tracking measurements with Lk= L = 30 nm
(fig. S4), ensuring robustness in the process.
In fluorophore tracking, the successive small
changes in fluorophore position inherently
allow for the continual use of L = 30 nm and
thus for the maximal use of the N photons
detected.
The tracking accuracy of our MINFLUX
system was highlighted by moving an indi-
vidual ATTO 647N fluorophore on a periodic
stepping trajectory along the x axis of a
piezoelectric stage (Fig. 1D). The steps were
fitted with an algorithm based on an itera-
tive change-point search (22) that was used
throughout our study. Our analysis showed
that ~70 photons recorded within 607 ms
clearly identified the steps with s = 2 nm in
both the x and y directions.
MINFLUX observes substeps and stalk
rotation of kinesin
Under consumption of an ATP molecule, the
catalytic motor domains (heads) of kinesin
take hand-over-hand steps of 16 nm (regular
steps), amounting to twice the tubulin dimer
spacing. Their conjoining coiled-coil stalk
domain is thus translocated in discrete 8-nm
steps (1, 3, 6, 23). However, it is still debated
(24–26) whether kinesin walks “like a human,”
i.e., with one head passing the stalk on the left
and the other one on the right (asymmetric),
or if it walks with both heads passing on the
same side (symmetric). Camera localization–
based fluorescence imaging with one-nanometer
accuracy (FIONA) (5) resolved regular kinesin
steps using a single fluorophore label at one
of the heads, but its time resolution of sev-
eral hundreds of milliseconds required slow-
ing down movement by administering ATP
concentrations that were ~1000 times lower
than in a cell (6). In fact, addressing steps at
physiological ATP concentrations has so far
required the use of beads that are orders of
magnitude larger than the kinesin heads.
For example, an optical trap study recently
observed force-dependent substeps by track-
ing a germanium bead of ~70 nm diameter
attached to the kinesin stalk (27). Thus, as in
all optical trap experiments, only the move-
ments of the protein center of mass could be
examined, not those of the individual heads.
Although a ≥30-nm gold bead bound to a
kinesin head allowed tracking the heads, dif-
ferent studies came to opposing results re-
garding the long-standing question of when
ATP is bound (15, 16). In fact, simulations
(28, 29) suggested that this discrepancy is
caused by the different labeling positions
because the beads are >200 times larger in
volume than the kinesin head.
Harnessing MINFLUX, we investigated the
stepping of different cysteine-light, truncated,
Wolff et al., Science 379, 1004–1010 (2023)
10 March 2023
2 of 7
RES EARCH | R E S E A R C H A R T I C L E
and cargo-free kinesin constructs labeled with
a fluorophore at various protein positions
through maleimide coupling. The kinesin
molecules were introduced into a flow cell in
which biotinylated and fluorescently labeled
(Alexa Fluor 488) microtubules were attached
through neutravidin to a PLL-PEG-biotin
polymer–coated coverslip. For kinesin center-
of-mass tracking, we labeled construct N356C
at its solvent-exposed cysteine introduced into
the stalk (Fig. 2A).
We recorded one-dimensional (1D) traces of
individual kinesin dimers labeled with a single
fluorophore (degree of labeling of 1, DOL1) on
the stalk walking along the microtubule axis
(on-axis displacement) with a temporal reso-
lution of ~1 ms and a precision of s ≈ 1.7 nm
(Fig. 2, B and C). These initial measurements
were performed at a 10 mM ATP concentration,
providing a walking speed of ~280 nm/s. The
traces were recorded with run lengths up to
~180 nm. On the basis of the residual noise
and the number of localizations between steps,
we determined a median precision of the mea-
sured step size of 0.57 nm (Fig. 2D). A histo-
gram of all kinesin center-of-mass steps revealed
a size range of ~3 to 11 nm (Fig. 2E), with
equally high peaks centered at 8 and 4 nm,
corresponding to expected regular steps and
substeps, respectively. The latter have so far
not been observed without attaching a much
bigger bead to the protein (27).
With the same kinesin construct, we also
recorded traces at a physiological 1 mM ATP
concentration. Despite the now increased walk-
ing speed of ~550 nm/s, both regular steps and
substeps of the kinesin center of mass were
resolved (fig. S5). The substantially smaller frac-
tion of detected substeps indicated a reduced
detection efficiency caused by the shorter sub-
step dwell times (fig. S6). The step-size histo-
gram did not exhibit its maximum at 8 nm,
as would be expected for regular center-of-
mass steps, but rather showed an unexpected-
ly high occurrence of 6- and 10-nm steps (Fig.
3A, middle). Plotting the sequence of con-
secutive step sizes in a 2D histogram revealed
that their sum frequently matched 16 nm,
indicating that these unusual steps typically
occurred sequentially (Fig. 3A, right). The
nonzero radius of the stalk (~1.0 nm, inferred
from PDB 1D7M) and the distance between
the maleimide and the fluorophore (up to
~1.0 nm; fig. S7) added up to a total fluoro-
phore displacement of up to ~2 nm from the
stalk axis of kinesin. Therefore, assuming that
the fluorophore displacement vector had a
component parallel to the walking direction,
we reasoned that the observed stepping asym-
metry was caused by a rotation of the stalk
during a regular step (Fig. 3A, left).
To test this hypothesis, we labeled the same
construct with an excess of fluorophores, en-
suring that the cysteines at amino acid posi-
A
N356C
Microtubule
8 nm
+
B
)
m
n
(
n
o
i
t
i
s
o
p
i
s
x
a
-
n
O
160
144
128
112
96
80
64
48
32
16
0
1 72
σ=1.72 nm
σ=1 72 nmm
σ=1 84 nm
σ 1.84 nm
σ=1.84 nm
C
)
m
n
(
n
o
i
t
i
s
o
p
i
s
x
a
-
n
O
D
e
c
n
e
r
r
u
c
c
O
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Time (s)
80
64
48
32
16
mm
4 nm4
4 nm
4 nm
4 nm
4 nm
4 nm
4
4 nm
4 nm
4 nm
4 nm
4 nm4
4 nm
n
8 nm8
8 nm
6 nm
6 nm
0
0.36
0.2
0.1
0.0
0.0
0.38
0.40
Time (s)
0.42
0.44
←0.57 nm
0.5
1.0
1.5
Step size precision (nm)
2.0
2.5
3.0
e
c
n
e
r
r
u
c
c
O
0.2
0.1
0.0
-12 -8
-4
0
4
Step size (nm)
8
12 16 20
Fig. 2. MINFLUX tracking of kinesin exhibits
4-nm center-of-mass substeps. (A) Scheme of
kinesin walking on a microtubule indicating the
labeling position of the fluorophore in the stalk.
(B) Exemplary position traces recorded along the
microtubule axis at 10 mM ATP. The data are
overlaid with the detected step function shown as
thick darker lines. (C) Magnification of the traces
shown in (B) between 0.36 and 0.44 s as highlighted
by gray shading. (D) Histograms of the step-size
precision (top) and the step sizes (bottom) for
1821 steps. The median step-size precision is
0.57 nm. Orange dashed lines highlight 4-nm-sized
substeps and 8-nm-sized regular steps.
tion 356 (aa356) of both monomers each
carried a fluorophore (degree of labeling of
2, DOL2) (figs. S7 and S8). As a result, we
found stepping symmetry reinstated, because
MINFLUX inherently localized the midpoint
between two adjacent identical fluorophores,
which by design coincided with the stalk axis
(Fig. 3B, left). To ensure that the histogram of
the DOL2 experiment exclusively represents
steps of kinesins with both fluorophores emit-
ting, only DOL2 tracking data (characterized
by a photon rate >167 kHz, as determined
from the DOL1 data; fig. S9) were plotted.
Supporting our hypothesis of a stalk rotation,
the resulting step-size histogram indeed shows
a rather narrow peak centered at 8 nm (Fig. 3B,
middle), and the 2D histogram of consecutive
step sizes indicates the dominance of suc-
cessive 8-nm steps (Fig. 3B, right). In a trace
(fig. S10) in which one of the fluorophores
bleached, a clear difference in the step sizes
before and after bleaching became apparent:
~8 nm (before) and alternating ~10 and 6 nm
(after). We conclude that the stalk rotates
when kinesin steps. Whether consecutive steps
cause a unidirectional (26, 30) or a back-and-
forth (24, 25) rotation cannot be deduced from
this experiment alone.
ATP binds in one-head-bound state
Next, we explored whether ATP binds to kine-
sin in its one-head-bound state (1HB, only
leading head bound) or its two-head-bound
state (2HB, leading and trailing head asso-
ciated with their binding site), a longstanding
open question concerning the kinesin mecha-
nochemical cycle (15, 16, 31–35). We used con-
struct T324C labeled at its solvent-exposed
cysteine (DOL1) located at the C-terminal end
of the a6 helix adjacent to the neck linker on
the head. When the head is microtubule bound,
the label is in the center on the right side of the
motor domain with respect to the walking
direction and the microtubule surface is
beneath the head. We recorded 2D traces along
(on-axis) and perpendicular to (off-axis) the
microtubule axis at ATP concentrations of
10 mM, 100 mM, and 1 mM. By tracking one
of the heads rather than the kinesin center of
mass, we observed traces with regular steps of
16 nm, the distance between every second bind-
ing site on the microtubule, and substeps
of ~8 nm resulting from the labeled head oc-
cupying an unbound intermediate state (Fig.
4A). Accordingly, the on-axis step-size histo-
gram shows a fraction of regular steps peaking
at 16 nm and a fraction of substeps distributed
at ~8 nm. In good agreement with the results
obtained from construct N356C, the fraction of
detected substeps decreased with increasing
ATP concentration, indicating an ATP depen-
dence of the unbound state (Fig. 4B). Note the
unexpectedly broad distribution of substep
sizes for T324C, which is discussed below.
Wolff et al., Science 379, 1004–1010 (2023)
10 March 2023
3 of 7
RES EARCH | R E S E A R C H A R T I C L E
To explicitly identify the bound states (B),
in which the labeled kinesin head is located
at its microtubule-binding site, and unbound
states (U), in which it is located in between,
we applied a hidden Markov model (HMM)
based solely on the existence of ~8-nm sub-
steps and the possibility of kinesin detaching
and reattaching to the microtubule. The model
identifies five different state transitions from
the sequence of detected steps (see materials
and methods section 1.4.7). All substeps cor-
respond to the labeled head transitioning
between the two states (B→U and U→B).
However, when they are unpaired, related to a
rare and not directly observable “slip state,”
the head effectively transitions between bound
states (B→B, see next section). Transitions be-
tween microtubule-binding sites (B→B), dur-
ing which the intermediate unbound state was
too short and thus missed, were identified as
the most likely source of the ~16-nm steps.
Potential ~16-nm transitions between unbound
states (U→U) of the labeled head comprise a
series of the transitions explained above (la-
beled head U→B, unlabeled head B→U and
U→B, and labeled head B→U), so missing
these states was deemed as highly unlikely
(see supplementary text section 2.4).
On the basis of these premises, we assigned
bound and unbound states to all traces of the
kinesin construct T324C with the HMM (see
representative trace in Fig. 4C). Using this
assignment, the dwell times in the bound and
unbound states were determined for each ATP
concentration. To obtain the average dwell
times t1HB and t2HB of the underlying 1HB and
2HB states, respectively, the histograms of res-
idence time in the bound and unbound states
were fitted simultaneously (fig. S11). The un-
bound state (1HB with labeled head unbound)
data were fitted with a monoexponential de-
cay function. For the bound state (2HB with
labeled head leading, 1HB with labeled head
bound, and 2HB with labeled head trailing), a
combination of three exponential decay func-
tions was used under the assumption of equal
binding kinetics for both heads. Matching the
trend deduced from the step-size histograms
in Fig. 4B, t1HB substantially increased with
decreasing ATP concentrations from ~8 ms
for 1 mM ATP to >30 ms for 10 mM ATP (Fig.
4D). By contrast, t2HB did not exhibit any ATP
dependence, displaying a dwell time of ~8 ms
for all ATP concentrations. We concluded that
ATP binds [presumably to the microtubule-
bound leading head (31)] when kinesin is in its
1HB state and the unbound head is in between
previous and next microtubule-binding sites.
ATP hydrolyzes in the 2HB state
Subsequently, we recorded traces of the
kinesin construct T324C with 1 mM ATPgS, a
slowly hydrolyzing ATP analog. The results
revealed that the use of ATPgS did not really
A
B
DOL1
6 nm
10 nm
-
-
-
DOL2
8 nm
8 nm
-
-
-
e
c
n
e
r
r
u
c
c
O
e
c
n
e
r
r
u
c
c
O
0.3
0.2
0.1
0.0
0.3
0.2
0.1
0.0
+
+
+
+
+
+
)
m
n
(
e
z
s
i
i
1
+
p
e
t
S
)
m
n
(
e
z
s
i
i
1
+
p
e
t
S
16
12
8
4
0
16
12
8
4
0
DOL1
0
4
8
12
16
20
Step size (nm)
DOL2
0
4
8
12
16
20
Step size (nm)
DOL1
0
4
8
12
Stepi size (nm)
16
DOL2
0
8
4
12
Stepi size (nm)
16
1
0
1
0
N
o
r
m
.
f
r
e
q
u
e
n
c
y
N
o
r
m
.
f
r
e
q
u
e
n
c
y
Fig. 3. Rotation of the stalk during kinesin stepping. (A) Left: Suggested model of stalk rotation
explaining the alternating sequence of larger and smaller steps at DOL1. Middle: Step-size histogram for
1D kinesin center-of-mass tracking with a single fluorophore (DOL1, Number of steps SDOL1 = 1810) at 1 mM
ATP concentration. Right: 2D histogram of consecutive step sizes showing alternating 6- and 10-nm steps
for DOL1. (B) Left: Suggested model of stalk rotation demonstrating the true 8-nm stalk displacement
at DOL2. Middle: Step-size histogram for 1D kinesin center-of-mass tracking with two fluorophores on both
protein-labeling sites of the kinesin dimer (DOL2, SDOL2 = 630) at 1 mM ATP concentration. Right: 2D
histogram of consecutive step sizes showing predominantly successive 8-nm steps for DOL2.
10 µM
100 µM
1000 µM
A
)
m
n
(
n
o
i
t
i
s
o
p
i
s
x
a
-
n
O
231
198
165
132
99
66
33
0
-33
T324C
0.0
0.1
0.2
0.3
Time (s)
0.4
0.5
0.6
B
e
c
n
e
r
r
u
c
c
O
0.2
0.1
0.0
0.1
0.0
0.1
0.0
C
m
n
(
)
n
o
i
t
i
s
o
p
i
s
x
a
-
f
f
O
33
0
-33
0
0.0 s.
.
0.1 s
0
0.2 s
0
0.3 s
0.4 s
33
66
99
132
165
188
On-axis position (nm)
D
)
s
m
(
e
m
i
t
l
l
e
w
D
35
30
25
20
15
10
5
10 µM
100 µM
1000 µM
0
8
16
24
Step size (nm)
τ
1HB
τ
2HB
10
100
ATP (µM)
1000
Fig. 4. Kinesin awaits ATP binding in 1HB state. (A) Exemplary traces recorded at 10 mM (blue),
100 mM (orange), and 1 mM (yellow) ATP concentrations with distinct plateaus spaced by ~16 nm. Exemplary
substep plateaus between two 8-nm steps are highlighted by black arrowheads. The inlay shows construct
T324C with the fluorophore at the labeling position on one kinesin head. (B) Histogram of detected step
sizes for each ATP concentration showing ~8-nm substeps (darker filling; S10mM = 1152, S100mM = 230, and
S1mM = 905) and 16-nm regular steps (lighter filling; S10mM = 557, S100mM = 254, and S1mM = 1110) as assigned
by the HMM. (C) 2D representation of the 10 mM trace shown in (A) with time stamps. Plateaus identified
by the HMM as unbound states are highlighted in red. For improved visibility, the raw data points are overlaid
with a 5-ms moving median filter. (D) Comparison of t1HB and t2HB for different ATP concentrations. Black
lines show the fitted Michaelis-Menten kinetics (Km = 28 ± 2 mM, kATP = 4.2 ± 0.4 s−1 mM−1) for the 1HB
state and a constant fit (t2HB = 8.5 ms) for the 2HB state. Error bars and parameter uncertainties denote the
64% confidence intervals of the fits.
Wolff et al., Science 379, 1004–1010 (2023)
10 March 2023
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
A
B
K28C
E215C
T324C
K28C
i
I
I
i
i
i
.
II
II
-
)
-
)
III
III
IV
IV
8
0
0
8
8
0
8
0
0
+
+
p
o
T
24
16
16
24
24
16
C
Side
Side
D
-50
10 µM
10 µM
m
n
(
100 µM
100 µM
m
o
t
t
o
B
1000 µM
1000 µM
Plus end
16 24
s
x
a
-
n
O
1
+
p
e
t
S
m
n
(
e
z
s
e
z
s
p
e
t
s
f
r
e
q
u
e
n
c
y
e
c
n
e
r
r
u
c
c
O
1 N
o
r
m
Center
Between
Step size (nm)
Step size (nm)
8
0
24
16
Stepi size (nm)
0.3
0.2
0.1
0.0
0.2
0.1
0.0
0.2
0.1
0.0
Fig. 5. Rotation of the unbound head recon-
structed by tracking various labeling sites on the
kinesin head. (A) Left: Construct E215C with the
fluorophore-labeling site on the front end of the
kinesin head and histograms of the detected step
sizes of this construct for 10 mM, 100 mM, and
1 mM ATP concentrations, respectively, featuring a
dominant peak at 16 nm (S10mM = 666, S100mM = 310,
and S1mM = 431) and a small fraction of ~8-nm steps
(S10mM = 280, S100mM = 50, and S1mM = 195). Right:
Construct K28C with labeling position at the back of
the kinesin head and histograms of detected step
sizes of this construct for 10 mM, 100 mM, and
1 mM ATP concentration, respectively, featuring an
increasing peak at 16 nm (S10mM = 154, S100mM =
236, and S1mM = 529) and a decreasing fraction of
~8-nm steps (S10mM = 718, S100mM = 217, and S1mM =
361) for increasing ATP concentrations. (B) 2D
histograms of consecutive step sizes for constructs
T324C and K28C, both showing successive regular
steps (I), regular steps followed by substeps (II),
successive substeps (III), and substeps followed by
regular steps (IV). For construct K28C, transitions
involving the unbound state generate symmetric
successive steps of ~8 nm. For construct T324C,
these transitions exhibit an asymmetry of
~6-nm substeps followed by ~10-nm substeps.
(C) Schematic of a surface-immobilized microtubule and
the assignment of protofilament classes. The assumed
true sideward displacement of the kinesin head
(magenta) in these classes is indicated by black arrows.
Magnification visualizes the projection of the 3D
displacement onto the imaging plane. (D) Scatter plot
of four kinesin traces recorded on a single microtubule
(blue circles) with one central trace highlighted for the
detected bound (dark magenta dots) and unbound
(orange dots) states. Red lines display the microtubule outline inferred from the traces. (E) Sideward displacement of all detected unbound states and their respective position
on the microtubule (blue dots). The orange dots correspond to the sideward displacement of the central trace shown in (E). The red line displays an ellipse fit to the data with
a major axis diameter of 30.6 nm and a minor axis diameter of 9.2 nm. (F) Proposed 3D orientation of the unbound state of the trailing head (orange; PDB: 1MKJ) on the
microtubule (alpha-tubulin is shown in dark gray, beta-tubulin in light gray, PDB: 6DPU). Colored arrows depict the displacement of the different labeling positions from the
microtubule-bound state of the trailing head (dark magenta; PDB: 3J8Y). Uncertainties of the positions for aa28 and aa324 are displayed by the shaded regions representing the
Gaussian width of the step-size distribution. The leading head is shown in the apo state (cyan; PDB: 4ATX).
T324C
forward~6 nm
sideward<2 nm
K28C
forward~8 nm
sideward~5 nm
150
On-axis position (nm)
Position on MT (nm)
Rightward
Plus end
Plus end
Leftward
-20 -10
n
o
i
t
i
s
o
p
ΔOff-axis
s
x
a
-
f
f
Right
Right
m
n
(
300
250
100
200
-10
O
Δ
20
10
50
50
10
E
F
s
x
a
-
f
f
O
0
0
0
0
)
i
i
affect the unbound state duration but did sub-
stantially increase the time spent in the bound
state (fig. S11). Therefore, we conjectured that
ATP hydrolysis did not take place in the 1HB
state, but rather after the unbound head moved
to its next binding site. This hypothesis was
further supported by comparing the run length
(actual distance traveled) and run fraction
(ratio of run length and distance from starting
point to end of microtubule) determined from
kymographs of total internal reflection fluo-
rescence (TIRF) microscopy measurements for
construct T324C under ATPgS and under ATP
consumption. For 1 mM ATPgS, kinesin walked
36 times more slowly but with an average run
length nearly as long and an average run frac-
tion nearly as large as for 1 mM ATP. (The ob-
served slight reduction of run length was likely
caused by increased bleaching over the much
longer run time.) Whereas walking speeds were
comparable for 1 mM ATPgS and 1 mM ATP,
the run length was substantially shorter and
the run fraction substantially smaller for
1 mM ATP (fig. S12). Because the 1HB state is
known to be the one that is most vulnerable
to kinesin detachment from the microtubule
(36), the combined results disqualify 1HB as
being the ATP-hydrolyzing state.
Further examination of the traces revealed a
rare occurrence of an uneven number of ~8-nm-
sized substeps between regular ~16-nm-sized
steps (fig. S13) without concurrent side step-
ping. We reasoned that the lack of a second
corresponding substep probably arose from
kinesin detaching into a weakly bound slip
state and subsequently reattaching to the mi-
crotubule before or after an unpaired substep.
When the heads change binding positions on
the microtubule, an uneven number of inter-
mediate substeps is expected in the traces.
Such a slip state has so far only been reported
for kinesin under load (27, 37), not for freely
walking kinesin.
Reconstruction of unbound head orientation
during stepping
To explore the 3D orientation of the unbound
head, we repeated the 2D-MINFLUX tracking
experiments with two additional kinesin con-
structs: E215C labeled at its solvent-exposed
cysteine located at the C-terminal end of the
b6-sheet (Fig. 5A, top left) and K28C labeled
at its solvent-exposed cysteine at aa28 (Fig. 5A,
top right). In the bound state, the labeling po-
sitions of E215C and K28C are located in the
very front and back of the head, respectively,
relative to the walking direction. The fraction
of detected pairs of substeps (B→U and U→B)
of construct E215C did not differ substantially
for 10 mM, 100 mM, and 1 mM ATP. Unlike in
Wolff et al., Science 379, 1004–1010 (2023)
10 March 2023
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
previous experiments with construct T324C
(Fig. 4B), these fractions always amounted to
only ~10% of the entire population of substep
pairs and regular steps (Fig. 5A, bottom left).
In fact, compared with construct T324C (espe-
cially for 10 mM ATP), this fraction was sub-
stantially smaller and the unbound state was
substantially shorter (fig. S14). By contrast,
construct K28C (Fig. 5A, right) did not notably
differ in these aspects from construct T324C,
indicating that the observed substeps of E215C
do not represent the real 1HB state of the la-
beled head (for details, see supplementary text
section 2.5). The substep peak of the step-size
histogram of construct K28C was sharper than
that of T324C, with a clear peak centered at
8 nm. Plotting the sequence of consecutive
step sizes in 2D histograms revealed an asym-
metry of 6-nm B→U substeps and subsequent
10-nm U→B substeps of construct T324C,
which is in contrast to the symmetric substep-
ping behavior of the K28C construct (Fig. 5B).
We cannot fully exclude that the modifi-
cation at aa215 hampered kinesin from enter-
ing a detectable 1HB state caused by altered
protein-protein interactions or steric hin-
drance. However, this scenario disagrees with
earlier Förster resonance energy transfer
(FRET)–based observations of the 1HB state
of a kinesin construct that was point mutated
and labeled at aa215 (38). Therefore, we advance
the following reasoning. Because MINFLUX
tracks the fluorophore position, our traces
actually represent displacements of individual
amino acids. Specifically, they map projections
of the 3D amino acid trajectories onto the fo-
cal plane. Differences in substeps between in-
dividual kinesin constructs can be attributed
to different trajectories of aa28, aa215, and
aa324 in space, pertaining to the “back,” “front,”
and “middle” part of the head, respectively.
This approach allowed us to reconstruct the
3D orientation of the labeled head in its un-
bound state. Under the assumption that each
kinesin construct entered the 1HB state with
its unbound head in between previous and
next microtubule-binding sites, the nonappar-
ent displacement of aa215 along the on-axis
actually indicates a rotation of the unbound
head around its front during a substep. Like-
wise, the symmetry of substep pairs of con-
struct K28C implies a forward rotation of aa28
by ~8 nm, and the asymmetry of substep pairs
of construct T324C suggests a displacement of
aa324 by ~6 nm along the microtubule axis
when entering the unbound state.
Next, we investigated the off-axis displace-
ment of the unbound head, which has been
reported to be rightward (16) or nonexistent
(15) by different bead-tracking studies using
an inherently artifact-prone label (28). We
found off-axis displacements in the unbound
state of <2 nm for T324C and up to ~5 nm for
K28C (fig. S15). For the entire substep popu-
lation of K28C, rightward, near-zero, and left-
ward off-axis displacements appeared. The
displacement was always consistent in mag-
nitude and direction within a single trace, as
shown by Pearson correlation analysis [r =
0.65 ± 0.02 (meanTSD)].
To correlate the observed off-axis displace-
ments with individual protofilament classes
of a single microtubule (“sides,” “center,” and
“between”) (Fig. 5C), we recorded trace sets
(137 traces in total) of construct K28C on 19
different microtubules using actively stabi-
lized samples (for details, see materials and
methods section 1.1.3 and fig. S16). The outer-
most traces of the individual sets were spaced
by ~30 nm, which agrees well with the micro-
tubule diameter of ~25 nm (39, 40) plus twice
the ~2.5-nm distance between the labeling po-
sition of construct K28C and the microtubule
surface (inferred from PDB 3J8Y). Thus, these
traces were assigned to kinesins walking along
side protofilaments and were used as refer-
ences for the remaining trace assignment.
Located between two side traces, the center
traces exhibit pronounced rightward displace-
ment into the unbound state (Fig. 5D). After
aligning all trace sets along their central on-
axes, the substep off-axis displacements were
plotted against the lateral offsets of the cor-
responding traces from the microtubule center
axis (Fig. 5E). Imperfections in the alignment
of all trace sets may introduce minor errors in
position, but the traces of kinesins walking
along side protofilaments display a near-zero
off-axis displacement of their unbound states,
ruling out a considerable displacement of aa28
away from the microtubule surface. Because
protofilaments at the bottom of microtubules
were mostly blocked by neutravidin and poly-
mer (Fig. 5C), virtually all of the center traces
can be attributed to kinesins walking on the
microtubule’s top. These traces showed max-
imum and predominantly rightward off-axis
displacements into the unbound state. This
finding is corroborated by further detailed
analysis of the between traces generated by
kinesins walking both on top and bottom
protofilaments, leading to both rightward
(for top) and leftward (for bottom) off-axis
displacements for simple geometric reasons
(see supplementary text section 2.7). We con-
clude that upon entering the unbound inter-
mediate state, aa28 of the trailing head is
displaced up to ~5 nm rightward with respect
to the kinesin coordinate system.
The combination of on-axis substep sizes
and associated off-axis displacements of the
investigated amino acid positions allowed us
to derive an approximate average 3D orienta-
tion of the unbound head during kinesin mo-
tion (Fig. 5F and movies S1 and S2), improving
and specifying the ones derived from FRET
studies (35, 38). In conjunction with stalk ro-
tation (Fig. 3), which is expected to resolve
torsion and thus decrease asymmetry caused
by a head’s full step, our data showing sys-
tematic rightward displacement of the un-
bound head indicate that the hand-over-hand
stepping is symmetric.
Our model of the mechanochemical cycle
of kinesin (visualized in fig. S17) is based on
MINFLUX tracking with 3 nm/0.63 ms STR
per individual localization using just 20 de-
tected photons, allowing the identification of
>12,000 kinesin steps with 1-nm precision (fig.
S18). We believe that this performance sets
a benchmark for protein tracking. Fluoro-
phores with higher emission rates will increase
the STR even further. The confocal arrange-
ment of the MINFLUX system, because it
strongly suppresses background, also allows
for protein investigations in living cells (41).
In fact, one can readily envisage applying
MINFLUX to any nanometer-scale changes of
fluorescently labeled biomolecules. We thus
believe that our study establishes MINFLUX
as a next-generation tool for recording confor-
mational changes of single proteins with mini-
mal invasiveness.
REFERENCES AND NOTES
1. K. Svoboda, C. F. Schmidt, B. J. Schnapp, S. M. Block, Nature
365, 721–727 (1993).
2. K. Svoboda, S. M. Block, Annu. Rev. Biophys. Biomol. Struct.
23, 247–285 (1994).
3. M. J. Schnitzer, S. M. Block, Nature 388, 386–390 (1997).
4. Y. Seol, A. E. Carpenter, T. T. Perkins, Opt. Lett. 31, 2429–2431
(2006).
5. A. Yildiz et al., Science 300, 2061–2065 (2003).
6. A. Yildiz, M. Tomishige, R. D. Vale, P. R. Selvin, Science 303,
676–678 (2004).
7. R. E. Thompson, D. R. Larson, W. W. Webb, Biophys. J. 82,
2775–2783 (2002).
8. A. R. Dunn, J. A. Spudich, Nat. Struct. Mol. Biol. 14, 246–248 (2007).
9. F. Ruhnow, D. Zwicker, S. Diez, Biophys. J. 100, 2820–2828 (2011).
10. J. Ortega-Arroyo, P. Kukura, Phys. Chem. Chem. Phys. 14,
15625–15636 (2012).
11. E. Reid et al., Am. J. Hum. Genet. 71, 1189–1194 (2002).
12. M. Fichera et al., Neurology 63, 1108–1110 (2004).
13. K. Poirier et al., Nat. Genet. 45, 639–647 (2013).
14. G. Chandrasekaran, P. Tátrai, F. Gergely, Br. J. Cancer 113,
693–698 (2015).
15. K. J. Mickolajczyk et al., Proc. Natl. Acad. Sci. U.S.A. 112,
E7186–E7193 (2015).
16. H. Isojima, R. Iino, Y. Niitani, H. Noji, M. Tomishige,
Nat. Chem. Biol. 12, 290–297 (2016).
17. F. Balzarotti et al., Science 355, 606–612 (2017).
18. Y. Eilers, H. Ta, K. C. Gwosch, F. Balzarotti, S. W. Hell,
Proc. Natl. Acad. Sci. U.S.A. 115, 6117–6122 (2018).
19. K. C. Gwosch et al., Nat. Methods 17, 217–224 (2020).
20. R. Schmidt et al., Nat. Commun. 12, 1478 (2021).
21. T. A. Klar, E. Engel, S. W. Hell, Phys. Rev. E 64, 066613 (2001).
22. R. Killick, P. Fearnhead, I. A. Eckley, J. Am. Stat. Assoc. 107,
1590–1598 (2012).
23. H. Kojima, E. Muto, H. Higuchi, T. Yanagida, Biophys. J. 73,
2012–2022 (1997).
24. W. Hua, J. Chung, J. Gelles, Science 295, 844–848 (2002).
25. B. Gutiérrez-Medina, A. N. Fehr, S. M. Block, Proc. Natl. Acad.
Sci. U.S.A. 106, 17007–17012 (2009).
26. A. Ramaiya, B. Roy, M. Bugiel, E. Schäffer, Proc. Natl. Acad. Sci.
U.S.A. 114, 10894–10899 (2017).
27. S. Sudhakar et al., Science 371, eabd9944 (2021).
28. K. J. Mickolajczyk, A. S. I. Cook, J. P. Jevtha, J. Fricks,
W. O. Hancock, Biophys. J. 117, 331–345 (2019).
29. S. Hasnain, M. L. Mugnai, D. Thirumalai, J. Phys. Chem. B 125,
10432–10444 (2021).
30. J. Howard, Annu. Rev. Physiol. 58, 703–729 (1996).
31. R. Takaki, M. L. Mugnai, Y. Goldtzvik, D. Thirumalai, Proc. Natl.
Acad. Sci. U.S.A. 116, 23091–23099 (2019).
32. W. O. Hancock, Biophys. J. 110, 1216–1225 (2016).
Wolff et al., Science 379, 1004–1010 (2023)
10 March 2023
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
33. J. Qin, H. Zhang, Y. Geng, Q. Ji, Int. J. Mol. Sci. 21, 6977 (2020).
34. E. Mandelkow, K. A. Johnson, Trends Biochem. Sci. 23,
429–433 (1998).
35. S. Verbrugge, Z. Lansky, E. J. Peterman, Proc. Natl. Acad. Sci.
U.S.A. 106, 17741–17746 (2009).
36. K. J. Mickolajczyk, W. O. Hancock, Biophys. J. 112, 2615–2623 (2017).
37. A. Toleikis, N. J. Carter, R. A. Cross, Biophys. J. 119, 1984–1994
(2020).
38. T. Mori, R. D. Vale, M. Tomishige, Nature 450, 750–754 (2007).
39. W. Brinkley, J. Struct. Biol. 118, 84–86 (1997).
40. M. C. Ledbetter, K. R. Porter, Science 144, 872–874 (1964).
41. T. Deguchi et al., Science 379, 1010–1015 (2023).
42. J. O. Wolff et al., Matlab scripts for: MINFLUX dissects the
unimpeded walking of kinesin-1, Zenodo (2022);
https://doi.org/10.5281/zenodo.7442902.
43. J. O. Wolff, L. Scheiderer, Raw data for: MINFLUX dissects
the unimpeded walking of kinesin-1, Zenodo (2022);
https://doi.org/10.5281/zenodo.7565676.
ACKN OW LEDG MEN TS
We thank M. Tarnawski and the Protein Core Facility at the MPI
Heidelberg for performing the QuikChange site-directed
mutagenesis and protein expression; our colleague M. Remmel for
access to the TIRF microscope; the MS Core Facility at the MPI
Heidelberg for recording the mass spectroscopy data; R. Vale
(University of California, San Francisco and Howard Hughes
Medical Institute) for providing plasmids K560CLM T324C and
CLM RP HTR; A. Yildiz (University of California, Berkeley) for
providing plasmid K560CLM E215C and the protein preparation
and purification protocol; and M. Kulp (Optical Facility, MPI
Göttingen) for imprinting aluminum gratings on coverslips. Author
contributions: J.O.W. and T.E. built the microscope and together
with J.E. wrote the software for controlling the setup and
performing measurements. J.E. designed the interferometric
system with input from S.W.H. and provided technical supervision.
J.O.W. built the active stabilization unit, wrote software for
MINFLUX track evaluation, and performed measurements initially
together with T.E. L.S. scrutinized the kinesin tracking literature,
designed the kinesin experiments, and prepared the related
samples advised by J.M. L.S. and J.O.W. processed, evaluated, and
interpreted kinesin measurements co-supervised by J.M., who was
also responsible for the project administration. S.W.H. proposed
and initiated the evaluation of MINFLUX spatiotemporal resolution
for protein tracking and was responsible for overall supervision
and steering. L.S., J.O.W., J.M., and S.W.H. wrote the manuscript
with input from all authors. All authors contributed to the
results through critical discussions throughout the course of the
project. Competing interests: S.W.H. is inventor on patent
applications WO 2013/072273 and WO 2015/097000 submitted
by the Max Planck Society that cover basic principles and
arrangements of MINFLUX, including single-molecule tracking.
J.E. and S.W.H. are inventors on patent application WO 2020/
064108 submitted by the Max Planck Society that covers
principles and arrangements of the phase/amplitude modulator for
shifting the intensity minimum. S.W.H. is a cofounder of the
company Abberior Instruments, which commercializes MINFLUX
microscopes. The remaining authors declare no competing
interests. Funding: This work was funded by intramural funds from
the Max Planck Society. Code and data availability: Matlab
scripts used for the sliding-curvature and fixed-curvature
estimator, as well as the step-finder and the hidden Markov model
and DNA sequences, have been deposited at Zenodo (42). Raw
data and Matlab scripts for displaying and processing the data
can be downloaded from Zenodo (43). License information:
Copyright © 2023 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science. No
claim to original US government works. https://www.science.org/
about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade2650
Materials and Methods
Supplementary Text
Figs. S1 to S20
Tables S1 to S4
References (44–47)
Movies S1 and S2
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
Submitted 4 August 2022; accepted 23 January 2023
10.1126/science.ade2650
Wolff et al., Science 379, 1004–1010 (2023)
10 March 2023
7 of 7
|
10.1126_science.ade3232
|
RES EARCH
HUBBARD MODEL
The Wiedemann-Franz law in doped Mott insulators
without quasiparticles
Wen O. Wang1,2*, Jixun K. Ding1,2, Yoni Schattner2,3,4, Edwin W. Huang5,6,7,
Brian Moritz2, Thomas P. Devereaux2,8,9*
Many metallic quantum materials display anomalous transport phenomena that defy a Fermi liquid
description. Here, we use numerical methods to calculate thermal and charge transport in the doped
Hubbard model and observe a crossover separating high- and low-temperature behaviors. Distinct
from the behavior at high temperatures, the Lorenz number L becomes weakly doping dependent and
less sensitive to parameters at low temperatures. At the lowest numerically accessible temperatures,
L roughly approaches the Wiedemann-Franz constant L0, even in a doped Mott insulator that lacks
well-defined quasiparticles. Decomposing the energy current operator indicates a compensation between
kinetic and potential contributions, which may help to clarify the interpretation of transport experiments
beyond Boltzmann theory in strongly correlated metals.
been hindered by the use of an assumed
Boltzmann-like theory and reductive conclu-
sions on the nature of transport in terms of
Drude-like single-particle concepts. This great-
ly amplifies the need for deeper analysis that
avoids oversimplifications, but there is very
little known from exact methods about the
nature of transport in strongly interacting
systems. Many advanced numerical calcula-
tions have focused on characterizing ground-
state properties (24, 25), but a picture of
transport is incomplete without an understand-
ing of the excited states in these materials.
Analytical approaches are hampered by the fact
that properly evaluating transport involves
calculating many higher-order correlation func-
tions without relying on the simplifying assump-
tions of quasiparticles and Boltzmann theory,
which only punctuates the need for more accu-
rate and precise determinations of transport.
Here, we numerically study the DC longitu-
dinal thermal conductivity k in the doped two-
dimensional (2D) t (cid:1) t′ (cid:1) U Hubbard model,
which exhibits strange metallic electric trans-
port over a wide hole doping p and tempera-
ture T range (26–29). We evaluate the many-body
Kubo formula using the determinant quan-
tum Monte Carlo (DQMC) (30, 31) algorithm,
which is numerically exact, unbiased, and non-
perturbative, and maximum entropy analy-
tic continuation (MaxEnt) (32, 33), which is
typically reliable in systems with strong in-
teractions that lack sharp features in fre-
quency [see supplementary materials of (26)].
L andau’s notion of quasiparticles greatly
simplified the language of transport in
systems with a macroscopic number of
interacting degrees of freedom in terms
of “free” objects with renormalized prop-
erties that participate in transport through
a semi-classical or Boltzmann framework. As
such, transport behavior of Fermi liquids is
governed solely by kinematic constraints of a
Fermi surface and collisions between other-
wise free particles. Yet in many correlated me-
tals, including the high–transition temperature
(or critical temperature, Tc) cuprates, anoma-
lous transport phenomena have been uncov-
ered that violate these rules: strange metal
resistivity that increases linearly with temper-
ature, not saturating as the quasiparticle mean-
free-path approaches the lattice spacing (1–3);
inconsistency with Kohler’s rule, which gov-
erns the scaling behavior of magnetoresistance
from Boltzmann theory (4–6); and violations
of the Wiedemann-Franz law, which constrains
the ratio of thermal to electrical conductivity
(7–18).
The ubiquity of behavior that violates no-
tions of the Fermi liquid has led to tremendous
interest in determining how heat and charge
currents propagate in such systems (19–23).
Analysis of the large body of experimental
transport results in correlated materials has
1Department of Applied Physics, Stanford University, Stanford,
CA 94305, USA. 2Stanford Institute for Materials and Energy
Sciences, SLAC National Accelerator Laboratory, 2575 Sand Hill
Road, Menlo Park, CA 94025, USA. 3Department of Physics,
Stanford University, Stanford, CA 94305, USA. 4AWS Center for
Quantum Computing, Pasadena, CA 91125, USA. 5Department
of Physics and Institute of Condensed Matter Theory, University
of Illinois at Urbana-Champaign, Urbana, IL 61801, USA.
6Department of Physics and Astronomy, University of Notre
Dame, Notre Dame, IN 46556, USA. 7Stavropoulos Center for
Complex Quantum Matter, University of Notre Dame, Notre
Dame, IN 46556, USA. 8Department of Materials Science and
Engineering, Stanford University, Stanford, CA 94305, USA.
9Geballe Laboratory for Advanced Materials, Stanford University,
Stanford, CA 94305, USA.
*Corresponding author. Email: wenwang.physics@gmail.com
(W.O.W.); tpd@stanford.edu (T.P.D.)
Fig. 1. Temperature and doping dependence of thermal and charge conductivity. (A) DC thermal
conductivity k. (B) k=T focused on the low-temperature regime. (C) DC charge conductivity s multiplied by
temperature T. (D) s focused on the low-temperature regime. The high-temperature dotted lines in (A) and
(C) are infinite-temperature limits calculated via a moments expansion (26, 35). Parameters: U=t ¼ 8 and
t0=t ¼ (cid:1)0:25. A crossover temperature Txo (cid:3) t separates low- and high-temperature regimes in (A) and (C).
Error bars are shown but may be smaller than the size of the data markers.
Wang et al., Science 382, 1070–1073 (2023)
1 December 2023
1 of 4
RES EARCH | R E S E A R C H A R T I C L E
We define k as the linear response of the heat
current hJQi induced by a parallel tempera-
ture gradient and normalized by system size
N, k ≡ (cid:1)hJQ;xi=ðN @xT Þ, under the condition
of zero charge current. Distinct from the in-
coherent behavior at high temperatures, we
observe that the Lorenz number, the ratio
between the thermal and charge conductivity
L ≡ k= T sð
Þ, has a weak doping and parameter
dependence in the low-temperature regime
and roughly approaches the Wiedemann-Franz
law prediction L0 ¼ p2=3 as temperature de-
creases down to the lowest accessible value, even
in the absence of long-lived quasiparticles. Meth-
odological details, including a systematic anal-
ysis of finite size and Trotter errors, as well as
extensive supporting data, can be found in (34).
Thermal and charge conductivity
The DC longitudinal thermal conductivity k Tð Þ
is shown in Fig. 1A; for comparison, the DC
longitudinal charge conductivity s Tð Þ (26) (mul-
tiplied by T ) is shown in Fig. 1C. In the infinite-
temperature limit, k º 1=T 2 and s º 1=T
(26, 27, 35, 36). As T decreases from this limit,
we observe a crossover at roughly Txo∼t, sep-
arating distinct behavior in two regimes for
both k and s. k decreases with doping at high
temperatures, whereas it increases with doping
at low temperatures. Although s generally in-
creases with doping at all temperatures, the tem-
perature dependence of T s displays kinks, or
even nonmonotonic behavior, at roughly Txo .
Below Txo, k=T and s display similar doping and
temperature dependences (Fig. 1, B and D), sug-
gesting persistent correlations between thermal
and charge transport even for a strange metal
phase where quasiparticles are not well-defined.
Lorenz number and its temperature and
parameter dependence
The Lorenz number L Tð Þ highlights the cor-
relation between thermal and charge trans-
port (Fig. 2). Aside from the half-filled Mott
insulator, where L diverges with decreasing
temperature, in the doped metals L shows a
crossover similar to that in k and s. At high
temperatures, high-energy excited states be-
come important (36, 37), such that quasi-
particles are not well-defined and electrons
have extraordinarily short mean-free-paths.
L has a nonmonotonic temperature depen-
dence and decreases with increasing doping.
Below Txo , L displays substantially reduced
doping dependence, collapsing roughly onto
a single set of curves. This set of curves in-
creases monotonically with decreasing tem-
peratures, approaching a constant that roughly
corresponds to L0 ¼ p2=3—the Lorenz num-
ber as predicted by the Wiedemann-Franz law.
In the Hubbard model, relaxation primarily
occurs through Umklapp scattering. To test
its impact on the conductivities and L, we
modulate Umklapp scattering by modifying
Þ normalized
Fig. 2. Lorenz number.
Symbols: Calculated
L ¼ k= Tsð
by L0. The lines are guides
to the eye. At low tem-
peratures, below Txo (cid:3) t,
L=L0 approaches roughly 1,
marked by the black star.
Parameters: U=t ¼ 8 and
t0=t ¼ (cid:1)0:25. Cartoons:
At high temperatures,
high-energy excited states
are important (36, 37) and
transport is incoherent;
electrons are strongly
correlated and have an
extraordinarily short
mean-free-path. At low
temperatures, the elec-
trons are on their way
toward some sort of
“coherence”; electrons
have a longer mean-free-
path, although not long
enough for well-defined long-lived quasiparticles. Although single-particle and individual transport properties
show signatures of anomalous strange metal and non-Fermi liquid behavior, the Lorenz number still roughly
approaches the Wiedemann-Franz law’s prediction as temperature decreases.
Fig. 3. Parameter dependence of the Lorenz number. L=L0 for (A) U=t ¼ 6 and t0=t ¼ (cid:1)0:25; (B) U=t ¼ 6
and t0=t ¼ 0; (C) U=t ¼ 8 and t0=t ¼ 0; (D) U=t ¼ 10 and t0=t ¼ 0. The black stars mark the value 1.
The lowest temperatures are lower for smaller U owing to a better behaved fermion sign.
Wang et al., Science 382, 1070–1073 (2023)
1 December 2023
2 of 4
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Kinetic and potential decomposition of the Lorenz number. (A) Normalized kinetic contribution
LK=L0. The black star marks the value 1. (B) Normalized potential contribution LP=L0. The black dotted line marks
the value 0. Parameters: U=t ¼ 8 and t0=t ¼ (cid:1)0:25.
the Hubbard U and next-nearest-neighbor
hopping t0, with the results shown in Fig. 3.
The high-temperature peak position of L is
largely controlled by U , increasing with in-
creasing U, similar to the behavior of the spe-
cific heat [see fig. S9 in (34)]. For temperatures
below the crossover, there is no strong de-
pendence of L on either U or t0, suggesting that
the low-temperature behavior is generic to the
strongly correlated Hubbard model: Chang-
ing the shape of the Fermi surface (t0) or the
strength of Umklapp scattering (U) does not
appreciably alter L at the temperatures ac-
cessible through DQMC.
Decomposing the Lorenz number
To better understand the behavior below Txo, it
is useful to look at the operator contributions to
the conductivities. Determining k in the Hubbard
model using the Kubo formula requires one to
consider the two-particle term in the energy
current operator arising from electron-electron
interactions, as opposed to Boltzmann theory
that relies entirely on single-particle proper-
ties. The energy current operator JE consists
of a single-particle kinetic energy contribution,
JK , similar to that appearing in the charge cur-
rent operator J, plus an additional term JP ,
which we call the potential energy current that
depends explicitly on the interaction and no-
tably contains a two-particle current [see eq. S2,
eq. S3, and the relevant discussion of the For-
malism in (34)]. The heat current JQ , from
which we obtain k, itself contains an additional
term (cid:1)mJ, where m is the chemical potential.
However, under the condition of zero charge
current hJi ¼ 0
Þ, terms proportional to hJi will
ð
not contribute to hJQi, leaving only hJKi and
hJPi. In this way, we separate k into kinetic
and potential contributions kK=P ≡ (cid:1)hJK=P;xi=
ðN @xT Þ. Similarly, we can express the Lorenz
number L as a sum of its kinetic and poten-
tial contributions, with L ¼ LK þ LP , where
LK=P ≡ kK=P= T sð
Þ (Fig. 4, A and B).
At high temperatures, the kinetic energy
contribution LK is relatively small and doping
independent, whereas the potential energy
contribution LP is large at small doping and
decreases for increasing hole concentration
owing to the reduction of double occupancies.
This doping dependence is imparted to the
combined L (as already shown in Fig. 2). Be-
low the crossover temperature Txo and at large
doping, LP is relatively small and LK and L
approach L0: At low doping, LK increases with
decreasing temperature, while LP decreases
and changes sign at roughly Txo. The separate
contributions from the kinetic and potential
terms show opposing behavior, which be-
comes more pronounced for lower doping, and
effectively compensate one another, result-
ing in L that approaches L0. Thus unexpected-
ly, the ratio L displays a relative insensitivity
to doping, and Hubbard model parameters
[see fig. S10 in (34)], at the lowest accessible
temperatures.
Discussion and outlook
The congruence between charge and thermal
transport in the Hubbard model is unexpected.
For scattering dominated by elastic processes,
such as disorder or quasi-elastic phonon scat-
tering above the Debye temperature, the thermal
and charge conductivity are correlated through
the Wiedemann-Franz law (13, 21, 38, 39), such
that for T much lower than the Fermi temper-
ature, one obtains the Lorenz number L ¼
L0 ¼ p2=3. For both Fermi liquids and non-
Fermi liquids without disorder, L deviates
substantially from L0 (21, 39, 40). Despite our
lack of knowledge about the exact behavior
of the Hubbard model at lower temperatures
(Fermi liquid or not), caused by the fermion
sign problem, the result that L approaches
a weakly doping and Hubbard parameter–
dependent constant very close to L0 indicates
a surprisingly universal behavior. This behavior
is observed only when both single- and two-
particle contributions are properly accounted
for in the heat-current operator.
Our results may be understood in three pos-
sible ways. First, although the temperatures in
our study are below the magnetic exchange
energy scale J, our results may not yet be in the
asymptotic low-temperature regime to assess
the T →0 limit. Second, one might expect the
approximate Wiedemann-Franz ratio to emerge
in a system where both charge and thermal
currents relax predominantly through Umklapp
scattering in our temperature regime. Lastly,
it may be that such a compensation effect be-
tween kinetic and potential energy contributions
to L cannot be cast in the usual Boltzmann-
like formulation for strongly interacting, aniso-
tropic systems such as the Hubbard model.
Finally, what can our results say about the
strong violation of the Wiedemann-Franz law
that has been observed in cuprates at room
temperature, with L larger than L0 by a factor
of 3 or more (7, 10, 18, 38)? One explanation
for this is that the strong interaction enhances
the electronic contribution to thermal trans-
port, whereas another explanation would rely
on a substantial phonon contribution to the
heat current. Our observation over the exper-
imentally relevant temperature range that the
electronic contribution L roughly approaches
L0 from below would be consistent with sce-
narios in which the large L in cuprates re-
quires an appreciable phonon contribution
to heat transport.
REFERENCES AND NOTES
1. O. Gunnarsson, M. Calandra, J. E. Han, Rev. Mod. Phys. 75,
1085–1099 (2003).
2. N. E. Hussey, K. Takenaka, H. Takagi, Philos. Mag. 84,
2847–2864 (2004).
3. P. W. Phillips, N. E. Hussey, P. Abbamonte, Science 377,
eabh4273 (2022).
J. M. Harris et al., Phys. Rev. Lett. 75, 1391–1394 (1995).
J. Ayres et al., Nature 595, 661–666 (2021).
4.
5.
6. T. Kimura et al., Phys. Rev. B Condens. Matter 53, 8733–8742
(1996).
7. P. B. Allen, X. Du, L. Mihaly, L. Forro, Phys. Rev. B Condens.
Matter 49, 9073–9079 (1994).
8. R. W. Hill, C. Proust, L. Taillefer, P. Fournier, R. L. Greene,
Nature 414, 711–715 (2001).
9. C. Proust, E. Boaknin, R. W. Hill, L. Taillefer, A. P. Mackenzie,
Phys. Rev. Lett. 89, 147003 (2002).
10. H. Minami et al., Phys. Rev. B Condens. Matter 68, 220503
(2003).
11. S. Nakamae et al., Phys. Rev. B Condens. Matter 68, 100502
(2003).
12. C. Proust, K. Behnia, R. Bel, D. Maude, S. I. Vedeneev,
Phys. Rev. B Condens. Matter Mater. Phys. 72, 214511
(2005).
13. X. F. Sun et al., Phys. Rev. B Condens. Matter Mater. Phys. 80,
104510 (2009).
14. G. Grissonnanche et al., Phys. Rev. B 93, 064513 (2016).
15. J. Zhang et al., Proc. Natl. Acad. Sci. U.S.A. 114, 5378–5383
(2017).
16. B. Michon et al., Phys. Rev. X 8, 041010 (2018).
17. G. Grissonnanche et al., Nature 571, 376–380 (2019).
18. J. Zhang et al., Phys. Rev. B 100, 241114 (2019).
19. S. A. Hartnoll, Nat. Phys. 11, 54–61 (2015).
20. S. A. Hartnoll, A. P. Mackenzie, Rev. Mod. Phys. 94, 041002 (2022).
Wang et al., Science 382, 1070–1073 (2023)
1 December 2023
3 of 4
RES EARCH | R E S E A R C H A R T I C L E
21. R. Mahajan, M. Barkeshli, S. A. Hartnoll, Phys. Rev. B Condens.
36. S. Mukerjee, V. Oganesyan, D. Huse, Phys. Rev. B Condens.
Matter Mater. Phys. 88, 125107 (2013).
22. T. Hartman, S. A. Hartnoll, R. Mahajan, Phys. Rev. Lett. 119,
Matter Mater. Phys. 73, 035113 (2006).
37. R. E. Prange, L. P. Kadanoff, Phys. Rev. 134, A566–A580
141601 (2017).
(1964).
23. M. Ulaga, J. Mravlje, P. Prelovšek, J. Kokalj, Phys. Rev. B 106,
38. C. H. Mousatov, S. A. Hartnoll, NPJ Quantum Mater. 6, 81
245123 (2022).
(2021).
24. D. P. Arovas, E. Berg, S. A. Kivelson, S. Raghu, Annu. Rev.
Condens. Matter Phys. 13, 239–274 (2022).
25. M. Qin, T. Schäfer, S. Andergassen, P. Corboz, E. Gull, Annu.
Rev. Condens. Matter Phys. 13, 275–302 (2022).
26. E. W. Huang, R. Sheppard, B. Moritz, T. P. Devereaux, Science
366, 987–990 (2019).
27. P. T. Brown et al., Science 363, 379–382 (2019).
28. M. A. Nichols et al., Science 363, 383–387 (2019).
29. W. Xu, W. R. McGehee, W. N. Morong, B. DeMarco, Nat.
Commun. 10, 1588 (2019).
30. R. Blankenbecler, D. J. Scalapino, R. L. Sugar, Phys. Rev. D Part.
Fields 24, 2278–2286 (1981).
31. S. R. White et al., Phys. Rev. B Condens. Matter 40, 506–516
(1989).
32. M. Jarrell, J. E. Gubernatis, Phys. Rep. 269, 133–195
(1996).
33. O. Gunnarsson, M. W. Haverkort, G. Sangiovanni, Phys. Rev. B
Condens. Matter Mater. Phys. 82, 165125 (2010).
34. See supplementary materials.
35. W. O. Wang, J. K. Ding, B. Moritz, E. W. Huang, T. P. Devereaux,
Phys. Rev. B 105, L161103 (2022).
39. E. Tulipman, E. Berg, NPJ Quantum Mater. 8, 66 (2023).
40. A. Lavasani, D. Bulmash, S. Das Sarma, Phys. Rev. B 99,
085104 (2019).
41. W. O. Wang, Source code for “The Wiedemann-Franz law in
doped Mott insulators without quasiparticles,” Zenodo (2023).
https://doi.org/10.5281/zenodo.7976147.
42. W. O. Wang, Data for “The Wiedemann-Franz law in doped
Mott insulators without quasiparticles,” Zenodo (2023).
https://doi.org/10.5281/zenodo.7976153.
ACKN OWLED GMEN TS
We acknowledge helpful discussions with A. Auerbach, D. Belitz,
E. Berg, S. A. Hartnoll, N. E. Hussey, S. A. Kivelson, P. A. Lee, R. T. Scalettar,
Z. X. Shen, and E. Tulipman. Funding: This work was supported by
the US Department of Energy (DOE), Office of Basic Energy
Sciences, Division of Materials Sciences and Engineering. E.W.H.
was supported by the Gordon and Betty Moore Foundation EPiQS
Initiative through the grants GBMF 4305 and GBMF 8691. Y.S.
was supported by the Gordon and Betty Moore Foundation’s EPiQS
Initiative through grants GBMF 4302 and GBMF 8686. Y.S.’s
contribution to the work was done prior to joining AWS Center for
Quantum Computing. Computational work was performed on the
Sherlock cluster at Stanford University and on resources of the
National Energy Research Scientific Computing Center, supported
by the US DOE, Office of Science, under Contract no. DE-AC02-
05CH11231. Author contributions: W.O.W. and T.P.D. conceived
the study. W.O.W. performed numerical simulations and conducted
data analysis and interpretation. J.K.D., Y.S., E.W.H., B.M., and
T.P.D. assisted in data interpretation. W.O.W., B.M., and T.P.D.
wrote the manuscript with input from all coauthors. Competing
interests: The authors declare no competing interests. Data and
materials availability: Code and data presented in this study are
deposited in Zenodo (41, 42). License information: Copyright ©
2023 the authors, some rights reserved; exclusive licensee
American Association for the Advancement of Science. No claim to
original US government works. https://www.sciencemag.org/
about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade3232
Materials and Methods
Supplementary Text
Figs. S1 to S15
References (43–47)
Submitted 8 August 2022; resubmitted 8 February 2023
Accepted 27 October 2023
10.1126/science.ade3232
Wang et al., Science 382, 1070–1073 (2023)
1 December 2023
4 of 4
|
10.1126_science.abq7361
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
HUMAN FERTILITY
The mechanism of acentrosomal spindle
assembly in human oocytes
Tianyu Wu†, Jie Dong†, Jing Fu†, Yanping Kuang†, Biaobang Chen, Hao Gu, Yuxi Luo, Ruihuan Gu,
Meiling Zhang, Wen Li, Xi Dong, Xiaoxi Sun*, Qing Sang*, Lei Wang*
INTRODUCTION: Spindle assembly is essential
for ensuring accurate chromosome transmission
in both meiosis and mitosis. In somatic cells,
mitotic spindle assembly is mediated by dupli-
cated centrosomes, but canonical centrosomes
are absent in the oocytes of many species. In
rodents, acentriolar microtubule organizing
centers (aMTOCs) are responsible for meiotic
spindle assembly, but it has long been sup-
posed that human oocytes lack prominent
aMTOCs on the meiotic spindle, and the exact
mechanism of acentrosomal spindle assembly
in human oocytes has remained unclear.
RATIONALE: Microtubule nucleation and en-
suring spindle assembly are core events reg-
ulating oocyte nuclear maturation. To identify
the potential proteins driving spindle micro-
tubule nucleation in human oocytes, we sys-
tematically localized 86 human centrosome and
microtubule-related proteins by immunofluo-
rescence or three-dimensional high-resolution
live cell imaging in more than 2000 human
oocytes. We then tracked the dynamic migra-
tion of identified microtubule nucleators at
different time points before and after nuclear
envelope breakdown (NEBD). We further down-
regulated corresponding proteins to confirm
their role in microtubule nucleation and spin-
dle assembly. Given that spindle microtubule
nucleation defects result in impaired spindle
assembly and abnormal oocyte maturation, we
screened for mutations in genes encoding com-
ponents of microtubule nucleators in a cohort
of 1394 infertile female patients characterized
by oocyte maturation arrest.
RESULTS: First, we found that in human oocytes
the nucleation of spindle microtubules is in-
itiated from kinetochores from 2 to 4 hours
after NEBD. We showed the process of spindle
microtubules nucleating from kinetochores
in human oocytes. We then found that there
are 43 proteins localized in the meiotic spin-
dle, among which four proteins—centriolar
coiled-coil protein 110 (CCP110), cytoskeleton-
associated protein 5 (CKAP5), disrupted in
schizophrenia 1 (DISC1), and transforming acid-
ic coiled-coil–containing protein 3 (TACC3)—
exhibited both kinetochore and spindle micro-
tubule localization. The localization of the four
proteins was notably different from their local-
ization in human mitotic cells and in mouse
oocytes. Together, the four proteins formed an
unusual structure that was surrounded by mi-
crotubules in human germinal vesicle (GV)
oocytes just before NEBD. We refer to this po-
tential nucleating structure as the human oo-
cyte microtubule organizing center (huoMTOC).
We found that a single huoMTOC is formed at
the cortex of human GV oocytes and migrates
to the nuclear envelope before NEBD. After
NEBD, the huoMTOC becomes fragmented
and is recruited to kinetochores to initiate spin-
dle microtubule nucleation. Down-regulation
of huoMTOC components caused considera-
bly impaired spindle microtubule nucleation
and spindle assembly in human oocytes. This
structure was not detected in the oocytes of
other mammalian species such as mice and pigs.
We finally identified two oocyte maturation
arrest patients with compound heterozygous
mutations in the key huoMTOC component
TACC3. All mutations disrupted the normal
function of TACC3, resulting in the absence
of the huoMTOC structure and completely im-
paired spindle assembly in the patients’ oocytes.
CONCLUSION: Our study shows that human
oocytes possess an aMTOC-like structure, the
huoMTOC, that serves as a major site of mi-
crotubule nucleation and is required for spin-
dle assembly. The huoMTOC shows drastically
different characteristics in terms of number,
localization, and composition compared with
aMTOCs in mouse oocytes. These findings
suggest that a distinct mechanism for the in-
itiation of microtubule nucleation and spindle
assembly has evolved in human oocytes. We
found that mutations in TACC3 cause defects
in spindle assembly by disrupting the struc-
ture of the huoMTOC, which leads to clinical
oocyte maturation arrest. This suggests that
the huoMTOC might be an important bio-
marker for evaluating the quality of human
oocytes.
Our discovery of huoMTOC provides in-
sights into the physiological mechanism of
microtubule nucleation and spindle assembly
in human oocytes. These findings also im-
prove our understanding of the pathological
mechanisms of oocyte maturation arrest.▪
The list of author affiliations is available in the full article online.
*Corresponding author. Email: wangleiwanglei@fudan.edu.cn (L.W.);
sangqing@fudan.edu.cn (Q.S.); xiaoxi_sun@aliyun.com (X.S.)
†These authors contributed equally to this work.
Cite this article as T. Wu et al., Science 378, eabq7361
(2022). DOI: 10.1126/science.abq7361
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.abq7361
Merge
Tubulin
TACC3
20 µm
20 µm
The huoMTOC structure in a human oocyte. The human GV oocyte shown here was matured for ~5 hours and fixed for immunofluorescence before NEBD. The huoMTOC
(TACC3, magenta) was surrounded by numerous microtubules (green) on the nuclear envelope. The dashed square shows the magnification region. The arrow highlights the huoMTOC.
Wu et al., Science 378, 745 (2022)
18 November 2022
1 of 1
RES EARCH
R E S E A R C H A R T I C L E
◥
HUMAN FERTILITY
The mechanism of acentrosomal spindle
assembly in human oocytes
Tianyu Wu1†, Jie Dong1†, Jing Fu2†, Yanping Kuang3†, Biaobang Chen4, Hao Gu1, Yuxi Luo1,
Ruihuan Gu2, Meiling Zhang5, Wen Li5, Xi Dong6, Xiaoxi Sun2*, Qing Sang1*, Lei Wang1*
Meiotic spindle assembly ensures proper chromosome segregation in oocytes. However, the mechanisms
behind spindle assembly in human oocytes remain largely unknown. We used three-dimensional
high-resolution imaging of more than 2000 human oocytes to identify a structure that we named the
human oocyte microtubule organizing center (huoMTOC). The proteins TACC3, CCP110, CKAP5, and
DISC1 were found to be essential components of the huoMTOC. The huoMTOC arises beneath the oocyte
cortex and migrates adjacent to the nuclear envelope before nuclear envelope breakdown (NEBD).
After NEBD, the huoMTOC fragments and relocates on the kinetochores to initiate microtubule
nucleation and spindle assembly. Disrupting the huoMTOC led to spindle assembly defects and oocyte
maturation arrest. These results reveal a physiological mechanism of huoMTOC-regulated spindle
assembly in human oocytes.
S pindle assembly is essential for ensuring
accurate chromosome transmission in
both mitosis and meiosis (1, 2). In somat-
ic cells, mitotic spindle assembly is me-
diated via duplicated centrosomes, which
consist of two centrioles surrounded by the
pericentriolar material (PCM) (3, 4). Centro-
somes are the major microtubule organizing
centers (MTOCs) in mitotic cells, and they are
responsible for microtubule nucleation and
spindle pole organization of the centrosomal
spindle (5, 6).
Unlike somatic cells, canonical centrosomes
are absent in the oocytes of many species
(7–11). Instead, acentriolar MTOCs (aMTOCs)
are observed in frog (12) and mouse (13–15)
oocytes, whereas no MTOC structures are
found in the Drosophila (16) or Caenorhabditis
elegans (17) oocytes. The difference suggests
that the mechanisms of female meiotic spindle
assembly are not conserved between species.
The aMTOC-directed meiotic spindle assembly
was only elucidated in mouse oocytes (15, 18, 19).
1Institute of Pediatrics, Children’s Hospital of Fudan
University, State Key Laboratory of Genetic Engineering,
Institutes of Biomedical Sciences, Shanghai Key Laboratory
of Medical Epigenetics, Fudan University, Shanghai 200032,
China. 2Shanghai Ji Ai Genetics and IVF Institute, Obstetrics
and Gynecology Hospital, Fudan University, Shanghai
200011, China. 3Department of Assisted Reproduction,
Shanghai Ninth People’s Hospital, Shanghai Jiao Tong
University School of Medicine, Shanghai 200011, China.
4NHC Key Lab of Reproduction Regulation, Shanghai
Institute for Biomedical and Pharmaceutical Technologies,
Fudan University, Shanghai 200032, China. 5Center for
Reproductive Medicine and Fertility Preservation Program,
International Peace Maternity and Child Health Hospital,
School of Medicine, Shanghai Jiao Tong University, Shanghai
200030, China. 6Reproductive Medicine Center, Zhongshan
Hospital, Fudan University, Shanghai 200032, China.
*Corresponding author. Email: wangleiwanglei@fudan.edu.cn (L.W.);
sangqing@fudan.edu.cn (Q.S.); xiaoxi_sun@aliyun.com (X.S.)
†These authors contributed equally to this work.
Mouse aMTOCs lack centrioles but contain
partial centrosomal proteins such as pericen-
trin (14), g-tubulin (20), CEP192 (18), and NEDD1
(21). Upon meiotic resumption after prophase I
arrest, multiple aMTOCs initiate microtubule
nucleation around the nuclear envelope in
mouse germinal vesicle (GV) oocytes. After
nuclear envelope breakdown (NEBD), aMTOCs
are clustered and then concentrated at the
spindle poles for bipolar spindle organiza-
tion (15, 18, 19).
A mechanistic understanding of meiotic spin-
dle assembly in human oocytes, on the other
hand, remains elusive (2). It is only known
that the spindle microtubule nucleation in
human oocytes is mediated by chromosomes
and promoted by guanosine triphosphate (GTP)–
bound Ran (RanGTP) (22). Subsequently, a
multipolar spindle is assembled as an interme-
diate, and then the spindle poles are merged
to form a bipolar spindle (22). However, hu-
man oocytes lack detectable aMTOCs at the
meiotic spindle poles (22, 23), and thus the
exact mechanism of acentrosomal spindle as-
sembly in human oocytes remains unclear.
Results
Spindle microtubules are nucleated from
kinetochores in human oocytes
A previous study implied that spindle micro-
tubules emanate from the kinetochores in hu-
man oocytes (22), and this phenomenon was
also observed in our immunofluorescent re-
sults in early prometaphase oocytes (fig. S1).
This suggests that kinetochores may serve as
the microtubule nucleation sites in human
oocytes. However, the dynamic process of how
spindle microtubules were originally nucleated
from kinetochores was not delineated. Thus,
we first observed the process of microtubule
nucleation in live human oocytes by using
three-dimensional (3D) high-resolution time-
lapse imaging (Fig. 1A). Human GV oocytes
were co-injected with mRNA encoding fluo-
rescently fused histone H2B (mCherry) and
centromere protein B (CENPB) (mClover3)
to visualize chromosomes and kinetochores,
respectively. Combined with fluorescent pro-
teins, SiR-tubulin (a dye with far-red fluores-
cence) was used to label microtubules before
oocyte maturation (24). According to our
three-channel time-lapse images, a micro-
tubule cluster (dashed square) derived from a
GV oocyte was observed proximal to chromo-
somes upon NEBD and disassembled in the
first few hours of meiosis I (Fig. 1A). Apart
from the microtubules (light-gray curve, bot-
tom panel) derived from the GV oocyte, a few
nascent microtubules (dark-gray curve, bottom
panel) were detected by microscopy after NEBD
(Fig. 1A).
To confirm that the nascent microtubules
were polymerized from kinetochores, we eval-
uated the tubulin intensity around kineto-
chores over time after NEBD. The intensity of
microtubules (tubulin) and chromosomes (H2B)
was measured around the representative kineto-
chores (Fig. 1A, arrows). The nascent micro-
tubules (dark-gray curve) were hardly observed
at the beginning of meiosis in human oocytes
(0 to 2 hours after NEBD). With the disassem-
bly of the derived microtubule cluster (light-
gray curve), the fluorescent tubulins began
to concentrate at the kinetochores starting at
~2 hours after NEBD (Fig. 1, A and B). Accord-
ing to three-channel time-lapse imaging, the
fluorescence of nascent microtubules (dark-
gray curve) overlapped primarily with kineto-
chores (red curve) from 2 to 2.5 hours after
NEBD and consistently throughout the begin-
ning of meiosis I (2 to 4 hours after NEBD)
(Fig. 1A). However, the fluorescence of nascent
microtubules (dark-gray curve) did not consis-
tently overlap with chromosomes (blue curve),
suggesting that the nascent microtubules pri-
marily polymerized from the kinetochores rath-
er than other regions of the chromosomes
(Fig. 1A). Along with meiotic maturation, mi-
crotubules were nucleated slowly but con-
tinuously (Fig. 1, A and B), and the stable
microtubules with high tubulin intensity could
be observed on kinetochores starting at 3 hours
after NEBD (Fig. 1, A and B). These results
suggest that in human oocytes, the nucleation
of spindle microtubules is initiated from kinet-
ochores starting at 2 to 4 hours after NEBD.
Next, to further confirm the nucleation of
spindle microtubules on kinetochores, we
treated live human metaphase I (MI) oocytes
with a reversible microtubule inhibitor (noco-
dazole) and tracked the dynamic recovery
of microtubule nucleation immediately after
nocodazole was washed out (Fig. 1C). Micro-
tubules and kinetochores were marked by
Wu et al., Science 378, eabq7361 (2022)
18 November 2022
1 of 12
RES EARCH | R E S E A R C H A R T I C L E
A
.
i
(
i
l
)
.
n
1.0
1.0
1.0
0.5
1.0
U
A
t
s
H
3:30
3:00
2:30
4:00
1:00
2:00
u
b
u
T
NEBD
e
g
r
e
M
y
t
i
s
n
e
t
n
I
B
P
N
E
C
B
2
H
e
n
o
Fig. 1. The dynamic process of
spindle microtubules nucleating
from kinetochores in human
oocytes. (A) Representative time-
lapse images showing microtubule
nucleation in human oocytes
after NEBD. Blue, chromosomes
(H2B-mCherry); gray, microtubules
(SiR-tubulin); magenta, kineto-
chores (mClover-CENPB). The
microtubule cluster was marked at
0 to 2 hours after NEBD (dashed
squares). Representative kineto-
chores of each time point are shown
in some z-sections. The intensity of
chromosomes and microtubules
was measured and plotted around
representative kinetochores. The
graphs in the bottom panel are the
fluorescence profiles of H2B, tubu-
lin, and CENPB across the repre-
sentative kinetochores along the
direction of the arrows in the
merged images at each time point.
The microtubules derived from GV
oocytes (light-gray curves) and
nascent microtubules (dark-gray
curves) were distinguished at
2 hours after NEBD. The plotted
intensity curves indicate the relative
location of each fluorescent protein.
Blue, chromosomes; gray, micro-
tubules; magenta, kinetochores.
A.U., arbitrary units. Scale bar,
5 mm. (B) Representative images
showing kinetochore nucleated
microtubules at different time points.
The intensity of the microtubules on
kinetochores measured in (A) was
compared between 2 and 4 hours after
NEBD (n > 24 kinetochores, signifi-
cance of the differences in intensity
was calculated by ANOVA and indi-
cated on the graph). Scale bar, 2 mm.
(C) Representative time-lapse
images (z-stack) showing microtubule
nucleation after nocodazole washout
in live human oocytes. Gray, micro-
tubules (GFP-MAP4); magenta,
kinetochores (mScarlet-CENPB). Flow
diagram shows the sequence of
experiments. Arrow indicates time-
lapse imaging initiation. Time is given
as hours:minutes after nocodazole
washout. Scale bar, 5 mm. (D) Single-slice images of boxed areas in (C). Time is given as hours:minutes after nocodazole washout. Scale bar, 2 mm. Intensity of kinetochores
and microtubules is indicated in the graphs. The bottom graphs are the fluorescence profiles of MAP4 (gray) and CENPB (magenta) proximal to the representative kinetochores
along the direction of the yellow arrow. (E) The intensity of microtubules on kinetochores in (D) was measured and compared after nocodazole washout (significance was
calculated by ANOVA). (F) Mechanistic model for microtubule nucleation in human oocytes. Time is after NEBD.
4
Derived from GV oocyte
(
e
r
o
h
c
o
e
n
k
n
o
y
t
i
s
n
e
n
Time from release (min
Time from NEBD (hours
Wash out and imaging
3
Nascent
Time from NEBD
2 3 4
Distance (µm)
Distance (µm)
Chromosome
Kinetochore
Nocodazole
P 0.0001
P<0.0001
n
o
i
t
c
e
s
-
z
Mature for
2- 4 hours
P=0.002
P<0.01
y
t
i
s
n
e
t
n
I
1 2 3
1 2 3
y
t
i
s
n
e
t
n
B
P
N
E
C
CENPB
1 2 3
1 2 3
Tubulin
Tubulin
4
P
A
M
h
5
.
3
u
b
u
T
D
u
b
u
T
C
B
0:05
0:15
0:10
0:20
10 h
0:15
0:10
0:05
0:20
E
F
U
A
U
A
U
A
0 .6
0 .8
0 .2
0 .0
0 .4
1 .0
1.0
1.0
0.5
0.5
1 h
1.0
0.5
0.5
0.5
0.5
0.0
0.5
1.0
0.0
0.5
0.0
0.5
1.0
1.0
0.0
1.0
0.5
1.5
0.5
0.0
1.0
0
1
20
10
15
)
.
)
.
)
.
2
h
3
h
h
4
2
0
1
3
0
0
0
0
1
2
3
0
0
0
4
2
0
3
1
3
0
1
0
1
1
2
2
3
1
0
2
0
4
0
0
2
3
0
n
5
0
0
n
i
l
i
l
(
(
.
.
.
t
t
I
i
i
fluorescent microtubule-associated protein
4 (MAP4) and CENPB proteins, respectively.
Initially, nocodazole completely disrupted the
microtubules in human MI oocytes, and re-
polymerized microtubules were not detected on
kinetochores until ∼10 min after nocodazole was
washed out. At that time, several nascent mi-
crotubules were observed near the kinetochores,
and some of them were emanating outward
(Fig. 1C). The trajectory of microtubules was
also shown in a single slice (Fig. 1D), and the
representative kinetochore was nucleating
Wu et al., Science 378, eabq7361 (2022)
18 November 2022
2 of 12
RES EARCH | R E S E A R C H A R T I C L E
microtubules slowly and continuously, suggest-
ing that nascent microtubules were nucleated
from kinetochores (Fig. 1, D and E). Together,
these results show the dynamic process of
acentrosomal spindle microtubules nucleating
from kinetochores in human oocytes (Fig. 1F).
Discovery of a specific microtubule nucleator in
human oocytes
To identify the specific factors driving spindle
microtubule nucleation from kinetochores in
human oocytes, we localized 86 human centro-
some and microtubule-related proteins by per-
forming immunofluorescence in >1000 fixed
human oocytes (Fig. 2A and figs. S2 and S3).
These proteins were classified according
to their function or localization in somatic
cells (21, 25, 26), but their localization in
human oocytes was largely unknown. These
proteins included 34 centrosomal proteins,
25 microtubule-associated proteins, 12 dynein-
related proteins, five regulatory kinases or
substrates, five spindle assembly factors, and
five nuclear pore–related proteins (fig. S2). A
total of 36 proteins were specifically local-
ized on spindle microtubules, two proteins
were concentrated on spindle poles, and one
protein showed spindle periphery localiza-
tion (Fig. 2A and fig. S3). Unexpectedly, four
proteins—centriolar coiled-coil protein 110
(CCP110), cytoskeleton-associated protein 5
(CKAP5), disrupted in schizophrenia 1 (DISC1),
and transforming acidic coiled-coil–containing
protein 3 (TACC3)—exhibited both kineto-
chore and spindle microtubule localization
(Fig. 2B), and such localization was consistent
until the MII stage (fig. S4), which was notably
different from their localization in human mi-
totic cells and in mouse oocytes (6, 21, 27–30).
CCP110 and DISC1 are centrosomal proteins
that are involved in spindle assembly at the
centrosomes of human mitotic cells and
aMTOCs of mouse oocytes (6, 21, 27). CKAP5
and TACC3 are microtubule-associated pro-
teins concentrated at centrosomes and micro-
tubules in human mitotic cells during mitosis
(28–30). Importantly, both CKAP5 and TACC3
have been reported to nucleate microtubules
in vitro and in mitotic cells, respectively (31–34).
These observations suggested that CCP110,
CKAP5, DISC1, and TACC3 are potential
Fig. 2. Identification of a micro-
tubule nucleator in human
oocytes. (A) Schematic of
the metaphase I spindle
in human oocytes (chromosomes
in blue, microtubules in green,
spindle poles in magenta, kineto-
chores in yellow, and spindle
periphery in gray). (B) Immuno-
fluorescence images of
metaphase I spindle in human
MI oocytes showing protein
localization relative to both
microtubules and chromosomes.
Gray, chromosomes; green,
microtubules. Scale bar, 5 mm.
(C) Immunofluorescence images
of the nucleus in human GV
oocytes before NEBD. The yellow
squares indicate the newly
identified structure surrounded
by microtubules. Scale bar,
10 mm. The yellow arrows indicate
intensity measurements in the
images at the top. The localization
of huoMTOC and microtubule
distribution of human GV oocytes
are shown below. Gray, chromatin
(Hoechst); green, microtubules
(tubulin); magenta, huoMTOC.
(D) Representative images of live
human GV oocytes. mScarlet-
TACC3 was used as the standard
for colocalization. Green, TACC3;
magenta, CCP110, CKAP5, or
DISC1; gray, chromosomes
(Hoechst). The yellow arrows
indicate the colocalization. Scale
bar, 10 mm.
A
C
)
.
U
A
.
(
y
t
i
s
n
e
t
n
I
B
Merge
CCP110
Tubulin
Chromosome
Merge
CKAP5
Tubulin
Chromosome
Spindle microtubules
Merge
DISC1
Tubulin
Chromosome
AURKA
BUGZ CETN2 CETN3 CENPJ
AKAP450
CEP120 CEP192 CEP250 CLIP1
DCTN1
DCTN2 DLGAP5 DYNLT1 GTSE1 HAUS4 HAUS6
HAUS8
CLTC
HMMR KANSL3 KIF11
NDE1 NDEL1
MYO10
PRC1
PLK4
PLK1
KIZ
KIF20A
NEK2 NUSAP1
TUBG1
TPX2
LIS1
PCM1
Merge
TACC3
Tubulin
Chromosome
Kinetochores and spindle microtubules
CCP110 CKAP5
DISC1
TACC3
Spindle poles
Spindle periphery
NUMA1
KIF2A
HOOK3
CCP110
CKAP5
DISC1
TACC3
Merge
CCP110
TACC3
D
Tubulin
Hoechst
1.0
0.5
0.0
0 10 20 30 40
1.0
0.5
0.0
0 10 20 30 40
1.0
0.5
0.0
1.0
0.5
0.0
0 10 20 30 40
0 10 20 30 40
Distance (µm)
s
e
t
y
c
o
o
V
G
n
a
m
u
h
e
v
L
i
Merge
CKAP5
TACC3
Merge
DISC1
TACC3
Wu et al., Science 378, eabq7361 (2022)
18 November 2022
3 of 12
RES EARCH | R E S E A R C H A R T I C L E
candidates for regulating microtubule nucle-
ation and polymerization in human oocytes.
Subsequently, these proteins were monitored
by immunofluorescence in human GV oocytes
just before NEBD (fig. S5 and movie S1). Nota-
bly, each of the four proteins showed an un-
usual structure surrounded by microtubules.
This structure was proximal to the nuclear en-
velope in human GV oocytes ∼0 to 2 hours
before NEBD (Fig. 2C), which was consistent
with our observations of a microtubule clus-
ter proximate to chromatin in human oocytes
after NEBD (movie S2). To test whether these
four proteins were colocalized, we overex-
pressed TACC3 (labeled with mScarlet) and the
other three proteins (labeled with mClover3).
As indicated in Fig. 2D, CCP110, CKAP5, and
DISC1 were all colocalized with TACC3, imply-
ing that the four proteins belong to the same
structure. To examine the interactions among
these proteins, coimmunoprecipitation was
performed in human embryonic kidney 293T
(HEK293T) cells transfected with plasmids
containing the corresponding genes. As a re-
sult, CCP110, CKAP5, and DISC1 all interacted
directly with TACC3, whereas the negative
control GDF9 had no interaction with TACC3
(fig. S6). This suggests that the four proteins
are all components of the same structure. In
addition, a dense microtubule cluster was ob-
served around this structure in human GV
oocytes, which caused asymmetrical microtu-
bule distribution around the nuclear envelope
prior to NEBD (Fig. 2C). Given these features,
we refer to this potential nucleating structure
as the human oocyte microtubule organizing
center (huoMTOC). This structure was not
detected in the oocytes of other mammalian
species such as mice and pigs (fig. S7), sug-
gesting the specificity of this structure in
human oocytes.
The huoMTOC is essential for
microtubule nucleation
According to the transcriptional landscape
data of human oocytes in public databases
(23), TACC3 shows overwhelmingly higher
expression (more than 29-fold) than CCP110,
CKAP5, and DISC1, implying its potential key
role in the huoMTOC. We therefore tried to
disrupt the huoMTOC by knocking down
TACC3 by injecting the corresponding short-
interfering RNAs (siRNAs) into human GV
oocytes (fig. S8). The integrity of the huoMTOC
was evaluated by immunofluorescence for
CCP110, CKAP5, DISC1, and TACC3. As ex-
pected, no nucleating structures were detected,
demonstrating the complete disruption of the
huoMTOC upon TACC3 depletion (Fig. 3A).
To determine whether huoMTOC is the main
microtubule nucleator, we assessed the mi-
crotubule distribution in oocytes under the
condition of huoMTOC deficiency. The asym-
metrical distribution of microtubules was
drastically diminished, implying the essen-
tial role of the huoMTOC for microtubule
nucleation in human oocytes (Fig. 3A). In ad-
dition, different degrees of disruption in the
asymmetrical distribution of microtubules
around the nuclear envelope were also ob-
served after specific depletion of endogenous
CCP110, CKAP5, or DISC1 (Fig. 3, B and C, and
fig. S8). Of note, depletion of TACC3 resulted
in the most severe disruption of microtubule
asymmetrical distribution in most of the ana-
lyzed human GV oocytes (Fig. 3, B and C).
These results indicate that the huoMTOC is
essential for the asymmetrical nucleation of
microtubules around the nuclear envelope
and that all four proteins are indispens-
able for maintaining normal function of the
huoMTOC, in which TACC3 presumably plays
a leading role.
The huoMTOC is fragmented and recruited to
kinetochores for the initiation of spindle
assembly in human oocytes
Unlike the mechanism in mouse MI oocytes in
which aMTOCs were aggregated on the spin-
dle poles during spindle assembly (15), the
components of the huoMTOC in human MI
oocytes were localized on kinetochores. To
reveal the dynamic process of huoMTOC-
regulated spindle assembly, human GV oocytes
were fixed for immunofluorescence at NEBD
or at 2, 4, or 6 hours before NEBD. Initially, the
huoMTOC appeared beneath the oocyte cor-
tex. Slowly, the huoMTOC migrated from the
cortex to the nuclear envelope of human GV
oocytes (Fig. 4, A and B). The huoMTOC then
expanded, and its nucleated microtubules grew
rapidly during the resumption of meiosis (Fig.
4, A and C).
We next visualized the dynamic changes of
the huoMTOC after NEBD by 3D time-lapse
imaging. At the beginning of NEBD, the
huoMTOC localized proximal to the chromo-
somes and became fragmented within the first
hour (fig. S9A and movie S3). The huoMTOC
microtubules were disrupted and were barely
observable in the first few hours after NEBD
(Fig. 4D and movie S4). The fragmented
huoMTOC was then relocated to chromosomes,
primarily to kinetochores (Fig. 4D, fig. S9B,
and movie S5). The spindle microtubules were
initially observed when the huoMTOC was re-
built (Fig. 4, D and E; fig. S9C; and movie S4).
Next, to determine the role of the huoMTOC
in spindle assembly, each huoMTOC compo-
nent was down-regulated in human GV oocytes
that were cultured to the MI stage. Down-
regulation of these components significantly
impaired spindle microtubule nucleation and
spindle assembly in human MI oocytes com-
pared with the control group (P < 0.001, Fisher’s
exact test), and stable microtubules were great-
ly decreased (Fig. 4, F and G). Compared to
other components, TACC3 depletion had the
most obvious effects on the microtubule nu-
cleation and spindle assembly (Fig. 4, F and
G), further suggesting that TACC3 plays a
leading role in the huoMTOC. To directly test
whether huoMTOC is essential for spindle as-
sembly, the huoMTOC marked by fluorescent
TACC3 was disrupted by laser ablation (fig.
S10, A and B). Similar to TACC3 depletion, the
spindle microtubule polymerization and spin-
dle assembly were significantly impaired by
the laser ablation of huoMTOC (P = 0.015,
Fisher’s exact test) (fig. S10, C and D). It has
been demonstrated that RanGTP is also re-
quired for spindle assembly in human oocytes
(22). Disruption of both TACC3 and RanGTP
aggravated the spindle microtubule polymer-
ization defects (fig. S11), indicating combined
effects of TACC3 and RanGTP on microtubule
nucleation and spindle assembly. These re-
sults suggest that the huoMTOC is required
for spindle microtubule polymerization and
spindle assembly of human oocytes.
huoMTOC deficiency interrupts normal
spindle assembly and causes clinical oocyte
maturation arrest
Oocyte maturation requires microtubule nu-
cleation and spindle assembly (22). In the
clinic, a number of infertile patients with re-
current failed in vitro fertilization (IVF) or in-
tracytoplasmic sperm injection (ICSI) attempts
have been diagnosed with oocyte maturation
arrest. Considering the key role of the huoMTOC
in human spindle assembly, we hypothesized
that a disrupted huoMTOC resulting from mu-
tations in CCP110, CKAP5, DISC1, or TACC3
may cause impaired spindle assembly and
abnormal oocyte maturation in patients.
We thus screened for likely pathogenic mu-
tations in a cohort of 1394 infertile female
patients characterized by oocyte maturation
arrest by analyzing their whole exome sequenc-
ing datasets (data deposited in the Genome
Variation Map of the National Genomics Data
Center under accession number GVM000402).
Each of these patients had undergone several
IVF attempts, all of which failed because of
oocyte maturation arrest. We identified two
patients with compound heterozygous muta-
tions in the key huoMTOC component TACC3
(Fig. 5A and table S1). According to the Ge-
nome Aggregation Database (gnomAD) and
our in-house controls (data deposited in the
Genome Variation Map of the National Geno-
mics Data Center under accession number
GVM000394), all four TACC3 variants from
patients are rare variants (table S2). Impor-
tantly, these two patients showed very similar
phenotypes, in which most of their retrieved
oocytes were immature after in vitro matu-
ration (table S1), and polarization microscopy
images showed no visible spindles in the live
oocytes (Fig. 5B). Immunofluorescence in fixed
NEBD oocytes also demonstrated that the
Wu et al., Science 378, eabq7361 (2022)
18 November 2022
4 of 12
A
)
.
U
A
.
(
y
t
i
s
n
e
t
n
I
B
RES EARCH | R E S E A R C H A R T I C L E
TACC3
CKAP5
CCP110
DISC1
Control siRNA
TACC3 siRNA
Control siRNA
TACC3 siRNA
Control siRNA
TACC3 siRNA
Control siRNA
TACC3 siRNA
Hoechst
Tubulin
1.0
0.5
0.0
0
10
30
50
1.0
0.5
0.0
0
10
30
50
1.0
0.5
0.0
1.0
0.5
1.0
0.5
1.0
0.5
0
10
30
0.0
50
0
10
30
0.0
0
50
10
Length (µm)
30
0.0
50
0
10
30
50
Control siRNA
TACC3 siRNA
CKAP5 siRNA
CCP110 siRNA
DISC1 siRNA
e
g
r
e
M
n
i
l
u
b
u
T
)
.
U
A
.
(
y
t
i
s
n
e
t
n
I
1.0
0.5
0.0
Hoechst
1.0
0.5
0.0
0
10
30
0
10
30
50
1.0
0.5
1.0
0.5
0.0
50
0
10
30
Length (µm)
50
0.0
0
10
30
50
1.0
0.5
0.0
1.0
0.5
0.0
1.0
0.5
0
10
30
0.0
50
0
10
30
50
C
)
%
(
n
o
i
t
u
b
i
r
t
s
d
i
l
e
u
b
u
t
o
r
c
m
i
l
a
c
i
r
t
e
m
m
y
s
A
P<0.001
100
(16)
80
60
40
20
0
(17)
(15)
(13)
(13)
Control
TACC3
CKAP5
CCP110
DISC1
0
10
30
50
siRNAs
Fig. 3. The huoMTOC is required for microtubule nucleation in human oocytes.
(A) Immunofluorescence images of human GV oocytes injected with TACC3
siRNAs. Green, microtubules (tubulin); magenta, huoMTOC (yellow squares);
blue, chromatin. Scale bar, 10 mm. The measurements of tubulin intensity along the
direction of the yellow arrows are shown at the bottom. Green, tubulin; blue,
chromosomes. (B) Immunofluorescence images of human GV oocytes injected with
specific siRNAs. Green, microtubules (tubulin); gray, chromosomes (Hoechst). The
yellow arrows indicate the direction of the tubulin intensity measurements shown
at the bottom. Scale bar, 10 mm. (C) The percentages of human oocytes with
asymmetrical microtubule distribution in (B). P < 0.001, Fisher’s exact test. Data
were from three independent experiments. The number of oocytes analyzed is
specified in parentheses.
meiotic spindle was completely disrupted
(Fig. 5C). We also determined the stability of
the huoMTOC and microtubule distribution
in the patients’ GV oocytes. Compared to normal
human GV oocytes, the huoMTOC was miss-
ing, and the asymmetrical distribution of mi-
crotubules around the nucleus was impaired
in the patients’ GV oocytes (Fig. 5D).
In addition, supplementing the wild-type
TACC3 mRNA successfully rescued the pheno-
type of spindle disruption resulting from TACC3
depletion, whereas supplementing the mutant
TACC3 mRNAs could not rescue the phenotype
(Fig. 5, E and F), suggesting that the mutations
had loss-of-function effects on TACC3. Thus,
TACC3 deficiency caused female infertility and
oocyte maturation arrest by disrupting the in-
tegrity of the huoMTOC. These results suggest
that disruption of the huoMTOC impaired the
nucleation of microtubules in the GV oocytes
of patients, further highlighting the critical
role of the huoMTOC in regulating microtu-
bule nucleation and the initiation of acentro-
somal spindle assembly.
Discussion
Here, we report a structure that we named the
huoMTOC, which serves as a major site of
microtubule nucleation and is required for
spindle assembly in human oocytes. A single
huoMTOC is formed near the cortex of human
GV oocytes at the time of meiosis resumption,
and it migrates to the nuclear envelope before
Wu et al., Science 378, eabq7361 (2022)
18 November 2022
5 of 12
RES EARCH | R E S E A R C H A R T I C L E
l
i
.
t
.
I
(
(
)
0
i
l
n
0
0
0
)
.
)
.
n
o
t
-2
-6
-6
-2
-4
-4
8 0
2 0
4 0
6 0
1.2
0.6
1.0
0.0
0.5
E
A
B
C
D
U
A
U
A
m
µ
(
4:00
5:00
3:10
n=3
n=3
n=3
y
t
i
s
n
e
u
b
u
T
3
C
C
A
T
s
u
e
c
u
n
e
g
r
e
M
y
t
i
s
n
e
t
n
I
e
c
n
a
t
s
D
NEBD (0 h)
Tubulin
TACC3
Tubulin
TACC3
Prophase (-4 h)
Time from NEBD (hours)
Time from NEBD (hours)
Late prophase (-2 h)
Early prophase (-6 h)
Fig. 4. The huoMTOC is
recruited from the
cortex to kinetochores
for spindle microtubule
polymerization in
human oocytes.
(A) Immunofluorescence
images of human GV
oocytes at NEBD and at
2 hours (late prophase),
4 hours (prophase), and
6 hours (early prophase)
before NEBD. Green,
microtubules (tubulin);
blue, chromatin
(Hoechst); magenta,
huoMTOC (TACC3). Green
arrows indicate the
microtubule cluster, and
magenta arrows indicate
the huoMTOC. Scale bar,
20 mm. (B) The distance
between the huoMTOC
and the nucleus was
measured in (A). Time is
after NEBD. (C) The
intensity of the micro-
tubule cluster (tubulin)
and the huoMTOC
(TACC3) was measured
at different time points in
(A). Time is after NEBD.
(D) Representative time-
lapse images showing
the relationship between
the huoMTOC and micro-
tubule nucleation in
live human oocytes.
Green, microtubules (SiR-
tubulin); blue, chromo-
somes (H2B-mClover);
magenta, huoMTOC
(mScarlet-TACC3). Scale
bar, 5 mm. (E) The inten-
sity of microtubules
(SiR-tubulin) and
huoMTOC (mClover-
TACC3) close to chromo-
somes was measured in
(D). Time is after NEBD.
The number of oocytes
analyzed in experiments
is indicated. The mean
and standard error were
calculated on the basis of two independent experiments. Error bars are standard deviations. (F) Immunofluorescence images of human MI spindles from control
and TACC3, CKAP5, CCP110, and DISC1 siRNA-injected human oocytes. Green, microtubules (tubulin); gray, chromosomes (Hoechst). Scale bar, 5 mm. (G) The spindle
assembly percentage measured and collected from (F). The number of oocytes analyzed in three independent experiments is indicated. P < 0.001, Fisher’s exact test.
Time from NEBD (hours)
CCP110
CKAP5
Control
TACC3
DISC1
e
m
o
s
o
m
o
r
h
C
CCP110
P<0.001
CKAP5
Hoechst
Control
TACC3
y
b
m
e
s
s
a
DISC1
e
g
r
e
M
siRNAs
e
d
n
p
S
siRNAs
3
C
C
A
T
u
b
u
T
12:10
11:40
11:20
u
b
u
T
5:50
8:20
7:00
(18)
(18)
(24)
(16)
(20)
G
100
F
40
20
60
80
%
n
i
l
5
3
4
6
0
n
i
l
)
(
i
l
l
NEBD. After NEBD, the huoMTOC becomes
fragmented and is recruited to chromo- somes
and kinetochores for spindle microtubule nucle-
ation (Fig. 6). With the microtubule poly-
merization, the huoMTOC proteins are also
recruited to the spindle microtubules (Fig. 6
and fig. S9C). Ablation of the huoMTOC re-
sults in microtubule loss and defective spindle
assembly. In addition, we demonstrated that
TACC3, CCP110, CKAP5, and DISC1 are essential
components of the huoMTOC and that muta-
tions in TACC3 cause clinical oocyte maturation
arrest and female infertility.
Distinct aMTOCs have been identified and
investigated in mouse oocytes as microtubule
Wu et al., Science 378, eabq7361 (2022)
18 November 2022
6 of 12
RES EARCH | R E S E A R C H A R T I C L E
C
A
l
I
I
II
II
1
2
1
2
1
2
1
F
I-1
I-2
I-1
I-2
II-2
II-1
II-1
B
D
t
h
g
L
Normal
e
g
r
e
M
e
g
r
e
M
Norma
3
C
C
A
T
Normal
Tubulin
Hoechst
Family 1
Family 2
C A C C G G A C C T A
Family 2 II-1
Family 1 II-1
C A C C G G/T A C C T A
A A A G C G G A G A C
C C C C C C/G C A T G C
T G G C A G C C C T G
C A C C G G A C C T A
C A C C G G/T A C C T A
T G G C A G/C C C C T G
T G G C A G/C C C C T G
amily 2 II-1
C C C C C C/G C A T G C
C C C C C C C A T G C
Family 2 II-1
Family 1 II-1
Family 1 II-1
c.673_708del
[c.530G>C];[WT]
[c.1892G>T];[WT]
[c.1184C>G];[WT]
[c.673_708del];[WT]
[c.673_708del];[WT]
A A A G C G G A G A C
C G G C A C G G T G G
A A A G C G G A G A C
C G G C A C G G T G G
A A A G C G G A G A C
C G G C A C G G T G G
[c.530G>C];[c.1184C>G]
c.1892G>T [p.Gly631Val]
[p.Lys225_Cys236del]
[c.673_708del];[c.1892G>T]
c.1892G>T
[p.Gly631Val]
c.530G>C
[p.Ser177Thr]
c.1184C>G
[p.Pro395Arg]
c.673_708del [p.Lys225_Cys236del]
c.530G>C [p.Ser177Thr]
c.1184C>G [p.Pro395Arg]
Fig. 5. Disruption of
the huoMTOC in human
oocytes impairs micro-
tubule nucleation and
spindle assembly.
(A) Pedigrees of the two
families with TACC3
mutations with Sanger
sequencing confirmation.
Squares denote male
family members, circles
denote female members,
black solid circles denote
probands, and the equal
sign denotes infertility.
(B) Human oocytes from
donors (normal) and
patients (family 1 II-1,
family 2 II-1) were
examined by light and
polarization microscopy.
The arrow indicates an MI
spindle. (C) Immuno-
fluorescence images of
human MI oocytes from
donors (normal) and
patients (family 1 II-1,
family 2 II-1). Green,
microtubules (tubulin);
blue, chromosomes
(Hoechst); magenta,
TACC3. Scale bar, 5 mm.
(D) Immunofluorescence
images of human GV
oocytes from donors
(normal) and patients
(family1 II-1, family2 II-1).
Green, microtubules
(tubulin); blue, chromo-
somes (Hoechst);
magenta, TACC3. Scale
bar, 10 mm. The yellow
square shows the
huoMTOC in normal
human oocyte. The
microtubule distribution
was measured as
previously described.
(E) Immunofluorescence
images of human MI
oocytes injected with
TACC3 siRNAs and wild
type or patient-derived
mutant mRNAs. Green,
microtubules (tubulin); blue, chromosomes (Hoechst), magenta, FLAG-TACC3. Scale bar, 5 mm. (F) The percentage of human oocyte maturation measured in (E).
(Fisher’s exact test, P < 0.05). The number of oocytes analyzed is specified in parentheses.
G631V
P395R
WT
S177T
K225_C236del
10
Length (µm)
K225_C236del
TACC3 siRNA
n
o
i
t
a
z
i
r
a
o
P
H2O
P<0.05
y
t
i
s
n
e
t
n
I
Hoechst
G631V
Tubulin
P395R
S177T
Tubulin
3
C
C
A
T
3
C
C
A
T
e
g
r
e
M
H2B
H2O
(10)
WT
e
t
a
r
(11)
100
U
A
E
a
r
u
1.0
0.5
0.0
0.5
1.0
1.0
0.5
0.0
0.0
a
M
F
(6)
(9)
(9)
(7)
50
50
30
60
80
10
30
10
30
20
40
n
o
%
)
.
0
0
0
0
i
t
)
(
(
.
t
i
l
50
nucleators for meiotic spindle assembly (15).
However, the primary spindle microtubule nu-
cleator of human oocytes has remained un-
known. The following factors might be reasons
why the huoMTOC was not identified in pre-
vious investigations: (i) Unlike mouse aMTOCs,
the huoMTOC is not concentrated on the spin-
dle poles in human MI and MII oocytes (22).
(ii) The classic aMTOC marker pericentrin is
not among the components of the huoMTOC
and therefore cannot label the structure. (iii) Only
a single huoMTOC is formed in late prophase,
and it is fragmented immediately after NEBD,
making it difficult to capture in human oo-
cytes by immunofluorescence.
In previous investigations, spindle assem-
bly of human oocytes was reported to be me-
diated by chromosomes and dependent on
Wu et al., Science 378, eabq7361 (2022)
18 November 2022
7 of 12
RES EARCH | R E S E A R C H A R T I C L E
Before NEBD
~ -6 hours
Early prophase
~ -4 hours
Prophase
~ -2 hours
Late prophase
~ 0 hours
NEBD
After NEBD
~ 0-2 hours
~ 3~5 hours
~ 6-8 hours
~ 9-12 hours
Nucleus
Chromosome
Microtubule
huoMTOC
Kinetochore
Fig. 6. Mechanistic model for huoMTOC migration and microtubule nucleation in human oocytes.
The huoMTOC (magenta) assembles near the cortex of GV oocytes and migrates to the nuclear envelope
before NEBD. The huoMTOC expands and the surrounding microtubules (green) keep growing until NEBD. After
NEBD, the huoMTOC fragments, and the surrounding microtubules are disassembled. The fragmented huoMTOC is
then recruited to kinetochores and initiates microtubule nucleation for meiotic spindle assembly.
RanGTP (22). It has been demonstrated that
RanGTP inhibition impairs spindle assembly in
human oocytes (22). In mouse or Drosophila oo-
cytes, spindles have defects but still assemble if
RanGTP is inhibited, suggesting that the RanGTP
pathway is not essential for meiotic spindle as-
sembly in oocytes of these two species (35, 36).
In the mouse oocytes lacking centrosomes, the
processes of spindle assembly and spindle mi-
crotubule nucleation are mainly achieved by
aMTOCs and facilitated by liquid-like meiotic
spindle domain (21, 37). It has been argued for
a long time that, unlike mouse oocytes, human
oocytes lack prominent aMTOCs (2, 22, 38).
However, in the present study, the spindle mi-
crotubule in human oocytes was observed to
be primarily polymerized from the kinetochores
by an aMTOC-like structure that we named
huoMTOC. In the Drosophila oocytes that also
lack prominent aMTOCs, the spindle assembly
is dominated by the chromosomes, which re-
cruit and/or nucleate the microtubules (39).
The spindle microtubules are organized around
the chromosomes into fibers of two types,
the interpolar and kinetochore microtubules,
to assemble the meiotic spindle of Drosophila
(39). Although the spindle assembly is directed
by chromosomes in both human and Drosophila
oocytes, the mechanisms involved could be
different. The meiotic spindle assembly in
Drosophila oocytes requires the chromosomal
passenger complex (16), but that in human
oocytes requires the huoMTOC.
In our study, the localization of 86 centro-
some and spindle-related proteins was per-
formed in human MI oocytes by systematic
immunofluorescent staining. We identified
the components (TACC3, CKAP5, CCP110, and
DISC1) of the huoMTOC on kinetochores in
MI oocytes and then verified their localiza-
tion in GV oocytes using high-resolution imag-
ing in live human oocytes. Emerging evidence
suggests that the human TACC3 plays an im-
portant role in microtubule growth and spin-
dle assembly during mitosis (30, 40–42). The
microtubule nucleation is blocked in ovarian
cancer cells when TACC3 expression is affect-
ed, suggesting that TACC3 is required for
centrosome-involved microtubule nucleation
(43). TACC3 depletion impairs the g-tubulin
ring complex assembly (44) and centrosome
integrity (45) of human somatic cells. CKAP5
was previously identified as a microtubule nu-
cleation factor in vitro (32, 33, 46). CKAP5 was
observed functioning synergistically with the
g-tubulin ring complex for de novo microtu-
bule nucleation (32) and catalyzing numer-
ous rounds of tubulin subunit addition at the
microtubule plus-end for spindle microtubule
assembly in vitro (33). In addition, CKAP5 was
also implicated in the importin-regulated mi-
crotubule nucleation as a microtubule poly-
merase in vitro (46). In mitosis, TACC3 and
CKAP5 are RanGTP-regulated spindle assem-
bly factors for spindle microtubule nucleation
and stabilization (47–49). Collectively, we in-
ferred that TACC3 and CKAP5 may act as
organizers of huoMTOC assembly and micro-
tubule nucleation in human oocytes. The
functions of CCP110 and DISC1 in micro-
tubule polymerization have not yet been deter-
mined in either mitotic or meiotic cells.
Apart from the four huoMTOC components,
we also identified 39 other proteins that
showed spindle-related localization. We there-
fore cannot exclude the possibility that some
of these proteins may also play a role in
huoMTOC formation and function. The spe-
cific components of huoMTOC and the mech-
anism for its nucleation of microtubules are
worth investigating in the future. In addition,
according to our observations, the localization
of most centrosome and microtubule-related
proteins in the MI spindle of human oocytes is
obviously different from that in mouse oocytes
(21). Thus, future investigations on the func-
tions of these proteins in human oocytes should
shed more light on the mechanism of human
oocyte spindle assembly.
In a recent study, a distinct mechanism of
spindle pole organization was discovered in
human oocytes, suggesting that loss of kinesin
superfamily protein C1 (KIFC1) induces meiotic
spindle instability (23). In addition, our pre-
vious investigations suggest that tubulin beta 8
class VIII (TUBB8) is the main isotype of spin-
dle b-tubulin in human oocytes but is not found
in mice or other nonprimate species (50). These
findings suggest that a distinct mechanism for
the initiation of microtubule nucleation and
spindle assembly has evolved in human oocytes,
which may contribute to a series of physio-
logical characteristics including increasing
chromosome segregation errors, high spindle
instability, and aneuploidy in human oocytes.
In the clinic, many infertility patients ex-
perience recurrent failure of IVF or ICSI at-
tempts owing to oocyte maturation arrest.
However, the genetic factors involved remain
unknown for most patients. In this study, we
found that mutations in TACC3 cause defects
in spindle assembly by disrupting the struc-
ture of the huoMTOC, which leads to oocyte
maturation arrest in patients. It is worth per-
forming mutational screening for both known
and potential genes that might participate in
the maintenance of huoMTOC integrity. The
results of such screening will help in precision
diagnosis for these patients and will provide
therapeutic targets for future clinical treat-
ments. Our findings provide not only insights
into the physiological mechanism of micro-
tubule nucleation and spindle assembly in
human oocytes but also improve our under-
standing of pathophysiological mechanisms of
human oocyte maturation arrest.
Materials and methods
Human oocyte collection and culture
Human GV oocytes were donated by patients
undergoing ICSI as part of their assisted re-
production treatment at Shanghai Ji Ai Genetics
and IVF Institute affiliated with the Obstetrics
and Gynecology Hospital of Fudan University,
Wu et al., Science 378, eabq7361 (2022)
18 November 2022
8 of 12
RES EARCH | R E S E A R C H A R T I C L E
Center for Reproductive Medicine and Fertility
Preservation program affiliated with International
Peace Maternity and Child Health Hospital of
Shanghai Jiao Tong University. Only imma-
ture oocytes that were unable to be used for
assisted reproduction treatment were col-
lected for this research, and the use of these
human GV oocytes was clearly explained to
the patient donors. All these female donors
were receiving ICSI treatment because of male
factor-induced infertility. The age range of the
donors was 25 to 38 and their mean age was
34.3. The BMI range of these donors was 20 to
22. The collected human oocytes were mixed
thoroughly and distributed randomly to the
control and experimental groups. Only mor-
phologically normal human GV oocytes were
used in this investigation. G-MOPS medium
(Vitrolife) with milrinone (2 mM, HY-14252,
MedChemExpress) was used to maintain hu-
man oocyte prophase arrest. The human GV
oocytes were matured in Multipurpose Hand-
ling Medium-Complete (MHM-C) (FUJIFILM
Irvine Scientific) or G-MOPS medium at 37°C
on a heating block.
Animals and oocyte culture
The porcine ovaries were collected from local
slaughterhouses and transported in warm
0.9% NaCl. Porcine GV oocytes were collected
in TCM199 medium at 39°C, and milrinone
(2 mM) was added to the medium to main-
tain oocyte prophase arrest. Only fully grown
oocytes were used in our experiments. For
maturation, GV oocytes were washed free from
milrinone and cultured in fresh TCM199
medium at 39°C and 5% CO2. Porcine GV
and MI oocytes were fixed for immunofluo-
rescence at 6 or 15 hours after oocyte isola-
tion, respectively.
Female C57Bl/6 mice (3- to 4-week-old) were
purchased from Beijing Vital River Laboratory
Animal Technology Co., Ltd. (Beijing, China).
All mice were used in accordance with insti-
tutional guidelines, and the experiments were
approved by the Animal Care and Use Com-
mittee of Fudan University, China. Mouse GV
oocytes were released from the ovaries at 44
to 52 hours after injection of 10 IU pregnant
mare serum gonadotropin. For maturation,
denuded GV oocytes were cultured in fresh
M2 medium at 37°C on a heat block. The pH
value of culture medium was between 7.2 and
7.4 (S210, Mettler Toledo). Mouse GV and MI
oocytes were fixed for immunofluorescence at
0.5 or 6 hours after in vitro maturation initia-
tion, respectively.
Inhibitor treatment
For microtubule depolymerization, nocodazole
(10 mM, HY-13520, MedChemExpress) was
added to human oocytes at ∼10 hours after
NEBD and 1 hour before time-lapse imaging.
The drug was dissolved in dimethyl sulfoxide
(Sigma-Aldrich) and used at a concentration
of 0.1% in G-MOPS.
Immunofluorescence microscopy
Human GV oocytes were fixed for immunofluo-
rescence at 2, 4, or 6 hours after milrinone
washout. Human MI oocytes were fixed at 5 to
12 hours after NEBD (prometaphase or meta-
phase). Human oocytes were fixed for 30 min
in phosphate-buffered saline (PBS) containing
2% formaldehyde and 0.1% Triton-X at 37°C
on a heat block and were then permeabilized
in PBS containing 0.5% Triton-X (PBT) at 4°C
overnight. Oocytes were extensively washed
with PBS between stages and then blocked
at room temperature in a blocking buffer of
3% bovine serum albumin (BSA) in 0.3% PBT.
All antibody incubations were performed in
blocking buffer at 4°C overnight for primary
antibodies and at 37°C for 1 hour for secondary
antibodies. Primary antibodies were anti-
centromere antibody (HCT-0100; Immuno-
Vision; 1:500), anti-AKAP450 antibody (611518;
BD Biosciences; 1:50), anti-ASPM antibody
(26223-1-AP; Proteintech; 1:50), anti-AURKA
antibody (NBP2-50041; Novus Biological; 1:50),
anti-AURKA antibody (10297-1-AP; Protein-
tech; 1:20), anti-AURKB antibody (ab45145;
Abcam; 1:50), anti-BBS4 antibody (12766-1-AP;
Proteintech; 1:50), anti-beta tubulin antibody
(ab204686; Abcam; 1:50), anti-beta tubulin anti-
body (ab11309; Abcam; 1:100), anti-BUGZ anti-
body (A20177; ABclonal; 1:20), anti-CAMSAP2
antibody (17880-1-AP; Proteintech; 1:50), anti-
CCP110 antibody (12780-1-AP; Proteintech; 1:50),
anti-CCP110 antibody (PA5-58775; Thermo Fisher
Scientific; 1:100), anti-CDK5RAP2 antibody
(06-1398; Merck; 1:100), anti-CENPJ antibody
(11517-1-AP; Proteintech; 1:50), anti-CEP57 anti-
body (24957-1-AP; Proteintech; 1:50), anti-CEP63
antibody (16268-1-AP; Proteintech; 1:50), anti-
CEP72 antibody (19928-1-AP; Proteintech; 1:50),
anti-CEP120 antibody (PA5-55985; Thermo Fisher
Scientific; 1:100), anti-CEP135 antibody (24428-
1-AP; Proteintech; 1:50), anti-CEP152 antibody
(21815-1-AP; Proteintech; 1:50), anti-CEP164
antibody (22227-1-AP; Proteintech; 1:50), anti-
CEP170 antibody (27325-1-AP; Proteintech;
1:50), anti-CEP192 antibody (18832-1-AP;
Proteintech; 1:50), anti-CEP250 antibody (14498-
1-AP; Proteintech; 1:50), anti-CEP290 antibody
(22490-1-AP; Proteintech; 1:50), anti-CETN2 anti-
body (A5397; ABclonal; 1:20), anti-CETN3
antibody (A8111; ABclonal; 1:50), anti-CKAP5
antibody (PA5-59150; Thermo Fisher Scienti-
fic; 1:100), anti-CKAP5 antibody (CL488-67631;
Proteintech; 1:50), anti-CLIP1 antibody (23839-
1-AP; Proteintech; 1:50), anti-CLTC antibody
(610500; BD Biosciences; 1:50), anti-CNTROB
antibody (26880-1-AP; Proteintech; 1:50), anti-
DCTN1 antibody (55182-1-AP; Proteintech; 1:50),
anti-DCTN2 antibody (A2200; ABclonal; 1:50),
anti-DYNC1H1 antibody (12345-1-AP; Proteintech;
1:50), anti-DISC1 antibody (A4678; ABclonal;
1:50), anti-DLGAP5 antibody (A13575; ABclonal;
1:50), anti-DYNLT1 antibody (11954-1-AP; Pro-
teintech; 1:50), anti-GTSE1 antibody (A302-
425A; Bethyl Laboratories; 1:50), anti-HAUS4
antibody (20104-1-AP; Proteintech; 1:50), anti-
HAUS6 antibody (A4797; ABclonal; 1:50), anti-
HAUS8 antibody (PA5-21331; Thermo Fisher
Scientific; 1:100), anti-HMMR antibody (15820-
1-AP; Proteintech; 1:50), anti-HOOK2 antibody
(ab133691; Abcam; 1:50), anti-HOOK3 anti-
body (A15536; ABclonal; 1:50), anti-HOOK3
antibody (15457-1-AP; Proteintech; 1:50), anti-
KANSL3 antibody (HPA035018; Merck; 1:100),
anti-KIF11 antibody (HPA010568; Merck; 1:100),
anti-KIF20A antibody (15911-1-AP; Proteintech;
1:50), anti-KIF20A antibody (CL594-67190; Pro-
teintech; 1:50), anti-KIF22 antibody (A19881;
ABclonal; 1:50), anti-KIF2A antibody (13105-
1-AP; Proteintech; 1:50), anti-KIF2B antibody
(A6480; ABclonal; 1:20), anti-KIFC1 antibody
(A3304; ABclonal; 1:20), anti-KIZ antibody (21177-
1-AP; Proteintech; 1:50), anti-LIS1 antibody
(H00005048-M03; Abnova; 1:50), anti-LMNB
antibody (A1910; ABclonal; 1:20), anti-LRRC45
antibody (PA5-54777; Thermo Fisher Scientif-
ic; 1:100), anti-MCRS1 antibody (HPA039057;
Merck; 1:100), anti-MYO10 antibody (sc-23137;
SantaCruz Biotechnology; 1:50), anti-NDE1 anti-
body (10233-1-AP; Proteintech; 1:50), anti-NDEL1
antibody (H00081565-D01P; Abnova; 1:50), anti-
NEDD1 antibody (13993-1-AP; Proteintech; 1:50),
anti-NEK2 antibody (14233-1-AP; Proteintech;
1:50), anti-NIN antibody (A8215; ABclonal; 1:50),
anti-NUMA1 antibody (A0527; ABclonal; 1:50),
anti-NUP107 antibody (A13110; ABclonal;
1:50), anti-NUP160 antibody (PRS4707; Merck;
1:100), anti-NUP62 antibody (13916-1-AP; Pro-
teintech; 1:50), anti- NUP85 antibody (19370-1-
AP; Proteintech; 1:50), anti-NUSAP1 antibody
(12024-1-AP; Proteintech; 1:50), anti-ODF2 anti-
body (12058-1-AP; Proteintech; 1:50), anti-PCNT
antibody (611815; BD Biosciences; 1:200), anti-
PCM1 antibody (19856-1-AP; Proteintech; 1:50),
anti-PLK1 antibody (A2548; ABclonal; 1:50),
anti-PLK3 antibody (10977-1-AP; Proteintech;
1:50), anti-PLK4 antibody (12952-1-AP; Pro-
teintech; 1:50), anti-PRC1 antibody (15617-1-
AP; Proteintech; 1:50), anti-RAE1 antibody
(20491-1-AP; Proteintech; 1:50), anti-RAN anti-
body (10469-1-AP; Proteintech; 1:50), anti-
SASS6 antibody (21377-1-AP; Proteintech; 1:50),
anti-SKAP2 antibody (12926-1-AP; Proteintech;
1:50), anti-SNF2H antibody (A2000; ABclonal;
1:50), anti-SPDL1 antibody (PA5-99285; Thermo
Fisher Scientific; 1:100), anti-SSX2IP antibody
(13694-1-AP; Proteintech; 1:50), anti-TACC3
antibody (A18641; ABclonal; 1:50), anti-TACC3
antibody (ab134154; Abcam; 1:100), anti-TOP2A
antibody (20233-1-AP; Proteintech; 1:50), anti-
TPX2 antibody (A18327; ABclonal; 1:20), anti-
TUBG1 antibody (A9657; ABclonal; 1:20).
Secondary antibodies were Alexa Fluor 647-
conjugated anti-rabbit IgG (4414S; Cell Sig-
naling; 1:200), Alexa Fluor 647-conjugated
Wu et al., Science 378, eabq7361 (2022)
18 November 2022
9 of 12
RES EARCH | R E S E A R C H A R T I C L E
anti-mouse IgG (4410S; Cell Signaling; 1:200),
Cy3-conjugated anti-rabbit IgG (AS008; ABclonal;
1:500), and Atto 488-conjugated anti-human
IgG (52526; Merck; 1:500). Chromatin was
briefly counterstained with Hoechst 33342
(20 mg·ml−1, HY-15559, MedChemExpres) be-
fore imaging. The samples were in PBS and
imaged with an LSM 880 confocal laser scan-
ning microscope (ZEISS) with a 63×/1.4 NA
Plan Apochromat oil immersion lens at room
temperature.
RNA interference
The siRNAs were provided by GenePharma or
Tsingke Biotechnology. The sequences of siRNA
for TACC3 down-regulation were 5′-GGU UCG
AAG AGG UUG UGU A-3′ and 5′-GCA UGC
ACG GUG CAA AUG A-3′. The siRNA sequen-
ces against CKAP5 were 5′-GGA AAT AGC TGT
TCA CAT A-3′ and 5′-GGC CAA AGC TCC AGG
ATT A-3′. The siRNA sequences targeting CCP110
were 5′-CAC UCU ACU GCA GCA AAG C-3′ and
5′-AUG UUC UUC UCC AAG GUG C-3′. The
siRNA sequence targeting DISC1 was 5′-GGA
UUU GAG AAU AGU UUC A-3′. The negative
control siRNAs were provided by the same com-
pany. To increase the efficiency of RNA inhibi-
tion, mixed siRNAs were microinjected into
human GV oocytes. The final concentration of
siRNAs was 40 mM.
Microinjection of human oocytes
Human GV oocytes were microinjected in
G-MOPS with milrinone (2 mM) on the stage of
an inverted microscope (Leica) with microma-
nipulators (Eppendorf). A 0.1 to 0.3% volume of
mRNA was injected using a timed pulse, and
the final concentration of mRNA was 1 mg/ml.
The injected GV oocytes were arrested in pro-
phase for 2–4 hours for mRNA expression.
mRNA synthesis
mRNA was transcribed in vitro from purified
linear double-stranded DNA templates.
mMessage T7 or T3 RNA polymerase kits
(New England Biolabs) were used for the
in vitro transcription reaction. The constructs
of mClover3-CENPB, mScarlet-CENPB, mClover3-
TACC3, mScarlet-TACC3, mClover3-CKAP5,
mClover3-CCP110, and mClover3-DISC1 were
made and used for mRNA production.
Live cell imaging
For high-resolution time-lapse imaging, time
points were acquired at 10-min intervals using
an LSM 880 confocal laser scanning micro-
scope (ZEISS) fitted with sensitive detectors,
an environmental chamber set to 37°C, and a
long-distance 40×/1.1 NA C-Apochromat water
immersion lens. A volume of 30 mm by 30 mm
by 15 mm centered around the chromosomes
was typically imaged. The chromosomes were
tracked automatically by MyPiC on LSM880.
Microtubules in live human oocytes were
visualized by SiR-tubulin (1 mM) staining in
G-MOPS. To reduce background noise, some
images were passed through a Gaussian filter
of 2 sigma in Fiji (NIH).
immunoglobulin G (IgG) (1:5000 dilution;
Abmart) or goat anti-mouse IgG (1:5000 di-
lution; Abmart) conjugated to horseradish
peroxidase.
Fluorescent intensity measurement
Clinical samples
The pattern of microtubule distribution was
defined by the intensity of tubulin around the
nuclear envelope. The direction of intensity
measurements is shown in the figures. The
intensity was measured by Image J (NIH). The
integrated intensity of kinetochore foci was
measured with the Foci_Picker3D plugin in
Image J. The same threshold was applied to
each focus within an oocyte.
Laser ablation
To examine the role of huoMTOC in microtu-
bule polymerization of human oocytes, the
huoMTOC was directly disrupted by laser in
live human GV oocytes. The human GV oo-
cytes expressing mClover3-TACC3 were ro-
tated with an unbroken microinjection pipette
to obtain huoMTOC. The square regions of
interest were marked and photobleached using
a 488-nm laser line at the maximum power.
The laser ablation was performed at 37°C.
Cell culture and transfection
HEK293T cells were obtained from the Cell Bank
of Shanghai Institute for Biological Sciences,
the Chinese Academy of Sciences (Shanghai,
China). Cells were cultured in Dulbecco’s mod-
ified Eagle’s medium supplemented with 10%
fetal bovine serum and 1% penicillin-streptomycin
(Gibco, Waltham, MA, USA) in an atmosphere
of 5% CO2 at 37°C to between 70 and 80% con-
fluence. Plasmids were transfected into HEK293T
cells using the PolyJet In Vitro DNA Transfec-
tion Reagent (SignaGen) according to the man-
ufacturer’s instructions.
Immunoblots and immunoprecipitation
HEK293T cells were harvested after transfec-
tion for 36 hours and washed with PBS. Cells
were lysed in radioimmunoprecipitation assay
lysis buffer (Shanghai Wei AO Biological Tech-
nology, Shanghai, China) with 1% protease in-
hibitor cocktail (Bimake, Houston, TX, USA).
After quantification with the bicinchoninic acid
assay (Shanghai Biocolor BioScience & Tech-
nology Co.), the supernatant was subjected to
immunoprecipitation with affinity beads (Sigma).
After incubation at 4°C for 4 hours, the beads
were washed with lysis buffer four times.
The bead-bound proteins were eluted using
sodium dodecyl sulfate (SDS) sample buffer,
resolved by SDS–polyacrylamide gel electro-
phoresis (SDS-PAGE), transferred to nitrocellu-
lose membranes (Pall Corporation), and probed
with rabbit anti-FLAG (1:3000 dilution; Cell
Signaling Technology) or mouse anti-vinculin
(1:5000 dilution; Sigma-Aldrich) antibodies.
The secondary antibodies were goat anti-rabbit
A cohort of 1394 infertile female patients with
oocyte maturation arrest recruited from the
Ninth Hospital affiliated with Shanghai Jiao
Tong University and Shanghai Ji Ai Genetics
and IVF Institute affiliated with the Obstetrics
and Gynecology Hospital of Fudan University
participated in this study. Written informed
consent was provided by patients. The recruit-
ment of patients was performed as follows: (i)
female patients were younger than 45 years
old, failing to conceive after 1 year (or longer)
of regular unprotected sex; (ii) had undergone
≥2 failed attempts of IVF/ICSI, characterized
by oocyte maturation arrest; (iii) female patients
with other known causes of infertility, includ-
ing male factors, chromosome anomalies, radio-
therapy, or chemotherapy, were excluded.
Peripheral blood samples were taken for DNA
extraction.
The GV and MI oocytes from two patients
with compound heterozygous mutations in
TACC3 were obtained as part of their assisted
reproduction treatment at the Shanghai Ninth
Hospital affiliated to Shanghai Jiao Tong
University.
This study was approved by the Ethics Com-
mittee of the Medical College of Fudan Univ-
ersity and the Reproductive Study Ethics
Committees of the hospitals.
Genetic studies
Genomic DNA was extracted from peripheral
blood using the QIAamp DNA Blood Mini Kit
(Qiagen). Whole-exome capture was performed
using the SeqCap EZ Exome Kit (Roche), and
sequencing was performed on the Illumina
NovaSeq 6000 platform (Illumina). Sequenc-
ing analysis was compared with the human
reference sequence (NCBI Genome build
GRCh37). Mutations were annotated with
GRCh37 and the dbSNP (version 138) and
gnomAD along with our in-house exome data-
base (data deposited in public database, acces-
sion numbers GVM000402 and GVM000394).
Statistical analysis
The investigators were not blinded during ex-
periments and outcome assessment. The ex-
periments were not randomized. The statistical
methods were not used to determine sample
size. Sample means were compared with either
Student’s t test or one-way analysis of variance
(ANOVA) with a post-hoc test as stated (two-
sided). Dichotomous data were compared
using Fisher’s exact test (two-tailed). All data
are from at least two independent experiments.
All tests were performed using GraphPad
Prism 7 (GraphPad Software).
Wu et al., Science 378, eabq7361 (2022)
18 November 2022
10 of 12
RES EARCH | R E S E A R C H A R T I C L E
RE FE RENCES AND N OT ES
1. S. L. Prosser, L. Pelletier, Mitotic spindle assembly in
animal cells: A fine balancing act. Nat. Rev. Mol. Cell Biol.
18, 187–201 (2017). doi: 10.1038/nrm.2016.162;
pmid: 28174430
2. B. Mogessie, K. Scheffler, M. Schuh, Assembly and positioning
of the oocyte meiotic spindle. Annu. Rev. Cell Dev. Biol. 34,
381–403 (2018). doi: 10.1146/annurev-cellbio-100616-060553;
pmid: 30028643
3. S. Doxsey, Re-evaluating centrosome function. Nat. Rev.
Mol. Cell Biol. 2, 688–698 (2001). doi: 10.1038/35089575;
pmid: 11533726
4. S. Petry, Mechanisms of mitotic spindle assembly. Annu. Rev.
Biochem. 85, 659–683 (2016). doi: 10.1146/annurev-biochem-
060815-014528; pmid: 27145846
5. A. Vertii, H. Hehnly, S. Doxsey, The centrosome, a multitalented
renaissance organelle. Cold Spring Harb. Perspect. Biol. 8,
a025049 (2016). doi: 10.1101/cshperspect.a025049;
pmid: 27908937
J. L. Badano, T. M. Teslovich, N. Katsanis, The centrosome in
human genetic disease. Nat. Rev. Genet. 6, 194–205 (2005).
doi: 10.1038/nrg1557; pmid: 15738963
6.
7. D. Szollosi, P. Calarco, R. P. Donahue, Absence of centrioles in
the first and second meiotic spindles of mouse oocytes.
J. Cell Sci. 11, 521–541 (1972). doi: 10.1242/jcs.11.2.521;
pmid: 5076360
8. A. T. Hertig, E. C. Adams, Studies on the human oocyte and
its follicle. I. Ultrastructural and histochemical observations on
the primordial follicle stage. J. Cell Biol. 34, 647–675 (1967).
doi: 10.1083/jcb.34.2.647; pmid: 4292010
9. T. Mikeladze-Dvali et al., Analysis of centriole elimination
during C. elegans oogenesis. Development 139,
1670–1679 (2012). doi: 10.1242/dev.075440;
pmid: 22492357
10. A. Pimenta-Marques et al., A mechanism for the elimination of
the female gamete centrosome in Drosophila melanogaster.
Science 353, aaf4866 (2016). doi: 10.1126/science.aaf4866;
pmid: 27229142
11. G. Manandhar, H. Schatten, P. Sutovsky, Centrosome reduction
during gametogenesis and its significance. Biol. Reprod. 72,
2–13 (2005). doi: 10.1095/biolreprod.104.031245;
pmid: 15385423
12. D. L. Gard, Microtubule organization during maturation of
Xenopus oocytes: Assembly and rotation of the meiotic
spindles. Dev. Biol. 151, 516–530 (1992). doi: 10.1016/0012-
1606(92)90190-R; pmid: 1601183
13. M. Łuksza, I. Queguigner, M. H. Verlhac, S. Brunet, Rebuilding
MTOCs upon centriole loss during mouse oogenesis. Dev. Biol.
382, 48–56 (2013). doi: 10.1016/j.ydbio.2013.07.029;
pmid: 23954884
21. C. So et al., A liquid-like spindle domain promotes
acentrosomal spindle assembly in mammalian oocytes. Science
364, eaat9557 (2019). doi: 10.1126/science.aat9557;
pmid: 31249032
22. Z. Holubcová, M. Blayney, K. Elder, M. Schuh, Error-prone
chromosome-mediated spindle assembly favors
chromosome segregation defects in human oocytes.
Science 348, 1143–1147 (2015). doi: 10.1126/science.
aaa9529; pmid: 26045437
23. C. So et al., Mechanism of spindle pole organization and
instability in human oocytes. Science 375, eabj3944 (2022).
doi: 10.1126/science.abj3944; pmid: 35143306
24. T. Wu, S. I. R. Lane, S. L. Morgan, F. Tang, K. T. Jones,
Loss of centromeric RNA activates the spindle assembly
checkpoint in mammalian female meiosis I. J. Cell Biol.
220, e202011153 (2021). doi: 10.1083/jcb.202011153;
pmid: 34379093
25. J. R. A. Hutchins et al., Systematic analysis of human protein
complexes identifies chromosome segregation proteins.
Science 328, 593–599 (2010). doi: 10.1126/science.1181348;
pmid: 20360068
26. H. V. Goodson, E. M. Jonasson, Microtubules and microtubule-
associated proteins. Cold Spring Harb. Perspect. Biol. 10,
a022608 (2018). doi: 10.1101/cshperspect.a022608;
pmid: 29858272
27. Z. Chen, V. B. Indjeian, M. McManus, L. Wang, B. D. Dynlacht,
CP110, a cell cycle-dependent CDK substrate, regulates
centrosome duplication in human cells. Dev. Cell 3, 339–350
(2002). doi: 10.1016/S1534-5807(02)00258-7;
pmid: 12361598
28. F. Gergely et al., The TACC domain identifies a family of
centrosomal proteins that can interact with microtubules.
Proc. Natl. Acad. Sci. U.S.A. 97, 14352–14357 (2000).
doi: 10.1073/pnas.97.26.14352; pmid: 11121038
29. F. Gergely, V. M. Draviam, J. W. Raff, The ch-TOG/XMAP215
protein is essential for spindle pole organization in human
somatic
cells. Genes Dev. 17, 336–341 (2003). doi: 10.1101/gad.245603;
pmid: 12569123
30. C. H. Lin, C. K. Hu, H. M. Shih, Clathrin heavy chain mediates
TACC3 targeting to mitotic spindles to ensure spindle stability.
J. Cell Biol. 189, 1097–1105 (2010). doi: 10.1083/
jcb.200911120; pmid: 20566684
31. S. Petry, R. D. Vale, Microtubule nucleation at the centrosome
and beyond. Nat. Cell Biol. 17, 1089–1093 (2015). doi: 10.1038/
ncb3220; pmid: 26316453
32. A. Thawani, R. S. Kadzik, S. Petry, XMAP215 is a microtubule
nucleation factor that functions synergistically with
the g-tubulin ring complex. Nat. Cell Biol. 20, 575–585
(2018). doi: 10.1038/s41556-018-0091-6; pmid: 29695792
14. M. J. Carabatsos, C. M. H. Combelles, S. M. Messinger,
33. G. J. Brouhard et al., XMAP215 is a processive microtubule
D. F. Albertini, Sorting and reorganization of centrosomes
during oocyte maturation in the mouse. Microsc. Res.
Tech. 49, 435–444 (2000). doi: 10.1002/(SICI)1097-
0029(20000601)49:5<435::AID-JEMT5>3.0.CO;2-H;
pmid: 10842370
15. M. Schuh, J. Ellenberg, Self-organization of MTOCs replaces
centrosome function during acentrosomal spindle assembly in
live mouse oocytes. Cell 130, 484–498 (2007). doi: 10.1016/
j.cell.2007.06.025; pmid: 17693257
16. S. J. Radford, A. L. Nguyen, K. Schindler, K. S. McKim, The
17.
chromosomal basis of meiotic acentrosomal spindle
assembly and function in oocytes. Chromosoma 126,
351–364 (2017). doi: 10.1007/s00412-016-0618-1;
pmid: 27837282
I. D. Wolff, M. V. Tran, T. J. Mullen, A. M. Villeneuve,
S. M. Wignall, Assembly of Caenorhabditis elegans
acentrosomal spindles occurs without evident microtubule-
organizing centers and requires microtubule sorting by
KLP-18/kinesin-12 and MESP-1. Mol. Biol. Cell 27,
3122–3131 (2016). doi: 10.1091/mbc.e16-05-0291;
pmid: 27559133
18. D. Clift, M. Schuh, A three-step MTOC fragmentation
mechanism facilitates bipolar spindle assembly in mouse
oocytes. Nat. Commun. 6, 7217 (2015). doi: 10.1038/
ncomms8217; pmid: 26147444
19. T. Wu, S. I. R. Lane, S. L. Morgan, K. T. Jones, Spindle tubulin
and MTOC asymmetries may explain meiotic drive in oocytes.
Nat. Commun. 9, 2952 (2018). doi: 10.1038/s41467-018-
05338-7; pmid: 30054463
20. C. Gueth-Hallonet et al., g-Tubulin is present in acentriolar
MTOCs during early mouse development. J. Cell Sci. 105,
157–166 (1993). doi: 10.1242/jcs.105.1.157; pmid: 8360270
polymerase. Cell 132, 79–88 (2008). doi: 10.1016/
j.cell.2007.11.043; pmid: 18191222
34. W. Fu et al., Self-assembly and sorting of acentrosomal
microtubules by TACC3 facilitate kinetochore capture during
the mitotic spindle assembly. Proc. Natl. Acad. Sci. U.S.A.
110, 15295–15300 (2013). doi: 10.1073/pnas.1312382110;
pmid: 24003142
35. J. Cesario, K. S. McKim, RanGTP is required for meiotic spindle
organization and the initiation of embryonic development in
Drosophila. J. Cell Sci. 124, 3797–3810 (2011). doi: 10.1242/
jcs.084855; pmid: 22100918
36. J. Dumont et al., A centriole- and RanGTP-independent spindle
assembly pathway in meiosis I of vertebrate oocytes.
J. Cell Biol. 176, 295–305 (2007). doi: 10.1083/jcb.200605199;
pmid: 17261848
37. C. So, S. Cheng, M. Schuh, Phase separation during
germline development. Trends Cell Biol. 31,
254–268 (2021). doi: 10.1016/j.tcb.2020.12.004;
pmid: 33455855
38. A. Webster, M. Schuh, Mechanisms of aneuploidy in human
eggs. Trends Cell Biol. 27, 55–68 (2017). doi: 10.1016/
j.tcb.2016.09.002; pmid: 27773484
39. S. Doubilet, K. S. McKim, Spindle assembly in the oocytes of
mouse and Drosophila—similar solutions to a problem.
Chromosome Res. 15, 681–696 (2007). doi: 10.1007/s10577-
007-1148-8; pmid: 17674154
40. I. Peset et al., Function and regulation of Maskin, a TACC
family protein, in microtubule growth during mitosis. J. Cell
Biol. 170, 1057–1066 (2005). doi: 10.1083/jcb.200504037;
pmid: 16172207
I. Peset, I. Vernos, The TACC proteins: TACC-ling microtubule
dynamics and centrosome function. Trends Cell Biol. 18,
41.
379–388 (2008). doi: 10.1016/j.tcb.2008.06.005;
pmid: 18656360
42. D. G. Booth, F. E. Hood, I. A. Prior, S. J. Royle, A TACC3/ch-TOG/
clathrin complex stabilises kinetochore fibres by inter-microtubule
bridging. EMBO J. 30, 906–919 (2011). doi: 10.1038/
emboj.2011.15; pmid: 21297582
43. R. Yao et al., A small compound targeting TACC3 revealed its
different spatiotemporal contributions for spindle assembly in
cancer cells. Oncogene 33, 4242–4252 (2014). doi: 10.1038/
onc.2013.382; pmid: 24077290
44. P. Singh, G. E. Thomas, K. K. Gireesh, T. K. Manna, TACC3
protein regulates microtubule nucleation by affecting g-tubulin
ring complexes. J. Biol. Chem. 289, 31719–31735 (2014).
doi: 10.1074/jbc.M114.575100; pmid: 25246530
45. T. V. Suhail, P. Singh, T. K. Manna, Suppression of centrosome
protein TACC3 induces G1 arrest and cell death through
activation of p38-p53-p21 stress signaling pathway. Eur. J.
Cell Biol. 94, 90–100 (2015). doi: 10.1016/j.ejcb.2014.12.001;
pmid: 25613365
46. J. Roostalu, N. I. Cade, T. Surrey, Complementary activities of
TPX2 and chTOG constitute an efficient importin-regulated
microtubule nucleation module. Nat. Cell Biol. 17, 1422–1434
(2015). doi: 10.1038/ncb3241; pmid: 26414402
47. P. R. Clarke, C. Zhang, Spatial and temporal coordination of
mitosis by Ran GTPase. Nat. Rev. Mol. Cell Biol. 9, 464–477
(2008). doi: 10.1038/nrm2410; pmid: 18478030
48. O. J. Gruss et al., Ran induces spindle assembly by reversing
the inhibitory effect of importin a on TPX2 activity. Cell
104, 83–93 (2001). doi: 10.1016/S0092-8674(01)00193-3;
pmid: 11163242
49. P. Kalab, R. Heald, The RanGTP gradient – a GPS for the
mitotic spindle. J. Cell Sci. 121, 1577–1586 (2008).
doi: 10.1242/jcs.005959; pmid: 18469014
50. R. Feng et al., Mutations in TUBB8 and human oocyte meiotic
arrest. N. Engl. J. Med. 374, 223–232 (2016). doi: 10.1056/
NEJMoa1510791; pmid: 26789871
AC KNOWLED GME NTS
We thank all the patients and volunteers who participated in
this study. We thank J. Dai from the Shanghai Academy of
Agricultural Sciences for providing the porcine ovaries. We
thank C. Zhang from Peking University for the gift of the human
TACC3 construct. We thank K. Qiao from the Cell Biological
Imaging Core Facility of IMIB at Fudan University for assistance
with imaging. We are grateful to X. Li (ZEISS) for technical
support with confocal live-oocyte time-lapse imaging. Funding:
This work was supported by the National Key Research and
Development Program of China (2021YFC2700100), the Basic
Science Center Program of the National Natural Science
Foundation of China (82288102), the National Natural Science
Foundation of China (81725006, 32130029, 82101737,
82171643, 81971450, and 81971382), the Shanghai Municipal
Science and Technology Major Project (2017SHZDZX01), the
Project of the Shanghai Municipal Science and Technology
Commission (21XD1420300), the Capacity Building Planning
Program for the Shanghai Women and Children’s Health Service,
and the collaborative innovation center project construction
for Shanghai Women and Children’s Health. Ethics statement:
The studies involving human participants were reviewed and
approved by the ethics committee of the Shanghai Ji Ai Genetics
and IVF Institute (JIAI E2020-23 and JIAI E2022-05), the
ethics committee of the International Peace Maternity and Child
Health Hospital (GKLW 2021-18), and the ethics committee of
Fudan University (FE21198 and 2020-012). The patients and
participants provided their written informed consent to
participate in this study. The animal study was reviewed and
approved by the animal welfare and ethics group of the
Department of Experimental Animal Science of Fudan University
(2017 1350 A265). Author contributions: T.W., J.D., J.F., and
Y.K. contributed equally to this work. L.W., Q.S., and X.S.
conceived of and designed the research study. T.W. designed
the experiments and the methods for data analysis. T.W., J.D.,
Y.K., and H.G. performed the experiments. Y.L. and R.G. helped
with the molecular work and genome sequencing. J.F., M.Z.,
and W.L. helped with human oocyte sample collection and
manipulation. X.D. helped with human oocyte in vitro
maturation. B.C. organized the medical records and analyzed
the whole-exome data. L.W., Q.S., and T.W. drafted and revised
the manuscript. Competing interests: The authors declare no
competing interests. Data and materials availability: All data
in the paper are present in the paper or the supplementary
materials. In addition, the variation data reported in this paper
Wu et al., Science 378, eabq7361 (2022)
18 November 2022
11 of 12
RES EARCH | R E S E A R C H A R T I C L E
have been deposited in the Genome Variation Map of the
National Genomics Data Center, China National Center for
Bioinformation/Beijing Institute of Genomics, Chinese Academy
of Sciences, under accession numbers GVM000402 (patients)
and GVM000394 (controls). License information: Copyright ©
2022 the authors, some rights reserved; exclusive licensee
American Association for the Advancement of Science. No claim
to original US government works. https://www.science.org/
about/science-licenses-journal-article-reuse
MDAR Reproducibility Checklist
Movies S1 to S5
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abq7361
Figs. S1 to S11
Tables S1 and S2
View/request a protocol for this paper from Bio-protocol.
Submitted 27 April 2022; resubmitted 24 August 2022
Accepted 14 October 2022
10.1126/science.abq7361
Wu et al., Science 378, eabq7361 (2022)
18 November 2022
12 of 12
|
10.1126_science.abq5209
|
RES EARCH
R E S E A R C H A R T I C L E S U M M A R Y
◥
CELL BIOLOGY
Endosomal lipid signaling reshapes the endoplasmic
reticulum to control mitochondrial function
Wonyul Jang, Dmytro Puchkov, Paula Samsó, YongTian Liang, Michal Nadler-Holly, Stephan J. Sigrist,
Ulrich Kintscher, Fan Liu, Kamel Mamchaoui, Vincent Mouly, Volker Haucke*
INTRODUCTION: Cells need to react appropri-
ately to nutritional cues. Defects in the rewiring
of metabolism in response to alterations in
nutrient supply have been linked to human
diseases ranging from diabetes to muscle
atrophy. Starvation represses anabolic path-
ways and facilitates catabolic ones, such as the
degradation of macromolecules by autophagy
and endolysosomes. Starvation also promotes
the b-oxidation of fatty acids in mitochondria
to produce adenosine triphosphate (ATP).
Within cells, organelles including lysosomes
and mitochondria undergo changes in shape
and dynamics. These processes are often regu-
lated by phosphoinositide lipids. Phosphoino-
sitides are also involved in the formation of
membrane contacts between organelles and
in the response of cells and tissues to growth
and nutrient signals. How the adaptive changes
that protect mammalian cells and tissues from
starvation-induced damage are coordinated on
a cell-wide scale is unknown.
RATIONALE: Endolysosomal membrane dynam-
ics and function are controlled by phospho-
inositide signaling lipids, most notably by the
synthesis and turnover of phosphatidylinositol
3-phosphate [PI(3)P]. Patients carrying muta-
tions in the gene encoding the lipid phosphatase
MTM1, an enzyme that mediates endosomal
PI(3)P turnover, suffer from X-linked centro-
nuclear myopathy (XLCNM), a severe neuro-
muscular disease characterized by muscle
atrophy, disorganization of mitochondria, and
defects in the organization of the muscle endo-
plasmic reticulum (ER). Given that PI(3)P is
a hallmark of endosomes, we hypothesized
that the control of early endosomal PI(3)P by
MTM1 might serve to orchestrate adaptive
changes in the dynamics of the ER and mito-
chondria in response to altering nutrient
supply.
RESULTS: Working with XLCNM patient–derived
myoblasts and engineered cell lines, we found
that nutrient starvation (for example, lack of
amino acids) induced the hydrolysis of PI(3)P
by endosomal recruitment of MTM1. Concom-
itantly, tubular ER membranes were observed
to be converted into ER sheets by live super-
resolution light microscopy. Mechanistically,
loss of early endosomal PI(3)P upon starvation
was found to reduce membrane contacts be-
tween peripheral ER tubules and early endo-
Role of MTM1-mediated endosomal
PI(3)P signaling in mitochondrial
metabolic rewiring through
reshaping of the ER in response
to starvation. In fed cells, early
endosomes form contacts with ER
tubules. Tubular ER membranes
facilitate mitochondrial fission and
serve as a source for lipid droplet
formation. Nutrient starvation–
induced hydrolysis of endosomal
PI(3)P by MTM1 reduces mem-
brane contacts between the
tubular ER and early endosomes.
The resulting loss of peripheral ER
tubules induces mitochondrial
network formation and the delivery
of fatty acids to mitochondria to
sustain cellular energy supply.
EE, early endosome; MT, microtubule;
FA, fatty acid; LD, lipid droplet.
Tubular
ER
MT
FA
Fed
Starved
PI3P
EE
Sheet
ER
MTM1
Tubular
ER
Mito
fission
Sheet ER
FA
β-oxidation↑
LD from
tubular ER
somes. These contacts function as physical
tethers that may transmit pulling forces from
highly motile peripheral endosomes to the
tubular ER. Using proximity labeling proteomic
and functional cell biological experiments we
demonstrated that the ER–endosome contacts
were mediated by binding of the related ER
membrane proteins RRBP1 and kinectin 1 to
PI(3)P on endosomes. To study the role of
starvation-induced reshaping of tubular ER
membranes into sheets on mitochondrial form
and function, we combined live imaging with
three-dimensional focused ion beam milling
scanning electron microscopy (FIB-SEM) and
proteomic analysis. We found that starvation-
induced ER reshaping by MTM1 reduced the
rate of mitochondrial fission and promoted
the formation of a hyperfused mitochondrial
network. Genetic manipulations that resulted
in ER sheet expansion caused the formation
of an enlarged mitochondrial network even
in fed cells. Conversely, impaired ER reshap-
ing and reduced mitochondrial network for-
mation were observed in starved myoblasts
from XLCNM patients. Mitochondrial net-
work formation appeared to be critical for
the delivery of fatty acids from lipid droplets
to mitochondria and for oxidative ATP pro-
duction to sustain energy supply in nutrient-
deprived cells.
CONCLUSION: Our data unravel a crucial role for
early endosomal lipid signaling in controlling
ER morphology and, thereby, mitochondrial
form and function to orchestrate the adaptive
response of cells to alterations in nutrient (e.g.,
amino acid) supply. This mechanism operates
independent of autophagy, a cellular self-eating
process typically induced by prolonged starva-
tion. Rather, it resembles an organellar con-
veyor belt, in which the tubular ER serves as a
membrane conduit that transmits nutrient-
triggered changes in endosomal PI(3)P levels
to metabolic organelles to enable metabolic
rewiring. How early endosomal PI(3)P levels
and MTM1 function are controlled by cellular
nutrient status is currently unknown. Defects
in ER shape, mitochondrial morphogenesis,
and cellular ATP depletion caused by loss of
MTM1 function can explain the observed myo-
fiber hypotrophy and defective ER organi-
zation in animal models of XLCNM and in
human patients who often appear under-
nourished. We therefore hypothesize that
dysregulated organelle remodeling may un-
derlie XLCNM caused by MTM1 mutations
in humans.▪
The list of author affiliations is available in the full article online.
*Corresponding author. Email: haucke@fmp-berlin.de
Cite this article as W. Jang et al., Science 378, eabq5209
(2022). DOI: 10.1126/science.abq5209
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.abq5209
Jang et al., Science 378, 1188 (2022)
16 December 2022
1 of 1
Corrected 7 September 2023. See full text. RES EARCH
R E S E A R C H A R T I C L E
◥
CELL BIOLOGY
Endosomal lipid signaling reshapes the endoplasmic
reticulum to control mitochondrial function
Wonyul Jang1, Dmytro Puchkov1, Paula Samsó1, YongTian Liang2, Michal Nadler-Holly1,
Stephan J. Sigrist2, Ulrich Kintscher3, Fan Liu1,3, Kamel Mamchaoui4,
Vincent Mouly4, Volker Haucke1,2,3*
Cells respond to fluctuating nutrient supply by adaptive changes in organelle dynamics and in metabolism. How
such changes are orchestrated on a cell-wide scale is unknown. We show that endosomal signaling lipid
turnover by MTM1, a phosphatidylinositol 3-phosphate [PI(3)P] 3-phosphatase mutated in X-linked
centronuclear myopathy in humans, controls mitochondrial morphology and function by reshaping the
endoplasmic reticulum (ER). Starvation-induced endosomal recruitment of MTM1 impairs PI(3)P-dependent
contact formation between tubular ER membranes and early endosomes, resulting in the conversion of
ER tubules into sheets, the inhibition of mitochondrial fission, and sustained oxidative metabolism. Our
results unravel an important role for early endosomal lipid signaling in controlling ER shape and, thereby,
mitochondrial form and function to enable cells to adapt to fluctuating nutrient environments.
T he ability of cells to react appropriately
to nutritional cues is of fundamental im-
portance for cell physiology, and defects
in the cellular response to altered nutri-
ent supply have been linked to human
diseases ranging from diabetes to muscle
atrophy (1–4). Among the early changes elic-
ited by nutrient stress are the suppression of
anabolic programs such as protein translation
(5, 6) and the concomitant induction of cat-
abolic processes involving the proteasome,
autophagy [e.g., lipophagy, reticulophagy (7)],
endolysosomal turnover of proteins (6, 8), and
increased mitochondrial b-oxidation of fatty
acids (9). How these adaptive responses that
protect mammalian cells and tissues from
starvation-induced damage and the induction
of apoptotic cell death are coordinated is un-
known. Endolysosomal membrane dynamics
and function are controlled by the spatiotem-
porally regulated synthesis and turnover of
phosphatidylinositol 3-phosphate [PI(3)P] and
related signaling lipids by phosphatidylinositol
(PI) 3-kinases and 3-phosphatases (10, 11).
MTM1, the founding member of the myotubu-
larin family of PI 3-phosphatases, is crucially
involved in PI(3)P homeostasis on endosomes
(12–15). Patients carrying mutations in the
MTM1 gene suffer from X-linked centronu-
clear myopathy (XLCNM), a severe, often fatal
disease characterized by muscle weakness
due to myofiber atrophy, disorganization of
mitochondria, and structural defects in the
1Leibniz-Forschungsinstitut für Molekulare Pharmakologie
(FMP), 13125 Berlin, Germany. 2Department of Biology,
Chemistry, and Pharmacy, Freie Universität Berlin, 14195
Berlin, Germany. 3Charité-Universitätsmedizin Berlin,
10117 Berlin, Germany. 4Centre de Recherche en Myologie,
Institut de Myologie, Inserm, Sorbonne Université, 75013
Paris, France.
*Corresponding author. Email: haucke@fmp-berlin.de
organization of the sarcoplasmic reticulum, a
specialized form of the endoplasmic reticulum
(ER) found in muscle tissue that is important
for excitation-contraction coupling (16, 17).
How XLCNM-linked mutations in endosomal
MTM1 cause such pleiotropic defects in the
organization of the ER and other organelles
has remained elusive but may relate to a thus
far unexplored function of endosomes in
orchestrating adaptive changes in organelle
dynamics.
Results
Nutrient-regulated reshaping of the ER is
controlled by endosomal MTM1
To address the question how XLCNM-linked
mutations in endosomal MTM1 cause defects
in ER organization, we analyzed ER morphol-
ogy (18) in myoblast cell lines from healthy
controls or XLCNM patients (table S1) (19–21)
suffering from pronounced or complete loss of
MTM1 protein (fig. S1A) and a resulting in-
crease in PI(3)P levels (15). The ER membrane
is composed of interconnected uniform flat
cisternal sheets, fenestrated sheets with nano-
holes (22, 23), and peripheral dynamic narrow
tubules (∼30 to 60 nm in diameter). ER sheets
and peripheral ER tubules were initially char-
acterized by confocal light microscopy (24)
and have recently been resolved at the nano-
scale by super-resolution imaging (22, 25). We
determined the morphology of the peripheral
ER by semiautomated image analysis of cells
expressing mEmerald- or Halo-tagged ver-
sions of the ER membrane protein Sec61b or
cells stained for the ER marker calreticulin
(fig. S1B). This analysis revealed a prominent
accumulation of tubular versus sheet ER in
myoblasts from XLCNM patients compared
with cells from healthy controls (Fig. 1, A and
B). Accumulation of tubular ER in XLCNM
myoblasts was confirmed by time-gated stimu-
lated emission depletion (gSTED) nanoscopy
imaging at 50-nm resolution [consistent with
(22)] (Fig. 1A). Depletion of MTM1 in healthy
controls phenocopied XLCNM myoblasts with
respect to the accumulation of tubular ER and
the concomitant reduction in sheet ER (Fig. 1C
and fig. S1C). These data suggest that the de-
fects in ER morphology observed in XLCNM
patient myoblasts are a consequence of MTM1
loss of function and are consistent with earlier
in vivo data indicating a possible role for MTM1
in the control of ER morphology (17).
A hallmark of MTM1 loss is the accumula-
tion of its substrate lipid PI(3)P (15). We rea-
soned that the observed defects in ER shape in
XLCNM myoblasts are a consequence of ele-
vated PI(3)P levels and, thus, might be rescued
by pharmacological inhibition of PI(3)P syn-
thesis. Consistently, specific inhibition of the
endosomal PI 3-kinase VPS34 in the presence
of VPS34-IN1 or reexpression of active MTM1
sufficed to restore normal ER morphology in
XLCNM patient myoblasts (Fig. 1D and fig. S1,
D and E). An elevated ER sheet-to-tubule ratio
was also observed in VPS34-IN1–treated hu-
man HeLa cells imaged by confocal light mi-
croscopy or gSTED nanoscopy at steady state
(Fig. 1E and fig. S1, F and G). In contrast, de-
pletion of phosphatidylinositol 3,5-bisphosphate
[PI(3,5)P2], another potential MTM1 substrate
lipid (26), did not affect ER shape (fig. S1, F
and H). Previous data show that PI(3)P is
involved in the cellular response to altered
nutrient supply (27–29). We therefore tested
whether PI(3)P metabolism might be regulated
by nutrients. We found endosomal PI(3)P levels
to decline upon cellular nutrient deprivation
[consistent with (29)] (Fig. 1F and fig. S1I),
and this was accompanied by progressive ER
sheet expansion in live or fixed cells imaged
by super-resolution STED or confocal micros-
copy (Fig. 1, G and H; fig. S1, J and K; and
Movies 1 and 2). The levels of major ER-
shaping proteins remained unaltered (fig. S3, J
and K). ER sheet expansion was also observed
in cells treated with the potent catalytic mech-
anistic target of rapamycin (mTOR) inhibitor
Torin 1 (fig. S1L), which is often used to mimic
conditions of starvation. Further analysis re-
vealed that deprivation of amino acids (gluta-
mine in particular), rather than growth factors
and glucose, was responsible for ER reshaping
during starvation (Fig. 1I and fig. S1M). In con-
trast, amino acid replenishment failed to rescue
the starvation-induced reduction in cell area
(fig. S1, N and O), suggesting that ER shape and
cell size are controlled by distinct mechanisms.
To probe whether reduced PI(3)P levels
causally underlie the starvation-induced re-
modeling of ER membranes, we depleted cells
of MTM1, a condition under which PI(3)P
accumulates (15), by RNA interference or
Jang et al., Science 378, eabq5209 (2022)
16 December 2022
1 of 16
Corrected 7 September 2023. See full text. RES EARCH | R E S E A R C H A R T I C L E
A
STED
Inset
STED
Inset
Confocal
l
l
E
Inset
Inset
5
1
L
N
O
S
M
D
0
9
1
1
B
A
8
8
2
1
M
K
Confocal
Halo-Sec61β
mEmerald-Sec61β
t
s
a
b
o
y
m
y
h
t
l
a
e
H
t
s
a
b
o
y
m
M
N
C
L
X
Fig. 1. Nutrient-regulated
reshaping of the ER is
controlled by endosomal
MTM1. (A) Confocal and STED
live images of Halo-Sec61b in
fed myoblasts from healthy
and XLCNM patients. (B) Ratio
of sheet/total ER area stained
for calreticulin. AB1190 (n =
206), KM1288 (n = 222), NL15
(n = 102). (C) Ratio of sheet/
total ER area in fed healthy
(AB1190) myoblasts at steady-
state treated with control
(siCo/siCo) or MTM1 siRNA
(siMTM1_1/_SP). siCo/siCo
(n = 181), siMTM1_1/_SP (n =
167). (D) Ratio of sheet/total
ER area in XLCNM patient
myoblasts treated with DMSO
(control) or VPS34-IN1 (5 mM,
2 hours). KM1288, DMSO
(n = 113), VPS34-IN1 (n = 148);
NL15, DMSO (n = 80), VPS34-
IN1 (n = 86). (E) Confocal
and STED images of DMSO or
VPS34-IN1-treated (5 mM,
2 hours) HeLa cells expressing
mEmerald-Sec61b. (F) PI(3)P
levels in fed (n = 820) or
starved (EBSS 2 hours; n =
725) HeLa cells. (G) STED
images of fed or starved live
HeLa cells stably expressing
Halo-Sec61b. See Movies 1
and 2. (H) Ratio of sheet/total
ER area in fed or starved
(EBSS 2 hours) HeLa cells
stained for calreticulin. Fed
(n = 312), Starved (n = 320).
(I) Ratio of sheet/total ER area
of HeLa cells expressing
mEmerald-Sec61b exposed to
different nutrients (2 hours). Full (n = 128), − (i.e., EBSS only, n = 148), +Dialyzed
FBS (n = 160), +Glucose (4.5 g/liter) (n = 111), all amino acids (+ All AA, n = 171).
(J) Ratio of sheet/total ER area in fed or starved HeLa WT or MTM1 KO cells.
Fed WT (n = 136), KO (n = 123); Starved WT (n = 156), KO (n = 207).
(K) (Top left) Electron microscopy images of starved WT or MTM1 KO HeLa cells.
Purple, ER cisternal cross sections. (Right) Length of ER cross sections (# of
objects: WT = 1018, KO = 1070) from three cells. (Bottom left) 3D FIB-SEM
analysis. 3D reconstruction of ER (purple) and mitochondria (brown). ER sheets
S
Full -
A
g/liter)
B
A
F
All
Dialyzed
+
(4.5
Glucose
1
n
I
4
3
S
P
V
EBSS
t
e
e
h
S
t
e
e
h
S
o
i
t
a
r
o
i
t
a
r
****
1.0
0.2
0.8
0.4
0.6
0.0
1.0
0.6
0.8
0.4
0.2
0.0
R
E
R
E
+
+
ns
ns
J
I
****
WTKOWTKO
Fed Starved
B
(steady fed)
****
1.0
****
o
i
t
a
r
R
E
t
e
e
h
S
0.8
0.6
0.4
0.2
0.0
AB1190
NL15
KM1288
Healthy XLCNM
C
1.5
1.0
0.5
o
i
t
a
r
R
E
t
e
e
h
S
****
0.0
siCo./siCo.
siMTM1_1/_SP
AB1190
****
****
D
o
i
t
a
r
R
E
t
e
e
h
S
1.0
0.8
0.6
0.4
0.2
0.0
DMSO
DMSO
VPS34In1
VPS34In1
KM1288 NL15
G
Live-STED
Halo-Sec61β
Inset
F
y
t
i
s
n
e
t
n
i
P
)
3
(
I
P
.
m
r
o
N
**
1.1
1.0
0.9
0.8
0.7
0.6
Fed
Starved
K
Starved-WT
Starved-KO
M
E
S
****
H
o
i
t
a
r
R
E
t
e
e
h
S
1.5
1.0
0.5
0.0
Fed
Starved
)
m
(
h
t
g
n
e
l
R
E
4
1
1.0
0.5
0.0
****
WTKO
Starved
M
E
S
B
F
-
I
are fenestrated in starved KO cells. See fig. S2G and movie S1. Scale bars:
10 mm (white), 2 mm (yellow), and 1 mm (black). n, total number of cells analyzed
from two (H) or three [(B) to (D), (F), (I), and (J)] independent experiments.
One-way ANOVA with Dunnett’s multiple comparisons test (B); two-tailed
Mann-Whitney test [(C), (D), (H), (J), and (K)]; one-sample t test (F);
Kruskal-Wallis test with two-sided Dunn’s multiple comparison test (I).
**P ≤ 0.01, ****P ≤ 0.0001. Data are median ± interquartile range [(B) to (D),
and (H) to (K)] or mean ± SD (F).
CRISPR-Cas9–mediated knockout (KO). Loss
of MTM1 potently antagonized the starvation-
induced conversion of peripheral ER tubules
into sheets (Fig. 1J and fig. S2, A to E), a pheno-
type that was rescued by reexpression of cat-
alytically active mCherry-MTM1 (fig. S2F). Further
ultrastructural analysis by three-dimensional
(3D) focused ion beam scanning electron mi-
croscopy (FIB-SEM) showed that even the peri-
nuclear ER, which appeared as flat uniform
sheets in starved wild-type (WT) cells, was
highly fenestrated in starved cells lacking
MTM1 (Fig. 1K, FIB-SEM; fig. S2G; and movie S1).
The total ER volume fraction was unaltered
(fig. S2H). Accumulation of highly fenestrated
sheet ER and ER tubules in MTM1 KO cells
was further evidenced by the reduced average
length of ER profiles determined by SEM anal-
ysis of 2D cross sections (Fig. 1K, SEM).
We also tested whether other members of
the myotubularin family of MTM1-related phos-
phatases affect ER shape. On the basis of their
mRNA expression levels in HeLa cells, we ana-
lyzed four myotubularin-related proteins (MTMRs)
and found that, of these four, only MTMR1, the
family member most closely related to MTM1,
affected ER shape (fig. S2, I to K). These data
suggest a close functional relationship between
endosomal PI(3)P and ER morphology, in par-
ticular the presence of ER tubules in cells. In-
creased levels of endosomal PI(3)P caused by
MTM1 loss of function prevent the starvation-
induced remodeling of ER membranes, and
Jang et al., Science 378, eabq5209 (2022)
16 December 2022
2 of 16
Corrected 7 September 2023. See full text.
RES EARCH | R E S E A R C H A R T I C L E
Movie 1. Live imaging of ER dynamics in fed or
starved cells. Live-cell spinning disk confocal
imaging of fed or starved HeLa cells expressing
mEmerald-Sec61b. Videos were acquired at
1 frame/5 min for 150 min and correspond to
fig. S1K.
Movie 2. Live STED imaging of tubular or sheet
ER from fed or starved HeLa cells expressing
Halo-Sec61 b. Live-cell STED imaging of starved
HeLa cells expressing Halo-Sec61b. Videos were
acquired at 1 frame/3 s for 30 s and are related
to Fig. 1G.
this may underlie the structural defects in ER
morphology found in muscle cells and tissue
from XLCNM patients (16, 17, 30, 31).
Starvation-induced PI(3)P hydrolysis by
MTM1 at ER–early endosome contacts mediates
ER reshaping
We next sought to understand how PI(3)P hy-
drolysis by MTM1 mediates the starvation-
induced conversion of peripheral ER tubules
into flat uniform sheets. As the ER is subject
to turnover by means of autophagy (32), we
probed whether blockade of autophagosome
formation interferes with starvation-induced
ER sheet expansion. However, ER sheet expan-
sion in starved cells was unaffected by phar-
macological or genetic inhibition of autophagy
in the presence of VPS34-IN1 (15) or knock-
down of the essential autophagy factor ATG5
(fig. S3, A to D). Moreover, loss of MTM1, that
is, a condition in which starvation-induced ER
sheet expansion is perturbed, did not alter the
ability of starved cells to form microtubule-
associated protein 1 light chain 3 (LC3)–positive
autophagosomes (fig. S3, E and F), the over-
all levels of the autophagy marker LC3-II (fig.
S3G), or the subcellular localization and ac-
tivation of transcription factor EB (TFEB), a
master regulator of autophagy gene expres-
sion (33) (fig. S3, L and M). MTM1 KO cells
displayed slightly reduced levels of the sheet-
localized ER-phagy receptor FAM134B under
fed and starved conditions (fig. S3, H and I).
Reduced levels of FAM134B have been shown
to result in sheet ER expansion (32), a pheno-
type opposite to the expansion of the tubular
ER observed in MTM1 KO cells (Fig. 1). The
starvation-induced increase in the sheet ER
ratio thus appears to be independent of FAM134B-
mediated ER-phagy. Although mechanistic
target of rapamycin complex 1 (mTORC1) ac-
tivity was elevated in fed MTM1 KO cells
[consistent with observations in MTM1 KO
mice (34)], loss of MTM1 did not affect sup-
pression of mTORC1 activity in starved cells
(fig. S3N). Finally, lumenal ER calcium levels
were not significantly altered (fig. S3O), and
no signs of an ER stress response were de-
tected in MTM1 KO cells (fig. S3P). We con-
clude that PI(3)P hydrolysis by MTM1 in starved
cells controls ER shape independently of au-
tophagy, the ER stress response, and ER cal-
cium homeostasis.
Given that PI(3)P is a hallmark of early endo-
somes (10, 35) and that the ER makes extensive
contacts with other organelles including the
plasma membrane (36), endosomes, lysosomes,
and mitochondria (37), we hypothesized that
MTM1 specifically acts on peripheral early
endosomes in starved cells and thereby con-
trols ER morphology (Fig. 2A), for example,
through membrane contacts. To test this, we
generated a cell line stably expressing a chimera
between the early endosomal protein Rab5A
and the biotinylating enzyme ascorbate per-
oxidase 2 (APEX2) (38) under the control of
a doxycycline-inducible promoter (fig. S4A).
Proximity labeling and affinity capture re-
vealed a prominent starvation-induced enrich-
ment of endogenous MTM1 on early endosomes,
while early endosomal antigen 1 (EEA1), a PI(3)P-
binding scaffold protein, was depleted in early
endosomes (Fig. 2B and fig. S4B). Nutrient
starvation thus induces the recruitment of
MTM1 to Rab5-containing early endosomes,
likely resulting in PI(3)P hydrolysis and ER
sheet expansion. We further probed this model
by a chemical genetic approach that capitalizes
on the FRB/FKBP system, which enables the
artificial tethering of organelles via rapalog-
induced heterodimerization of chimeras be-
tween FK506-binding protein (FKBP) and the
FKBP-rapamycin binding (FRB) domain of
mTOR (39). We found that acute rapalog-
induced recruitment of active FKBP-MTM1
to early endosomes tagged with FRB-Rab5A
resulted in massive ER tubule-to-sheet con-
version in fed cells (Fig. 2C and fig. S4C), phe-
nocopying starvation-induced PI(3)P depletion.
Recruitment of active MTM1 to lysosomes (fig.
S4, D and E) or early endosomal recruitment
of catalytically inactive mutant MTM1 (CS) did
not affect ER shape (Fig. 2C). We conclude that
recruitment of catalytically active MTM1 to
early endosomes drives ER sheet expansion
to mount the cellular response to nutrient
starvation.
The role of endosomal MTM1 in controlling
ER shape might depend on the formation of
hitherto molecularly undefined membrane con-
tacts between the tubular ER and highly motile
early endosomes that could provide a force
that aids in keeping ER tubules under tension.
In support of this hypothesis, we observed
PI(3)P and the early endosome marker Rab5A
to tightly colocalize with the tubular ER net-
work in the cell periphery (Fig. 2D). Moreover,
tracking early endosomes and the ER network
in live cells by high-speed spinning disk con-
focal imaging showed that forming ER tubules
are tightly associated with motile early endo-
somes (Fig. 2E and Movie 3), a conclusion fur-
ther supported by the close association of ER
tubules with early endosomes marked by in-
ternalized bovine serum albumin–gold (BSA-
gold) in electron micrographs (Fig. 2F). Hence,
the large population of peripheral early endo-
somes may serve to promote ER tubules. Con-
sistently, we found that acute depletion of early
endosomes from the cell periphery by rapalog-
induced endosomal recruitment of the micro-
tubule minus end–directed dynein adaptor
BICD2 caused a near complete loss of the tu-
bular ER network and a concomitant accumu-
lation of perinuclear sheet ER (Fig. 2, G and H;
fig. S4F; and movie S2).
We reasoned that MTM1-mediated hydrolysis
of PI(3)P at early endosomes during starvation
may control ER shape either by (i) controlling
early endosome motility or (ii) altering the
number or stability of hitherto unknown phys-
ical contacts between early endosomes and
the ER (fig. S4G). As early endosome motility
was unaffected by nutrient starvation (fig. S4,
H and I), we followed the alternative hypoth-
esis that nutrient regulation of early endo-
somal PI(3)P controls membrane contact sites
between early endosomes and the ER and,
thereby, ER shape. To monitor such contacts,
we determined the fractional overlap between
the peripheral ER and early endosomes by
multicolor STED microscopy (fig. S4J). Starva-
tion reduced the fractional overlap of the ER
with early endosomes (Fig. 2I). The total con-
tact area between the ER and early endosomes
marked by internalized BSA-gold was also
found to be significantly reduced in starved
cells analyzed by electron microscopy (Fig. 2J).
To confirm these findings by another inde-
pendent approach, we generated a stable cell
line that coexpresses equimolar ratios of an
ER membrane targeting domain (ERM) fused
Jang et al., Science 378, eabq5209 (2022)
16 December 2022
3 of 16
Corrected 7 September 2023. See full text. RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Starvation-induced PI(3)P hydrolysis by MTM1 at ER–early
endosome contacts mediates ER reshaping. (A) PI(3)P-positive early endo-
somes (EE) may control the tubular ER. (B) Starvation-induced recruitment
of MTM1 to EE. EE-associated MTM1 (left) and EEA1 (right) in fed versus starved
HeLa cells. n = 3 independent experiments. (C) Ratio of sheet/total ER in
HeLa cells coexpressing FRB-iRFP-Rab5A and mRFP-FKBP-MTM1 WT (wild-type)
or CS (inactive mutant) ± rapalog. FKBP-MTM1 WT − (n = 162), rapalog (n =
265), FKBP-MTM1 CS − (n = 138), rapalog (n = 187). (D) Confocal images of HeLa
cells stably coexpressing Halo-Rab5A (EE) and mEmerald-Sec61b (ER) and
stained for PI(3)P. (E) Time-lapse confocal images of HeLa cells coexpressing
mEmerald- Sec61b (ER) and mCherry-Rab5 (EE). Yellow arrowheads mark motile
EE. (F) Electron micrographs illustrating tubular ER (blue) contacts with EE
(orange) marked by internalized BSA-gold. (G) Rapalog-induced acute depletion
of EE from the periphery. See movie S2. (H) Ratio of sheet/total ER in mock-
(n = 177) versus rapalog-treated (n = 141) HeLa cells as in (G). (I) Fractional
overlap between the peripheral ER and EE in fed or starved cells determined
by STED microscopy. Fed = 118 ROIs; Starved = 166 ROIs (40 to 50 cells from
four experiments). (J) Morphometric analysis of contact length (nanometers)
between ER and BSA-gold labeled EE in fed (42 endosomes) versus starved
cells (43 endosomes). (K) Reconstitution of split-GFP fluorescence by ER-EE
contacts. (L) Normalized number of ER/EE contacts in fed (n = 238) versus
starved (n = 257) HeLa cells. (M) Rescue of ER/EE contacts in starved MTM1-
depleted HeLa cells. siCo Fed (n = 123), Starved (n = 77), siMTM1_1 Fed
(n = 111), Starved (n = 85). (N) ERM-FRB(cid:129)rapalog(cid:129)FKBP-Rab5A ER/EE
synthetic tether. (O) Ratio of sheet/total ER in fed or starved cells from
(N). Endosomal FKBP Rab5 Fed − (n = 68), rapalog (n = 70), Starved − (n =
92), rapalog (n = 86); Cytosolic Rab5-FKBP Fed − (n = 97), rapalog (n = 90),
Starved − (n = 104), rapalog (n = 120). (P) Confocal images from (O) stained
for HA to mark the ER, gray. Scale bars: 2 mm (yellow), 1 mm (black), and
100 nm (white). n, number of cells analyzed from two [(M) and (O)] or three
[(C), (H), and (L)] independent experiments. One-samplet
unpaired t test [(C), (I), (L), and (O)]; two-tailed Mann-Whitney test [(H) and (J)];
*P ≤ 0.05, **P ≤ 0.01, ****P ≤ 0.0001; ns, nonsignificant. Data are mean ± SD [(B),
(I), (L), and (M)] or median ± interquartile range [(C), (H), (J), and (O)].
test (B); two-tailed
Jang et al., Science 378, eabq5209 (2022)
16 December 2022
4 of 16
Corrected 7 September 2023. See full text. RES EARCH | R E S E A R C H A R T I C L E
Movie 3. Live imaging of early endosome motility and the tubular ER. Live-cell high-speed spinning
disk confocal imaging of fed HeLa cells expressing mEmerald-Sec61b (labeled gray) and mCherry-Rab5
(labeled yellow). Videos were acquired at 1 frame/0.4 s for 1 min and correspond to Fig. 2E.
to GFP11 (40) along with a GFP1-10-Rab5A chi-
mera under the control of a doxycycline-inducible
promoter (fig. S4, K and L). Formation of mem-
brane contacts between early endosomes and
the ER in fed cells reconstitutes GFP fluores-
cence (Fig. 2K). ER–early endosome membrane
contacts were greatly reduced in starved cells
(Fig. 2L and fig. S4M), and this effect was re-
verted completely by depletion of MTM1 (Fig.
2M). These data suggest that MTM1-mediated
hydrolysis of PI(3)P at early endosomes reduces
the contacts between these organelles and the
ER (Fig. 2K). We directly tested this model by
examining the effects of semisynthetic tether-
ing of early endosomes to the ER using chem-
ical genetics. Acute rapalog-induced formation
of ER–early endosome tethers mediated by
ERM-FRB•rapalog•FKBP-Rab5A potently sup-
pressed starvation-induced ER tubule-to-sheet
conversion (Fig. 2, N to P; fig. S5B; and movie
S3) but had no effect on early endosome
motility (fig. S5A). In contrast, soluble, non-
lipidated Rab5A (Rab5-mRFP-FKBP) or cyto-
solic mRFP-FKBP had no effect on ER shape
(Fig. 2O and fig. S5, B and C). Extending the
length of the semisynthetic tether well beyond
the typical distance of 20 to 30 nm observed
for organelle contacts in vivo (36, 37) did not
affect its ability to prevent starvation-induced
ER reshaping (fig. S5, D and E), suggesting
that the exertion of a physical pulling force is
critical for the regulation of ER shape by mem-
brane contacts with early endosomes. Further-
more, the capability of ERM-FRB•rapalog•
FKBP-Rab5A tethers to prevent starvation-
induced ER tubule-to-sheet conversion was
unaffected by depletion of PI(3)P, indicating
that PI(3)P acts upstream of ER–early endo-
some membrane contact site formation (fig. S5,
F and G). Other organelles—late endosomes
and lysosomes, in particular—have also been
shown to form numerous contacts with the ER
(41, 42). Early endosomes display an even dis-
tribution throughout cells and into the periphery
where the tubular ER is located and outnumber
lysosomes by up to an order of magnitude. In
contrast, most lysosomes are concentrated in
the perinuclear area at steady state (fig. S6, A
to D). As a consequence, the total number of
membrane contacts between the ER and early
endosomes greatly exceeds that of the ER with
lysosomes (fig. S6, E to H), in spite of the high
relative fraction of lysosomes in touch with the
ER (42, 43). Redistribution of lysosomes to the
perinuclear area either through acute rapalog-
induced dynein adaptor recruitment or deple-
tion of PI(3,5)P2 did not affect ER shape (fig.
S6, I to L; fig. S1H; and movie S4). Sustained
loss of the lysosomal kinesin adaptor Arl8b,
a protein essential for lysosome dispersion,
only marginally increased the sheet ER frac-
tion [fig. S6, M to O; see also (44)], whereas
depletion of the ER-lysosome contact site pro-
tein protrudin (41) was without effect. Hence,
the tubular ER is largely controlled by its mem-
brane contacts with early endosomes (as
demonstrated in this study) and a smaller con-
tribution from lysosomes, possibly dependent
on cell type or conditions (44). Taken together,
our findings unravel a crucial role for MTM1-
mediated PI(3)P hydrolysis in the reduction of
membrane contacts between the ER and early
endosomes to reshape the ER in response to
changing nutrient levels.
Contacts formed by ER membrane
protein–mediated recognition of early
endosomal PI(3)P controls the tubular ER
To identify the molecular machinery that teth-
ers ER tubules to early endosomes in fed cells,
we capitalized on the observation that in WT
but not in MTM1 KO cells ER–early endosome
contacts are reduced under conditions of nu-
trient starvation (Fig. 3A). Starved WT or MTM1
KO HeLa cells inducibly expressing APEX2-
Rab5A were subjected to combined proximity
labeling and affinity purification–mass spec-
trometry to probe the molecular environment
of Rab5A-containing early endosomes. Subse-
quent biochemical fractionation to enrich for
ER membrane proteins (fig. S7, A to C, and
table S2) and comparison with previously iden-
tified ER surface proteins (45) identified sev-
eral putative ER transmembrane proteins that
might serve as early endosome tethers (fig.
S7D). Analysis of the phenotypic consequences
of cellular depletion of these factors revealed
that only loss of ER ribosome-binding protein
1 (RRBP1) promoted ER tubule-to-sheet con-
version in fed cells (fig. S7E). RRBP1 is local-
ized exclusively to ER membranes (Fig. 3, B
and C) and has been associated with changes
in ER morphology, although no consensus re-
garding its precise function exists (46–48). ER
tubule-to-sheet conversion was exacerbated by
concomitant loss of RRBP1 and its close para-
log kinectin 1 (KTN1) (Fig. 3D and fig. S8, A to
C). Moreover, loss of RRBP1 and KTN1 potent-
ly reduced the number of ER–early endosome
membrane contacts in fed cells (Fig. 3E), sug-
gesting that RRBP1 and KTN1 regulate ER
shape by acting as tethers for early endosomes.
KTN1 and RRBP1 harbor functionally unchar-
acterized lysine-rich regions (LR1, -2, and -3)
in their cytoplasmic domains (Fig. 3F and fig.
S8, A and B) that could conceivably attach to
PI(3)P. Notably, KTN1 was found to be in con-
tact with PI(3)P-enriched early endosomes by
spatiotemporally resolved interaction proteo-
mics using 2xFYVE-APEX as a probe (49). We
found recombinant RRBP1-LR1-3 (fig. S8D) to
bind to PI(3)P with preference over both PI(4)
P and PI(3,4)P2 (Fig. 3, G and H). These data
suggest that RRBP1 might recognize PI(3)P on
early endosomes to form tethers with the ER.
Consistent with this model, we observed that
expression of a mini-RRBP1 truncation pro-
tein variant harboring only the lysine-rich
cytoplasmic domain fused to its ER transmem-
brane anchor sufficed to restore a normal
tubular ER network in HeLa cells depleted of
endogenous RRBP1 and KTN1 (Fig. 3I, WT).
Deletion of lysine-rich regions 1 or 3 or of
the transmembrane anchor rendered trun-
cated mini-RRBP1 inactive (Fig. 3I and fig. S8,
E and F). These data suggest that RRBP1 and
KTN1 mediate recognition of early endosomal
PI 3-phosphates and, possibly, additional fac-
tors, to facilitate contact site formation be-
tween the ER and early endosomes, which
control the tubular ER network in human cells.
MTM1-dependent ER reshaping is required
for mitochondrial network formation
during starvation
Several mechanisms have been shown to con-
tribute to mitochondrial morphogenesis, includ-
ing membrane contacts between mitochondria
and lysosomes (50), late Golgi-derived vesicles
(51), and the tubular ER (37, 52). The tubular ER
also directly promotes mitochondrial fission
(53) and acts as a donor organelle for the for-
mation of lipid droplets (LDs) (54, 55) that
serve as an energy reservoir in fed cells and
tissues. Under conditions of starvation (e.g.,
deprivation of glutamine and other amino
acids), mitochondria undergo hyperfusion into
tubular networks to protect themselves from
mitophagy (56) and to enable efficient utiliza-
tion of fatty acids (57).
On the basis of these prior works, we hy-
pothesized that the MTM1-mediated starvation-
induced reshaping of the tubular ER into
sheets (Fig. 1) might serve to enable cells to
metabolically adapt to altering nutrient envi-
ronments, for example, by altering mitochon-
drial organization. To test this hypothesis, we
examined the effect of MTM1 loss on mito-
chondrial morphology and respiratory func-
tion. We found mitochondria to become
hyperfused in starved WT but not in MTM1 KO
cells or in WT cells depleted of endogenous
Jang et al., Science 378, eabq5209 (2022)
16 December 2022
5 of 16
Corrected 7 September 2023. See full text. RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Membrane contacts formed by ER ribosome-binding
protein 1 and kinectin 1–mediated recognition of early
endosomal PI(3)P control the tubular ER. (A) Starvation-
induced dissociation of ER–early endosome (EE) contacts in
WT but not in MTM1 KO cells. (B) Volcano plot of proximity
biotinylated interactome of APEX2-Rab5A isolated from ER
membranes of starved MTM1 KO versus WT HeLa cells. Dark
purple: RRBP1 and KTN1. (C) Confocal images of HeLa cells
coexpressing mEmerald-Sec61b (ER) and RRBP1 full length
(FL)–mCherry. (D) Ratio of sheet/total ER in fed HeLa cells
treated with control (siCo/siCo = 174) or KTN1+RRBP1 siRNAs
(siKTN1/siRRBP1 = 221). (E) Normalized number of ER/EE
contacts in control or KTN1/RRBP1-depleted fed HeLa cells.
n = 3 independent experiments. (F) Schematic illustrating
truncated mini-RRBP1 (amino acids 1 to 150). TM, trans-
membrane anchor; LR1-3, cytoplasmic lysine-rich regions 1 to 3.
(G) GST-RRBP1-LR1-3 binds to PI(3)P liposomes. Supernatant
(S) and liposomal pellet (P) fractions were analyzed by
immunoblotting for GST. No PIP, liposomes lacking phospho-
inositides. (H) Quantified data as in (G) from n = 3 independent
experiments. (I) Ratio of sheet/total ER in control (siCo/siCo)
or KTN1/RRBP1-depleted stable doxycycline-inducible HeLa
cells expressing the indicated truncated RRBP1 protein
(“mini-RRBP1”) variants [see (F)]. WT, wild-type V5-tagged
mini-RRBP1; delLR1, mutant RRBP1 lacking lysine-rich region 1;
delLR3, mutant RRBP1 lacking lysine-rich region 3; delTM,
mutant RRBP1 lacking its transmembrane domain. siCo/siCo −
(n = 111), WT (n = 129), delLR1 (n = 102), delLR3 (n = 120),
delTM (n = 130); siKTN1/siRRBP1 − (n = 274), WT (n = 271),
delLR1 (n = 326), delLR3 (n = 263), delTM (n = 288). Scale bars:
2 mm (yellow). n indicates the total number of cells analyzed
from three independent experiments. Two-tailed Mann-Whitney
test (D); one-sample t test between Starved and Fed siCo/siCo,
two-tailed t test between Fed siCo/siCo and Fed siKTN1/siRRBP1
(E); one-way ANOVA with Dunnett’s multiple comparisons test (H);
Kruskal-Wallis test with two-sided Dunn’s multiple comparison
test (I). *P ≤ 0.05, **P ≤ 0.01, ****P ≤ 0.0001. Data are median ±
interquartile range [(D) and (I)]; data are mean ± SD [(E) and (H)].
MTM1 (Fig. 4, A and B, and fig. S9A). Further
ultrastructural analysis by 3D FIB-SEM re-
vealed the formation of an extensive cell-wide
network of hyperfused mitochondria in starved
WT cells. In contrast, starved MTM1 KO cells
displayed an accumulation of small spherical
mitochondria suggestive of exacerbated mito-
chondrial fission and defects in cristae mor-
phology (Fig. 4, C to E, and fig. S9, B to D),
whereas the total mitochondrial volume frac-
tion was unchanged (Fig. 4F). Reduced mito-
chondrial network formation and defective
ER reshaping were also observed in starved
XLCNM patient–derived myoblasts (fig. S9,
G and H). Defective mitochondrial morpho-
genesis was not a consequence of the altered
expression of mitochondrial fusion- or fission-
related proteins (58) such as mitofusin 1/2,
OPA1 (59), and DRP1 or its hyperactive form
(pS616-DRP1) in MTM1 KO cells (fig. S3, J
and K). Moreover, acute rapalog-induced for-
mation of ER–early endosome tethers to inhibit
loss of ER tubules prevented the starvation-
induced formation of a hyperfused mitochon-
drial network in WT cells (Fig. 4G and fig. S9,
E and F), thereby phenocopying MTM1 loss.
Conversely, reducing membrane contacts be-
tween the ER and early endosomes by deple-
tion of RRBP1 and KTN1 led to the formation
of hyperfused mitochondrial networks and
expansion of the sheet ER in MTM1 KO cells
(fig. S9, I to K). RRBP1 and KTN1 therefore
act downstream of MTM1-mediated PI(3)P
hydrolysis.
If reshaping of the tubular ER into sheets
causally underlies the formation of a hyper-
fused mitochondrial network, one would expect
experimental manipulations that reshape the
ER to affect mitochondrial morphology. Con-
sistently, we found that cellular depletion of
either Rab10, a factor required for maintenance
of the tubular ER (60); RRBP1 and KTN1; or the
tubular ER-shaping protein reticulon 4 (61)—
conditions that result in ER sheet expansion
(fig. S10, A, C, D, and F)—cause the formation
of an enlarged mitochondrial network in fed
cells (fig. S10, B, C, E, and F). Mitochondrial
hyperfusion was further induced by overex-
pression of the ER sheet–inducing membrane
protein Climp63 (46) (fig. S10, G and H). Final-
ly, we observed increased rates of mitochon-
drial fission in starved MTM1 KO cells (fig.
S10, I to K, and movie S5), that is, conditions
under which ER tubules accumulate. Col-
lectively, these findings establish that the
starvation-induced conversion of ER tubules
to sheets by MTM1-mediated hydrolysis of
PI(3)P at ER–early endosome contacts facili-
tates the formation of a functional mitochon-
drial network.
Defective ER morphogenesis in absence of
endosomal MTM1 impairs mitochondrial
metabolic rewiring during starvation
To probe the physiological consequences of
defective mitochondrial morphogenesis, we
monitored mitochondrial oxygen consumption
and mitochondria-driven adenosine triphos-
phate (ATP) production in WT and MTM1 KO
cells using Seahorse technology. Starved
MTM1 KO cells displayed severely reduced
basal as well as maximal oxygen consumption
Jang et al., Science 378, eabq5209 (2022)
16 December 2022
6 of 16
Corrected 7 September 2023. See full text. C
Starved
WT
KO
0
.
0
0
3
RES EARCH | R E S E A R C H A R T I C L E
siCo.
siMTM1_1
Fed
Starved
Fed
Starved
A
0
2
M
O
T
t
e
s
n
I
d
e
t
n
e
m
g
e
S
D
l
e
v
i
t
a
u
m
u
C
)
³
l
m
µ
(
e
m
u
o
v
a
i
r
d
n
o
h
c
o
t
i
m
Starved- WT
Starved- KO
E
n
o
i
r
d
n
o
h
c
o
t
i
m
l
i
a
u
d
v
d
n
I
i
15
10
5
0
0
100
200
300
400
# of mitochondria
B
n
o
i
r
d
n
o
h
c
o
t
i
m
/
a
e
r
a
n
a
e
M
8
6
4
)
³
m
µ
(
e
m
u
o
v
l
1.0
0.5
)
.
U
A
.
(
400 ****
****
***
300
200
100
0
Fed
Starved
Starved
Fed
Fed
Starved
siCo.
siMTM1_1
KO
69
mito
310
mito
62
%
Top 3%
18
%
15
10
5
0
Top 3%
l
a
i
r
d
n
o
h
c
o
t
i
m
f
o
m
u
S
)
³
m
µ
(
e
m
u
o
v
l
F
n
o
i
t
c
a
r
f
e
m
u
o
v
l
l
a
i
r
d
n
o
h
c
o
t
i
M
ns
(EM)
0.15
l
m
s
a
p
o
t
y
c
r
e
p
0.10
0.05
0.00
WT KO
Starved
G
n
o
i
r
d
n
o
h
c
o
t
i
m
/
a
e
r
a
n
a
e
M
0.0
WT KO
Starved
l
V
o
u
m
e
(
µ
m
³
)
7
.
2
8
0
(ER-EE tether)
600
*
400
)
.
U
A
.
(
200
0
Rapalog
EtOH
Starved
Fig. 4. MTM1-dependent ER reshaping is required for mitochondrial
network formation during starvation. (A) Confocal images of mitochondria
(TOM20) in fed or starved control (siCo) and MTM1-depleted (siMTM1_1) HeLa
cells. (Bottom) Segmented mitochondria in ROI (yellow, 15 mm by 15 mm).
(B) Mean area of individual mitochondrion per ROI from fed or starved WT or
MTM1 KD or KO HeLa cells. siCo Fed = 345 ROIs, Starved = 418 ROIs; siMTM1_1
Fed = 354 ROIs, Starved = 440 ROIs; KO Fed = 343 ROIs, Starved = 322 ROIs
from four independent experiments. Each ROI represents a single cell. (C) 3D
rendering of mitochondria by FIB-SEM in starved WT and MTM1 KO HeLa
cells. Heat bar reflects individual mitochondrial volumes. Images are represen-
tative of three cells. (D) Cumulative plot of mitochondrial volume distribution
in starved WT versus MTM1 KO cells derived from FIB-SEM. (E) 3D volumes
of individual mitochondria as in (C) were plotted. (Inset) The largest 3% of
mitochondria occupied 62% of the total mitochondrial volume in WT (gray)
but only 18% in KO (red) cells. The total volume of mitochondria in WT cells
(69 mito = 12.95 mm3) versus KO cells (310 mito = 13.25 mm3) was unchanged.
(F) Mitochondrial volume fraction in starved WT and MTM1 KO HeLa cells (n =
10) analyzed by stereological analysis of thin-sectioned electron micrographs.
(G) Artificial ER–early endosome tethering prevents mitochondrial hyperfusion.
Mean area of individual mitochondria per ROI from starved HeLa cells ±
rapalog (Fig. 2N). EtOH = 251 ROIs, rapalog = 250 ROIs from three independent
experiments. Each ROI represents a single cell. Scale bars: 10 mm (white),
2 mm (yellow), 1 mm (black). One-way ANOVA with Tukey’s multiple comparisons
(B); two-tailed unpaired t test [(F) and (G)]; *P ≤ 0.05, ***P ≤ 0.001,
****P ≤ 0.0001; ns, nonsignificant. Data are mean ± SD [(B) and (G)] or
median ± interquartile range (F).
rates resulting in significantly impaired mito-
chondrial ATP synthesis (Fig. 5, A and B, and
fig. S11, A and B). As a consequence, starved
MTM1 KO cells suffered from a pronounced
reduction of total cellular ATP (Fig. 5C). Sim-
ilar results were observed in cells, in which the
ER was artificially tethered to early endo-
somes (Fig. 5D). No significant differences in
mitochondrial basal oxygen consumption or
mitochondrial ATP synthesis were detected
in fed MTM1 KO cells, whereas the maximal
oxygen consumption rate was marginally de-
creased (fig. S11, C to E). The mitochondrial
membrane potential was unaffected by MTM1
loss, irrespective of the nutritional status of
the cells (fig. S11F). As defective mitochondrial
morphogenesis and cellular ATP depletion
are associated with reduced cell viability, we
probed whether MTM1 KO cells might be
poised to undergo apoptosis under conditions
of limited nutrient availability. Starved MTM1
KO cells indeed suffered from increased levels
of cleaved caspase 3 and poly(ADP ribose)
polymerase (PARP), common indicators of
apoptotic cell death (fig. S11, G to I). How-
ever, defective mitochondrial morphogenesis
in MTM1 KO cells was not a secondary con-
sequence of increased apoptosis, as treatment
of KO cells with the pan-caspase inhibitor
Z-VAD-FMK effectively blocked apoptosis (fig.
S11J) but failed to rescue defects in mitochon-
drial morphology (fig. S11K). We note that sim-
ilar apoptotic phenotypes have been reported
in starved OPA1 and mitofusin 1/2 KO cells
defective in mitochondrial fusion (62).
These results indicate that MTM1-mediated
reshaping of the tubular ER into sheets is
required for the formation of a functional mito-
chondrial network and mitochondria-driven
ATP production in starved cells. Previous work
has shown that the effective transfer and uti-
lization of fatty acids (FAs), a major substrate
for mitochondrial ATP synthesis via b-oxidation,
in starved cells requires mitochondria to be
organized into a highly tubulated hyperfused
network (56, 57), although the exact underly-
ing molecular mechanism is unclear. We there-
fore hypothesized that impaired mitochondrial
ATP production in MTM1 KO cells might be a
consequence of defective FA trafficking. We
directly tested this by monitoring the fate of
FAs in WT and MTM1 KO cells under different
nutrient conditions. Depending on metabolic
state, cytosolic FAs can be metabolized in mito-
chondria (e.g., during starvation) or stored in
LDs (54, 57, 63), which exclusively form from
the tubular ER (55, 64). Consistently, we found
that in WT cells, the number of LDs inter-
mittently declined at the onset of starvation
(≤2 hours), likely as a result of increased mito-
chondrial b-oxidation of FAs and blocked LD
formation upon loss of the tubular ER, before
eventually rising (Fig. 5E and fig. S11, L to N)
owing to autophagy-promoted lipid buildup
during sustained long-term (≥6 hours) starva-
tion (57). In contrast, the number and total
volume of LDs increased in MTM1 KO cells at
the onset of starvation (Fig. 5E and fig. S11O).
LD accumulation persisted upon treatment of
MTM1 KO cells with the pan-caspase inhibitor
Z-VAD-FMK (fig. S11P). These results suggest that
starvation-induced ER tubule-to-sheet conversion
Jang et al., Science 378, eabq5209 (2022)
16 December 2022
7 of 16
Corrected 7 September 2023. See full text.
RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. Defective ER morphogenesis in absence
of endosomal MTM1 impairs mitochondrial
metabolic rewiring during starvation. (A) Basal
mitochondrial oxygen consumption rate (mito-OCR)
of starved WT or MTM1 KO HeLa cells.
(B) Mitochondria-dependent ATP production of
starved WT or MTM1 KO HeLa cells. (C) Normalized
total cellular ATP levels of starved WT (set to 1) or
MTM1 KO HeLa cells. n = 3 independent
experiments. (D) Normalized total cellular ATP levels
of fed versus starved HeLa cells ± rapalog (Fig. 2N).
Data for fed cells (− rapalog) were set to 1. n = 5
independent experiments. (E) Number of BODIPY
493/503–labeled lipid droplets (LDs) in fed or
starved (2 hours) WT or MTM1 KO HeLa cells. WT
Fed (n = 157), Starved (n = 181); KO Fed (n = 153),
Starved (n = 148). (F) Schematic depicting the pulse-
chase assay to monitor FA mobilization. (G) Confocal
images of WT and MTM1 KO HeLa cells pulse-labeled
with RedC12 and chased for 2 hours in EBSS and
stained for TOM20. (H) Number of RedC12-labeled
LDs in WT or MTM1 KO HeLa cells chased for 0
or 2 hours in EBSS. 0h: WT (n = 126), KO (n = 112);
2h: WT (n = 139), KO (n = 118). (I) Pearson
correlation coefficient of RedC12-labeled FAs and
mitochondria (TOM20) from randomly selected
100 pixel by 100 pixel ROIs in WT or MTM1 KO cells
as in (H). WT = 169 ROIs, KO = 203 ROIs from
three independent experiments. (J) Schematic
of trafficking of pulse-labeled FA RedC12 in WT and
MTM1 KO cells during starvation. (K) Gene Ontology
(GO) analysis of proteins depleted in starved
MTM1 KO compared with WT HeLa cells. Terms
related to fatty acid catabolism and mitochondrial
function are highlighted in red. All of these
proteins were unaltered in fed MTM1 KO cells.
(L) Volcano plot of proteins depleted in starved
MTM1 KO compared with WT HeLa cells. Red dots,
proteins enriched from GO analysis shown in (K).
Scale bars: 10 mm. n indicates the total number of
cells analyzed from three independent experiments.
Two-tailed Mann-Whitney test [(B), (D) (Fed
vs. Starved), (E), (H), and (I)]; one-way ANOVA with
Dunnett’s multiple comparisons test for all other
comparisons of (D); two-tailed unpaired t test (A);
one-sample t test (C); ns, nonsignificant, *P ≤ 0.05,
**P ≤ 0.01, ***P ≤ 0.001, ****P ≤ 0.0001. Data are median ± interquartile range [(A), (B), (E), (H), and (I)] or mean ± SD [(C) and (D)].
mediated by MTM1 orchestrates the reflux of
FAs to mitochondria for b-oxidation while
counteracting their storage in LDs. We directly
probed this model by monitoring FA trafficking
during early stages of starvation using the
fluorescent FA analog BODIPY 558/568 C12
(Red C12) (Fig. 5F). Pulse-labeled Red C12 was
efficiently transported to mitochondria in starved
WT cells (Fig. 5G and fig. S11Q), consistent with
prior data (57). In contrast, in starved MTM1 KO
cells, Red C12 failed to accumulate in mito-
chondria and instead was targeted to LDs
(Fig. 5, G to J). Impaired mitochondrial lipid
and fatty acid catabolism was also clearly shown
by the unbiased quantitative proteomic anal-
ysis of fed or starved WT and MTM1 KO cells
using tandem mass tag (TMT) labeling. These
analyses revealed a down-regulation of pro-
teins involved in mitochondrial respiration and
transport (Fig. 5K, fig. S11R, and tables S3 and
S4), for example, mitochondrial very long-chain
specific acyl–coenzyme A dehydrogenase, car-
nitine palmitoyltransferase 2, NADH:ubiquinone
oxidoreductase, and the mitochondrial protein
import factor TIMM17B (Fig. 5L), possibly as
an indirect consequence of the observed struc-
tural mitochondrial defects in starved MTM1
KO cells (Fig. 4). None of these proteins were
altered in KO cells under fed conditions (fig.
S11S), indicative of a specific defect of MTM1 KO
cells to appropriately respond to altered nu-
trient supply. These findings indicate that
MTM1 mediates reshaping of the ER at the
onset of starvation to drive the formation of a
functional mitochondrial network and facili-
tate mitochondrial b-oxidation, which sustains
ATP production. Conversely, defective ER mor-
phogenesis in the absence of endosomal MTM1
impairs mitochondrial metabolic rewiring dur-
ing starvation.
Discussion
How cells and tissues orchestrate adaptive
changes in organelle dynamics and metabo-
lism on a cell-wide scale has remained unclear.
Jang et al., Science 378, eabq5209 (2022)
16 December 2022
8 of 16
Corrected 7 September 2023. See full text. RES EARCH | R E S E A R C H A R T I C L E
Here we reveal a key role for early endosomal
lipid signaling mediated by MTM1, a lipid
phosphatase mutated in XLCNM in humans
(19, 30, 31), in controlling the tubular ER and,
thereby, mitochondrial morphology and meta-
bolic function in the acute response to fluctuating
nutrient conditions. This mechanism operates
independent of ER-phagy, a process typically
induced by prolonged starvation (65). We dem-
onstrate that starvation-induced PI(3)P hydro-
lysis by endosomal MTM1 reduces previously
undescribed membrane contacts between pe-
ripheral ER tubules and early endosomes. These
contacts act as physical tethers that may trans-
mit pulling forces from highly motile periph-
eral endosomes to the ER and are mediated by
endosomal PI(3)P binding to RRBP1, a large ER
membrane protein overexpressed in colorectal
cancer (66), and its close paralog kinectin 1.
Whether ER–early endosome membrane con-
tacts also enable material exchange in vivo as
shown for contact sites between the ER and
the plasma membrane, the trans-Golgi network,
or lysosomes (37) remains to be determined.
Loss of ER tubules, possibly in conjunction
with ER-independent mechanisms, drives mito-
chondrial network formation and, directly or
indirectly, facilitates FA transfer to mitochon-
dria to fuel b-oxidation and, thereby, mitochon-
drial ATP production to sustain cellular energy
homeostasis. The precise relationship between
mitochondrial morphogenesis, FA mobilization
to mitochondria, and mitochondrial ATP pro-
duction remains to be defined. Interestingly,
ER sheets are favored over tubules from an
energetic perspective (67, 68) and, hence,
should prevail under conditions of nutrient
starvation when cellular energy levels are low.
Consistent with this, it has recently been shown
that the hepatic ER in obese mice is character-
ized by disorganized ER sheets and a predomi-
nance of ER tubules and accompanying defects
in lipid metabolism (69). Our findings thus
identify an organellar conveyor belt, in which
the tubular ER serves as a membrane conduit
that transmits nutrient-triggered changes (i.e.,
in glutamine and other amino acids) in early
endosomal PI(3)P levels to metabolic organ-
elles such as LDs and mitochondria [in agree-
ment with (56)] to enable metabolic rewiring
under conditions of limited nutrient supply
and, possibly, in cancer [e.g., when RRBP1 is
overexpressed (66)]. Defects in ER shape, mito-
chondrial morphogenesis, and cellular ATP de-
pletion caused by loss of MTM1 function can
explain the observed myofiber hypotrophy and
defective sarcoplasmic reticulum organization
in animal models of XLCNM (16, 17) and in
human patients who often appear undernour-
ished (19, 30, 31). Furthermore, it is conceivable
that reduced contact formation between early
endosomes and ER tubules due to MTM1-
mediated PI(3)P hydrolysis, in addition to its
effects on ER shape and mitochondrial func-
tion, may facilitate endosomal exocytosis of
b-integrins, a mechanism shown to be defec-
tive in XLCNM (15). How precisely early endo-
somal PI(3)P levels and MTM1 function are
controlled by cellular nutrient status remains
poorly understood. VPS34, the main PI(3)P-
synthesizing enzyme on endosomes has been
reported to be stimulated by fed signals (70),
that is, conditions in which MTM1 activity
is repressed.
The endosomal signaling lipid–based path-
way to control oxidative cell metabolism un-
covered in this work may synergize with other
cellular mechanisms that impinge on the dy-
namics of metabolically active organelles. For
example, it has been shown that the function
and localization of lysosomes depend on motor
proteins (8) as well as on their association with
the ER (41) and are regulated by cellular nu-
trient status (71), which in turn affects nutrient
signaling (28, 72). Late endosomes (i.e., organ-
elles distinct from the Rab5-positive early en-
dosomes described here) have been shown to
undergo fission at sites of contact with the ER
that are molecularly and functionally distinct
(73) from the ER–early endosome contacts
identified in this study. Finally, mitochondria–
lysosome contacts have been shown to regu-
late mitochondrial fission (50). Whether any
of these contacts are subject to nutrient regu-
lation and impact on cell metabolism will need
to be addressed in future studies. Conceivably,
lipid phosphatases including other members of
the myotubularin family (26), many of which
are linked to human disease, may play crucial
physiological roles in the regulation of these
and other membrane contacts.
Materials and methods
Materials
Plasmids
Plasmids used: mEmerald-Sec61b -C1 (Addgene
#90992), mCh-Rab5 (Addgene #49201), ER-
GCaMP6-1-150 (Addgene #86918), tdTomato-
BICD2-FKBP (Addgene #64205), and mCh-Climp63
(mouse) (Addgene #136293). HA-BICD2-FRB was
kindly provided by G. G. Farías. Plasmids for
transient transfection (e.g., mCherry MTM1
WT, mRFP-FKBP-empty, mRFP-FKBP-MTM1
WT, mRFP-FKBP-MTM1 C375S, FRB-iRFP-
Rab5A, mRFP-FKBP-Rab5A, TMEM192-3xHA-
FRB, ERM-2xHA-FRB) were generated with
the pcDNA3.1(+) vector and polymerase chain
reaction (PCR) or ligation-based cloning. Note
that ERM is the N-terminal ER membrane
targeting sequence of residues 1 to 27 of ER-
resident P450 oxidase 2C1. Full-length RRBP1
was amplified from pcDNA4 HisMax-V5-GFP-
RRBP1 (Addgene #92150) and inserted into
pcDNA3.1(+)-based mCherry expression vec-
tor with tags at its C terminus. PCR-amplified
RRBP (amino acids 1–150) was inserted in
pGEX4T-1 vector by ligation-based cloning.
Plasmids for lentivirus transduction [e.g.,
mEmerald-Sec61b, mScarlet-Sec61b, 3xHA-
APEX2-Rab5A, Halo-Rab5A, ERM-GFP11-p2a-
GFP1-10-Rab5A, ERM-2xHA-FRB, ERM-2xHA-
(EAAAK)x9-FRB, mRFP-FKBP-Rab5A, Rab5A-
mRFP-FKBP, V5-RRBP1 1-150aa WT, V5-RRBP1
del PB1, V5-RRBP1 del PB3, V5-RRBP1 del TM]
were generated with the pLVX-TetONE puro
vector (Takara, Catalog 631849) by PCR, re-
petitive oligo annealing, or inverse PCR with
Gibson assembly. All constructs were verified
by double-stranded DNA sequencing.
Primary antibodies for immunoblots
Anti-GAPDH (glyceraldehyde-3-phosphate de-
hydrogenase; mouse, Sigma-Aldrich, G8795,
1:1000), anti-MTM1 (goat, Invitrogen, PA5-
17972, 1:250), anti-Calnexin (rabbit, Abcam,
ab75801, 1:2000), anti-calreticulin (rabbit,
Thermo Fisher, PA 3-900, 1:1000), anti-Reticulon4
(Nogo) (mouse, Santa Cruz, sc-271878, 1:1000),
anti-RPL26 (rabbit, Proteintech, 17619-1-AP,
1:1000), anti-RPL7 (rabbit, Proteintech, 14583-
1-AP, 1:1000), anti-EEA1 (mouse, BD biosciences,
610456, 1:500), anti-Histone H1 (rabbit, Abcam,
ab17729, 1:1000), anti-GST (mouse, Thermo
Fisher, MA4-004, 1:1000), anti-KTN1 (rabbit,
Proteintech, 19841-1-AP, 1:500), anti-RRBP1
(rabbit, Proteintech, 22015-1-AP, 1:500), anti-
Clim63 (mouse, Enzo Life Sciences, ENZ-
ABS669-0100, 1:2000), anti-TOM20 (mouse,
Santa Cruz, sc-17764, 1:200), anti-p-DRP1 (S616)
(rabbit, Cell signaling, 3455S, 1:500), anti-
DRP1 (rabbit, Abcam, ab184247, 1:500), anti-
MFN1 (rabbit, Proteintech, 13798-1-AP, 1:500),
anti-MFN2 (rabbit, Proteintech, 12186-1-AP,
1:500), anti-OPA1 (rabbit, Proteintech, 27733-
1-AP, 1:500), anti- GRP78/BIP (rabbit, Proteintech,
11587-1-AP, 1:3000), anti-CHOP (rabbit, Pro-
teintech, 15204-1-AP, 1:500), anti-LC3-II (mouse,
MBL, M152-3, 1:200), anti-Cleaved PARP(Asp214)
(rabbit, Cell Signaling, 9541S, 1:1000), anti-
Cleaved Caspase3(Asp175) (rabbit, Cell Signal-
ing, 9661T, 1:250), anti-V5 (mouse, Invitrogen,
P/N 46-0705, 1:1000), anti-Phospho-p70 S6
Kinase (Thr389) (rabbit, Cell signaling, 9205L,
1:500), anti-p70 S6 Kinase Antibody (rabbit, Cell
Signaling, 9202L, 1:1000), anti-ATG5 (rabbit,
Proteintech, 10181-2-AP, 1:1000).
Primary antibodies for
immunofluorescence
Anti-calreticulin (rabbit, Thermo Fisher, PA
3-900, 1:200), anti-calreticulin (rabbit, Abcam,
ab92516, 1:200), anti-Rab5A (mouse, BD
biosciences, 610724, 1:100), anti-GFP (mouse,
Invitrogen, A-11120, 1:500), anti-LC3-II (mouse,
MBL, M152-3, 1:100), anti-HA (mouse, Santa
Cruz, sc-7392, 1:500), anti-RFP (rabbit, MBL,
PM005, 1:400), anti-LAMP1 (mouse, BD bio-
sciences, 555798, 1:1000), anti-LAMP1 (rabbit,
Cell signaling, 9091P, 1:300), anti-V5 (mouse,
Invitrogen, P/N 46-0705, 1:200), anti-TOM20
(mouse, Santa Cruz, sc-17764, 1:500), anti-
TOM20 (rabbit, Abcam, ab186734, 1:500),
Jang et al., Science 378, eabq5209 (2022)
16 December 2022
9 of 16
Corrected 7 September 2023. See full text. RES EARCH | R E S E A R C H A R T I C L E
anti-TFEB (rabbit, Biomol, A303-673A, 1:200),
Streptavidin, Alexa Fluor 647 conjugate (Thermo
Fisher, S21374).
siRNAs
The small interfering RNAs (siRNAs) used
were: Scrambled siControl 5′-CGUACGCG-
GAAUACUUCGA-3′, or Sigma MISSION Uni-
versal Negative Control (SIC001), siMTM1-1
5′-GATGCAAGACCCAGCGTAA-3′, siMTM1-2
5′-TATGAGTGGGAAACGAAATAA-3′, siMTM1-
SP (ON-TARGETplus SMARTpool, Dharmacon,
L-008036-00-0005), siMTMR1-1 5′-GAGA-
TAGTGTGCAAGGATA-3′, siMTMR1-2 5′-CG-
CTGATACCAACAAGACAAA-3′, siMTMR2-1
5′-GGACATCGATTTCAACTAA-3′, siMTMR2-2
5′-CGGCCAAGTGTTAATGCTGTT-3′, siMTMR2-
3 5′-GTAGAAAGTCTTCGGAATTTA-3′ ,
siMTMR6-1 5′-GGACTACAAGATTTGTGAA-3′,
siMTMR6-2 5′-CGGGACTACAAGATTTGTGAA-
3′, siMTMR12-1 5′-CCAGGTGAACAGCTGCTTT-3′,
siMTMR12-2 5′-GCAAGAGAATTAGCAAACTTA-
3′, siMTMR12-3 5′- CGCTTCAAACATCAACGA-
CAA-3′, siKTN1 (ON-TARGETplus SMARTpool,
Dharmacon, L-010605-00-0005), siERLIN1
(ON-TARGETplus SMARTpool, Dharmacon,
L-015639-01-0005), siERLIN2 (ON-TARGETplus
SMARTpool, Dharmacon, L-017943-01-0005),
siRRBP1 (ON-TARGETplus SMARTpool,
Dharmacon, L-011891-02-0005), siOSBPL8
(ON-TARGETplus SMARTpool, Dharmacon,
L-009508-00-0005), siITPR2 (ON-TARGETplus
SMARTpool, Dharmacon, L-006208-02-0005),
siARL8B (ON-TARGETplus SMARTpool,
Dharmacon, L-020294-01-0005), siRTN4
(ON-TARGETplus SMARTpool, Dharmacon,
L-010721-00-0005), siRab10 (ON-TARGETplus
SMARTpool, Dharmacon, L-010823-00-0005),
siProtrudin 5′-CTTCTTGATCCAGCTGGAGG-
3′, siATG5 (ON-TARGETplus SMARTpool,
Dharmacon, L-004374-00-0005).
Cell culture
HeLa, human embryonic kidney 293 T
(HEK293T), and Cos7 cells were obtained
from ATCC. Cells were cultured in Dulbecco’s
modified Eagle’s medium (DMEM) high
(4.5 g/liter glucose, Thermo Fisher, 41965062)
containing 10% fetal bovine serum (FBS;
Thermo Fisher, 10270106) and 50 units/ml
penicillin, 50 mg/ml streptomycin (Thermo
Fisher, 15070063). Cells were routinely tested
for mycoplasm contamination. The human
myoblast cell line KM1288 was derived from
the deltoid muscle of a patient with XLCNM
and carries a missense genomic mutation
(c.205<T) in exon 4 of the MTM1 gene (20).
The human myoblast cell strain NL15 was
derived from the quadriceps of a patient with
XLCNM and carries the missense genomic
mutation R241C in the MTM1 gene (21). The
control myoblast cell line AB1190 was derived
from the paravertebral muscle of a healthy
individual. AB1190 and KM1288 were immor-
talized as described before (74). Myoblasts were
cultured in the homemade medium: 1 volume
medium 199 (Thermo Fisher, 41150020) + 4
volumes DMEM high, supplemented with
20% FBS, Fétuin (25 mg/ml, Thermo Fisher,
10344026), human epidermal growth factor
(5 ng/ml, Thermo Fisher, PHG0311), basic
fibroblast growth factor (0.5 ng/ml, Thermo
Fisher, PHG0026), insulin (5 mg/ml, Sigma,
91077C-1G), and dexamethasone (0.2 mg/ml,
Sigma, D4902-100mg). To induce starvation, cells
were washed with prewarmed Earle’s balanced
salt solution (EBSS; Thermo Fisher, 24010043)
five times for 10 s each and then incubated
in EBSS in 5% CO2 at 37°C for 2 hours, unless
indicated otherwise. This washing step is im-
portant to fully remove remaining nutrients
and growth factors. For steady-state (fed) con-
ditions, cells were incubated overnight in fresh
complete DMEM medium supplemented with
10% FBS. For the experiment in fig. S1M, EBSS
was supplemented, as indicated, with 5% di-
alyzed FBS (One Shot format, Gibco, A3382001)
or insulin (final at 5 mg/ml) or glucose (final at
4.5 g/liter) or sodium pyruvate (final at 1 mM)
(Gibco, 11360070) or MEM essential amino
acids (50x) solution (Sigma, M5550) (final 1x
dilution) or MEM solution or nonessential
amino acids (100x) (Gibco, 11140050) (final 1x
dilution) or L-glutamine (final at 4 mM) (Gibco,
25030081).
CRISPR-Cas9–mediated genome engineering
Guide RNAs targeting genomic human MTM1
exon 2 (sgMTM1: 5′ AGTTGATGCAGAAGC-
CATCC 3′) was cloned into Lenti-CRISPRv2
(Addgene plasmid # 52961). HeLa cells were
transfected using FuGene-6 as a transfection
reagent. Cells were t selected with puromycin
(2 mg/ml) for 72 hours. Surviving cells were
diluted and plated into 96-well plates with a
density of 1 cell per well. Expanded colonies
were screened by immunoblotting using anti-
MTM1 antibodies.
Generation of doxycycline-inducible stable
cell lines
Lentivirus was generated by transient trans-
fection of HEK293T cells seeded in 10-cm cell
culture plates at 80 to 90% confluency with
pCMV delta R8.2 (3.5 mg), VSV-G (0.5 mg), and
pLVX-TetONE puro-based constructs (4 mg)
combined with 16 ml of JetPrime in 400 ml
JetPrime buffer. After 16 hours of transfection,
cells were replenished with 7 ml fresh DMEM.
After 48 hours, the supernatant was collected,
and cellular debris was removed by centri-
fugation (3000 rpm, 20 min). Viral super-
natant (4 to 5 ml) was added to the cells with
polybrene (Merck, Cat.#TR-1003-G) at 10 mg/ml.
Cells were incubated with virus for 16 hours
and then replenished with fresh DMEM con-
taining puromycin (2 mg/ml) followed by selec-
tion for 2 to 3 days. After selection, cells were
stabilized for 3 to 4 days in culture without
puromycin. For the induction of protein ex-
pression, doxycycline (1 mg/ml) was added for
16 hours.
Transient transfection
HEK293T and HeLa cells and myoblast cells
were transfected with plasmids using jetPRIME
(PolyPlus, 101000001) or FuGene-6 (Promega,
E2691) or ViaFect (Promega, E4981) according
to the manufacturer’s protocol, respectively
[e.g., DNA (micrograms): reagent (microliters)
ratio of 1:2]. After 18 to 24 hours of transfection,
cells were further treated and analyzed. For
siRNA transfection, cells were transfected with
the indicated siRNA (20 nM) using jetPRIME
according to the manufacturer’s protocol. After
48 hours of transfection, cells were further
treated and analyzed.
Dyes and pharmacological inhibitors
Dyes
TMRE (tetramethylrhodamine ethyl ester per-
chlorate): Cell signaling, Mitochondrial mem-
brane potential assay kit, #13296, 200 nM).
Lysotracker Red DND-99: Thermo Fisher,
L7528, 1 mM for 45 min. Note that for fixed cell
samples, buffer containing detergents should
not be used, in order to preserve lysotracker
staining. BODIPY 493/503: Thermo Fisher,
D3922, 2 mM for 20 min in serum free media.
No detergent-containing buffers should be used.
Before treatment, cells were washed twice with
serum-free media. BODIPY 558/568 C12: Thermo
Fisher, D3835, 1 mM for 20 min in serum-free
media. Before labeling, cells were washed twice
with serum-free media. For TOM20 antibody
co-staining, fixed samples were permeabilized
with 20 mM digitonin for 5 min.
MitoTracker Deep Red FM: Thermo Fisher,
M22426, 100 nM for 30 min in full medium.
CellMask Deep Red Plasma membrane Stain:
Thermo Fisher, C10046, fixed cells were stained
for 20 min at 1:2000 dilution. Halo-tag li-
gands (CA-JF646) were chemically synthesized
as described in (75) and used at 100 nM for 10
to 16 hours.
Pharmacological inhibitors
VPS34-IN1 (Selleckchem, S7980, 1mM), apilimod
(Echelon Biosciences, B-0308, 50nM), rapalog
(Takara, A/C Heterodimerizer 635056, 0.5 mM),
nocodazole (Sigma, M1404, 5mM), dimethyl sulf-
oxide (DMSO; Sigma, D2650), thapsigargin
(Sigma, T9033, 3mM), Torin1 (Tocris, Cat. No.
4247, 1 mM), Z-VAD-FMK (Tocris, Cat. No.
2163, 25 mM).
RNA isolation, RT-PCR, and qRT-PCR
Total RNA was isolated using the RNeasy Plus
Mini Kit (Qiagen) according to the manufac-
turer’s instructions. Quantification of RNA
yields was done with a multimode microplate
reader (SPECTROstar Nano from BMG Labtech).
Jang et al., Science 378, eabq5209 (2022)
16 December 2022
10 of 16
Corrected 7 September 2023. See full text. RES EARCH | R E S E A R C H A R T I C L E
A reverse transcription kit (SuperScript IV;
Invitrogen) was used to reverse transcribe
RNA (800 ng) in a 20 ml reaction using oligo(dT)
and random hexamer Primers. Reverse tran-
scription polymerase chain reaction (RT-PCR)
was performed with specific primers set against
each gene with indicated cycles. For quan-
titative reverse transcription polymerase chain
reaction (qRT-PCR), SYBR Green Master Mix
(BioRad) was used according to the manufac-
turer’s protocol with BioRad CFX Connect.
qRT-PCR primers
GAPDH Fwd: 5′-CTTCGCTCTCTGCTCCTCCT-3′;
Rev: 5-GTTAAAAGCAGCCCTGGTGA-3′, MTM1
Fwd: 5′-GTTTGAGATCCTCACGAGATACG-3′;
Rev:5′-GTCCATCCATCCACGTTAAACTT-3′,
MTMR1 Fwd: 5′-CCTTGATGTTCCCCTTGGAGT-
3′; Rev:5′-GTGCCTGTCCGTTAGAAAGAG-3′,
MTMR2 Fwd: 5′-GTGGAAAGCGAAGCAAA-
GAAG-3′; Rev:5′-CTTGGCCGGGCATCAAATATAA-
3′, MTMR3 Fwd: 5′-GACTGAACAACGCAATCC-
GAC-3′; Rev:5′-CCTTGAAGTTACATGCTCCCC-3′,
MTMR4 Fwd: 5′-CCAAGCCAAGGATCTGTTCCC-
3′; Rev:5′-GCCGGTAGTTAGAGATGGCAA-3′,
MTMR5 Fwd: 5′-CGACCACACGGAGGTGTTC-
3′; Rev:5′-GGTTCCCAATCACGTTCTCCA-3′,
MTMR6 Fwd: 5′-GTTCCCCGGATAGCAAGCAAA-
3′; Rev:5′-GTGGCTGACTACATCGACAAAT-3′,
MTMR7 Fwd: 5′-TCCGCTTGGTAGATCGAGTGT-
3′; Rev:5′-TTTTCCACGAATATGACATGGGT-3′,
MTMR8 Fwd: 5′-TGCACTCCATCACATTGCCA-
3′; Rev:5′-GGCACACAAGGTCAGAATCTAA-3′,
MTMR9 Fwd: 5′-TGAAGCTCTTCGGAAGG-
TAGC-3′; Rev:5′-GTGGCTGACCACTTCGCATAA-3′,
MTMR10 Fwd: 5′-ATCCACTTGCCTTCCAGAA-
TACA-3′; Rev:5′-CACAAGAGCACTGCCGTTAGA-3′,
MTMR11 Fwd: 5′-GCTGCTCAGAGTTGGTTTTGA-
3′; Rev:5′-CCCCGAATACTGTTGGGCTT-3′,
MTMR12 Fwd: 5′-GGCTCCTAAACTGCTTAAA-
CGA-3′; Rev:5′-GTTGCCTTTGGTCCGTTCCA-3′,
MTMR13 Fwd: 5′-TCATCGTGGTAGGCTAT-
GACC-3′; Rev:5′-CCAGGCTGACAAAACAACTCA-3′,
MTMR14 Fwd: 5′-GGAGTTCTCCCGGACTCAGTA-
3′; Rev:5′-AACAGTAGTCTCGGCCAAACA-3′.
Immunoblot analysis
Cells were lysed with radioimmunoprecipita-
tion assay (RIPA) buffer (50 mM Tris–Cl pH 7.5,
150 mM NaCl, 1% NP-40, 0.5% sodium deoxy-
cholate, 0.1% SDS) containing protease and
phosphatase inhibitors. Lysates were incu-
bated for 10 min on ice before centrifugation
at 17,000g for 10 min at 4°C. Protein concen-
tration was measured by Bradford or bicin-
choninic acid (BCA) assays. Cell lysates in
Laemmli sample buffer were boiled for 10 min.
Between 20 and 40 mg of protein was resolved
by SDS–polyacrylamide gel electrophoresis
(SDS-PAGE). Immunoblotting was done on
nitrocellulose membranes. Membranes were
incubated with the indicated primary anti-
bodies at 4°C overnight. The next day, bound
primary antibodies were detected by incu-
bation with IRDye 680/800CW-conjugated or
horseradish peroxidase (HRP)–conjugated sec-
ondary antibodies via the Odyssey Fc Imaging
system (LI-COR Biosciences).
Light microscopy
Immunocytochemistry
Cells were seeded on Matrigel-coated coverslips
(BD/Corning), fixed in 4% paraformaldehyde
(PFA) for 15 to 20 min at room temperature (RT).
Cells were washed three times with phosphate-
buffered saline (PBS) before incubation with
3% bovine serum albumin (w/v) in 0.3% PBST
(Triton X-100) for 20 min. Using the same buffer,
the cells were incubated with primary antibodies
for 2 hours at RT, washed three times with 0.3%
PBST, and then incubated with secondary anti-
bodies for 1 hour at RT. After washing with 0.3%
PBST, cells were mounted with Hoechst 33258
(Invitrogen, H3569, 1:2000) to counter stain
nuclei. Images were acquired on Zeiss 710 or
780 Laser Scanning Confocal Microscopes
using ZEN.
Note for the specific staining conditions:
For calreticulin staining, cells were fixed with
37°C 4% PFA for 34 min at 37°C. For Rab5A
staining, cells were fixed with 37°C 4% PFA for
8 min at 37°C. For LC3-II staining, cells were
permeabilized with 20 mM digitonin for 5 min
at RT. After permeabilization, all subsequent
procedures were done in PBS.
PI(3)P staining
Cells were fixed in 2% PFA for 15 min at RT,
washed twice in PBS with 50 mM NH4Cl, and
permeabilized with 20 mM digitonin in buffer
A (20 mM PIPES pH 6.8 with NaOH, 137 mM
NaCl, 2.7 mM KCl) for 5 min. Note that per-
meabilization is critical for successful PI(3)P
staining, so depending on the batch of digitonin,
one might need to optimize the concentration
and incubation time.
Cells were washed three times in buffer A
before addition of purified GFP (or mCherry)-
2xFYVEHrs at 0.25 mg/ml in buffer A with 5%
normal goat serum, 50 mM NH4Cl for 1 hour.
Samples were washed and decorated with anti-
bodies against GFP or RFP in buffer A with 5%
normal goat serum, 50 mM NH4Cl for 2 hours.
Cells were washed three times in buffer A, in-
cubated for 1 hour with secondary antibodies
in buffer A with 5% normal goat serum, 50 mM
NH4Cl. Samples were washed three times with
buffer A, cells were postfixed for 5 min in 2%
PFA, washed three times in PBS, and mounted
with Hoechst 33258. Fluorescent sum inten-
sity was measured from individual cells and
normalized to the level of PI(3)P detected in
fed cells set to 1.
Live-cell imaging
Cells (6 × 103 to 8 × 103) were seeded on
Matrigel-coated 8-well chamber slides (ibidi,
80827). The next day, live-cell imaging was
carried out using Nikon-CSU Yokogawa Spin-
ning disk (CSU-X1) microscope equipped with
an EMCCD Camera (Andor AU-888) at 37°C in
the presence of 5% CO2. Images were acquired
using Nikon Elements.
Image analysis and quantification
Acquired images were first segmented to re-
solve individual cells by manually drawing each
cell boundary using the Fiji freehand selec-
tion tools combined with Clear Outside option.
Individual cell image files were saved in the in-
put folder and batch processed in CellProfiler
[v4.1.3 (76)] or in Fiji. CellProfiler pipeline
modules used in this study were posted at
https://cellprofiler.org/published-pipelines.
Sheet-to-tubular ER ratio
Total ER (calreticulin or mEmerald-Sec61b)
segmentation was done using a three-class
Otsu threshold followed by conversion into
binary images. ER area was then measured
using the MeasureObjectSizeShape module
in CellProfiler. From the same image, sheet
ER was segmented in a similar manner except
using the Minimum Cross-Entropy threshold.
Depending on signal-to-noise ratio and im-
age quality, threshold settings (e.g., Thresh-
old smoothing scale, correction factor) were
adjusted using the IdentifyObjects module in
CellProfiler. Sheet ER area was then divided
by total ER area using CalculateMath module
in the same pipeline of CellProfiler. Calculated
sheet ER area/total ER area ratio was defined
as “Sheet ER ratio.”
Mean area and number of mitochondria
per ROI
Regions of interest (ROIs; e.g., 15 mm by 15 mm)
were randomly generated from individual
cells in Fiji, exported to CellProfiler, and
filtered with a Gaussian filter module with
1 or 2 sigma. ROIs were then processed using
the EnhanceOrSuppressFeatures module with
the following options: Operation = “Enhance,”
Feature type = “Neurites,” Enhancement
method = “Tubeness,” Smoothing scale =
“1.” After binarization, mitochondria were
segmented using the Robust Background
Threshold in the IdentifyObjects CellProfiler
module. The number and area of individual
mitochondria per ROI were calculated using
MeasureObjectSizeShape module in CellProfiler.
Lysosome position
Images were processed in CellProfiler using
the EnhanceOrSuppressFeatures module with
the following options: Operation = “Enhance,”
Feature type = “Speckles,” Feature size = “4-6.”
Lysosome were then segmented using Robust
Background threshold in the IdentifyObjects
module. The distance between individually
identified lysosomes from the cell centroid
was calculated using the RelateObjects module
Jang et al., Science 378, eabq5209 (2022)
16 December 2022
11 of 16
Corrected 7 September 2023. See full text. RES EARCH | R E S E A R C H A R T I C L E
in CellProfiler. The standard deviation of the
distance of lysosomes from the centroid define
the extent of lysosomal dispersion or clustering.
Early endosome distribution
Rectangular ROIs (80 pixels by width pixels rang-
ing from cell center to cell boundary) were ran-
domly selected from individual cells in Fiji and
processed using the EnhanceOrSuppressFeatures
module with the following options: Operation =
“Enhance,” Feature type = “Speckles,” Feature
size = “4-6.” A Gaussian filter with sigma = 1 or
2 was applied in CellProfiler. Early endosomes
and lysosomes were segmented using two-
classes Otsu and robust background thresholds
using the IdentifyObjects module. To compare
the partitioning of early endosomes and lyso-
somes between the cell center and the periphery,
segmented individual objects were shrunken to
a dot using the ExpandOrShrinkObjects module
with the “Shrink objects to a point” option in
CellProfiler and converted into binary image
files. The MeasureObjectIntensityDistribution
module was then used to measure the inten-
sity distribution from each object’s center to
its boundary within a set of bins with the
following options: Bin = “5,” Measurement =
“Fraction at distance.” The relative intensity
fraction among the five bins was calculated
from the binary images in CellProfiler. This
intensity value indirectly represents the frac-
tion of endosomes per bin.
Fluorescent imaging of ER–early
endosome contacts and dynamics
Tubular ER segmentation of 100 pixel by
100 pixel ROIs from the cell periphery were
generated using the Minimum Cross-Entropy
threshold option in the IdentifyObjects mod-
ule, and the ExpandOrShrinkObjects module
with Operation = “Skeletonize each object,”
and subsequent ExpandOrShrinkObjects
module with Operation = “Expand objects
by a specified number of pixels” setting with
1-pixel expansion in CellProfiler. Early endo-
somes and lysosomes from the same image
were segmented as described above. Segmented
individual objects were then shrunken to a dot
using the ExpandOrShrinkObjects module with
“Shrink objects to a point” option, followed
by Operation = “Expand objects by a specified
number of pixels” setting with 1-pixel expansion.
To calculate the number of endosomes that
contact the ER, the MeasureObjectNeighbors
module in CellProfiler was operated with
the following options: Method to determine
neighbors = “Within a specified distance” set-
ting distance = 1 pixel. If the processed tubular
ER and endosomes overlapped with each other
at least by 1 pixel, a contact was scored.
Split GFP ER–early endosome contact sensor
To quantitatively determine the number ER–
early endosome membrane contacts by measur-
ing the number of GFP puncta, images were
blurred with a Gaussian filter (sigma = 1 or 2,
depending on the signal-to-noise ratio), back-
ground was subtracted with a rolling ball ra-
dius of 50 pixels, and segmented using the Find
Maxima tool with prominence = 10 in Fiji.
STED image analysis of ER–early
endosome contact
ROIs (8 mm by 8 mm) were randomly gen-
erated from individual cells in Fiji, ex-
ported to CellProfiler, processed using the
EnhanceOrSuppressFeatures module with
the following options: Operation = “Enhance,”
Feature type = “Speckles,” with subsequent
Gaussian filter module with 1. This allowed the
segmentation of early endosomes via robust
background thresholds using the IdentifyObjects
module. Segmented early endosomes were
then converted to binary images using the
ConvertObjectsToImage module. ER images
were filtered with a Gaussian filter module with
1 and converted to binary images using the
Threshold module. The MeasureImageOverlap
module was used to analyze the fractional
overlap between the ER and early endosomes.
Lipid droplets (BODIPY 493/503, Red C12)
Lipid droplet images were blurred with a
Gaussian filter (sigma = 1), enhanced using
the EnhanceOrSuppressFeatures module
(Operation = “Enhance”, Feature type =
“Speckles” with size = 10), and then segmented
by the three-classes Otsu threshold option in
CellProfiler. Segmented objects were further
filtered on the basis of their measured sizes
and intensities using the FilterObjects module
in CellProfiler to minimize spurious detec-
tions. To measure the volume of lipid droplets,
z-stacked confocal images were analyzed using
the 3D objects counter plug-in in Fiji.
Pearson coefficient
To quantitatively assess colocalization between
two channels, two or three randomly chosen
100 pixel by 100 pixel ROIs from the cell pe-
riphery were blurred with a Gaussian filter
(sigma = 1). Pearson’s coefficients were calcu-
lated using the Coloc2 plugin in Fiji or the
MeasureColocalization module with corre-
lation option in CellProfiler.
Super-resolution microscopy
Stimulated emission depletion (STED) images
were taken with either a Leica SP8 TCS STED
microscope (Leica Microsystems) for fixed
samples or a STEDYCON (Abberior Instruments
GmbH, Göttingen) for live cell imaging. Leica
SP8 TCS STED microscope was equipped with
a pulsed white-light excitation laser [WLL;
∼80-ps pulse width, 80 MHz repetition rate
(NKT Photonics)] and a STED laser for de-
pletion (775 nm). The system was controlled
by the Leica LAS X software. For images of
mEmerald-Sec61b, fixed cells were further
stained for GFP-Booster nanobody conjugated
to Atto647N (Chromotek, gba647n-100, 1:200).
Fluorophore specific excitation (Ex.) and emis-
sion filter (EmF.) settings: Atto647N (Ex.:
640 nm; EmF.: 650 to 700 nm). Time-gated
detection was set from 0.3 to 6 ns. The fluo-
rescence emission signal was collected by hy-
brid detectors (HyD). Images were acquired
with a HC PL APO CS2 100×/1.40 NA oil ob-
jective (Leica Microsystems) in a pixel size of
18.9 nm by 18.9 nm. STEDYCON was mounted
on the Nikon Eclipse Ti research microscope
equipped with a Plan APO 100x/1.45 NA oil
objective (Nikon). Excitation laser with a wave-
length of 640 nm was used for the Halo ligand
JF647 fluorescence. The wavelength of the
STED laser was 775 nm. Live cells expressing
Halo-Sec61b were imaged at 37°C degree with
a pixel size of 20 nm by 20 nm.
Focused ion beam milling scanning electron
microscopy (FIB-SEM)
Cells were fixed with 2% glutaraldehyde (GA)
in PBS, washed with 0.1 M cacodylate buffer,
and embedded in Durcupan following a mod-
ified rOTO protocol with 1% (w/v) OsO4/
1.5% (w/v) K3Fe(CN)6 in 0.1 M cacodylate buf-
fer (pH 7.4); 0.2% (w/v) thiocarbohydrazide in
water; 1% OsO4 in water, 1% aqueous uranyl
acetate with corresponding washes in be-
tween and subsequent acetone dehydration
and resin infiltration. For polymerization, cov-
erslips were mounted onto pre-polymerized
resin blocks and placed into a heating cupboard
for 48 hours. After polymerization, coverslips
were removed by liquid nitrogen treatment,
and blocks were glued to SEM aluminum stubs
with conductive silver epoxy and carbon sputter
coating (≈30 nm).
Helios 5CX FIB-SEM autoslice and view
workflow was used to section and image the
embedded cells. FIB was run at 30 kV, 0.23 nA,
and 10-nm milling step. Cross section images
were scanned at 1.5 to 2 kV and 86 pA to 0.17 nA.
Dwell time was 5 ms at 3.37-nm pixel resolution
and ICD detection. The resulting 3D stacks
were binned to isotropic 10-nm voxel resolution.
Alignment, segmentation of mitochondria, and
analysis of ER networks were performed using
Microscopy Image Browser, 3D visualization
was done using Imaris. To access how con-
tinuous or fenestrated the ER is, the length
of ER cisternae was measured in several dis-
tantly positioned 2D sections from FIB-SEM
stacks. Sections for measurements from each
cell were randomly chosen and were at least
1 mm apart from each other.
For 2D analysis of mitochondrial cristae
length, ultrathin sections of cells were collected
onto carbon-coated coverslips and imaged
with a Helios 5CX SEM. The length of cristae
was measured using Image J and normalized
per mitochondrial cross-section perimeter.
Jang et al., Science 378, eabq5209 (2022)
16 December 2022
12 of 16
Corrected 7 September 2023. See full text. RES EARCH | R E S E A R C H A R T I C L E
For 2D analysis of ER–early endosome mem-
brane contacts, 5 nm BSA-gold was allowed to
internalize into HeLa cells for 5 min to reach
early endosomes. Cells were washed before fix-
ation with 2% GA. Cells were processed for EM
analysis as described above. Ultrathin sections
mounted onto EM grids were imaged using a
Zeiss 900 transmission electron microscope
and ER–early endosome contact sites were
measured in Fiji.
Volume fraction analysis was performed
on ultrathin sections mounted on wafers and
scanned with a directional back-scatter de-
tector at 2 kV, 0.34 nA, 5-ms dwell time. A
measurement grid (1 mm2 per point) was super-
imposed over cellular profiles and the refer-
ence volume (cytoplasm), as well as the relative
volumes of the ER, and mitochondria were es-
timated according to stereological principles.
APEX2 proximity labeling experiments
HeLa cells stably expressing doxycycline-
inducible 3xHA-APEX2-Rab5A were pretreated
with doxycycline for one day. Cells were then
incubated with 500 mM phenol-biotin (Sigma,
SML2135) at 37°C for 30 min. To enable ef-
ficient biotinylation, cells were treated with
1 mM hydrogen peroxide and incubated at
RT for 1 min. The reaction was terminated by
removing the medium and the addition of ice-
cold quenching buffer (10 mM sodium azide,
10 mM sodium ascorbate, 5 mM Trolox in PBS).
Cells were washed with PBS, lysed with RIPA
buffer (see above section on immunoblotting),
and centrifuged at 17,000g for 10 min at 4°C.
The protein concentration of the resulting
supernatant was determined using the Bradford
assay. Then, 150- to 200-mg samples were saved
as whole cell lysate (WCL) control, while 1000
or 1500 mg of lysates were incubated with 40 ml
of Streptavidin Magnetic Beads (Thermo Fisher,
88817) overnight, nutating at 4°C. Magnetic
beads were washed twice with RIPA buffer,
once with 1 M KCl, once with 0.1 M Na2CO3,
once with 2 M Urea in 10 mM Tris-HCl, pH
8.0, two times with RIPA buffer, and finally
eluted with Laemmli sample buffer in the pres-
ence of 2 mM biotin. WCL and eluate samples
were analyzed by SDS-PAGE gels and immu-
noblotting. For quantification, the band inten-
sities of eluted endogenous MTM1 or EEA1
were normalized to the band intensity of self-
biotinylated Rab5A in Streptavidin-HRP blots
(Fig. 2B and fig. S4B).
ER isolation and biotinylated protein collection
for liquid chromatography combined with mass
spectrometry (LC-MS)
HeLa WT or MTM1 KO cell line stably ex-
pressing doxycycline-inducible 3xHA-APEX2-
Rab5A were seeded in 15-cm dishes with 17 ×
105 cells (WT = 6 dishes, KO = 3 dishes) in the
presence of doxycycline for 16 hours before
biotinylation. The next day, cells at ∼85% con-
fluency were treated with 500 mM phenol-biotin
and hydrogen peroxide to induce biotinylation,
quenched, washed once with isolation buffer
(225 mM mannitol, 75 mM sucrose, 30 mM
Tris-HCl pH 7.4, 0.1 mM EGTA), and, finally,
scraped into 1 ml of the same buffer resulting
in ∼2 ml total volume. Half a milliliter of this
sample was saved (WCL) as a control. The re-
maining 1.5 ml was homogenized using a 2 ml
Dounce homogenizer (∼100 to 120 strokes) on
ice in the presence of protease and phosphatase
inhibitors. To isolate light membranes con-
taining the ER, the material was centrifuged
twice at 600g for 2 min each. The pellet (F1)
including nuclei and cell debris was resus-
pended in 150 ml RIPA buffer. The remaining
supernatant was centrifuged twice at 7000g
for 20 min each. The resulting pellet (F2) in-
cluding the mitochondrial fraction was resus-
pended in 150 ml RIPA buffer and further
fractionated by centrifugation at 20,000g for
60 min to pellet light ER membranes (F3). This
ER-enriched membrane fraction was resus-
pended in150 ml RIPA buffer. Protein concen-
tration was determined using the Bradford
assay. To isolate biotinylated proteins, 130 mg
of WCL or F3 fraction was rotated in the pres-
ence of Streptavidin Magnetic Beads. Eluted
biotinylated proteins were reduced (5 mM
dithiothreitol, 30 min at 55°C), alkylated (15 mM
iodacetamide, 20 min at RT in the dark), and
submitted to LC-MS analysis (samples from
two independent experiments).
LC-MS analysis of affinity-purified samples
Proteins were loaded on SDS-PAGE and sub-
jected to in-gel digestion. In brief, gel bands
were excised, and protein digestion was carried
out using trypsin at an enzyme-to-protein ratio
of 1:100 (w/w) at 37°C overnight. LC-MS mea-
surement was achieved by reverse phase high-
performance liquid chromatography (RP-HPLC)
on a Thermo Scientific Dionex UltiMate 3000
system connected to a PepMap C-18 trap-
column [0.075 mm by 50 mm, 3-mm particle
size, 100-Å pore size (Thermo Scientific)] and a
200 cm mPAC column was used (PharmaFluidics,
Ghent, Belgium) with 750 or 350 nl/min flow
rate with a 120-min gradient. Samples were
analyzed on an Orbitrap Fusion mass spectrom-
eter. MS1 scan were acquired in the Orbitrap
with a range of 375 to 1500 m/z, mass resolution
of 120,000, automatic gain control (AGC) target
value of 4 × 105, and 50-ms maximum injection
time. MS2 scans were acquired in the ion trap
with an AGC target value of 1 × 104 and 35-ms
maximum injection time. Precursor ions with
charge states 2 to 4 were isolated with an iso-
lation window of 1.6 m/z and 40 s dynamic ex-
clusion. Precursor ions were fragmented using
higher-energy collisional dissociation with 30%
normalized collision energy. Analysis of the raw
data was done with MaxQuant (MQ) software
version 1.6.2.6. MaxQuant standard settings
were kept as default. In the search parameters,
two missed cleavage sites were included, the
fixed modification was set to cysteine carba-
midomethyl modification, and variable modifi-
cations to methionine oxidation and N-terminal
protein acetylation. The peptide mass tolerance
was set to 4.5 parts per million (ppm) for MS1
scans and 20 ppm for MS2 scans. Match be-
tween runs option was enabled. The database
search was done using Andromeda against
the Human UniProt/Swiss-Prot database with
common contaminants. The false discovery
rate (FDR) was set to 1% for both peptide and
protein level. Protein quantification was done
on the basis of at least two razor and unique
peptides. Label-free quantification and iBAQ
calculation were enabled. Statistical analysis
was done on the “ProteinGroups” table with
Perseus version 1.6.7.0. Proteomics data have
been deposited to the ProteomeXchange Con-
sortium via PRIDE and are available via
ProteomeXchange with identifier PXD033846.
Quantitative whole-cell proteomics with
TMT labeling
Fed or starved (2 hours) WT or MTM1 KO HeLa
cells (one 10-cm dish at 90% confluency) col-
lected from three independent experiments
were lysed in 300 ml 8 M urea-lysis-buffer in
50 mM triethylammonium bicarbonate (TEAB)–
containing protease inhibitor cocktail (Roche)
and 0.5 ml Benzonase Nuclease HC (Millipore).
Samples were applied to a Bioruptor Pico
(Diagenode) for 10 cycles (30 s on/30 s off) at
4°C. Samples were then incubated for 30 min
at 25°C. Proteins were reduced with 5 mM Tris
(2-carboxyethyl)phosphin-hydrochloride (TCEP)
and alkylated with 40 mM chloroacetamide
(CAA) for 60 min at 37°C in the dark. Protein
digestion was carried out using Lys C at an
enzyme-to-protein ratio of 1:100 (w/w) at 37°C
for 3 hours. After diluting to 2 M urea with
50 mM TEAB buffer, the digestion was con-
tinued with trypsin at an enzyme-to-protein
ratio of 1:100 (w/w) at 37°C and overnight.
Digestion was stopped by adding formic acid
to a final concentration of 1%. Samples were
desalted with C18 Sep-Pak cartridge (Waters)
and quantified with Pierce colorimetric pep-
tide assay (Thermo Fisher Scientific). Peptides
were dried under speed vacuum and stored
at −20°C.
TMT labeling
Peptides were reconstituted in 50 mM TEAB
buffer to a concentration of 2.1 mg/ml. TMT
10-plex reagent (Thermo Fisher Scientific)
was dissolved in 20 ml 100% acetonitrile to
reach 0.2 mg. For each TMT channel, peptides
(100 mg) were labeled with 0.2 mg TMT 10-plex
reagent (two multiplex—10plex for fed and
10plex for starved). For the internal standard
(IS), peptides from all samples were mixed to
reach 100 mg total peptides and labeled with
Jang et al., Science 378, eabq5209 (2022)
16 December 2022
13 of 16
Corrected 7 September 2023. See full text. RES EARCH | R E S E A R C H A R T I C L E
the TMT 131 channel to be able to compare
between the two plexes. The labeling reac-
tion was carried out for 60 min at RT and
quenched with 55 mM Tris pH 8.0 for 15 min
at RT. All TMT 10-plex labeled samples were
mixed and desalted with C18 Sep-Pak car-
tridge (Waters). The eluted peptides from the
Sep-Pak were dried under speed vacuum and
stored at −20°C.
High-pH prefractionation
The mixed TMT labeled peptides were re-
constituted in 10 mM NH4OH buffer with
1% acetonitrile to reach 200 mg. The Peptides
were fractionated by High-pH chromatogra-
phy using a Gemini column (3 mm, C18, 110 Å,
Phenomenex) on an Agilent 1260 Infinity II
system. An 85-min gradient was applied, and
72 fractions were collected and pooled into
12 fractions. Fractions were dried under speed
vacuum until analysis by LC-MS.
LC-MS analysis
Separation of the labeled TMT samples was
achieved by RP-HPLC on a Thermo Scientific
Dionex UltiMate 3000 system, as described
above. Samples were analyzed on an Orbitrap
Fusion Lumos mass spectrometer with FAIMS
Pro device (Thermo Scientific). MS1 and MS2
scans were acquired in the Orbitrap with a
mass resolution of 120,000 and 50,000, respec-
tively, MS1 scan range was set to between 400
and 1600 m/z, standard AGC target, and maxi-
mum injection time was set to auto. Precursor
ions with charge states 2 to 6 were isolated with
an isolation window of 0.7 m/z and dynamic
exclusion of 60 s. MS2 scans were set to custom
AGC target with normalized AGC target of
250%, and maximum injection time was set
to auto. Precursor ions were fragmented using
higher-energy collisional dissociation with 38%
normalized collision energy. Cycle time was set
to 2 s. An internal stepping of CVs −50, −65,
and −85 was used in all runs. Data acquisition
was done with Xcalibur software 4.4 and Instru-
ment Control Software version 3.4. Data analy-
sis for the TMT labeled samples was done in
Proteome Discoverer version 2.5. The TMT
10-plex was set as the quantification method,
and the 131 mass was set as the control chan-
nel. For the Sequest HT search, the following
parameters were applied: MS1 ion mass toler-
ance of 10 ppm and a MS2 mass tolerance of
0.02 Da. Tryptic digestion allowing two missed
cleavages, minimum peptide length of 6 amino
acids and maximum peptide length of 144
amino acids. The following modifications
were included: cysteine carbamidomethylation
(+57.021 Da) as static modification, methionine
oxidation (+15.995 Da) and N-terminal acet-
ylation (+42.011 Da) were set as dynamic modi-
fications. In addition, TMT 6-plex (229.163 Da)
was set as static modification for peptide
N-terminal and for lysine residue. Strict FDR
was set to 0.01, and relaxed FDR was set to
0.05. The search was performed against the
Human UniProt/Swiss-Prot database. Unique
and razor peptides were used for quantifica-
tion, co-isolation threshold was set to 50, and
average reporter S/N to 10. Data were normal-
ized against total peptide amount, and scaling
was done against the control channel average.
The result “Proteins” output table was exported,
and the statistical analysis was done in Perseus
version 1.6.15.0. Then, Gene Ontology analysis
was conducted using Metascape web tool (77)
(https://metascape.org/gp/index.html#/main/
step1). The proteomics data have been depos-
ited to the ProteomeXchange Consortium via
PRIDE and are available via ProteomeXchange
with identifier PXD033850.
Liposome co-sedimentation assays
Powdered lipids were individually resuspended
in chloroform then mixed together in a glass
sample vial and slowly evaporated with a dry
N2 stream. Liposome composition was 60%
phosphatidylcholine (Avanti #850375P), 19.8%
phosphatidylethanolamine (Avanti #850725P),
0.2% rhodamine-phosphatidylethanolamine
(Avanti #810150P), 10% cholesterol (Avanti
#700000), and 10% phosphatidylinositol-3-
sphosphate (Avanti #850150P) or 10% phos-
phatidylinositol 4-phosphate (Avanti #850151P)
or 10% phosphatidylinositol 3,4-biphosphate
(Avanti #850153P). The mixture was resus-
pended in 100 ml HEPES-buffered salt solution
(20 mM Hepes-NaOH pH 7.5, 150 mM NaCl)
for 40 min at 37°C, occasionally vigorously
vortexed, and sonicated in a 37°C water bath
five times, 45 s on, 1 min off. Liposomes (25 ml)
were gently mixed with 25 ml (3 mg suspended
with the same buffer) recombinant Escherichia
coli BL21 expressed GST-RRBP1 1-150aa WT
and incubated for 45 min at 25°C. Liposomes
were reisolated by centrifugation at 70,000g
for 15min at 25°C. Supernatant and pellet frac-
tions were dissolved in Laemmli sample buffer
and analyzed by SDS-PAGE.
Seahorse XFe96 analyzer
The Seahorse XFe96 sensor cartridge was hy-
drated, and 1 × 104 cells were seeded on 96-well
plates one day before the assay. The next day,
while cells were fed or starved for 100 min at
37°C incubator without CO2, the mitochon-
drial stress kit (Agilent, 103015-100) compound
was prepared in complete DMEM or EBSS and
loaded on the cartridge (OligomycinA1: 2 mM;
FCCP 1.7 mM, Rot/AA 1 mM). Assay media did
not include pyruvate. After instrument cali-
brations, cells were transferred to the XFe96
analyzer to record oxygen consumption rate at
37°C. All values were background-subtracted
(i.e., control well without cells). Finally, cells
were lysed with RIPA buffer and protein con-
tents was measured using the BCA assay for
data normalization.
Luminescent ATP assay
Cells (5 × 103 to 8 × 103) were seeded in the
black 96-well plate. On the next day, total
cellular ATP levels of fed or starved cells were
analyzed using ATPlite Luminescence Assay
System (PerkinElmer, 6016943) according to
the manufacturer’s protocol. Luminescence
was measured by TECAN Luminescence plate
reader, and the value was normalized by pro-
tein concentration measured via Bradford or
BCA assay.
Software
Cartoons and schematics were generated using
BioRender and Adobe illustrator.
Statistics and reproducibility
Statistical analysis and graphing were carried out
using Prism 8. Normality testing (D’Agostino-
Pearson) was conducted to determine whether
to use parametric or nonparametric statistical
tests. To compare two datasets, normally dis-
tributed data were analyzed by two-tailed un-
paired Student’s t test, whereas non-normally
distributed data were analyzed by two-tailed
Mann-Whitney test. For the comparison of
more than two normally distributed datasets,
we used ordinary one-way analysis of variance
(ANOVA) with either Dunnett’s multiple com-
parisons (to compare the mean of each column
with a control column) or Tukey’s multiple
comparisons test (to compare the mean of each
column with every other column). To com-
pare more than two non-normally distributed
datasets, we used the Kruskal-Wallis test with
two-sided Dunn’s multiple comparison test. All
statistical analyses were performed on samples
drawn from at least two or three independent
experiments.
REFERENCES AND NOTES
1. P. Bonaldo, M. Sandri, Cellular and molecular mechanisms of
muscle atrophy. Dis. Model. Mech. 6, 25–39 (2013).
doi: 10.1242/dmm.010389; pmid: 23268536
2. B. H. Goodpaster, L. M. Sparks, Metabolic flexibility in health
and disease. Cell Metab. 25, 1027–1036 (2017). doi: 10.1016/
j.cmet.2017.04.015; pmid: 28467922
3. M. Sandri et al., Foxo transcription factors induce the atrophy-
related ubiquitin ligase atrogin-1 and cause skeletal muscle
atrophy. Cell 117, 399–412 (2004). doi: 10.1016/S0092-8674
(04)00400-3; pmid: 15109499
4. S. Schiaffino, K. A. Dyar, S. Ciciliot, B. Blaauw, M. Sandri,
Mechanisms regulating skeletal muscle growth and atrophy.
FEBS J. 280, 4294–4314 (2013). doi: 10.1111/febs.12253;
pmid: 23517348
5. H. An, A. Ordureau, M. Körner, J. A. Paulo, J. W. Harper,
Systematic quantitative analysis of ribosome inventory during
nutrient stress. Nature 583, 303–309 (2020). doi: 10.1038/
s41586-020-2446-y; pmid: 32612236
6. A. Efeyan, W. C. Comb, D. M. Sabatini, Nutrient-sensing
mechanisms and pathways. Nature 517, 302–310 (2015).
doi: 10.1038/nature14190; pmid: 25592535
7. H. An et al., TEX264 is an endoplasmic reticulum-resident
ATG8-interacting protein critical for ER remodeling during
nutrient stress. Mol. Cell 74, 891–908.e10 (2019). doi: 10.1016/
j.molcel.2019.03.034; pmid: 31006537
8. A. Ballabio, J. S. Bonifacino, Lysosomes as dynamic regulators
of cell and organismal homeostasis. Nat. Rev. Mol. Cell Biol.
21, 101–118 (2020). doi: 10.1038/s41580-019-0185-4;
pmid: 31768005
Jang et al., Science 378, eabq5209 (2022)
16 December 2022
14 of 16
Corrected 7 September 2023. See full text. RES EARCH | R E S E A R C H A R T I C L E
9. S. M. Houten, S. Violante, F. V. Ventura, R. J. Wanders, The
biochemistry and physiology of mitochondrial fatty acid
b-oxidation and its genetic disorders. Annu. Rev. Physiol. 78,
23–44 (2016). doi: 10.1146/annurev-physiol-021115-105045;
pmid: 26474213
10. T. Balla, Phosphoinositides: Tiny lipids with giant impact on cell
regulation. Physiol. Rev. 93, 1019–1137 (2013). doi: 10.1152/
physrev.00028.2012; pmid: 23899561
11. Y. Posor, W. Jang, V. Haucke, Phosphoinositides as membrane
organizers. Nat. Rev. Mol. Cell Biol. 23, 797–816 (2022).
doi: 10.1038/s41580-022-00490-x; pmid: 35589852
12. C. C. Campa et al., Rab11 activity and PtdIns(3)P turnover
removes recycling cargo from endosomes. Nat. Chem. Biol. 14,
801–810 (2018). doi: 10.1038/s41589-018-0086-4;
pmid: 29915378
13. C. Cao, J. M. Backer, J. Laporte, E. J. Bedrick,
A. Wandinger-Ness, Sequential actions of myotubularin lipid
phosphatases regulate endosomal PI(3)P and growth factor
receptor trafficking. Mol. Biol. Cell 19, 3334–3346 (2008).
doi: 10.1091/mbc.e08-04-0367; pmid: 18524850
14. C. Cao, J. Laporte, J. M. Backer, A. Wandinger-Ness, M. P. Stein,
Myotubularin lipid phosphatase binds the hVPS15/hVPS34
lipid kinase complex on endosomes. Traffic 8, 1052–1067
(2007). doi: 10.1111/j.1600-0854.2007.00586.x;
pmid: 17651088
15. K. Ketel et al., A phosphoinositide conversion mechanism
for exit from endosomes. Nature 529, 408–412 (2016).
doi: 10.1038/nature16516; pmid: 26760201
16. L. Al-Qusairi et al., T-tubule disorganization and defective
excitation-contraction coupling in muscle fibers lacking
myotubularin lipid phosphatase. Proc. Natl. Acad. Sci. U.S.A.
106, 18763–18768 (2009). doi: 10.1073/pnas.0900705106;
pmid: 19846786
17. L. Amoasii et al., Myotubularin and PtdIns3P remodel the
sarcoplasmic reticulum in muscle in vivo. J. Cell Sci. 126,
1806–1819 (2013). doi: 10.1242/jcs.118505; pmid: 23444364
18. Y. Shibata, G. K. Voeltz, T. A. Rapoport, Rough sheets and
smooth tubules. Cell 126, 435–439 (2006). doi: 10.1016/
j.cell.2006.07.019; pmid: 16901774
19. J. Laporte et al., Mutations in the MTM1 gene implicated in
X-linked myotubular myopathy. Hum. Mol. Genet. 6, 1505–1511
(1997). doi: 10.1093/hmg/6.9.1505; pmid: 9305655
20. C. R. Pierson et al., Modeling the human MTM1 p.R69C
mutation in murine Mtm1 results in exon 4 skipping and a
less severe myotubular myopathy phenotype. Hum. Mol.
Genet. 21, 811–825 (2012). doi: 10.1093/hmg/ddr512;
pmid: 22068590
21. O. M. Dorchies et al., Normal innervation and differentiation
of X-linked myotubular myopathy muscle cells in a
nerve-muscle coculture system. Neuromuscul. Disord. 11,
736–746 (2001). doi: 10.1016/S0960-8966(01)00221-8;
pmid: 11595516
22. L. K. Schroeder et al., Dynamic nanoscale morphology of the
ER surveyed by STED microscopy. J. Cell Biol. 218, 83–96
(2019). doi: 10.1083/jcb.201809107; pmid: 30442642
23. G. E. Palade, Studies on the endoplasmic reticulum. II.
Simple dispositions in cells in situ. J. Biophys. Biochem.
Cytol. 1, 567–582 (1955). doi: 10.1083/jcb.1.6.567;
pmid: 13278367
24. G. K. Voeltz, W. A. Prinz, Y. Shibata, J. M. Rist, T. A. Rapoport,
A class of membrane proteins shaping the tubular endoplasmic
reticulum. Cell 124, 573–586 (2006). doi: 10.1016/
j.cell.2005.11.047; pmid: 16469703
25. J. Nixon-Abell et al., Increased spatiotemporal resolution
reveals highly dynamic dense tubular matrices in the
peripheral ER. Science 354, aaf3928 (2016). doi: 10.1126/
science.aaf3928; pmid: 27789813
26. M. A. Raess, S. Friant, B. S. Cowling, J. Laporte, WANTED -
Dead or alive: Myotubularins, a large disease-associated
protein family. Adv. Biol. Regul. 63, 49–58 (2017). doi: 10.1016/
j.jbior.2016.09.001; pmid: 27666502
27. E. L. Axe et al., Autophagosome formation from membrane
compartments enriched in phosphatidylinositol 3-phosphate
and dynamically connected to the endoplasmic reticulum.
J. Cell Biol. 182, 685–701 (2008). doi: 10.1083/jcb.200803137;
pmid: 18725538
28. Z. Hong et al., PtdIns3P controls mTORC1 signaling through
lysosomal positioning. J. Cell Biol. 216, 4217–4233 (2017).
doi: 10.1083/jcb.201611073; pmid: 29030394
29. M. J. Munson et al., mTOR activates the VPS34-UVRAG
complex to regulate autolysosomal tubulation and cell survival.
EMBO J. 34, 2272–2290 (2015). doi: 10.15252/
embj.201590992; pmid: 26139536
30. J. Laporte et al., A gene mutated in X-linked myotubular
myopathy defines a new putative tyrosine phosphatase family
conserved in yeast. Nat. Genet. 13, 175–182 (1996).
doi: 10.1038/ng0696-175; pmid: 8640223
31. M. W. Lawlor, J. J. Dowling, X-linked myotubular myopathy.
Neuromuscul. Disord. 31, 1004–1012 (2021). doi: 10.1016/
j.nmd.2021.08.003; pmid: 34736623
32. A. Khaminets et al., Regulation of endoplasmic reticulum
turnover by selective autophagy. Nature 522, 354–358 (2015).
doi: 10.1038/nature14498; pmid: 26040720
33. C. Settembre et al., TFEB links autophagy to lysosomal
biogenesis. Science 332, 1429–1433 (2011). doi: 10.1126/
science.1204592; pmid: 21617040
34. K. M. Fetalvero et al., Defective autophagy and mTORC1
signaling in myotubularin null mice. Mol. Cell. Biol. 33, 98–110
(2013). doi: 10.1128/MCB.01075-12; pmid: 23109424
35. A. L. Marat, V. Haucke, Phosphatidylinositol 3-phosphates-at
the interface between cell signalling and membrane traffic.
EMBO J. 35, 561–579 (2016). doi: 10.15252/embj.201593564;
pmid: 26888746
36. Y. Wu et al., Contacts between the endoplasmic reticulum
and other membranes in neurons. Proc. Natl. Acad. Sci. U.S.A.
114, E4859–E4867 (2017). doi: 10.1073/pnas.1701078114;
pmid: 28559323
37. H. Wu, P. Carvalho, G. K. Voeltz, Here, there, and everywhere:
The importance of ER membrane contact sites. Science
361, eaan5835 (2018). doi: 10.1126/science.aan5835;
pmid: 30072511
38. S. S. Lam et al., Directed evolution of APEX2 for electron
microscopy and proximity labeling. Nat. Methods 12, 51–54
(2015). doi: 10.1038/nmeth.3179; pmid: 25419960
39. A. Fegan, B. White, J. C. Carlson, C. R. Wagner, Chemically
controlled protein assembly: Techniques and applications.
Chem. Rev. 110, 3315–3336 (2010). doi: 10.1021/cr8002888;
pmid: 20353181
40. T. Calì, M. Brini, Quantification of organelle contact sites by
split-GFP-based contact site sensors (SPLICS) in living cells.
Nat. Protoc. 16, 5287–5308 (2021). doi: 10.1038/
s41596-021-00614-1; pmid: 34686857
41. C. Raiborg et al., Repeated ER-endosome contacts promote
endosome translocation and neurite outgrowth. Nature
520, 234–238 (2015). doi: 10.1038/nature14359;
pmid: 25855459
42. A. M. Valm et al., Applying systems-level spectral imaging
and analysis to reveal the organelle interactome. Nature
546, 162–167 (2017). doi: 10.1038/nature22369;
pmid: 28538724
43. A. A. Rowland, P. J. Chitwood, M. J. Phillips, G. K. Voeltz, ER
contact sites define the position and timing of endosome
fission. Cell 159, 1027–1041 (2014). doi: 10.1016/
j.cell.2014.10.023; pmid: 25416943
44. M. Lu et al., The structure and global distribution of the
endoplasmic reticulum network are actively regulated
by lysosomes. Sci. Adv. 6, eabc7209 (2020). doi: 10.1126/
sciadv.abc7209; pmid: 33328230
45. V. Hung et al., Proteomic mapping of cytosol-facing outer
mitochondrial and ER membranes in living human cells by
proximity biotinylation. eLife 6, e24463 (2017). doi: 10.7554/
eLife.24463; pmid: 28441135
46. Y. Shibata et al., Mechanisms determining the morphology of
the peripheral ER. Cell 143, 774–788 (2010). doi: 10.1016/
j.cell.2010.11.007; pmid: 21111237
47. P. Zheng et al., ER proteins decipher the tubulin code to
regulate organelle distribution. Nature 601, 132–138 (2022).
doi: 10.1038/s41586-021-04204-9; pmid: 34912111
48. C. S. Janota et al., Shielding of actin by the endoplasmic
reticulum impacts nuclear positioning. Nat. Commun. 13,
2763 (2022). doi: 10.1038/s41467-022-30388-3;
pmid: 35589708
49. B. T. Lobingier et al., An approach to spatiotemporally
resolve protein interaction networks in living cells. Cell 169,
350–360.e12 (2017). doi: 10.1016/j.cell.2017.03.022;
pmid: 28388416
50. Y. C. Wong, D. Ysselstein, D. Krainc, Mitochondria–lysosome
contacts regulate mitochondrial fission via RAB7 GTP
hydrolysis. Nature 554, 382–386 (2018). doi: 10.1038/
nature25486; pmid: 29364868
51. S. Nagashima et al., Golgi-derived PI(4)P-containing vesicles
drive late steps of mitochondrial division. Science 367,
1366–1371 (2020). doi: 10.1126/science.aax6089;
pmid: 32193326
in human cells. Science 353, aaf5549 (2016). doi: 10.1126/
science.aaf5549; pmid: 27418514
53. J. R. Friedman et al., ER tubules mark sites of mitochondrial
division. Science 334, 358–362 (2011). doi: 10.1126/
science.1207385; pmid: 21885730
54. T. C. Walther, J. Chung, R. V. Farese Jr., Lipid droplet
biogenesis. Annu. Rev. Cell Dev. Biol. 33, 491–510 (2017).
doi: 10.1146/annurev-cellbio-100616-060608; pmid: 28793795
55. A. Santinho et al., Membrane curvature catalyzes lipid droplet
assembly. Curr. Biol. 30, 2481–2494.e6 (2020). doi: 10.1016/
j.cub.2020.04.066; pmid: 32442467
56. A. S. Rambold, B. Kostelecky, N. Elia, J. Lippincott-Schwartz,
Tubular network formation protects mitochondria from
autophagosomal degradation during nutrient starvation.
Proc. Natl. Acad. Sci. U.S.A. 108, 10190–10195 (2011).
doi: 10.1073/pnas.1107402108; pmid: 21646527
57. A. S. Rambold, S. Cohen, J. Lippincott-Schwartz, Fatty acid
trafficking in starved cells: Regulation by lipid droplet lipolysis,
autophagy, and mitochondrial fusion dynamics. Dev. Cell 32,
678–692 (2015). doi: 10.1016/j.devcel.2015.01.029;
pmid: 25752962
58. M. Giacomello, A. Pyakurel, C. Glytsou, L. Scorrano, The cell
biology of mitochondrial membrane dynamics. Nat. Rev. Mol.
Cell Biol. 21, 204–224 (2020). doi: 10.1038/s41580-020-0210-7;
pmid: 32071438
59. T. Wai et al., Imbalanced OPA1 processing and mitochondrial
fragmentation cause heart failure in mice. Science 350,
aad0116 (2015). doi: 10.1126/science.aad0116;
pmid: 26785494
60. A. R. English, G. K. Voeltz, Rab10 GTPase regulates ER
dynamics and morphology. Nat. Cell Biol. 15, 169–178 (2013).
doi: 10.1038/ncb2647; pmid: 23263280
61. J. E. Lee, P. I. Cathey, H. Wu, R. Parker, G. K. Voeltz,
Endoplasmic reticulum contact sites regulate the dynamics of
membraneless organelles. Science 367, eaay7108 (2020).
doi: 10.1126/science.aay7108; pmid: 32001628
62. L. C. Gomes, G. Di Benedetto, L. Scorrano, During autophagy
mitochondria elongate, are spared from degradation and
sustain cell viability. Nat. Cell Biol. 13, 589–598 (2011).
doi: 10.1038/ncb2220; pmid: 21478857
63. W. M. Henne, M. L. Reese, J. M. Goodman, The assembly of
lipid droplets and their roles in challenged cells. EMBO J. 37,
e98947 (2018). doi: 10.15252/embj.201898947;
pmid: 29789390
64. M. A. Roberts, A. Segura-Roman, J. A. Olzmann, Organelle
biogenesis: ER shape influences lipid droplet nucleation.
Curr. Biol. 30, R770–R773 (2020). doi: 10.1016/
j.cub.2020.05.027; pmid: 32634419
65. J. R. Liang et al., A genome-wide ER-phagy screen highlights
key roles of mitochondrial metabolism and ER-resident
UFMylation. Cell 180, 1160–1177.e20 (2020). doi: 10.1016/
j.cell.2020.02.017; pmid: 32160526
66. Y. Pan et al., Endoplasmic reticulum ribosome-binding protein
1, RRBP1, promotes progression of colorectal cancer and
predicts an unfavourable prognosis. Br. J. Cancer 113, 763–772
(2015). doi: 10.1038/bjc.2015.260; pmid: 26196185
67. A. H. Bahrami, G. Hummer, Formation and stability of lipid
membrane nanotubes. ACS Nano 11, 9558–9565 (2017).
doi: 10.1021/acsnano.7b05542; pmid: 28873296
68. T. Shemesh et al., A model for the generation and interconversion
of ER morphologies. Proc. Natl. Acad. Sci. U.S.A. 111,
E5243–E5251 (2014). doi: 10.1073/pnas.1419997111;
pmid: 25404289
69. G. Parlakgül et al., Regulation of liver subcellular architecture
controls metabolic homeostasis. Nature 603, 736–742 (2022).
doi: 10.1038/s41586-022-04488-5; pmid: 35264794
70. D. S. Hirsch et al., Insulin activation of vacuolar protein sorting
34 mediates localized phosphatidylinositol 3-phosphate
production at lamellipodia and activation of mTOR/S6K1.
Cell. Signal. 26, 1258–1268 (2014). doi: 10.1016/
j.cellsig.2014.02.009; pmid: 24582588
71. V. I. Korolchuk et al., Lysosomal positioning coordinates
cellular nutrient responses. Nat. Cell Biol. 13, 453–460 (2011).
doi: 10.1038/ncb2204; pmid: 21394080
72. A. L. Marat et al., mTORC1 activity repression by late
endosomal phosphatidylinositol 3,4-bisphosphate. Science
356, 968–972 (2017). doi: 10.1126/science.aaf8310;
pmid: 28572395
73. M. J. Hoyer et al., A novel class of ER membrane proteins
regulates ER-associated endosome fission. Cell 175, 254–265.
e14 (2018). doi: 10.1016/j.cell.2018.08.030; pmid: 30220460
52. S. C. Lewis, L. F. Uchiyama, J. Nunnari, ER-mitochondria
contacts couple mtDNA synthesis with mitochondrial division
74. K. Mamchaoui et al., Immortalized pathological human
myoblasts: Towards a universal tool for the study of
Jang et al., Science 378, eabq5209 (2022)
16 December 2022
15 of 16
Corrected 7 September 2023. See full text. RES EARCH | R E S E A R C H A R T I C L E
neuromuscular disorders. Skelet. Muscle 1, 34 (2011).
doi: 10.1186/2044-5040-1-34; pmid: 22040608
75. F. Bottanelli et al., Two-colour live-cell nanoscale imaging of
intracellular targets. Nat. Commun. 7, 10778 (2016).
doi: 10.1038/ncomms10778; pmid: 26940217
76. C. McQuin et al., CellProfiler 3.0: Next-generation image
processing for biology. PLOS Biol. 16, e2005970 (2018).
doi: 10.1371/journal.pbio.2005970; pmid: 29969450
77. Y. Zhou et al., Metascape provides a biologist-oriented
resource for the analysis of systems-level datasets. Nat.
Commun. 10, 1523 (2019). doi: 10.1038/s41467-019-09234-6;
pmid: 30944313
ACKN OW LEDG MEN TS
We thank members of the Haucke lab for discussion and Y. Posor
for critical reading of the manuscript. We thank S. Zillmann,
D. Löwe, M. Mühlbauer, and C. Schmidt for expert technical
assistance and M. Lehmann, C. Schmied, and J. Eichhorst for aid
with microscopy and image analysis. We also thank G. G. Farías
and G. Voeltz for plasmids and protocols and the MyoLine platform
of the Institute of Myology in Paris for aid in the generation
of myoblast cell lines. Funding: This work was funded by a
Leibniz-German Academic Exchange Service (DAAD) Research
Fellowship (57423756) (W.J.), the Postdoctoral Fellowship
Program (Nurturing Next-generation Researchers) of the National
Research Foundation of Korea (NRF) (2018R1A6A3A03010583)
(W.J.), and Deutsche Forschungsgemeinschaft (TRR186/ A08)
(V.H.). Author contributions: Conceptualization: W.J. and
V.H. Investigation: all cell and molecular biology: W.J.; electron
microscopy: D.P. with W.J.; quantitative proteomics: W.J. with
M.N.-H. and F.L.; Seahorse analysis: W.J., Y.L., S.J.S., and U.K.;
XLCNM myoblast patient cells: K.M. and V.M.; CRISPR: W.J. and
P.S. Funding acquisition: W.J. and V.H. Project administration: V.H.
Supervision: V.H., F.L., U.K., and S.J.S. Writing – original draft:
W.J. and V.H. Writing – review & editing: all authors. Competing
interests: The authors declare no competing financial interests.
Data and materials availability: All data are available in the main
text or the supplementary materials. Proteomics data have been
deposited to the ProteomeXchange Consortium via PRIDE and
are available via ProteomeXchange with identifier PXD033846.
Materials and reagents are available from the corresponding author
upon request. License information: Copyright © 2022 the
authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.science.org/about/science-
licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abq5209
Figs. S1 to S11
Tables S1 to S4
MDAR Reproducibility Checklist
Movies S1 to S5
View/request a protocol for this paper from Bio-protocol.
Submitted 13 April 2022; resubmitted 23 September 2022
Accepted 25 October 2022
10.1126/science.abq5209
Jang et al., Science 378, eabq5209 (2022)
16 December 2022
16 of 16
Corrected 7 September 2023. See full text.
|
10.1126_science.ade3535
|
RES EARCH
CARBON OFFSETS
Corrected 31 January 2024. See full text.
Action needed to make carbon offsets from forest
conservation work for climate change mitigation
Thales A. P. West1,2*, Sven Wunder3,4, Erin O. Sills5, Jan Börner6,7, Sami W. Rifai8,
Alexandra N. Neidermeier1, Gabriel P. Frey6, Andreas Kontoleon2,9
Carbon offsets from voluntary avoided-deforestation projects are generated on the basis of performance in
relation to ex ante deforestation baselines. We examined the effects of 26 such project sites in six countries on
three continents using synthetic control methods for causal inference. We found that most projects have
not significantly reduced deforestation. For projects that did, reductions were substantially lower than claimed.
This reflects differences between the project ex ante baselines and ex post counterfactuals according to
observed deforestation in control areas. Methodologies used to construct deforestation baselines for carbon
offset interventions need urgent revisions to correctly attribute reduced deforestation to the projects, thus
maintaining both incentives for forest conservation and the integrity of global carbon accounting.
F or nearly two decades, the performance-
based payment mechanism for reduced
carbon emissions from deforestation and
forest degradation known as REDD+ has
been under intense debate (1). Although
regulations and capacity for national REDD+
programs are still under development (2, 3), many
standalone, voluntary REDD+ projects are op-
erational worldwide (4). These projects intend to
conserve forests through many activities, such as
improved monitoring and enforcement, promo-
tion of sustainable practices, and local stakeholder
engagement, often funded by the commercial-
ization of carbon offsets [each corresponding to
1 Mg of carbon dioxide (CO2) either removed from
or not emitted into the atmosphere]. In 2021,
two-thirds of the 227.7 million offsets from the
land-use sector (excluding agriculture) traded
in carbon markets, with a total value of USD
$1.3 billion, originated from REDD+ projects (5).
Numerous policy discussions and initiatives
focus on how to scale and integrate the carbon-
emission reductions claimed by voluntary carbon-
offset projects, particularly from REDD+ activities,
into climate policies and Nationally Determined
Contributions (NDCs) reported to the United
Nations Framework Convention on Climate
Change (3, 6–8). However, there is little rigorous
evidence on the contributions of these projects
(9, 10), with some studies suggesting that many
are associated with little or no actual emission
reductions (11–17).
Carbon offsets from REDD+ projects are is-
sued on the basis of comparison between the
observed forest cover in the project areas and
deforestation baseline scenarios expected to
have been realized in the absence of REDD+,
which are de facto unobservable (3, 17). Many
project baselines are formed through the ex-
trapolation of historical deforestation aver-
ages or trends, often spatially projected over a
reference region that encompasses the project
sites (17). These crediting baselines may be-
come unrealistic counterfactuals with exten-
sive changes in economic or political conditions
known to affect deforestation rates (18), com-
bined with questionable modeling decisions
underlying the spatial projections (19, 20). Base-
lines could also be opportunistically inflated by
profiteers seeking to maximize the volume of
offsets issued by a project (21). As a result, carbon
offsets may lack “additionality”—they may not
reflect actual emission reductions (22).
This study provides a pantropical comparison
between ex post deforestation counterfactuals,
informed by observable control areas, and
the ex ante baselines adopted by 27 voluntary
REDD+ projects in six tropical countries: Peru,
Colombia, Democratic Republic of Congo (DRC),
Tanzania, Zambia, and Cambodia (Fig. 1, figs.
S1 and S2, and tables S1 and S2) certified un-
der the Verified Carbon Standard (23). Because
some projects are composed of multiple dis-
connected sites, we evaluated those individually,
increasing our sample to 31 project sites. We
present both project-specific and cross-project
1Environmental Geography Group, Institute for Environmental
Studies (IVM), VU University Amsterdam, Amsterdam,
Netherlands. 2Centre for Environment, Energy and Natural
Resource Governance, University of Cambridge, Cambridge,
UK. 3European Forest Institute (EFI), Barcelona, Spain.
4Center for International Forestry Research (CIFOR), Lima,
Peru. 5Department of Forestry and Environmental Resources,
North Carolina State University, Raleigh, USA. 6Center for
Development Research (ZEF), University of Bonn, Bonn,
Germany. 7Institute for Food and Resource Economics (ILR),
University of Bonn, Bonn, Germany. 8School of Biological
Sciences, University of Adelaide, Adelaide, Australia.
9Department of Land Economy, University of Cambridge,
Cambridge, UK.
*Corresponding author. Email: t.a.pupowest@vu.nl
A
B
C
200 km
450 km
Fig. 1. Voluntary REDD+ project sites included in the study. (A) Peru and Colombia. (B) Democratic Republic of Congo (DRC), Tanzania, and Zambia.
(C) Cambodia. Study areas are indicated in red. Purple areas are the sites excluded from the analysis.
West et al., Science 381, 873–877 (2023)
25 August 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
analyses based, respectively, on the standard
and generalized versions of the synthetic con-
trol (SC) method for causal inference (24, 25)
that estimate the reductions in deforestation
in project sites attributable to the REDD+
interventions.
The SCs are constructed to be based on con-
trol areas (“donors”), selected from project-
specific donor pools, which had similar levels
of forest cover and other characteristics (table
S2) and were exposed to similar levels of de-
forestation pressure (as determined by com-
paring the average annual deforestation in the
projects’ and donors’ 1-km and 10-km buffer
zones before the project implementation). The
SC method selects and constructs a weighted
combination of control areas that has similar
characteristics as those of the REDD+ site
and follows a similar historical trajectory of
deforestation (supplementary materials). Before
interpreting individual project results, we con-
ducted project-specific “validation” tests to check
whether the standard SC method (24) was able
to construct SCs with deforestation rates similar
to that of project areas during the immediate
preproject period (17). Conservatively, we focus
the discussion of our results on projects with
SCs that performed well in the validation test
(fig. S5 and table S3).
Average impacts were estimated with the
generalized SC (GSC) method (25) on the basis
of two independent sets of control areas. Fur-
thermore, to address the concern that the se-
lected control areas may not represent potential
counterfactuals for REDD+ project sites, we also
estimated average project impacts through
comparison of operational project sites with
“yet-to-become” project areas throughout the
Corrected 31 January 2024. See full text.
study period using matching-based methods
for time-series cross-sectional analyses (26).
Because the evaluated projects span multiple
countries and contexts, our analyses shed light
on the robustness of the assumptions adopted
for the construction of REDD+ baselines under a
wide range of deforestation conditions (fig. S2).
Individual REDD+ project impacts
The individual SC analyses show mixed im-
pacts of the voluntary REDD+ projects on de-
forestation. Results from the validation tests
suggest that the SC method could replicate
pre-REDD+ deforestation trends in the “to-be”
project sites (fig. S6 and table S3). We discarded
only one project (1775-1) from the analyses be-
cause of the poor fit between the deforestation
in the SC and the REDD+ site before project
implementation in our validation test. Four
other projects (985, 1360-1, 1389, and 1748)
were also discarded out of an abundance of
caution because despite the agreement in pre-
project deforestation between the SCs and the
REDD+ sites, their buffer deforestation rates
suggested substantially different levels of de-
forestation pressure. Our final sample was
thus reduced to 26 project sites.
Eight of the remaining 26 project sites showed
some evidence of additional reductions in de-
forestation compared with their individual SCs
(figs. S7 and S8), although generally not to the
extent claimed by the projects based on their
ex ante crediting baselines. Additionality was
most likely in Peru, where half of the REDD+
sites had significantly less deforestation than
that of the ex post counterfactuals, with statis-
tical significance judged by means of placebo
tests. Three of the seven Colombian project
Peru
Colombia
Africa
1
0
-1
-2
1
0
-1
-2
1
0
-1
-2
-5
0
5
10
-5
0
5
-5
0
5
10
)
%
(
n
o
i
t
a
t
s
e
r
o
f
e
d
n
o
T
T
A
2.0
1.0
0.0
)
%
(
n
o
i
t
a
t
s
e
r
o
f
e
D
-10
0
2.0
1.0
0.0
2.0
1.0
0.0
10
Time relative to project implementation (years)
-10
-5
5
0
-10
0
10
Fig. 2. Estimated average impacts of REDD+ projects in Peru, Colombia, and Africa on annual
deforestation. Averages are based on the GSC method and the donor pool of control areas selected for the
individual project’s SCs. (Top) The average treatment effect on the treated (ATT) project sites. (Bottom)
Projects’ (solid red line) and counterfactuals’ (dashed blue line) deforestation averages. Shaded red areas
indicate bootstrapped 95% confidence intervals around the projects’ deforestation average.
sites and one of two Cambodian sites achieved
significant deforestation reductions according
to the SCs and placebo tests. No evidence of
avoided deforestation was found for the REDD+
sites in the DRC, Tanzania, and Zambia with
regard to their counterfactuals.
Average REDD+ project impacts
Average project impacts on deforestation [av-
erage treatment effects on the treated (ATT)]
in Peru, Colombia, and Africa (DRC, Tanzania,
and Zambia) were estimated with the GSC
method (Fig. 2, top). Cambodian projects were
excluded from this analysis because of the lim-
ited sample size. Unlike the individual project
evaluations, the GSC analyses were based ex-
clusively on annual deforestation rates and
time-variant covariates. The GSC analyses
were based on two independent sets of se-
lected control areas for each region: In the
first set, only the donors selected for the con-
struction of the individual SCs were consi-
dered, whereas in the second set, controls
were selected through cardinality matching
(27), independent of the SC analyses. Our con-
clusions are robust to all of these method-
ological variations.
For the first set of controls, the average im-
pact of the Peruvian projects on forest loss over
10 years was −0.24% or avoided deforestation
of 686 ha year−1 (table S7). This effect was sta-
tistically significant in the first 4 years of project
implementation (Fig. 2, bottom). An ATT of
−0.14% or 414 ha year−1 (table S8) was found for
the African projects, whereas a smaller effect
was associated with the Colombian projects
(−0.03% or 49 ha year−1) (table S9). Neither the
estimates for the Colombian nor the African
groups of projects were statistically significant.
Even assuming the estimated average reduc-
tions in deforestation to be significant in all
three regions (a plausible assumption given
our small sample sizes), they would still be
substantially lower than avoided deforestation
calculated from the average ex ante baseline
deforestation rates adopted by the projects in
Peru (3661 ha year−1), Colombia (2550 ha year−1),
and Africa (2700 ha year−1) through 2020.
These results are robust to using control
areas selected through cardinality matching.
On the basis of our second control set, we es-
timated ATTs of the Peruvian, Colombian, and
African REDD+ sites as −0.42% (1266 ha) year−1,
−0.01% (30 ha) year−1, and −0.21% (423 ha) year−1,
respectively (fig. S10, top, and tables S10 to
S12). Again, only the estimate for the Peruvian
projects was statistically significant, and only
in the first 4 years of project implementation
(fig. S10, bottom). Although the estimated
impacts in Peru were −0.18 percentage points
larger as compared with the first control set,
the translated absolute average reduction
still represents just one-third of the average
reductions claimed by the Peruvian projects.
West et al., Science 381, 873–877 (2023)
25 August 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
Corrected 31 January 2024. See full text.
Fig. 3. Cumulative deforestation from the baseline scenarios adopted by the REDD+ projects versus observed cumulative deforestation in the SCs. Orange, REDD+
projects; SCs, blue. Shaded blue areas are possibility intervals. Dotted red lines indicate the observed cumulative deforestation in the project sites. Dashed black lines indicate the
project implementation year. Asterisks indicate a significant reduction in deforestation in the REDD+ site compared with the SC based on placebo tests (scales differ).
Last, results from the analysis of already op-
erational versus “to-become” project sites cor-
roborate the findings from the GSC analyses.
The average REDD+ impacts on deforestation
ranged from −0.01% to 0.12% year−1 (or −113 to
121 ha year−1), across different model specifi-
cations, but the estimates were not significant
in any of the regions (fig. S14).
Carbon-offset implications
We investigated the implications of our findings
for the environmental integrity of the credits
issued by the REDD+ projects. These impli-
cations are based on the 18 out of 26 projects
with sufficient publicly available information
about baseline deforestation rates (Fig. 3 and
tables S5 and S6). Only one project baseline
was lower than its SC (project 958), and only one
project baseline was similar to its SC (project
1325). All other ex ante baselines posited more
deforestation than estimated ex post accord-
ing to SCs. According to the projects’ ex ante
estimates, up to 90.5 million carbon offsets could
potentially have been generated by these 18
REDD+ projects through 2020. Yet 61.7 million
of these offsets (68%) would have originated from
projects that have not significantly reduced
deforestation (and carbon emissions) compared
with their SCs. The remaining 28.8 million off-
sets (32%) would have originated from projects
likely associated with some avoided defores-
tation, but not to the extent expected by the
project developers. If we replace the ex ante
baselines adopted by the projects with the de-
forestation observed in the SCs, our esti-
mates suggest that only 6.4 million (7%) of the
West et al., Science 381, 873–877 (2023)
25 August 2023
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
90.5 million expected offsets from the REDD+
projects would be associated with additional
carbon emission reductions.
As of November 2021, those 18 REDD+ projects
had issued 62 million carbon-offset credits
(table S4). Out of those, at least 14.6 million
(24%) have already been used by individuals
or organizations around the world to offset
their greenhouse gas emissions. Thus, accord-
ing to our SC-based estimates, these projects
have already been used to offset almost three
times more carbon emissions than their actual
contributions to climate change mitigation—
with another 47.4 million carbon offsets being
readily available in the market.
Discussion
Overall, the weight of evidence suggests that
the voluntary REDD+ projects in our sample
across six tropical countries achieved much less
avoided deforestation than forecast by project
developers. Only a minority of projects achieved
statistically significant reductions in compar-
ison with ex post counterfactuals. Our findings
corroborate prior studies that questioned the
additionality, and thus environmental integ-
rity, of carbon-offset interventions (11–17). Ex
ante baselines that exaggerate the deforesta-
tion that would occur without REDD+, likely
facilitated by methodological flexibility in their
construction and exacerbated by adverse site
selection (14), are a major reason for the gap
between offsets projected ex ante and actual
offsets estimated ex post. Poor performance by
the REDD+ projects (failure to de facto reduce
deforestation) may also be a factor (21, 28).
In an evaluation of voluntary REDD+ projects
in the Brazilian Amazon, West et al. (17) pointed
to the potential confounding effect created by
Brazil’s post-2004 policy interventions to con-
trol deforestation, triggering a widespread re-
duction in forest loss between 2004 and 2012
(29). As a result, the high regional deforestation
rates observed before 2004, used to inform the
Brazilian project ex ante baselines, likely led to
an overestimation of the projects’ performance.
Yet unlike Brazil, the six countries in this study
did not experience a similar nationwide reduc-
tion in deforestation after the REDD+ projects
were implemented (fig. S3). Hence, the un-
realistic ex ante baselines adopted by many
projects likely resulted from the use of meth-
odologies that systematically fail to produce
credible counterfactuals for the REDD+ in-
terventions, compromising the evaluation of
the projects’ effect on mitigating deforestation
and thus carbon emissions. This may be due to
potentially four complementary reasons: poor
foresight, adverse site selection, limited room
for adjustments over time, and “gaming.” First,
projects may have (unintentionally) overesti-
mated future deforestation pressures by follow-
ing baseline methodologies that heavily rely
on the continuation of historical trends that
Corrected 31 January 2024. See full text.
no longer represent current conditions, often
combined with problematic spatial deforesta-
tion projections (19, 30). Second, projects may
have been preferentially sited where conditions
are conducive to reducing deforestation and
therefore where deforestation may have been
lower than suggested by ex ante baselines
constructed from deforestation trends in the
surrounding region (14). Third, rules adopted
by certification standards require baselines to
be fixed for a period of usually 10 years, restrict-
ing adjustments to reflect changes in defores-
tation drivers over time (23). Last, flexibility in
baseline methodologies may have been oppor-
tunistically exploited to maximize revenues
from offset sales.
By contrast, both standard and generalized
SC methods use pre-REDD+ information to
identify control areas but use contemporane-
ous ex post information on deforestation in the
project and control areas to measure addition-
ality. Following well-accepted guidelines for
rigorous impact evaluation (31), if properly
selected, such ex post counterfactuals can cap-
ture the effects of contemporaneous changes in
deforestation drivers and thus are less likely to
generate effect estimates confounded by ex-
ternal factors (10, 17). REDD+ projects adopting
a similar dynamic approach would likely reduce
the additionality problems with project ex ante
baselines and offsets identified in this study.
Despite the clear advantages (from a causal
inference perspective) of using ex post methods
such as SCs to construct deforestation base-
lines for REDD+ interventions, some imple-
mentation and monitoring challenges would
likely arise from their adoption. First, given
the biophysical heterogeneity of tropical re-
gions, and limited data, ideal control areas for
the project sites may not always exist or be
possible to identify (supplementary materials,
appendix A). Second, those control areas could
be manipulated (for example, intentionally de-
graded or conserved) to misleadingly improve
or reduce estimated project performance. Third,
dynamic baselines may still fail to account
for all relevant determinants of deforestation
owing to data constraints. Last, long-lasting vol-
untary REDD+ projects may eventually outlive
their SCs as new interventions or other regime
shifts occur in control sites.
One alternative would be to require projects
to adopt transparent ex ante jurisdictional
baselines that are preestablished by govern-
ment agencies. Although these baselines might
still fall short at demonstrating causal links
and thus additionality, they could be both more
transparent and updated more frequently than
individual project baselines to reflect emerg-
ing deforestation pressures and spatial pat-
terns. Reductions in deforestation relative to
jurisdictional baselines would still have to be
correctly attributed to either government ef-
forts to control forest loss or to private REDD+
interventions (10), in addition to changes in
external deforestation drivers (such as agricul-
tural commodity prices). Transferring the re-
sponsibility of baseline construction from project
developers to jurisdictions could reduce the
room for “baseline gaming,” although the risk
of adverse site selection could remain (14). On
balance, nesting voluntary projects into sub-
national jurisdictions appears to be a promis-
ing future pathway for REDD+, a practice that
is gaining traction globally (6). However, the
specifics of implementation are crucial and
require careful consideration.
Another possible explanation for limited ad-
ditionality is poor performance by the projects.
This would be likely if projects were failing to
control deforestation in project areas. We ob-
served generally low deforestation rates in
project areas, but we cannot compare that with
the projects’ targeted deforestation rates in those
areas because those are not always reported in
the publicly available information on projects.
However, it is clear that some projects have
struggled with on-the-ground implementation
and execution of envisioned conservation ac-
tivities; others may have promoted ineffective
actions, whether because of funding uncertain-
ties, slow commercialization of carbon cred-
its, lack of experience, or poor management
(21, 28, 32). Many projects claimed to have
started much earlier than the year they were
certified. Although this allows projects to is-
sue and sell retroactive offsets immediately
after certification (33), it also implies that they
did not have access to carbon funding during
their initial years, potentially compromising
the execution of planned conservation actions.
A recent evaluation on the effectiveness of the
same type of REDD+ interventions reported
significant reductions in average deforestation
rates (34). The study, based on satellite pixel
matching, estimated an average deforestation
rate of 0.2% year−1 in the REDD+ sites versus
0.4% year−1 for their matched control areas.
The size of these estimates is similar to some
of our estimates from GSC analyses; but in
our case, they were statistically insignificant,
potentially because of our smaller sample sizes
compared with the pixel-based samples from
the previous study. And from the offsetting per-
spective, the average reductions—significant or
not—were substantially lower than the addi-
tional reductions in deforestation projected and
claimed by the projects.
Our study provides further evidence on the
effectiveness of voluntary REDD+ projects and
questions their de facto additionality (35).
Only a minority of the projects significantly
reduced deforestation in the project areas com-
pared with the ex post counterfactuals, and
even those, with one exemption, did not reduce
deforestation to the extent claimed. Although
REDD+ payments are typically conditioned on
performance in project areas, only the offsets
West et al., Science 381, 873–877 (2023)
25 August 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
associated with additional reductions in de-
forestation relative to a counterfactual genu-
inely offset emissions of potential buyers in the
voluntary carbon markets (7, 21). Certification
schemes are allegedly in place to safeguard
the additionality of offsets, but our results in-
dicate that currently used baseline method-
ologies do not guarantee additionality. It is
critical to develop new and rigorous methods
for the construction of credible deforestation
baselines for voluntary REDD+ interventions
and to properly and regularly assess their con-
tribution to climate change mitigation.
Last, the evidence from this and other studies
indicates that some voluntary projects have
effectively reduced deforestation (34), partic-
ularly in Peru. For REDD+ to be scaled and
achieve its ambitious goals worldwide, it is
paramount that we better understand the fac-
tors that drive both mitigation performance
and impacts on local communities. Building on
this knowledge, academics, practitioners, and
policy-makers must form effective partnerships
to help REDD+ fulfill its original promise.
RE FE RENCES AND N OT ES
1. A. Angelsen, Rev. Dev. Econ. 21, 237–264 (2017).
2.
J. Börner et al., in Transforming REDD+: Lessons and new
directions, A. Angelsen et al., Eds. (CIFOR, 2018), pp. 105–116.
3. FAO, “From reference levels to results reporting: REDD+ under
the United Nations Framework Convention on Climate Change.
2019 update” (Food and Agriculture Organization, Rome, 2019).
4. S. S. Atmadja et al., Environ. Res. Lett. 17, 044038 (2022).
5. S. Donofrio, P. Maguire, C. Daley, C. Calderon, K. Lin, The Art of
Integrity: State of the Voluntary Carbon Markets 2022 Q3
(Forest Trends’ Ecosystem Marketplace, 2022).
6. D. Lee, P. Llopis, R. Waterworth, G. Roberts, T. Pearson,
Approaches to REDD+ Nesting: Lessons Learned from Country
Experiences (World Bank, 2018).
Corrected 31 January 2024. See full text.
7. Taskforce on Scaling Voluntary Carbon Markets, Phase 1—Final
Report (Institute of International Finance, 2021).
8. Voluntary Carbon Markets Integrity Initiative, Aligning
Voluntary Carbon Markets with the 1.5°C Paris Agreement
Ambition (Voluntary Carbon Markets Integrity Initiative,
2021).
9. A. E. Duchelle, G. Simonet, W. D. Sunderlin, S. Wunder,
Curr. Opin. Environ. Sustain. 32, 134–140 (2018).
10. E. O. Sills et al., Glob. Environ. Change 43, 148–160
(2017).
11. G. Badgley et al., Glob. Change Biol. 28, 1433–1445
(2022).
12. R. Calel, J. Colmer, A. Dechezleprêtre, M. Glachant, SSRN
Electron. J. 5709 (2021).
13. M. Cames et al., How Additional is the Clean Development
Mechanism? (Öko-Institut e.V., 2016).
14. P. Delacote, G. Le Velly, G. Simonet, Resour. Energy Econ. 67,
101277 (2022).
15. B. Haya et al., Clim. Policy 20, 1112–1126 (2020).
16. A. Kollmuss, L. Schneider, V. Zhezherin, Has Joint Implementation
Reduced GHG Emissions? Lessons Learned for the Design of
Carbon Market Mechanisms (Stockholm Environment Institute,
2015).
17. T. A. P. West, J. Börner, E. O. Sills, A. Kontoleon, Proc. Natl.
Acad. Sci. U.S.A. 117, 24188–24194 (2020).
18. J. Busch, K. Ferretti-Gallon, Rev. Environ. Econ. Policy 11, 3–23
(2017).
19. S. Sloan, J. Pelletier, Glob. Environ. Change 22, 440–453
(2012).
20. M. Ehara et al., For. Policy Econ. 129, 102474
(2021).
21. C. Seyller et al., Int. For. Rev. 18, 231–246 (2016).
22. S. Wunder, Conserv. Biol. 21, 48–58 (2007).
23. Verra, VCS Standard, version 4.0 (Verra, 2019).
24. A. Abadie, A. Diamond, J. Hainmueller, J. Stat. Softw. 42, 1–17
(2011).
25. Y. Xu, Polit. Anal. 25, 57–76 (2017).
26. K. Imai, I. S. Kim, E. H. Wang, Am. J. Pol. Sci. 0, 1–19
(2021).
27. J. R. Zubizarreta, R. D. Paredes, P. R. Rosenbaum, Ann. Appl. Stat.
8, 204–231 (2014).
28. M. S. Verra, Suruí Forest Carbon Project (2018); https://verra.
org/media-statement-surui-forest-carbon-project.
29. T. A. P. West, P. M. Fearnside, Land Use Policy 100, 105072
(2021).
30. O. Mertz et al., J. Land Use Sci. 13, 1–15 (2018).
31. P. J. Ferraro, M. M. Hanauer, Annu. Rev. Environ. Resour. 39,
495–517 (2014).
32. S. Wunder et al., Front. For. Glob. Change 3, 11 (2020).
33. N. Linacre, R. O’Sullivan, D. Ross, L. Durschinger, I. Deshmukh,
“REDD+ Supply and Demand 2015–2025” (US Agency for
International Development, 2015).
34. A. Guizar-Coutiño, J. P. G. Jones, A. Balmford, R. Carmenta,
D. A. Coomes, Conserv. Biol. 36, e13970 (2022).
35. A. Angelsen, Moving ahead with REDD: Issues, Options and
Implications (CIFOR, 2008), vol. 6.
36. T. A. P. West, Replication Data for: Action needed to make
carbon offsets from forest conservation work for climate
change mitigation. DataverseNL (2023); https://doi.org/
10.34894/IQC9LM.
AC KNOWLED GME NTS
We thank the German Development Agency (GIZ) for their support
during the planning phase of this study. Funding: This research
was supported by Norway’s International Climate and Forest
Initiative (NICFI), the Meridian Institute, CIFOR’s Global
Comparative Study on REDD+, and the European Forest Institute’s
BMEL-financed NewGo project. J.B. acknowledges partial funding
by the Deutsche Forschungsgemeinschaft (DFG, German Research
Foundation) under Germany’s Excellence Strategy, EXC 2070–
390732324. Author contributions: Conceptualization: T.A.P.W.,
S.W., E.O.S., J.B., and A.K. Methodology: T.A.P.W., S.W., E.O.S.,
J.B., and A.K. Data processing: T.A.P.W., S.W.R., A.N.N., and G.F.;
Formal analyses: T.A.P.W. Visualization: T.A.P.W. Writing: T.A.P.W.,
S.W., E.O.S., J.B., S.W.R., A.N.N., and A.K. Funding acquisition:
T.A.P.W, S.W., and J.B. Competing interests: The authors declare
no competing interests. Data and materials availability: The
data and codes used in this study can be accessed through
the DataverseNL repository (36). License information: Copyright
© 2023 the authors, some rights reserved; exclusive licensee
American Association for the Advancement of Science. No claim to
original US government works. https://www.science.org/about/
science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade3535
Materials and Methods
Figs. S1 to S14
Tables S1 to S12
Appendix A
References (37–46)
Submitted 15 August 2022; accepted 7 July 2023
10.1126/science.ade3535
West et al., Science 381, 873–877 (2023)
25 August 2023
5 of 5
|
10.1126_science.add8655
|
RES EARCH
R E S E A R C H A R T I C L E
◥
PLANT GENETICS
Dual domestications and origin of traits
in grapevine evolution
Yang Dong1,2†, Shengchang Duan1,2†, Qiuju Xia3†, Zhenchang Liang4†, Xiao Dong1,2†§,
Kristine Margaryan5,6‡, Mirza Musayev7‡, Svitlana Goryslavets8‡, Goran Zdunić9‡,
Pierre-François Bert10‡, Thierry Lacombe11‡, Erika Maul12‡, Peter Nick13‡, Kakha Bitskinashvili14‡,
György Dénes Bisztray15‡, Elyashiv Drori16,17‡, Gabriella De Lorenzis18‡, Jorge Cunha19,20‡,
Carmen Florentina Popescu21‡, Rosa Arroyo-Garcia22‡, Claire Arnold23‡, Ali Ergül24‡, Yifan Zhu1‡,
Chao Ma25‡, Shufen Wang1,2, Siqi Liu1,2, Liu Tang1,2, Chunping Wang1,2, Dawei Li1,2, Yunbing Pan1,2,
Jingxian Li1,2, Ling Yang1,2, Xuzhen Li1,2, Guisheng Xiang1,2, Zijiang Yang1,2, Baozheng Chen1,2,
Zhanwu Dai4, Yi Wang4, Arsen Arakelyan5,26,27, Varis Kuliyev28, Gennady Spotar8, Nabil Girollet10,
Serge Delrot10, Nathalie Ollat10, Patrice This11, Cécile Marchal29, Gautier Sarah11, Valérie Laucou11,
Roberto Bacilieri11, Franco Röckel12, Pingyin Guan13, Andreas Jung30, Michael Riemann13,
Levan Ujmajuridze14, Tekle Zakalashvili14, David Maghradze14, Maria Höhn15, Gizella Jahnke15,
Erzsébet Kiss15, Tamás Deák15, Oshrit Rahimi16, Sariel Hübner31, Fabrizio Grassi32,33,
Francesco Mercati34, Francesco Sunseri35, José Eiras-Dias19,20, Anamaria Mirabela Dumitru21,
David Carrasco22, Alberto Rodriguez-Izquierdo22, Gregorio Muñoz36, Tamer Uysal37, Cengiz Özer37,
Kemal Kazan38, Meilong Xu39, Yunyue Wang1, Shusheng Zhu1, Jiang Lu40, Maoxiang Zhao25,
Lei Wang25, Songtao Jiu25, Ying Zhang41, Lei Sun41, Huanming Yang42, Ehud Weiss43,
Shiping Wang25, Youyong Zhu1, Shaohua Li4*, Jun Sheng1,2*, Wei Chen1,2*
We elucidate grapevine evolution and domestication histories with 3525 cultivated and wild accessions
worldwide. In the Pleistocene, harsh climate drove the separation of wild grape ecotypes caused by
continuous habitat fragmentation. Then, domestication occurred concurrently about 11,000 years ago in
Western Asia and the Caucasus to yield table and wine grapevines. The Western Asia domesticates
dispersed into Europe with early farmers, introgressed with ancient wild western ecotypes, and
subsequently diversified along human migration trails into muscat and unique western wine grape
ancestries by the late Neolithic. Analyses of domestication traits also reveal new insights into selection
for berry palatability, hermaphroditism, muscat flavor, and berry skin color. These data demonstrate the
role of the grapevines in the early inception of agriculture across Eurasia.
T he cultivated grapevine (Vitis vinifera
ssp. vinifera, hereafter V. vinifera) shares
a close relationship with humans (1).
With unmatched cultivar diversity, this
food source (table and raisin grapes)
and winemaking ingredient (wine grapes)
became an emblem of cultural identity in
major Eurasian civilizations (1–3), leading to
intensive research in ampelography, archae-
obotany, and historical records to reveal its
history (4). Early work asserted that V. vinifera
originated from its wild progenitor Vitis
vinifera ssp. sylvestris (hereafter V. sylvestris)
~8000 years ago during the Neolithic agri-
cultural revolution in the Western Asia (5, 6).
In recent years, various genetic studies ex-
plored this proposition (6–13), but the crit-
ical details of grapevine domestication were
often inconsistent. Studies argued for the
existence of domestication centers in the
western Mediterranean (13), Caucasus (12, 14),
and Central Asia (12), which in turn cast doubt
on the popular notion of a single past domes-
tication event (10, 11). Three demographic
inferences yielded population split times be-
tween V. vinifera and V. sylvestris to dates
between 15,000 and 400,000 years ago, pre-
dating the historical consensus on domestica-
tion time (7–9). Because early domesticates
spread to other parts of Eurasia through poorly
defined migration routes in the ensuing millen-
nia (5), the single-origin theory also confounds
the origin order between table and wine grape-
vines. One view proposes a wine grapevine–
first model, with the two types diverging
~2500 years ago (7, 10, 11). Hybridization with
local V. sylvestris was common in creating ex-
tant European wine grapes (10, 11), but when
these introgression events occurred is unknown.
Several studies suggest that the earliest cul-
tivation of European wine grapes in France
and Iberia postdates 3000 years ago (10, 15).
These discrepancies primarily result from the
inadequate sampling of grapevine accessions
and the limited resolution of genetic data in
previous analyses. Therefore, we report the
genomic variation dataset from a global co-
hort to systematically delineate the structure
of grapevine genetic diversity, explore the
origin of V. vinifera, deduce a putative dis-
persal history, and investigate key domesti-
cation traits and diversification signatures.
Results
We constructed a chromosome-level reference
V. sylvestris genome assembly (VS-1 from
Tunisia) to attain genomic variations, which
shows a higher percentage of anchored chro-
mosomal lengths than PN40024 (fig. S1 and
tables S1 to S9) (16). From the 3304 assem-
bled accessions from a dozen Eurasian germ-
plasm and private collections, we obtained
good-quality Illumina paired-end sequenc-
ing data to an average 20× coverage for 3186
grapevine accessions (2237 V. vinifera and
949 V. sylvestris; tables S10 to S13). The
sample selection preferentially included old,
autochthonous, and economically important
varieties to maximize the spectrum of genetic
diversity. We also included genomic data for
339 previously sequenced accessions (266
V. vinifera and 73 V. sylvestris; table S14) in
the analyses (7, 8, 17), producing the final
cohort of 3525 grapevine accessions (2503
V. vinifera and 1022 V. sylvestris). The align-
ment of the Illumina reads to the VS-1 refer-
ence genome identifies 45,624,306 biallelic
single-nucleotide polymorphisms (SNPs) and
7,314,397 biallelic short Indels [≤40 base pairs
(bp); 73.2% shorter than 5 bp] (16), among
which rare alleles (minor allele frequency ≤1%)
accounted for the majority (fig. S2 and tables
S15 to S22).
Core accessions differentiate by eight distinct
genetic ancestries
Clones, mutants, synonyms, and homonyms
are common phenomena in grapevine germ-
plasm and collections (18). Using the identity-
by-state sharing pattern estimators, we found
1534 accessions sharing the genetic profile
with at least one other in the cohort, total-
ing 498 distinct genotypes (fig. S3 and table
S23) (16). We kept one accession for each
distinct genotype, corrected misidentified
accessions, and excluded interspecific hy-
brids for a core cohort of 2448 grapevines
(1604 V. vinifera and 844 V. sylvestris; fig.
S3), which remain representative of the major
viticultural regions (19) in the world (Fig. 1A
and fig. S3).
Principal component analysis (PCA) showed
that V. sylvestris and V. vinifera separately
spread out along the first two axes (total vari-
ance explained: PC1 7.56% and PC2 1.71%),
with both displaying a crude Western Asia to
Western Europe gradient (Fig. 1B and figs. S4
and S5). The PC3 axis (1.26% variance) sepa-
rates V. vinifera individuals according to their
utilization, agreeing with the main table and
wine grapevine clades in the maximum likeli-
hood phylogenetic tree and reticulate phylo-
genetic network (figs. S6 and S7). The V. vinifera
accessions show a weak isolation-by-distance
correlation (Fig. 1C), suggesting a disconnection
between the viticultural geographic pattern
and the genetic structures in the grapevine
Dong et al., Science 379, 892–901 (2023)
3 March 2023
1 of 10
RES EARCH | R E S E A R C H A R T I C L E
(20). This observation could be due to the
extensive exchange of superior cultivars across
regions and the subsequent interbreeding
throughout history.
Given the poor resolution of viticultural re-
gions in defining grapevine diversity, we lever-
aged genetic ancestry information from an
unsupervised ADMIXTURE analysis to cate-
gorize core accessions (Fig. 1D and fig. S8) (16).
At K = 2, all V. vinifera accessions contain a
majority east (red) ancestry that matches the
ancestry of the V. sylvestris accessions in the
East Mediterranean region. At K = 8, hierar-
chical clustering of ancestry components iden-
tifies four V. sylvestris groups from distinct
geographic regions: Western Asia (Syl-E1,
84.3% K2), the Caucasus (Syl-E2, 72.7% K6),
Central Europe (Syl-W1, 94.7% K1), and the
Iberian Peninsula (Syl-W2, 69.8% K8; Fig. 1, D
to F). V. sylvestris accessions collected from
other regions show admixed genetic struc-
tures (16). For cultivated grapevines (CGs),
six genetic ancestries could designate six
distinctive groups (CG1 to CG6), all covering
a broad range of viticultural regions (Fig. 1,
D to F) (16). Accessions with pure or close to
pure ancestries (fig. S9) (16) helped to ascribe
names to these groups as Western Asian table
grapevines (CG1, 73.9% K2), Caucasian wine
grapevines (CG2, 66.4% K6), muscat grape-
vines (CG3, 87.7% K5), Balkan wine grapevines
(CG4, 69.9% K4), Iberian wine grapevines
(CG5, 68.8% K7), and Western European
wine grapevines (CG6, 68.4% K3). The ad-
mixed V. vinifera accessions showed dif-
ferent combinations of genetic ancestries
(fig. S9). The four V. sylvestris and six V.
vinifera groups, supported by archetypal
analysis at K = 8 (fig. S10), formed identi-
fiable clusters in the PCA plots (Fig. 1G and
fig. S4) and were thus suitable for population
genomic investigations.
Separation of V. sylvestris ecotypes
in Pleistocene
According to the genetic ancestries and the oc-
cupied ecological niches in the western Eurasia
continent, we designate V. sylvestris accessions
in Western Asia and the Caucasus as the eastern
ecotype (V. sylvestris eastern ecotype, hereafter
Syl-E) and accessions in Central Europe and
the Iberian Peninsula as the western ecotype
(V. sylvestris western ecotype, hereafter Syl-W)
(Fig. 2A). The large between-ecotype fixation in-
dex values [e.g., Syl-E1 versus Syl-W1, pairwise
population fixation index (FST) = 0.340] and the
small within-ecotype fixation index values (Syl-E1
versus Syl-E2, FST = 0.101; Syl-W1 versus Syl-W2,
FST = 0.072; fig. S11 and table S26) support this
designation. Both nucleotide diversity (p) and
individual heterozygosity show that the west-
ern ecotype (especially Syl-W1) has significantly
reduced variation compared with its eastern
counterpart (fig. S11). Furthermore, the linkage
disequilibrium decay (LD, r2) was much slower
in Syl-W (1.0 to 1.6 Kb at half of maximum r2)
than in Syl-E (400 to 600 bp at half of maximum
r2; fig. S12). These data demonstrate that the
eastern ecotype retains more genetic diversity.
Demographic inference with folded SNP
frequency spectra reveals an ancient population
bottleneck in Syl-E ~400,000 to 800,000 years
ago and in Syl-W ~150,000 to 400,000 years ago
(Fig. 2B and fig. S13). This Pleistocene period,
characterized by changing climate cycles (21, 22),
also witnessed the deduced population split
(median time ~200,000 to 400,000 years ago)
between the two ecotypes (Fig. 2C). The slow
descent of the split line suggests that the
geographic isolation process was gradual (fig.
S13). At ~56,000 years ago, the population split
between Syl-E1 and Syl-E2 occurred during
the last glacial cycle (11,700 to 115,000 years
ago), when the global climate trended toward
dryer and colder conditions (23). Close to the
time of the Last Glacial Maximum (LGM;
~21,000 years ago), V. sylvestris subgroups
experienced a second population bottleneck
(~40,000 years ago), with effective population
sizes (Ne) reaching a minimum of 10,000 to
40,000 (Fig. 2B and fig. S13). After this result,
ecological niche modeling predicts that the
areas with suitable environmental conditions
for Syl-E and Syl-W (suitability > 0.75) remained
connected at the Pleistocene Last Interglacial
(~130,000 years ago) (fig. S14) but became
entirely separated at the LGM (Fig. 2D). The
post-bottleneck Ne rebound was steeper in the
Syl-W accessions, but the numbers decreased
to lower levels in recent times (Fig. 2B and fig.
S13). This result agrees with the reduced ge-
netic diversity in Syl-W and the abrupt pop-
ulation split between Syl-W1 and Syl-W2 at
~2500 years ago.
Dual origin of V. vinifera at the advent
of agriculture
The wet climate in the Early Holocene (11,700
to 8300 years ago) (24) facilitated the expan-
sion of suitable habitats for Syl-E, resulting in
a large geographic span from Central Asia
to the Iberian Peninsula (Fig. 2D). This ex-
pansion supports the eastern origin and sub-
sequent continental dispersal of V. vinifera.
Because CG1 shares the main ancestral com-
ponent with Syl-E1 and CG2 with Syl-E2 (Fig. 1,
D and F), the possibility of two domestication
events becomes evident. Indeed, both CG1 and
CG2 maintain the highest genetic diversity and
manifest the quickest LD decay among all CG
groups (figs. S11 and S12). Furthermore, they are
less differentiated from their corresponding
wild ecotypes (Fig. 3A and fig. S11). The Akaike
information criterion (AIC)–based phylogenetic
selection also prefers a dual origin tree model
(fig. S15), which agrees with the outgroup f3 sta-
tistics biplots that CG1 and CG2 are genetically
1State Key Laboratory for Conservation and Utilization of Bio-Resources in Yunnan, Yunnan Agricultural University, Kunming 650201, China. 2Yunnan Research Institute for Local Plateau
Agriculture and Industry, Kunming 650201, China. 3State Key Laboratory of Agricultural Genomics, BGI-Shenzhen, Shenzhen 518083, China. 4Beijing Key Laboratory of Grape Science and
Oenology and Key Laboratory of Plant Resources, Institute of Botany, the Chinese Academy of Sciences, Beijing 100093, China. 5Institute of Molecular Biology, NAS RA, 0014 Yerevan, Armenia.
6Yerevan State University, 0014 Yerevan, Armenia. 7Genetic Resources Institute, Azerbaijan National Academy of Sciences, AZ1106 Baku, Azerbaijan. 8National Institute of Viticulture and
Winemaking Magarach, Yalta 298600, Crimea. 9Institute for Adriatic Crops and Karst Reclamation, 21000 Split, Croatia. 10Bordeaux University, Bordeaux Sciences Agro, INRAE, UMR EGFV, ISVV,
33882 Villenave d’Ornon, France. 11AGAP Institut, University of Montpellier, CIRAD, INRAE, Institut Agro Montpellier, 34398 Montpellier, France. 12Julius Kühn Institute (JKI) – Federal Research
Center for Cultivated Plants, Institute for Grapevine Breeding Geilweilerhof, 76833 Siebeldingen, Germany. 13Botanical Institute, Karlsruhe Institute of Technology, 76131 Karlsruhe, Germany.
14LEPL Scientific Research Center of Agriculture, 0159 Tbilisi, Georgia. 15Hungarian University of Agriculture and Life Sciences (MATE), 1118 Budapest, Hungary. 16Department of Chemical
Engineering, Ariel University, 40700 Ariel, Israel. 17Eastern Regional R&D Center, 40700 Ariel, Israel. 18Department of Agricultural and Environmental Sciences, University of Milano, 20133 Milano,
Italy. 19Instituto Nacional de Investigação Agrária e Veterinária, I.P./INIAV-Dois Portos, 2565-191 Torres Vedras, Portugal. 20Green-it Unit, Instituto de Tecnologia Química e Biológica,
Universidade Nova de Lisboa, 2780-157 Oeiras, Portugal. 21National Research and Development Institute for Biotechnology in Horticulture, Stefanesti, 117715 Arges, Romania. 22Center for Plant
Biotechnology and Genomics, UPM-INIA/CSIC, Pozuelo de Alarcon, 28223 Madrid, Spain. 23University of Lausanne, 1015 Lausanne, Switzerland. 24Biotechnology Institute, Ankara University,
06135 Ankara, Turkey. 25Department of Plant Science, School of Agriculture and Biology, Shanghai JiaoTong University, Shanghai 200240, China. 26Armenian Bioinformatics Institute, 0014
Yerevan, Armenia. 27Biomedicine and Pharmacy, RAU, 0051 Yerevan, Armenia. 28Institute of Bioresources, Nakhchivan Branch of the Azerbaijan National Academy of Sciences, AZ7000
Nakhchivan, Azerbaijan. 29Vassal-Montpellier Grapevine Biological Resources Center, INRAE, 34340 Marseillan-Plage, France. 30Historische Rebsorten-Sammlung, Rebschule (K39), 67599
Gundheim, Germany. 31Galilee Research Institute (Migal), Tel-Hai Academic College, 12210 Upper Galilee, Israel. 32Department of Biotechnology and Biosciences, University of Milano-Bicocca,
20126 Milano, Italy. 33NBFC, National Biodiversity Future Center, 90133 Palermo, Italy. 34Institute of Biosciences and Bioresources, National Research Council, 90129 Palermo, Italy. 35Department
AGRARIA, University Mediterranea of Reggio Calabria, Reggio 89122 Calabria, Italy. 36IMIDRA, Alcalá de Henares, 28805 Madrid, Spain. 37Viticulture Research Institute, Ministry of Agriculture and
Forestry, 59200 Tekirdağ, Turkey. 38Queensland Alliance for Agriculture and Food Innovation, University of Queensland, St. Lucia, Queensland 4072, Australia. 39Institute of Horticulture, Ningxia
Academy of Agricultural and Forestry Sciences, Yinchuan 750002, China. 40Center for Viticulture and Oenology, School of Agriculture and Biology, Shanghai JiaoTong University, Shanghai
200240, China. 41Zhengzhou Fruit Research Institutes, CAAS, Zhengzhou 450009, China. 42BGI-Shenzhen, Shenzhen 518083, China. 43The Martin (Szusz) Department of Land of Israel Studies
and Archaeology, Bar-Ilan University, 5290002 Ramat-Gan, Israel.
*Corresponding author. Email: wchenntr@gmail.com (W.C.); shengjun@dongyang-lab.org (J.S.); shhli@ibcas.ac.cn (S.L.)
†These authors contributed equally to this work. ‡Institution contacts for biological samples. Ordered by country names. §Present address: Department of Chromosome Biology, Max Planck Institute for Plant
Breeding Research, 50829 Cologne, Germany.
Dong et al., Science 379, 892–901 (2023)
3 March 2023
2 of 10
RES EARCH | R E S E A R C H A R T I C L E
A
UK
France
Germany
Czechia
Moldova
Hungary
Serbia
Ukraine
Russia
V. vinifera
V. sylvestris
No. of
100
accessions 10
Map Data ©2021 Google
Austria
Slovenia
Switzerland
Romania
Bulgaria
B&H
Georgia
Italy
Croatia
Montenegro
Portugal
Spain
Morocco
Algeria
Tunisia
Cyprus
Albania
Greece
Turkey
Armenia
Syria
Lebanon
Jordan
USA
Egypt
Israel
Argentina
South Africa Australia
N.D.
Sudan
Yemen
Daghestan (RUS)
Azerbaijan
Uzbekistan
Kazakhstan
Tajikistan
Turkmenistan
Kyrgyzstan
China
Afghanistan
Iran
India
Pakistan
S. Korea
D
K=2
K=8
E
Syl-W1 (n=165)
F
Major Viticultural Regions
Western Asia
Turkey
Central Asia
Eastern Asia
Russia/Ukraine
Maghreb
Rest of World
Balkan
Western Euro
Caucasus
Central Euro
Eastern Euro
Iberia
Italy
W. Asia
Caucasus
Rus/Ukr
C. Asia
Turkey
E. Euro
C. Euro
Iberia
Rest.
World
Maghreb
Balkan
Italy
W. Euro
E. Asia
B
4
0
.
0
)
%
1
7
.
1
(
2
C
P
0
0
.
0
4
0
.
0
-
V. vinifera
V. sylvestris
s
i
r
t
s
e
v
y
s
l
.
V
n
r
e
t
s
e
W
e
p
y
t
o
c
E
i
d
e
x
m
d
A
i
)
x
m
d
A
-
l
y
S
(
n
r
e
t
s
a
E
e
p
y
t
o
c
E
0.000
0.025
PC 1 (7.56%)
V. sylvestris
Mantel r = 0.448
(P = 7 10-4)
V. vinifera
Mantel r = 0.280
(P = 0.014)
d
e
t
a
v
i
t
l
u
c
s
p
u
o
r
g
e
p
a
r
g
1 104
Distance (km)
0
Syl-W1
Syl-W2
2 104
Syl-E1
Syl-E2
a
r
e
f
i
n
v
i
.
V
j
r
o
a
m
x
S
i
-0.025
C
4
.
0
3
.
0
T
S
F
2
.
0
1
.
0
0
.
0
5
2
0
.
0
5
2
0
.
0
-
5
7
0
.
0
-
G
)
%
6
2
1
(
.
3
C
P
i
d
e
x
m
d
A
CG1
CG3
CG5
CG2
CG4
CG6
0.04
-0.04
-0.00
PC 2 (1.71%)
i
)
x
m
d
A
C
-
(
Syl-W1
Syl-W2
Syl-E2
Syl-E1
CG1
CG2
CG3
CG4
CG5
CG6
Japan
K1
K5
K2
K6
K3
K7
K4
K8
(%)
0
20
40
60
80 100
Hungary
Austria
Germany
Syl-W1
94.7%
Syl-W2 (n=112)
Syl-W2
69.8%
P
ortu
g
al
Spain
France
Syl-E1 (n=98)
Israel
Syl-E2 (n=81)
Armenia
Iran
Georgia
n
a
j
i
a
b
r
e
z
A
Syl-E1
84.3%
Syl-E2
72.7%
CG1
73.9%
CG2
CG3
CG4
CG5
66.4%
87.7%
69.9%
68.8%
CG6
68.4%
CG1 (n=343)
CG2 (n=96)
CG3 (n=117)
Ukr/Rus
Turkey
E. Asia
.
a
c
u
a
C
C. Asia
W. Asia
Maghreb
Balkan
E. Euro
Ukr/Rus
Caucasus
E. Euro
E. Asia
Other Regions
Rest. World
W. Euro
Italy
CG4 (n=246)
CG5 (n=137)
CG6 (n=129)
Turkey
Italy
Ukr/Rus
E. Euro
Balkan
Italy
Iberian
C. Euro
Iberian
W. Euro
Dong et al., Science 379, 892–901 (2023)
3 March 2023
3 of 10
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Genetic diversity of global core V. sylvestris and V. vinifera
accessions. (A) Geographical locations of the 2448 core grapevine
accessions. (B) PCA according to major viticultural regions. Large square/
circle highlights median position. Star shows VS-1 position. (C) Isolation-
by-distance test of V. sylvestris and V. vinifera accessions. Linear regression
with 95% confidence interval is shown. (D) ADMIXTURE clustering of the
accessions. (E) Geographic locations of the accessions in each group.
Gray represents minor locations. (F) Average proportion of major genetic
ancestries in grapevine groups. (G) PC2 versus PC3 projection according to
grapevine group.
A
Western Ecotype
Syl-W1
No. of
accessions
B
6
0
1
100
10
Syl-E1
e
N
5
0
1
4
0
1
6
0
1
e
N
5
0
1
Alps
Syl-W2
Black Sea
Syl-E2
Mediterranean Sea
Zagros Mountain
Map Data ©2021 Google
C
Population Split Time
200-400 kya
kya
e
e
n
n
e
e
c
c
o
o
t
t
s
s
e
e
P
P
i
i
l
l
56 kya
ya
Syl-E1
4
0
1
Syl-W1
Eastern Ecotype
1
2
4 6 810
20
40 6080100 200 400600800
3
( 10
years ago)
1000
103
(ya)
104
105
106
D
Western Ecotype
Last Glacial Maximum
~21 Kya
E
E
-
-
l
l
y
y
yyy
S
S
W
W
-
-
l
l
y
y
S
S
Syl-E1/W1
Syl-E1/W2
Syl-E2/W1
Syl-E2/W2
0.75-0.80
0.80-0.85
0.85-0.90
>0.90
<0.75
Eastern Ecotype
Early Holocene
~11.7 - 8.3 Kya
ya
2,500 ya
e
e
n
n
e
e
c
c
o
o
o
o
H
H
l
l
E2 E1 W1 W2
Syl-E1/E2
Syl-W1/W2
Fig. 2. Population history of V. sylvestris ecotypes. (A) Geographic isolation
and population separation of V. sylvestris ecotypes. Pie charts show mean
ancestry proportion at K = 8. Same color scheme as in Fig. 1B is used. (B)
Demographic histories of V. sylvestris populations deduced from Stairway Plot 2.
Lines indicate medians with 75% and 95% confidence intervals. (C) Population
split times among ecotypes with MSMC2. Red bars indicate medians with 95%
confidence intervals. (D) Ecological niche modeling of the suitable habitats for V.
sylvestris ecotypes. The color scale shows suitability score.
closer to Syl-E1 and Syl-E2, respectively (Fig.
3B, fig. S15, and table S27). The population
split lines of CG1/Syl-E2 and CG2/Syl-E1 pairs
resemble that of Syl-E1/Syl-E2 and differ from
those of CG1/Syl-E1 and CG2/Syl-E2 pairs (Fig.
3C and fig. S16). These data collectively support
a dual origin of V. vinifera and reject the
popular theory of a single primary domestica-
tion center (10, 11). Both CG1/Syl-E1 and CG2/
Syl-E2 population pairs separated quickly
(Fig. 3C), which is compatible with a clean-
split scenario. We estimate the median popu-
lation split time to be ~11,000 years ago (95%
confidence interval: ~10,500 to 12,500 years ago)
for both pairs, suggesting that the domestica-
tion events took place concurrently around the
advent of agriculture. Because CG1 and CG2
separately represent table and wine grapevine
ancient genetic backgrounds (K2 and K6;
fig. S9), the dual origin rejects the assump-
tion that wine grapevines predate table grape-
vines (7, 10, 11).
Dispersal of grapevine domesticates along
human migration routes
The geographic distributions of CG1 and CG2
cultivars across Eurasia and North Africa cor-
respond to vastly different human migration
routes for the two grapevine groups (Fig. 3D).
The CG2 cultivars were mainly confined to
both sides of the Caucasus Mountains, with a
limited dispersal into the Carpathian Basin
by the northern Black Sea. This result con-
trasts with previous models implying that CG2
played a central role in the formation of wine
grapevines in Europe (3). Instead, CG2 repre-
sents a local domestication effort that had a
minor impact on grapevine diversification. By
comparison, the dispersal of CG1 in four direc-
tions spanned Eurasia and North Africa. First,
the eastward expansion through Central Asia
into India and China follows the Inner Asian
Dong et al., Science 379, 892–901 (2023)
3 March 2023
4 of 10
RES EARCH | R E S E A R C H A R T I C L E
A
FST
CG1
CG2
CG3
CG4
CG5
CG6
B
1.68
1.67
1.66
1.65
1.64
)
d
n
u
o
R
t
;
X
,
2
G
C
(
3
f
0.4
T
S
F
0.2
0
4
8
o
i
t
a
r
12
0.6
0.4
T
S
F
0.2
o
i
t
a
r
0
4
8
12
0.3
0.2
0.1
0
C
1.0
Population Split Time
Syl-E
Syl-E1
E2
W1W2
0.8
Syl-E1/ CG1
Syl-E2/ CG2
Syl-E1/E2
CG1/ Syl-E2
Syl-E1/ CG2
Syl-E2/CG1
Syl-E1/CG2
Syl-E1/E2
Syl-E2/CG2
Years (g=3, µ=5.4 10-9)
Syl-E1/CG1
56 kya
e
n
e
c
o
i
t
s
e
P
l
Syl-E2
Syl-E1
11 Kya (Domestication)
CG2
CG1
advent of
farming
e
n
e
c
o
o
H
l
104
105
106
103 (ya)
104
105
Caucasus Western Asia
0.6
R
C
C
R
0.4
0.2
0.0
103
D
Syl-E2
CG4
CG3
CG6
CG5
Syl-W2
Syl-W1
Syl-E1
Dispersal Route
No. of V. vinifera
accessions
CG1
CG2
100
10
1
Caucasus
Domestication
Center
(Wine)
1.63
1.62
1.62 1.63 1.64 1.65 1.66 1.67 1.68
f3 (CG1, X; Rotund)
E
Map Data ©2021 Google
Western Asia
Domestication
Center
(Table)
0.6
Syl-E1/CG1
Chr 5:
UPL6
SWEET17
MecgoR
Chr 11:
PME
CCoAOMT
Chr 15:
BEAT
PPR
Chr 18:
PPR
Cyp71A
0.5
0.4
Chr1 2
3
4
5
6
7
8
9
10
11 12
13
14
15
16 17
18
19
Chr 2:
NPF
VvMybA
Sex Determination Region
Chr 8:
FER4
Chr 9:
GA2OX
Chr 17:
PPR
RNF181
UCH
NPF
SDH
Chr1 2
3
4
5
6
7
8
9
10
11 12
13
14
15
16 17
18
19
Chr 3:
UFGT
Chr 6:
PFKFB1
UCH
PPR
Chr 7:
ANKRD44
TR2-like
Syl-E2/CG2
Chr 17:
WAK
SSL
T
S
F
0.3
0.2
-3 -2 -1 0 1 2 3
Log2( -Syl-E1/ -CG1)
0.1
0.0
0.3
0.2
T
S
F
0.1
0.0
-2 -1 0 1 2
-3
Log2( -Syl-E2/ -CG2)
3
Fig. 3. Dual domestications of V. vinifera in Western Asia and the Caucasus.
(A) Pairwise fixation index of the major grapevine groups. (B) Outgroup
f3 statistics biplot measuring genetic similarity. Rotund, Muscadinia rotundifolia.
Stars mark the f3 statistics for CG1/CG2. (C) Estimated split times among
Syl-E1/2 and CG1/2 with MSMC2 (left). Red bars indicate medians with 95%
confidence intervals. (D) Geographic distribution of CG1 and CG2 in relation to the
domestication centers. Human dispersal routes are shown. (E) Shared (sky blue)
and unique domestication selective sweep regions (red and dark teal) in V. vinifera.
Mountain Corridor, a path that also witnessed
the exchange of other crops (i.e., wheat, barley,
and millet) between the West and the East
(25). Second, the northbound expansion could
mirror the early cultural contact of West-
ern Asia over the Zagros mountains with the
Caucasus (26, 27). Third, the northwest ex-
pansion through Anatolia into the Balkans
bespeaks the spread of farming into Europe
(28, 29). Finally, a westward expansion moved
across the North African coastline to reach
Morocco (30). Even though grapevine domes-
ticates followed the trails of past human migra-
tion, the timing and dispersal details require
paleogenomic data for delineation.
Shared and unique domestication signatures in
CG1 and CG2 grapevines
Given the dual origin scenario, we investigated
domestication signatures in both Syl-E1/CG1
and Syl-E2/CG2 group pairs by selecting geno-
mic regions that display increased nucleotide
diversity differences and population differen-
tiation (both top 5%; Fig. 3D). This method
yields 1140 domestication selective sweep genes
in 132 regions for CG1 and 887 genes in 137
regions for CG2 (table S28), among which only
189 genes in 31 regions exist in both groups
(table S29). Most shared signals are on chromo-
somes 2 and 17, confirming previous find-
ings that the selection on flower sexual morphs
(sex determination region, SDR), berry skin
color (VvMybA gene cluster), and berry de-
velopment (SDH gene cluster) were of great
importance during grapevine domestication
(8, 11). In addition, our analysis identifies
Dong et al., Science 379, 892–901 (2023)
3 March 2023
5 of 10
RES EARCH | R E S E A R C H A R T I C L E
A
Syl-E1
C
5 105
N
1.0e+03
4.0e+03
1.6e+04
6.4e+04
2.6e+05
Prob
50.0%
0.0%
28.9%
24.7%
50.0%
10,564 ya
Syl-E2
Syl-W1
Syl-W2
CG2
CG4
CG3
CG6
Migration
weight
0.5
)
a
y
(
e
m
T
i
105
104
103
N
1.0e+03
4.0e+03
1.6e+04
6.4e+04
2.6e+05
1.0e+06
Prob
50.0%
0.0%
N
1.0e+03
4.0e+03
1.6e+04
6.4e+04
2.6e+05
1.0e+06
Prob
50.0%
0.0%
N
1.0e+03
4.0e+03
1.6e+04
6.4e+04
2.6e+05
1.0e+06
Prob
50.0%
0.0%
16.6%
18.0%
41.5%
8,070 ya
7,740 ya
6,910 ya
CG5
CG1
0
10 s.e.
0.000
0.010
Drift parameter
0.020
P2
CG3 CG4
CG5 CG6
B
P1
CG1
CG1
Syl-W1
Syl-W2
O
P3
Syl-W1
Syl-W2
CG1
0.0
0.1
D-statistic
0.2
0.3
Syl-E1
CG1
CG3
Syl-W2
Syl-E1
CG1
CG4
Syl-W2
Syl-E1
CG1
CG5
Syl-W2
Syl-E1
CG1
CG6
Syl-W2
D
2nd Unique Introgression
from Syl-W into CG6
CG6-CG1 split
(~6,900 ya)
CG6
CG4
CG5
CG5-CG1 split
(~7,700 ya)
CG4-CG1 split
(~8,000 ya)
Map Data ©2021 Google
No. of
accessions
100
10
1
(Wine)
CG2
CG3
CG3-CG1 split
(~10,500 ya)
CG1
Shared Introgression
from Syl-W
(Table)
Fig. 4. Stepwise diversification of V. vinifera in Europe. (A and B) Introgression from Syl-W into European V. vinifera groups revealed by TreeMix (A) and confirmed by
D-statistic (B). (C) Four population simulation of split times and genetic introgression using Momi2. Median numbers from 100 bootstrap runs are shown. (D) Origination
of V. vinifera groups (CG3 to CG6) by the end of the Neolithic. Geographic distributions of CG groups are shown by colored circles. See fig. S24 for details on CG3.
shared domestication genes that possibly
underlie grapevine growth (e.g., NPF), phys-
iology (e.g., FER4), fruit set (e.g., the GA2OX
gene cluster), and resistance to biotic/abiotic
stress (e.g., FER4, the PPR gene cluster, and
the RNF181 gene cluster) [see (16) for gene
descriptions].
As expected for dual domestications, most
selective sweep signatures in CG1 and CG2 are
unique and target distinctive chromosomal re-
gions (Fig. 3E). Even though CG1 and CG2 cor-
respondingly represent table and wine grapevines,
many unique signatures seem to suggest a con-
vergent selection mechanism targeting different
aspects of common domestication traits. An
obvious example is the improvement of berry
palatability through the reduction of alkaloid
biosynthesis (the MecgoR gene cluster in CG1
and the TR2 and SSL gene clusters in CG2) and
the enhancement of carbohydrate metabolism
(SWEET17 in CG1 and PFKFB1 in CG2). Other
examples include perceived berry desirability
(the BEAT gene cluster for floral scent in CG1
and the UFGT gene cluster for berry color in
CG2) and response to environmental stresses
(UPL6 in CG1 and WAK in CG2). These find-
ings suggest that the initial cultivation of CG1
and CG2 may have been to serve early humans’
caloric and micronutrient needs. The selection
of genetic features suitable for winemaking in
CG2 could have been serendipitous, and the prac-
tice of winemaking with CG2 (e.g., 8000 years
ago) (14) possibly postdates grapevine domes-
tication. Because gene annotation depends
on homology-based inference, it should be noted
that many genes mentioned here need further
verification in grapevines.
Wine grapevine diversification in Europe
Because the CG1 early domesticates dispersed
into Europe through Anatolia, a crucial ques-
tion concerns the diversification history of
European wine grapevines in the ensuing
millennia. In particular, the shared areas
of suitable habitats for Syl-E and Syl-W in
the early Holocene (black area in Fig. 2D)
formed an ecological foundation for the
genetic exchange between CG1 and local
refugia Syl-W accessions in the coastal re-
gions of the northern Mediterranean Sea and
the southern Black Sea, the Iberian Peninsula,
and an area corresponding to present-day
western France. It is therefore important
to examine where and how distinct grape-
vine genetic ancestries (CG3 to CG6) formed
with relevance to Syl-W introgression (10, 11).
We have chosen cultivars in each group
with at least 75% major ancestry (and with an
average Syl-W ancestry in each V. vinifera
group <3%) to perform population analyses.
This selection rules out many old varieties
‘Lambrusco’ cultivars deriving about
(i.e.,
half of their ancestries from Syl-W; fig. S9),
which likely showcase secondary diversifica-
tion efforts after the distinct ancestries had
been established. The TreeMix analysis finds
one migration edge that points from Syl-W to
a population ancestral to CG3 to CG6 (esti-
mated weight, 0.114; Fig. 4A and fig. S17),
suggesting an ancient introgression event
occurred before the diversification of all
European grapevines. An additional migration
Dong et al., Science 379, 892–901 (2023)
3 March 2023
6 of 10
RES EARCH | R E S E A R C H A R T I C L E
A
)
o
i
t
a
r
(
2
g
o
L
-
2
0
-2
-4
-6
-8
Selective Sweep Region
Syl-E1/CG1
Syl-E2/CG2
0.6
0.4
FST
ST
0.2
0.0
Chr2: 14.25
14.30
14.35 (Mb)
FRO7 GDS
SKU5 BFRUCT/TRA TPPF
MUR3
PLATZ
WRKY
YABBY
INP
KAS3
FMO
Major haplotype
Minor haplotype
M
f
H1
H2
Mv
fv
H3
H4
H5
C
CG1
CG2
CG3
CG4
CG5
CG6
n=164 666 104 26
85
100%
75%
50%
25%
0%
f/f
H1/H1
H1/f
H1/H2
H2/f
D
1st
Recombination
B
M/f
M/f
f/f
H1/f
N.D.
H1/H1
H2/f
H1/H2
Mv/f
M/H1
M/H5
H1/fv
H5/f
H4/f
H2/H2
H2/H3
1 2
Recombination Site
3 45
0/0
1/0
1/1
-/-
M
f
E
Map Data ©2021 Google
site 4
site 3
Mv
site 2
site 1
fv
H1
H4
2nd
Recombination
site 5 site 4
site 5
H3
H5
H2
2nd Recombination
H2 haplotype
in V. v. samples
H4 haplotype
in V. v.?
H4 haplotype
H2 haplotype
1st Recombination
H4 haplotype
in V. s. samples (IS164, IS167, IS180)
Fig. 5. Selection and evolution of the SDR in the core grapevine accessions. (A) The SDR in VS-1. Red arrows indicate identified recombination sites. (B) SDR
genotypes from associated SNPs reveal five recombination sites (dashed lines) and genotype diversity (right). Major and minor haplotypes are shown on the left.
(C) Distribution of SDR genotypes in the six major grapevine groups. (D) Recombination history of all SDR haplotypes. (E) Putative dispersal route of the H4 haplotype
and the origination of H2 haplotype.
edge also points from Syl-W to CG6 (estimated
weight, 0.292), which implies an independent
introgression event unique to Western Euro-
pean wine grapevines in the past. Various
combinations of D-statistics testing the gene
flow from Syl-W into CG groups (Z score > 3.0,
adjusted P < 4.17 × 10−5; Fig. 4B and table S31)
support this introgression history. Additional-
ly, gene flow from Syl-W into CG3 to CG6
inferred from Momi2 align with their corre-
sponding divergence from CG1, further sup-
porting the introgression history (Fig. 4C). The
estimated median divergence times date the
creation of muscat grapes (CG3) to 10,500 years
ago, Balkan wine grapes (CG4) to 8070 years
ago, Iberian wine grapes (CG5) to 7740 years ago,
and Western European wine grapes to 6910 years
ago (Fig. 4D). These stepwise diversification
times agree with the historical migration of
Anatolian farmers into Europe (26, 29, 31, 32),
substantiating the role of viticulture in forming
Neolithic agricultural societies.
The migration edge weights, f4 ratio, and
Momi2 estimates collectively show that an-
cient introgression from Syl-W accounts for
~11.4 to 18.0% of the CG3 to CG6 genomes
(Fig. 4 and table S30). In addition, at least
one other independent introgression event
contributed ~25.0 to 30.0% additional Syl-W
to the CG6 ancestry. We have screened the
introgression tracts in CG3 to CG6 by choos-
ing the genomic windows with the top 1%
df and fdM values (fig. S18). Ten shared re-
gions among the CG3 to CG6 groups con-
tain genes that are putatively involved in
plant immunity (e.g., CYSK), abiotic stress
response (e.g., GBA), and carbohydrate me-
tabolism (e.g., TPS/TPP) (table S31). This
result agrees with the proposal that intro-
gression helps grapevines adapt to new en-
vironments and become more suitable for
winemaking (10, 11).
Genetic analyses of domestication and
diversification traits
Hermaphroditism: origin of H2 haplotype
The transition from dioecy in V. sylvestris (male,
M/f; female, f/f) to hermaphroditism in V. vinifera
Dong et al., Science 379, 892–901 (2023)
3 March 2023
7 of 10
RES EARCH | R E S E A R C H A R T I C L E
Geological
period
Changing Climate
Cycles
Dry
Climate
LGM
Pleistocene
Paleolithic
Archeological
Time (Kya)
200
Increasing Climate
Variations
Holocene
Mesolithic and Neolithic
Bronze and Iron Ages
to present
56
21
11.0
8.0 7.7
6.9
2.5
Human Population
Bottleneck
Agriculture Reaching
Eastern Europe
Agriculture Reaching
South France
Early Hominin
Expansion
Modern Human
Out of Africa
Advent of
Agriculture
Agriculture
Reaching
Iberia
Grapevine Reaching
Xinjiang
Grapevine
Grapevine
Population
Population
Bottleneck
Bottleneck
Introgression
Introgression
Introgression
Introgression
Separation of
Separation of
Syl-W1 and Syl-W2
Syl-W1 and Syl-W2
Syl-W2
Syl-W1
Syl-W1
alaaa
ers
ers
ispspspppppppers
erer
d D
rd DDD
d d
WWWWestwa
Westward Dispersal
WW
Syl-W
Syl-W
sylvestris
V. sylvestris
Separation of
Separation of
Syl-E and Syl-W
Syl-E and Syl-W
Syl-E
Syl-E
Separation of
Separation of
Syl-E1 and Syl-E2
Syl-E1 and Syl-E2
Domestication
Domestication
of CG1
of CG1
Domestication
Domestication
of CG2
rn Europe
CG6 in WesteWW
CG6 in Western Europe
CG5 in Iberia
CG5 in Iberia
CG4 in Balkan
CG4 in Balkan
CG3 Broad Distribution
CG3 Broad Distribution
CG1 Broad Distribution
CG1 Broad Distribution
Syl-E1
Syl-E1
Syl-E2
Syl-E2
CG2 In Caucasus
CG2 In Caucasus
V. sylvestris evolution
Domestication and
early diversification
Intensive diversification
and dispersal
Fig. 6. Schematic graph of grapevine evolutionary history. Key events in the evolutionary history of grapevines are shown alongside major events in global climate
change and human migration.
is the most prominent phenotypic change
during domestication (33). It involves recom-
bination events between M and f around a se-
lective sweep region on chromosome 2 known
as the SDR (Fig. 5A). Previous studies have
identified two major hermaphroditic haplo-
types (H1 and H2) and four hermaphroditic
genotypes (H1/f, H2/f, H1/H1, and H1/H2)
from select cultivars (33), but the recombi-
nation history remains unclear. The analysis
of our grapevine cohort reveals five recombi-
nation sites in the SDR (Fig. 5B), which not
only confirms known genotypes but also iden-
tifies new minor haplotypes (male variant Mv,
female variant fv, H3, H4, and H5) and geno-
types (Mv/f, M/H1, M/H5, H1/fv, H5/f, H4/f,
H2/H2, and H2/H3) in both wild and culti-
vated grapevines (Fig. 5B and table S32).
Among all SDR haplotypes, M and H1 mani-
fest the highest subtype diversity (figs. S19 to
S22). Furthermore, the SDR genotype statistics
reveal a distribution bias of the H2-containing
SDRs in the Iberian (CG5) and Western Euro-
pean (CG6) grapevines (Fig. 5C and fig. S23). To
investigate this observation, we constructed a
putative recombination history for all known
SDR haplotypes (Fig. 5D), which showed that a
first recombination event between the paren-
tal M and f haplotypes created Mv (site 4), fv
(site 3), H1 (site 2), and H4 (site 1). On this
basis, H1 experienced a second recombination
event with f to produce H3 (site 5) and H5
(site 4), whereas H4 recombined again with f
at site 5 to bring about H2. Because three Syl-E
V. sylvestris (IS164, IS167, and IS180) and 11
V. vinifera accessions in the cohort contain
H4 (Fig. 4G), a likely scenario supports a west-
ward dispersal of H4 after human selection to
reach the Iberian Peninsula [e.g., in extant old
Iberian cultivar ‘Malvasia Fina’ (PO153)], where
H2 originated from H4 through secondary
recombination and later became dominant
during the diversification of Iberian and West-
ern European cultivars.
Muscat flavor: Trait selection may reduce
grapevine fitness
Muscat grapevine is unique for its floral aromas,
which result from a hard-to-define concoction
of monoterpenoids in the fruit (34). Given the
broad geographic distribution (fig. S24) and
ancient history of muscat grapevines, it is not
easy to pinpoint the center of origin. However,
Momi2 estimate predicts a population split
from CG1 at ~10,564 years ago (Fig. 4C), sug-
gesting an origination site within the bound-
ary of Western Asia. This scenario agrees with
the relatively low FST values and sizeable gene
flow with CG1 (Fig. 4 and fig. S11). The CG3
group also shows low genetic diversity and
high LD extent compared with the others (figs.
S11 and S12). One possible reason is the grad-
ual loss of ancient CG3 cultivars in Anatolia and
the surrounding regions throughout history
(fig. S24). Even though the muscat aroma is a
complex trait, genome-wide association anal-
ysis based on a binary differentiation reveals
18 SNP signatures on chromosomes 5 and 18
(fig. S24 and table S33). This set includes a
nonsynonymous SNP Chr5:19419686 in the
VvDXS gene linked to the trait (34). Examina-
tion of the genotype at this locus shows that
108 of the 134 muscat grapevines (including
‘Muscat Hamburg,’ ‘Königin der Weingärten,’
and ‘Muscat of Alexandria,’ which are commonly
Dong et al., Science 379, 892–901 (2023)
3 March 2023
8 of 10
RES EARCH | R E S E A R C H A R T I C L E
used as parental cultivars) are heterozygous
(G/T), and only eight individuals are homo-
zygous (T/T) for the alternative SNP (exact
test for Hardy-Weinberg equilibrium, D =
20.68, P = 2.01 × 10−13). Additionally, most
grapevines without muscat aroma are homo-
zygous for the reference SNP (G/G; 1451 of
1468; exact test for Hardy-Weinberg equilib-
rium, D = 0.049, P = 1.00). This result sug-
gests that selection on this allele might have
put a constraint on grapevine fecundity, there-
by preventing the alternative SNP from reach-
ing fixation.
Berry skin color: New genes associated
with white grapes
The emergence of white grapes from their red-
berried congeners is an essential domestica-
tion episode in viticulture history. The color
change results from a reduction of anthocyanin
synthesis in berry skin cells, where the expres-
sion of proposed master regulators such as
VvMybA decreased significantly in select culti-
vars because of a Gret1 retrotransposon (35),
nonconservative exonic mutations (36), or large
deletions in the locus (37). We performed genome-
wide association analysis on this large grape-
vine cohort (fig. S25, A and B) and identified
multiple significant SNPs across the genome
(fig. S25C). The most prominent peak spans a
broad genomic region from 3.51 to 16.05 Mb on
chromosome 2, overlapping the VvMybA locus.
Among all significant exonic SNPs in this region
(table S34), nonsynonymous SNPs with the
smallest P values localize to two uncharacterized
genes outside the VvMybA locus (fig. S25D),
the putative protein functions of which are
acylaminoacyl-peptidase (Vvsyl02G000229) and
lysine-specific demethylase (Vvsyl02G001064).
These SNPs are overwhelmingly homozygous
for the reference allele in white grapes and are
heterozygous in red grapes (fig. S25E). We
validated the SNPs in red-berried V. sylvestris
accessions to account for possible false pos-
itives and confirmed their genotypes as being
predominantly heterozygous (fig. S25E and
table S34). By comparison, significant exonic
SNPs in VvMybA genes [including Chr2:5116947
G/T reported previously in (36)] show shared
genotypes between white grapes and the
V. sylvestris accessions (fig. S25E). It is unclear
how Vvsyl02G000229 and Vvsyl02G001064
might regulate anthocyanin synthesis, but
these results demonstrate that exonic muta-
tions in the two genes are better predictors
of berry skin colors. Furthermore, the heter-
ozygous SNP states in V. sylvestris accessions
suggest that the white berry alleles existed in
natural wild populations before grapevine
domestication.
Discussion
Our systematic genomic survey of V. sylvestris
and V. vinifera accessions paints a defined pic-
ture of grapevine evolutionary history, which
echoes key events in the history of world climate
change and human migration (Fig. 6). The Pleisto-
cene era witnessed the continuous fragmentation
of habitats, the decline of effective popula-
tion size, and the separation of ecotypes for
V. sylvestris. It is highly likely that modern
humans extensively used grapevines as an
energy source from the late Pleistocene, but
the harsh climate was not suited for agricul-
ture (38). As the climatic conditions amelio-
rated at the Pleistocene-Holocene transition,
the grapevine, with its relatively stable peren-
nial yield, unsurprisingly became one of the
earliest candidates for domestication. The
dual events underpin the model that plant
domestication occurs in large, culturally con-
nected areas over a long time (39), but the
domestication time gap remains between ge-
nomic inference and archaeological evidence
(table S35 and figs. S26 and S27) (16). The
diverse SDR haplotypes suggest that an early
goal could be the conscious selection (40) and
propagation of rare, naturally occurring her-
maphroditic individuals from the V. sylvestris
population because they allow mass planta-
tion without male plants. The selection on
phenotype, but not on genotype, also implies
that the different hermaphroditic haplotypes
were subject to strong genetic drift, which is
supported by the high frequency of H1 and
the almost extinct H4 in extant cultivars. The
Mesolithic and Neolithic periods also saw the
early dispersal and diversification of grape-
vines such that unique ancestries emerged in
the Balkans, Iberia, and Western Europe with
the help of V. sylvestris introgression into CG1.
This event mirrors early farmer migration in
Europe, consolidating the role of viticulture in
forming sedentary societies. A higher level of
cultural exchange characterizes the last stage
since the Bronze Age and the trading of su-
perior grapevine cultivars along trade routes.
This is especially evident in the plethora of
Italian cultivars with three or more genetic
ancestries, but unfortunately poses a chal-
lenge to disentangle the genealogical history
of each grapevine cultivar (20). Finally, genet-
ically reliable wild grapevines from Central
Asia, a region battered by climate change
and social instability for the past few mil-
lennia, are no longer available to test Vavilov’s
theory for a diversity center or a hypothet-
ical turnover of grapevine types caused by
Islam conversion in the region. Paleogenomic
data may help to resolve these questions in the
future.
RE FERENCES AND NOTES
1. P. E. McGovern, U. Hartung, V. R. Badler, D. L. Glusker,
L. J. Exner, Expedition 39, 3–21 (1997).
2. P. This, T. Lacombe, M. R. Thomas, Trends Genet. 22, 511–519
(2006).
3. F. Grassi, G. De Lorenzis, Int. J. Mol. Sci. 22, 4518 (2021).
4. D. Cantu, M. A. Walker, The Grape Genome (Springer Nature,
2019).
5. D. Zohary, M. Hopf, E. Weiss, Domestication of Plants in the Old
World: The Origin and Spread of Domesticated Plants in
Southwest Asia, Europe, and the Mediterranean Basin (Oxford
Univ. Press, 2012).
6. S. Myles et al., Proc. Natl. Acad. Sci. U.S.A. 108, 3530–3535
(2011).
7. Y. Zhou, M. Massonnet, J. S. Sanjak, D. Cantu, B. S. Gaut,
Proc. Natl. Acad. Sci. U.S.A. 114, 11715–11720 (2017).
8. Z. Liang et al., Nat. Commun. 10, 1190 (2019).
9. A. Sivan et al., Plants People Planet 3, 414–427 (2021).
10. S. Freitas et al., Sci. Adv. 7, eabi8584 (2021).
11. G. Magris et al., Nat. Commun. 12, 7240 (2021).
12. S. Riaz et al., BMC Plant Biol. 18, 137 (2018).
13. R. Arroyo-García et al., Mol. Ecol. 15, 3707–3714
(2006).
14. P. McGovern et al., Proc. Natl. Acad. Sci. U.S.A. 114,
E10309–E10318 (2017).
15. J. Ramos-Madrigal et al., Nat. Plants 5, 595–603 (2019).
16. See the supplementary materials.
17. M. J. Roach et al., PLOS Genet. 14, e1007807 (2018).
18. T. Lacombe et al., Theor. Appl. Genet. 126, 401–414
(2013).
19. R. Bacilieri et al., BMC Plant Biol. 13, 25–25 (2013).
20. F. Mercati et al., Front. Plant Sci. 12, 692661 (2021).
21. R. Hosfield, J. Cole, Quat. Sci. Rev. 190, 148–160 (2018).
22. A. Timmermann et al., Nature 604, 495–501 (2022).
23. E. C. Corrick et al., Science 369, 963–969 (2020).
24. M. Engel et al., Quat. Int. 266, 131–141 (2012).
25. C. J. Stevens et al., Holocene 26, 1541–1555 (2016).
26. I. Lazaridis et al., Nature 536, 419–424 (2016).
27. C.-C. Wang et al., Nat. Commun. 10, 590 (2019).
28. R. Pinhasi, J. Fort, A. J. Ammerman, PLOS Biol. 3, e410
(2005).
29. I. Mathieson et al., Nature 555, 197–203 (2018).
30. R. Fregel et al., Proc. Natl. Acad. Sci. U.S.A. 115, 6774–6779
(2018).
I. Olalde et al., Science 363, 1230–1234 (2019).
31.
32. S. Brunel et al., Proc. Natl. Acad. Sci. U.S.A. 117, 12791–12798
(2020).
33. C. Zou et al., Proc. Natl. Acad. Sci. U.S.A. 118, e2023548118
(2021).
34. F. Emanuelli et al., BMC Plant Biol. 10, 241–241
(2010).
35. S. Kobayashi, N. Goto-Yamamoto, H. Hirochika, Science 304,
982–982 (2004).
36. A. R. Walker et al., Plant J. 49, 772–785 (2007).
37. A. R. Walker, E. Lee, S. P. Robinson, Plant Mol. Biol. 62,
623–635 (2006).
38. P. J. Richerson, R. Boyd, R. L. Bettinger, Am. Antiq. 66,
387–411 (2001).
39. R. G. Allaby, C. J. Stevens, L. Kistler, D. Q. Fuller, Trends Ecol.
Evol. 37, 268–279 (2022).
40. R. S. Meyer, M. D. Purugganan, Nat. Rev. Genet. 14, 840–852
(2013).
41. Code for: Y. Dong et al., Dual domestications and origin of
traits in grapevine evolution, Zenodo (2023); https://doi.org/
10.5281/zenodo.7523647.
AC KNOWLED GME NTS
We thank F. Pelsy, L. Garmendia Auckenthaler, A.-F. Adam-Blondon,
C. Cornier, P. Kozma, O. Bachmann, F. Gillet, J.-M. Gobat,
S. Dedet, J. Daumann, K. Huber, V. Risovannaya, A. Polulyah,
B. Louis, M. Lafargue, G. Jean-Pascal, G. Melyan, D. I. Sumedrea,
A. Naqinezhad, M. Filipova, technical staff from EGFV and UEVB, and
the Danube-Auen National Park for assistance in the sample
collection and laboratory work and P. Kupfer, E. D. O. Roberson,
and D. Petkova for comments on the manuscript. Funding: This
work was supported by the Natural Science Foundation of
China (grant 32070599 to W.C.); Yunnan Agricultural University
(Research Fund A2032002519 to W.C.); China Agriculture
Research System of MOF and MARA CARS-29 (S.W.); the Science
Committee at the Ministry of SCS (RA 20APP-4E007 to K.M.);
Alliance of International Science Organization (ANSO-CR-PP-2020-
04-A to K.M.); Ministerio de Ciencia, Innovación y Universidades
and Agencia Estatal de Investigación of Spain (RTI2018-094470-R-
C21 to R.A.G.); Predoctoral Fellowship PRE2019-088446 (A.R.I.);
Israel Ministry of Science and Technology (90-23-020-12 to E.D.);
Fondation Giacomi and Swiss National Science Foundation (SNSF
43307 to C.A.); European Regional Fund (KK.05.1.1.02.0010 to
G.Z.); Georgian state budget (L.U., K.B., and T.Z.); TUBITAK
and Ministry of Agriculture and Forestry of Republic of Türkiye
(grant 105G078 to A.E.); and the Israel Science Foundation
(551/18 to E.W.). Author contributions: Conceptualization:
Dong et al., Science 379, 892–901 (2023)
3 March 2023
9 of 10
RES EARCH | R E S E A R C H A R T I C L E
Y.D., Z.L., S.W., J.S., W.C.; Formal analysis: S.D., Q.X., X.D.; Funding
acquisition: W.C.; Investigation: Y.Z., C.M., S.W., S.L., L.T., C.W., D.L.,
Y.P., J.L., L.Y., X.L., G.X., Z.Y., B.C., Y.W., P.G., M.R., O.R., A.R.I., Y.W.,
S.Z.; Resources: Z.L., K.M., M.M., S.G., G.Z., P.F.B., T.L., F.R., P.N.,
K.B., G.D.B., E.D., G.D.L., J.C., C.F.P., R.A.G., C.A., A.E., Z.D., V.K., G.S.,
N.G., S.D., N.O., P.T., C.M., V.L., A.J., L.U., T.Z., D.M., M.H., G.J.,
E.K., T.D., F.G., F.M., F.S., J.E.D., A.M.D., D.C., G.M., T.U., C.Ö., K.K.,
M.X., J.L., M.Z., L.W., S.J., Y.Z., L.S., S.L.; Supervision: Y.D., H.Y.,
Y.Z., S.W., J.S., W.C.; Validation: All authors participated in the
interpretation of the data; Visualization: S.D., Q.X., X.D.; Writing -
original draft: Y.D., S.D., Q.X., X.D., W.C.; Writing - review & editing:
W.C. with input from all coauthors. Competing interests: A.J.
is the founder and owner of Historische Rebsorten vineyard. The
remaining authors declare no competing interests. Data and
materials availability: The VS-1 genome assembly is available at the
Genome Warehouse in the National Genomics Data Center, China
National Center for Bioinformation, under accession numbers
CRA006898 and GWHBQCW00000000. The raw resequencing data
are available at the Genome Warehouse in the National Genomics
Data Center, China National Center for Bioinformation, under
accession number CRA006917. The code for the work can be
accessed at Zenodo (41). License information: Copyright © 2023
the authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original US
government works. https://www.science.org/about/science-licenses-
journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add8655
Materials and Methods
Supplementary Text
Figs. S1 to S27
Tables S1 to S35
References (42–153)
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
Submitted 11 July 2022; accepted 23 January 2023
10.1126/science.add8655
Dong et al., Science 379, 892–901 (2023)
3 March 2023
10 of 10
|
10.1126_science.ade3483
|
RES EARCH
NEUROMORPHIC CHIPS
Edge learning using a fully integrated neuro-inspired
memristor chip
Wenbin Zhang1†, Peng Yao1†, Bin Gao1*, Qi Liu1, Dong Wu1, Qingtian Zhang1, Yuankun Li1, Qi Qin1,
Jiaming Li1, Zhenhua Zhu2, Yi Cai2, Dabin Wu1, Jianshi Tang1, He Qian1, Yu Wang2, Huaqiang Wu1*
Learning is highly important for edge intelligence devices to adapt to different application scenes and
owners. Current technologies for training neural networks require moving massive amounts of data
between computing and memory units, which hinders the implementation of learning on edge devices.
We developed a fully integrated memristor chip with the improvement learning ability and low
energy cost. The schemes in the STELLAR architecture, including its learning algorithm, hardware
realization, and parallel conductance tuning scheme, are general approaches that facilitate on-chip
learning by using a memristor crossbar array, regardless of the type of memristor device. Tasks
executed in this study included motion control, image classification, and speech recognition.
H umans’ ability to learn plays a vital role
in the growth of intelligence and fast
adaptation to unseen scenes (Fig. 1A) or
dynamically changing environments.
Edge artificial intelligence (AI) appli-
cations also require hardware with such learn-
ing abilities to enable the associated devices to
adapt to new scenes or user habits (1). However,
deep neural network (DNN) training (2, 3) is
typically implemented with conventional hard-
ware based on the von Neumann computing
architecture and high-precision digital com-
puting paradigm (4). The extensive data move-
ment between the processor chip and off-chip
main memory incurs massive energy con-
sumption and accounts for most of the latency
of the whole training process (5, 6). Therefore,
although cloud computing platforms can han-
dle such energy-intensive training (4, 7), their
high energy consumption hinders the imple-
mentation of learning on power-limited edge
computing platforms (1). By contrast, memristor-
based neuro-inspired computing eliminates
this extensive data movement through its dis-
ruptive computation-in-memory architecture
and analog computing paradigm (6, 8–10). A
memristor crossbar array can store an analog
synaptic weight and perform in situ vector-
matrix multiplication (VMM) operations in
parallel in a single time step by exploiting Ohm’s
law and Kirchhoff’s law. A neuro-inspired com-
puting chip that integrates multiple memristor
crossbar arrays and complementary metal-oxide
semiconductor (CMOS) circuits can easily im-
plement DNN inference (11–14) and has great
potential to handle fully on-chip learning with-
out any assistance from off-chip memory (15–17).
1School of Integrated Circuits, Beijing National Research
Center for Information Science and Technology (BNRist),
Tsinghua University, Beijing, China. 2Department of
Electronic Engineering, Beijing National Research Center for
Information Science and Technology (BNRist), Tsinghua
University, Beijing, China.
*Corresponding author. Email: gaob1@tsinghua.edu.cn (B.G.);
wuhq@tsinghua.edu.cn (H.W.)
†These authors contributed equally to this work.
The substantial energy efficiency enhancement
provided by memristor-based neuro-inspired
computing makes this paradigm promising for
developing future chips that can enable low-
power learning devices.
Several studies (18–25) have experimentally
demonstrated learning using memristor cross-
bar arrays for in situ weight tuning although
using software or external digital processors to
implement the backpropagation (BP) algorithm
(2). However, realizing a complete fully inte-
grated memristor chip with strong learning abil-
ity and low energy costs remains challenging. The
key challenge lies in the inefficiency of mapping
the BP algorithm to on-chip hardware. First, an
in-memory implementation of the BP algorithm
requires costly conductance tuning operations
with write verification due to device nonidealities,
such as device variability and nonlinear conduc-
tance modulation (15, 26–29). Second, it is dif-
ficult to achieve efficient parallel conductance
tuning with write verification (19–21, 23), which
makes on-chip learning more time- and energy-
consuming. Third, the high-precision data pro-
cessing operations required during weight
update calculations require a large circuit area
and high energy consumption, leading to un-
acceptable overhead (26, 30).
In this work, we demonstrated a memristor-
based neuro-inspired computing chip that
enabled fully on-chip learning, for which a
memristor-featured sign- and threshold-based
learning (STELLAR) architecture was proposed.
In this architecture, the on-chip updating scheme
was first proposed to tune the memristor with-
out verification. This scheme saved excessive
write-and-read costs in the conductance tuning
operations when compared to the write verifi-
cation scheme, and moreover, it could accom-
modate device tuning issues of nonlinearity
and asymmetry to realize software-comparable
accuracies. Second, the on-chip calculation
module was designed to determine the weight
update direction, and this process solely in-
volved the signs of the inputs, outputs, and
errors instead of their high-precision formats.
This design reduced the circuit design burden
and avoided massive overhead during on-chip
learning. Third, a cycle-parallel conductance
tuning scheme was proposed, wherein con-
ductance tuning was performed in a row-by-
row parallel fashion. This scheme further
reduced the induced energy consumption and
latency and accommodated the limited endur-
ance of memristors.
The fabricated neuro-inspired computing
chip integrated two memristor crossbar ar-
rays (~160,000 cells in total) and all the nec-
essary circuit modules, including controllers
for configuration, drivers for computing and
programming, low-cost data converters, and
memristor-featured learning modules (Fig.
1B). On the basis of the obtained hardware-
measured results, the energy consumption of
the memristor chip was a factor of 35 lower
than that of a digital accelerator-based system.
We demonstrated several improvement learn-
ing tasks, including motion control for a light-
chasing car, image classification, and audio
recognition. The scalability of the STELLAR
schemes to large neural networks for improve-
ment learning tasks was also verified with the
simulation of a residual neural network on the
CIFAR-100 dataset (31). The memristor-based
neuro-inspired computing chip could facilitate
the development of edge AI devices that could
adapt to new scenes and users (Fig. 1B).
Memristor-featured architecture for
on-chip learning
To support on-chip learning with appreciable
energy efficiency, area efficiency, and accuracy,
we proposed the STELLAR architecture (Fig. 2A).
STELLAR architecture exploits the bidirec-
tional analog switching behavior of the mem-
ristor device (32). During the weight update
stage, only the weight update direction must
be calculated from the signs of the inputs, out-
puts, and errors. In addition, the architecture
predefines a threshold, which filters out the
small errors when calculating the error signs
and plays a vital role in the convergence of the
learning algorithm by avoiding updates that
are exceedingly sensitive and unnecessary. By
omitting these small updates, the memristor-
based gradient vectors under the STELLAR
update scheme could approximate conven-
tional BP gradient vectors more closely to ac-
commodate practical device nonideal factors
(such as asymmetric tuning of device conduc-
tance). The detailed analysis and simulation
can be found in materials and methods sec-
tion 2 (STELLAR update scheme under de-
vice asymmetric switching). This threshold
is hardware-reconfigurable to adapt to var-
ious learning tasks. The algorithmic details
of the STELLAR architecture can be found in
materials and methods section 1 (Algorithm
for the STELLAR architecture). Depending on
Zhang et al., Science 381, 1205–1211 (2023)
15 September 2023
1 of 7
RES EARCH | R E S E A R C H A R T I C L E
A
B
Dogs
Rabbits
Pre-acquired knowledge
Controllers for configuration
Control signals
Electronic synapses
t
s
o
c
-
w
o
L
s
r
e
t
r
e
v
n
o
c
a
t
a
d
In situ parallel
weight update
Update signals
l
y
p
p
u
s
e
g
a
t
l
o
V
g
n
i
t
u
p
m
o
c
r
o
f
s
r
e
v
i
r
D
i
g
n
m
m
a
r
g
o
r
p
d
n
a
Memristor-featured learning modules
Architecture design of
the fully integrated memristor chip
Dogs
Cats
Only a few new inputs
Dogs
Rabbits
Cats
Improvement learning
Improved knowledge
Trained in old scene
Adapting to new scene
Neuro-inspired chip
for improvement learning
Edge AI devices with
improvement learning ability
Fig. 1. Edge learning with a neuro-inspired memristor chip. (A) Illustration
of the improvement learning ability possessed by human brains. With preacquired
knowledge about old dogs and rabbits, the learning of new samples (i.e., new
dogs) and a new class (i.e., cats) is quickly realized with only a few new inputs.
(B) Design considerations and future applications of the memristor-based
neuro-inspired computing chip. The chip, designed for fully on-chip learning,
integrates all the necessary modules with memristor arrays. It equips edge AI
devices with the learning ability, enabling them to quickly adapt to new scenes.
the weight update direction, a corresponding
identical SET or RESET pulse is applied to the
memristor cells. With this scheme, we could
realize energy-efficient hardware by avoiding
the complex precise weight update calculation
and write verification processes as well as the
complex peripheral circuit design.
The learning performance of the STELLAR
architecture was compared with that of the
conventional methods through simulations on
the Modified National Institute of Standards
and Technology (MNIST) dataset (33). Here,
all memristors in the second layer were set to
random conductance states before the learn-
ing process started. Figure 2B shows the learn-
ing accuracies of the conventional BP method
without and with different write variations
(1% and 3%, given as the percentages of the
full conductance window of the memristor
device) in comparison with those of the pro-
posed method under various thresholds. The
simulation details can be found in the ma-
terials and methods section 3 (Comparison
between STELLAR and the conventional BP
algorithm). An appropriately selected threshold
yielded improved convergence and learning ac-
curacy (fig. S1A). An extremely small threshold
led to weight updates that were too frequent
and an oscillating state of the network, and a
threshold that was excessively large led to in-
adequate weight updates. The comparison of
the energy consumption among different meth-
ods is shown in Fig. 2C. Despite maintaining
almost the same accuracy, the energy consump-
tion of the STELLAR architecture was two orders
of magnitude lower than that of the conventional
BP method owing to the substantial reductions
in the precise weight update calculations and
the write verification overhead.
The STELLAR architecture realized positive
and negative weights with differential pairs
of memristor cells (20, 23, 34, 35) (Fig. 2D).
In conventional crossbar arrays with one-
transistor-one-resistor (1T1R) configurations,
the two memristor cells in a differential pair
are connected to different source lines (SLs),
and subtraction is accomplished in a digital
fashion. In crossbar arrays with two-transistor-
two-resistor (2T2R) configurations, the two
memristor cells in a differential pair are con-
nected to the same SL, and subtraction is ac-
complished directly in the current domain
(11). The 2T2R design greatly reduces the SL
current and thus, the IR drop issues, hence
enabling a larger array size. Here, we propose
a cycle-parallel conductance tuning scheme
for such a differential pair configuration (Fig.
2E). In this scheme, the SET and RESET op-
erations are performed alternately for the
learning iterations of arriving input samples
(e.g., images). If the SET (RESET) operation is
performed in the current learning iteration,
then the RESET (SET) operation is performed
in the next learning iteration. Only one oper-
ation is performed on the whole array in each
iteration; this is achieved by selecting which
cell of a given differential pair to operate on.
Taking the SET update mode as an example, if
the weight update direction is positive (i.e.,
DW > 0), the positive cell is updated with the
SET operation to increase the weight; if DW < 0,
the negative cell is updated with the SET oper-
ation to decrease the weight; and if DW = 0,
neither cell is updated. The conductance tuning
is performed in a row-by-row parallel fashion,
and the detailed schematics of the array oper-
ations are shown in fig. S3. The memristor de-
vices in the same row were selected through
their word line (WL) signals and tuned depend-
ing on the corresponding SL signals, and the bit
line (BL) signals remained constant. The cycle-
Zhang et al., Science 381, 1205–1211 (2023)
15 September 2023
2 of 7
Memristor
synapses
Activation
function
Memristor
synapses
Activation
function
Output
RES EARCH | R E S E A R C H A R T I C L E
A
Input
[
x1
x2
xl
[
X
B
100
C
)
%
(
y
c
a
r
u
c
c
A
80
60
40
20
0
)
n
o
i
t
a
r
e
t
i
/
J
n
(
y
g
r
e
n
E
3
10
2
10
1
10
0
10
W
1
Forward stage
Weight update stage
[
y11
y12
y1m
[
Y
SY
(signs of Y
)
96.7%
BP w/o var
96.7
±
0.1
96.4
±
0.1
55.8
±
16.1
96.6
±
0.1
53.4
±
6.8
Var=1% Var=3%
BP
Th=0
Th=5
STELLAR
Th=9
×97.4
×83.2
~100×
×1.3
×1.0
×0.7
Target
[
t1
t2
tn
[
T
Error
calculation
E
Tenray
Ternary
Reconfigurable
threshold
[
y21
y22
y2n
[
Y
Sign
SY
signs of Y )
(
SE
(signs of E)
W
2
Cycle-parallel
conductance
tuning
W
2
STELLAR
update calculation
D
E
w = g+ - g-
1T1R
2T2R
g+
g+
g-
g-
positive
negative
Subtraction
Subtraction
Subtraction in current
Alternating SET and RESET update for conductance tuning
SET update for iteration i
RESET update for iteration i+1
l
e
l
l
a
r
a
p
e
s
w
w
o
r
-
i
l
e
l
l
a
r
a
p
e
s
w
w
o
r
-
i
Var=1% Var=3%
BP
Th=0
Th=5
STELLAR
Th=9
g+
g-
SET
RESET
∆w>0 ∆w= 0
∆w<0
∆w>0 ∆w=0
∆w<0
Fig. 2. Design of the memristor-featured architecture for on-chip learning.
(A) Diagram of the STELLAR architecture used in the memristor chip. The
STELLAR algorithm features sign-based weight update calculation and reconfig-
urable threshold for the error sign calculation during the weight update stage.
(B and C) Comparison of the classification accuracies (B) and the weight update
energy consumptions (C) using the STELLAR and the conventional BP algorithm
on the MNIST dataset in simulations. (B) and (C) show the results of different
write variations when using the BP algorithm and the results of different
thresholds in the STELLAR learning approach. The bars and error bars show the
averages and standard deviations of the results of 10 repeated experiments,
respectively. The dashed line in (B) shows the average accuracy obtained by BP
without considering the write variations of the memristors. (D) Representations
of weights with a differential conductance pair (left) and with 1T1R (center)
and 2T2R (right) configurations. (E) Illustration of the cycle-parallel conductance
tuning scheme. The red and blue rectangles represent the positive and negative
memristor cells in each differential pair, respectively. The upward and downward
arrows represent the SET and RESET operations on the associated memristor
cell, respectively.
parallel conductance tuning scheme could be
applied to either 1T1R or 2T2R memristor ar-
rays. Because only one-half of the memristor
devices were updated during each on-chip learn-
ing iteration, the cycle-parallel conductance
tuning scheme reduced the induced energy
consumption and alleviated the requirement
regarding the memristor endurance. The de-
tailed analysis of endurance requirements can
be found in materials and methods section 4
(Endurance requirements for the cycle-parallel
conductance tuning scheme).
Chip design, fabrication, and measurements
Figure 3A shows the overall circuit implemen-
tation of the proposed STELLAR architecture.
This memristor chip consisted of controllers
for configuration; a 2T2R memristor array
(1568 × 100), a 1T1R memristor array [100 × 20;
see materials and methods section 5 (Weight
configuration based on the 1T1R memristor
array) for details]; BL, WL, and SL drivers for
computing and programming; low-cost analog-
to-digital converters (ADCs); modules for
Zhang et al., Science 381, 1205–1211 (2023)
15 September 2023
3 of 7
G
E
R
G
E
R
G
E
R
M5
M4
M3
M2
M1
RES EARCH | R E S E A R C H A R T I C L E
A
Array1 Controller
Array2 Controller
2T2R Memristor Array
r
e
f
f
u
B
t
u
p
n
I
r
e
v
i
r
D
L
B
&
r
e
v
i
r
D
L
W
BLP[0]
WLP[0]
WLN[0]
BLN[0]
BLP[783]
WLP[783]
WLN[783]
BLN[783]
]
0
[
L
S
]
1
[
L
S
]
9
9
[
L
S
r
e
v
i
r
D
L
S
C
D
A
A
R
-
C
D
A
A
R
-
C
D
A
A
R
-
B
C
1T1R Memristor Array
BL[0]
WL[0]
BL[1]
WL[1]
BL[99]
WL[99]
r
e
v
i
r
D
L
B
&
r
e
v
i
r
D
L
W
C
D
A
A
R
-
C
D
A
A
R
-
r
e
v
i
r
D
L
S
XOR
s
r
e
t
n
u
o
C
r
e
f
f
u
B
t
u
p
t
u
O
]
0
[
L
S
]
1
[
L
S
]
9
1
[
L
S
XOR
Weight Update Logic
Input Buffer
WL Driver & BL Driver
2T2R Memristor Array
RA-ADC × 100
1T1R Memristor Array
Weight Update Logic
RA-ADC × 20
Counters
Memristor pair
Memristor cell
TiN
TaOy
m
HfOx
n
0
TiN 5
Transistor pair
1 µm
D
100
95
90
)
%
(
y
c
a
r
u
c
c
A
G
0 1 2 3 4 5 6 7 8 9
E
100
)
%
(
y
c
a
r
u
c
c
A
95
90
85
)
%
(
y
c
a
r
u
c
c
a
n
i
e
g
n
a
h
C
1
0
−1
0 1 2 3 4 5 6 7 8 9
80
0
12
24
Time (days)
36
48
F
100
)
%
(
y
c
a
r
u
c
c
A
75
50
25
0
Energy
18.9%
0.2%
6.2%
1.9%
72.8%
Training set
Test set
1
2
Epochs
3
FWD ADC
FWD array
FWD other
Update array
Update other
Fig. 3. The memristor chip for on-chip learning. (A) Overview of the
architecture of the memristor chip. (B) Optical microscope image of the chip,
where several key components are labeled. (C) A cross-section transmission
electron microscopy (TEM) image showing 2T2R cells. The transistors in a 2T2R
cell share a common source terminal. Inset: Cross-section TEM image of the
memristor device. (D) On-chip classification accuracy obtained for each class by
using the off-chip trained weights for the MNIST dataset, with bars and error bars
showing the averages and standard deviations of the accuracies achieved over
five inference iterations, respectively. The standard deviations ranged from 0.07
to 0.27. (E) Changes in the classification accuracy over 48 days after weight
transfer, with each point showing the average accuracy over five inference results.
Inset: The average changes in the accuracy of each class after 48 days of weight
transfer. The standard deviations in the inset ranged from 0.1 to 0.32. (F) Changes in
the classification accuracy over three learning epochs in the on-chip learning task for
the MNIST dataset, with each epoch including 60,000 iterations. (G) Energy
breakdown of the memristor chip during the on-chip learning process.
memristor-featured on-chip learning (i.e., error-
circulating subtractors and weight update logic);
and input and output buffers. The first-layer
memristor array adopted a 2T2R configuration
to reduce the IR drop issues occurring in such a
large array, and the second-layer memristor array
adopted a 1T1R configuration to support more
flexible in situ weight tuning. The controllers
decoded the input stage selection signals and
provided the output configuration signals to
other circuit modules to switch the chips to dif-
ferent working stages [see materials and meth-
ods section 6 (Circuit design of the memristor
chip) and fig. S5]. Resolution-adjustable ADCs
(RA-ADCs) feature configurable resolutions and
support flexible threshold values (11). The error
calculation was accomplished with subtractors,
which were realized with counters. The weight
Zhang et al., Science 381, 1205–1211 (2023)
15 September 2023
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
update logic determined the weight update
direction and conductance tuning operations.
A micrograph of the fabricated chip is shown
in Fig. 3B. The chip area breakdown is described
in fig. S6B. The memristor device used a ma-
terial stack of TiN/HfOx/TaOy/TiN, and the
fabrication process was compatible with the
standard CMOS process [see materials and
methods section 7 (Fabrication of the mem-
ristor chip) and fig. S6A]. Consequently, the
memristors could be conveniently integrated
with complex CMOS circuits to produce an
excellent yield (almost 100% of all 160,000
cells). The cross-section transmission electron
microscopy (TEM) image in Fig. 3C showed
the integration of memristor cells with CMOS
circuitry. The fabricated memristors exhibited
uniform and repeatable bidirectional analog
switching with identical pulse trains (fig. S6D).
The ~160,000 total on-chip memristor cells
could be uniformly programmed to 32 con-
ductance states, with maximum, minimum,
and average success rates of 99.98, 99.69, and
99.90%, respectively [see materials and meth-
ods section 8 (Measurements of the memristor
devices) and fig. S6C].
The on-chip inference was first demonstrated
for MNIST handwritten digit classification. The
weights were off-chip trained and then trans-
ferred to the chip as memristor conductance
[see materials and methods section 9 (Off-chip
training and on-chip inference)]. The measured
classification accuracy of each class (0 to 9) is
presented in Fig. 3D; the average accuracy was
95.8%. The effect of cell conductance fluctua-
tions on the chip accuracy was also evaluated
(Fig. 3E). We monitored the accuracy for 48 days,
and no obvious accuracy degradation was ob-
served. We also demonstrated real-time hand-
written digit recognition with the memristor
chip (see movie S3).
An on-chip learning task, MNIST image clas-
sification, was further demonstrated to verify
the on-chip learning ability based on a 784-100-
10 multilayer perceptron (MLP). The weights in
the first layer were trained off-chip and then
transferred to the chip as memristor conduc-
tance. The memristors in the second layer were
first programmed to the high-resistance state
(HRS) and then updated using the STELLAR
scheme. All data processing and signal control
processes were executed on the chip. After three
epochs of on-chip learning with the training
set, the classification accuracies were increased
from 8.6 and 8.4% to 94.9 and 92.3% on the
training set and test set, respectively (Fig. 3F).
We then evaluated the energy consumption of
the on-chip learning with hardware-measured
results [see materials and methods section 11
(Energy consumption benchmark)]. We also
evaluated the energy consumption of a digital
accelerator-based system (36) for one training
iteration with the same MLP network. The en-
ergy breakdown of the memristor chip during
the on-chip learning process is presented in Fig.
3G. The energy consumption could be further
reduced by optimizing the ADC design (37–39).
On-chip improvement learning
The memristor chip was used to further dem-
onstrate four improvement learning tasks, in-
cluding the motion control task of learning new
samples, audio recognition task of learning
new samples, image classification task of learn-
ing a new class, and motion control task of
learning a new class. The improvement learn-
ing featured the fast learning of new knowl-
edge and maintaining preacquired knowledge.
As illustrated in Fig. 1A, the learning of new
knowledge (e.g., new samples or classes) was
quickly realized with only a few new inputs,
and this learning was done without losing pre-
acquired knowledge. The implementation of
improvement learning mainly included two
stages. First, a model was trained off-chip on the
base dataset, and then the model was trans-
ferred to the chip. Next, under the STELLAR
architecture, on-chip improvement learning
with new data was performed on the basis of
the transferred memristor chip. Improvement
learning is a specific learning format to imple-
ment lifelong learning. This learning scheme
aims to rapidly learn knowledge from the new
data without forgetting previous knowledge
on the old data, making it different from trans-
fer learning (40) that focuses on transferring
knowledge from original data to new data, re-
gardless of the accuracy drop on the original
data. We demonstrated these improvement learn-
ing tasks because we consider that recognizing
new classes or samples with a base model is a
practical and promising edge learning task.
We first demonstrated the learning of new
samples in a motion control task of a light-
chasing car (Fig. 4A). The car was designed to
pursue the location of a laser light spot; it was
equipped with a camera to capture environ-
mental images, steering motor for direction
control, and driving motor for throttle control.
The memristor chip received the input features
of the environmental images from the PC and
provided the output control signals for the
steering angles and driving throttles [see ma-
terials and methods section 12 (Motion con-
trol task of learning new samples) for details].
As illustrated in Fig. 4B, a convolutional neu-
ral network (CNN) that included six convolu-
tion layers and two fully connected (FC) layers
with the dimensions of 512 × 100 × 10 was first
trained off-chip with old scene data (i.e., dark
scene data), and then the weights of these two
FC layers were transferred to the two cor-
responding arrays of the memristor chip. Next,
improvement learning of the new scene (i.e., a
bright scene) was performed on the chip by
tuning the weights of the last FC layer. The
details can be found in materials and methods
section 16 (The hybrid system for running the
motion control task). Before improvement
learning, the car could have lost track of the
target (i.e., light spot) in the bright scene: It
deviated from the target or moved forward
even if no target was present. After improve-
ment learning, the car adapted well to the
bright scene and still performed well in the dark
scene (movie S1). Figure 4C shows the evolu-
tion of the scores during improvement learn-
ing, where a score of 1.0 denotes the best
performance [see materials and methods sec-
tion 12 (Motion control task of learning new
samples)]. The scores became stable after im-
provement learning with 500 training samples
from the new scene. After improvement learn-
ing with all the training samples, the average
score in the new scene significantly increased
from 0.605 to 0.912, and that of the old scene
increased from 0.951 to 0.963, showing that no
degradation occurred (Fig. 4D). Figure 4E
shows the transitions and changes in conduc-
tance weights before and after the improve-
ment learning.
Next, we demonstrated the learning of a new
class in an image classification task involving
the MNIST dataset (Fig. 4F). The base model
was trained to recognize images of the digits 0
and 2–9 (old classes) and then transferred to
the memristor chip. Next, the improvement
learning of the new class (i.e., the digit 1) was
performed on the chip. As shown in Fig. 4G,
the accuracy achieved for the new class in-
creased substantially during the improvement
learning with only a few training samples, and
the accuracy for the remaining nine old classes
was not significantly reduced. The accuracies
yielded for the new class and old classes sta-
bilized after 100 training samples. After im-
provement learning with 100 training samples,
the average accuracy for the new class increased
from 7.02 to 93.0%, and that achieved for the
old classes slightly decreased from 95.3 to
93.2% (Fig. 4H). The conductance weight dis-
tributions before and after improvement learn-
ing and their difference are shown in Fig. 4I.
In addition, the memristor chip was also
used to implement an audio recognition task
of learning new samples and a motion control
task of learning a new class. In the audio rec-
ognition task, the on-chip improvement learning
improved the recognition accuracy for female
audio samples based on the weights pretrained
with male audio samples (fig. S8A). In the mo-
tion control task of learning a new class, the
on-chip improvement learning of the “moving
backward” action helped the car to accomplish
the tough task based on the model only knowing
when to move forward or stop (fig. S9A). After
improvement learning, the accuracy achieved
for the new class (i.e., moving backward) in-
creased from 0 to 95.2%, and that for the old
classes (i.e., moving forward or stopping) de-
creased from 89.5 to 89.4%. The implemen-
tation details of these two tasks can be found
Zhang et al., Science 381, 1205–1211 (2023)
15 September 2023
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
A
B
F
Laser
pointer
Camera
Camera
images
PC
i
P
y
r
r
e
b
p
s
a
R
Steering
motor
Driving
motor
Steering
angle
Driving
throttle
Memristor
chip
Model for dark scene
1
Weight transfer
2
New
features
input
r
e
y
a
l
0
0
1
×
2
1
5
d
e
t
c
e
n
n
o
c
y
l
l
u
f
8 neurons
for
angle
regression
move
stop
Samples of
bright scene
Update
Model for images 0 & 2–9
1
Weight transfer
2
New
class
input
r
e
y
a
l
0
0
1
×
4
8
7
d
e
t
c
e
n
n
o
c
y
l
l
u
f
0
2
3
4
5
6
7
8
9
1
Class “1”
Update
1.0
0.9
0.8
0.7
0.6
0.5
0
C
e
r
o
c
S
E
100
75
50
25
0
0
G
)
%
(
y
c
a
r
u
c
c
A
I
D
e
r
o
c
S
1.0
0.9
0.8
0.7
0.6
0.5
0.951±0.005
0.963±0.004
0.912±0.005
0.605±0.070
Dark
Bright
Dark
Bright
Base
Improved
Dark scene
Bright scene
700
1400
2100
Iterations
6 µA
0 µA
-6 µA
6 µA
0 µA
-6 µA
H
100
95.3±0.2
93.2±0.5
93.0±0.4
)
%
(
y
c
a
r
u
c
c
A
Old classes
New class: "1"
50
100
150
Iterations
75
50
25
0
7.0±4.1
Old
New
Base
New
Old
Improved
6 µA
0 µA
-6 µA
6 µA
0 µA
-6 µA
Fig. 4. Improvement learning demonstrations on the memristor chip.
(A) Illustration of the motion control task (left) and its control system (right).
(B) Illustration of the learning of new samples in the light-chasing car task.
(C) Changes in the scores over the course of on-chip improvement learning of
bright scene samples. The lines and the error bars show the averages and
standard deviations of the scores of 10 repeated experiments with shuffled
training sets, respectively. (D) The initial and final scores in (C). (E) The evolution
of the weight distribution (100 × 10) before (top) and after (middle) the on-chip
improvement learning. Bottom: distribution of the weight changes. (F) Illustration
of the learning of a new class in the image classification task. (G) Changes in the
accuracies over the course of on-chip improvement learning of the new class,
showing the results from 10 repeated experiments with 150 randomly selected
images from the training set each time. (H) The initial accuracies and the accuracies
achieved after on-chip improvement learning of 100 iterations in (G). (I) The
evolution of the weight distribution (100 × 10) before (top) and after (middle) the on-
chip improvement learning. Bottom: Distribution of the weight changes.
Zhang et al., Science 381, 1205–1211 (2023)
15 September 2023
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
in materials and methods sections 14 and 15
(Audio recognition task of learning new sam-
ples, Motion control task of learning a new
class). Moreover, we performed another sim-
ulation with a memristor ResNet20 network
for CIFAR-100 image recognition, which re-
quired 20 unseen classes to be learned after
training on the remaining 80 categories. The
accuracies on old, new, and whole datasets
for this task were similar to software results
with the floating-point precision [see mate-
rials and methods section 17 (The scalability
of the method to larger neural networks) and
fig. S10]. These results illustrated that the pro-
posed STELLAR architecture could be scalable
to larger neural networks and could realize
efficient improvement learning with high-
precision software accuracies.
Conclusions
We developed a fully integrated memristor chip
with the improvement learning ability and low
energy cost. The schemes in the STELLAR ar-
chitecture, including its learning algorithm,
hardware realization, and parallel conduc-
tance tuning scheme, are general approaches
that facilitate on-chip learning by using a
memristor crossbar array, regardless of the
type of memristor device. We demonstrated
the improvement learning of both new sam-
ples and a new class across various tasks,
including motion control, image classification,
and speech recognition, which showed that
the STELLAR architecture accommodated the
device nonidealities and equipped the mem-
ristor chip with improvement learning ability
to adapt to new scenarios. With further cir-
cuitry engineering (41) based on advanced fab-
rication technology, the STELLAR architecture
could enable on-chip learning memristor chip
with an energy efficiency about 75 times higher
than that of the digital accelerator (36). More
details can be seen in materials and methods
section 18 (The energy efficiency estimation of
the memristor-based learning chip). This study
is an important step toward future chips with
high energy efficiency and extensive learning
capabilities. We hope our findings will accel-
erate the development of future smart edge
devices that can adapt to different application
scenarios and owners.
RE FERENCES AND NOTES
1. W. Shi, J. Cao, Q. Zhang, Y. Li, L. Xu, IEEE Internet Things J. 3,
637–646 (2016).
2. D. E. Rumelhart, G. Hinton, R. J. Williams, Nature 323,
533–536 (1986).
3. Y. LeCun, Y. Bengio, G. Hinton, Nature 521, 436–444 (2015).
4. A. Coates et al., in Proceedings of the 30th International
Conference on Machine Learning (PMLR, 2013), pp. 1337–1345.
5. M. Horowitz, “1.1 Computing’s energy problem (and what
we can do about it)” in 2014 IEEE International Solid-State
Circuits Conference Digest of Technical Papers (ISSCC) (IEEE,
2014), pp. 10–14.
6. H.-S. P. Wong, S. Salahuddin, Nat. Nanotechnol. 10, 191–194
(2015).
7. E. Strubell, A. Ganesh, A. McCallum, in Proceedings of the
AAAI Conference on Artificial Intelligence (AAAI, 2020),
pp. 13693–13696.
8. D. Ielmini, H.-S. P. Wong, Nat. Electron. 1, 333–343 (2018).
9. M. A. Zidan, J. P. Strachan, W. D. Lu, Nat. Electron. 1, 22–29 (2018).
10. M. Lanza et al., Science 376, eabj9979 (2022).
11. Q. Liu et al., in 2020 IEEE International Solid- State Circuits
Conference - (ISSCC) (IEEE, 2020), pp. 500–502.
12. W. Wan et al., in 2020 IEEE International Solid- State Circuits
Conference - (ISSCC) (IEEE, 2020), pp. 498–500.
13. C.-X. Xue et al., Nat. Electron. 4, 81–90 (2021).
14. W.-H. Chen et al., Nat. Electron. 2, 420–428 (2019).
15. S. Ambrogio et al., Nature 558, 60–67 (2018).
16. A. Sebastian, M. Le Gallo, R. Khaddam-Aljameh, E. Eleftheriou,
Nat. Nanotechnol. 15, 529–544 (2020).
17. W. Zhang et al., Nat. Electron. 3, 371–382 (2020).
18. M. Prezioso et al., Nature 521, 61–64 (2015).
19. P. Yao et al., Nat. Commun. 8, 15199 (2017).
20. C. Li et al., Nat. Commun. 9, 2385 (2018).
21. Z. Wang et al., Nat. Electron. 2, 115–124 (2019).
22. C. Li et al., Nat. Mach. Intell. 1, 49–57 (2019).
23. P. Yao et al., Nature 577, 641–646 (2020).
24. F. Cai et al., Nat. Electron. 2, 290–299 (2019).
25. E. J. Fuller et al., Science 364, 570–574 (2019).
26. G. W. Burr et al., IEEE Trans. Electron Dev. 62, 3498–3507 (2015).
27. H. Wu et al., in 2017 IEEE International Electron Devices Meeting
(IEDM) (IEEE, 2017), pp. 11.5.1–11.5.4.
28. P. Y. Chen, X. Peng, S. Yu, IEEE Trans. Comput. Aided Des.
Integrated Circ. Syst. 37, 3067–3080 (2018).
29. Y. Xi et al., Proc. IEEE 109, 14–42 (2021).
30. Q. Zhang et al., Neural Netw. 108, 217–223 (2018).
31. A. Krizhevsky, G. Hinton, “Learning multiple layers of features
from tiny images”; (Univ. of Toronto, 2009) https://www.cs.
toronto.edu/~kriz/learning-features-2009-TR.pdf.
32. W. Wu et al., in 2018 IEEE Symposium on VLSI Technology
(IEEE, 2018), pp. 103–104.
33. Y. Lecun, L. Bottou, Y. Bengio, P. Haffner, Proc. IEEE 86,
2278–2324 (1998).
35. L. Song, X. Qian, H. Li, Y. Chen, in 2017 IEEE International
Symposium on High Performance Computer Architecture
(HPCA) (IEEE, 2017), pp. 541–552.
36. D. Han et al., IEEE J. Solid-State Circuits 56, 2858–2869
(2021).
37. R. Khaddam-Aljameh et al., in 2021 Symposium on VLSI
Circuits (IEEE, 2021), pp. 1–2.
38. W. Wan et al., in 2020 IEEE Symposium on VLSI Technology
(IEEE, 2020), pp. 1–2.
39. J.-M. Hung et al., Nat. Electron. 4, 921–930 (2021).
40. K. Weiss, T. M. Khoshgoftaar, D. Wang, J. Big Data 3, 9 (2016).
41. W.-H. Huang et al., in 2023 IEEE International Solid- State
Circuits Conference (ISSCC) (IEEE, 2023), pp. 15–17.
42. W. Zhang, P. Yao, B. Gao, H. Wu, Data for Edge Learning Using
a Fully Integrated Neuro-Inspired Memristor Chip, Zenodo,
(2023); https://doi.org/10.5281/zenodo.8145521.
43. W. Zhang, P. Yao, B. Gao, H. Wu, Zenodo (2023); https://doi.org/
10.5281/zenodo.8151757.
AC KNOWLED GME NTS
Funding: This work was supported in part by the STI 2030-Major
Projects (2021ZD0201200), the National Natural Science
Foundation of China (92064001, 62025111), the XPLORER Prize,
the Shanghai Municipal Science and Technology Major Project, and
the Beijing Advanced Innovation Center for Integrated Circuits.
Author contributions: B.G. and H.W. supervised this project and
proposed the overall architecture. W.Z., P.Y., and B.G. contributed
to the whole experiment design and the writing of the manuscript.
P.Y., Q.L., Do.W., and Da.W. designed the circuits. B.G., J.T.,
H.Q., and H.W. contributed to the memristor device fabrication
and integration with CMOS. W.Z., P.Y., and Q.L. contributed
to the implementation of the test system. Q.Z. worked on the
development of AI models. W.Z. implemented the AI models on the
chip and conducted the chip and system measurements. W.Z.,
P.Y., Y.L., Q.Q., J.L., Z.Z., Y.C., and Y.W. performed the simulation
and benchmark. All authors discussed the results and reviewed the
manuscript. Competing interests: The authors declare that they
have no competing interests. Data and materials availability: All
data are available in the manuscript or the supplementary material
and can be found in Zenodo (42). All other software and data for
running the simulations and experiments is available through
Zenodo (43) and Github, including our hardware platform code,
neural networks for improvement learning tasks, and simulation
code to verify the advantages and scalability of the STELLAR
schemes. License information: Copyright © 2023 the authors,
some rights reserved; exclusive licensee American Association for
the Advancement of Science. No claim to original US government
works. https://www.sciencemag.org/about/science-licenses-
journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade3483
Materials and Methods
Figs. S1 to S11
Tables S1 to S3
References (44–53)
Movies S1 to S3
34. P. Chi et al., in Proceedings of the 43rd International
Symposium on Computer Architecture (ACM/IEEE, 2016),
pp. 27-39.
Submitted 11 August 2022; resubmitted 30 May 2023
Accepted 8 August 2023
10.1126/science.ade3483
Zhang et al., Science 381, 1205–1211 (2023)
15 September 2023
7 of 7
|
10.1126_science.ade4970
|
RES EARCH
STELLAR ASTROPHYSICS
An observed population of intermediate-mass helium
stars that have been stripped in binaries
M. R. Drout1,2*†, Y. Götberg2*†, B. A. Ludwig1, J. H. Groh3, S. E. de Mink4,5,
A. J. G. O’Grady1,6, N. Smith7
The hydrogen-rich outer layers of massive stars can be removed by interactions with a binary companion.
Theoretical models predict that this stripping produces a population of hot helium stars of ~2 to 8 solar
masses (M⊙), however, only one such system has been identified thus far. We used ultraviolet photometry
to identify potential stripped helium stars then investigated 25 of them using optical spectroscopy. We
identified stars with high temperatures (~60,000 to 100,000 kelvin), high surface gravities, and hydrogen-
depleted surfaces; 16 stars also showed binary motion. These properties match expectations for stars
with initial masses of 8 to 25 M⊙ that were stripped by binary interaction. Their masses fall in the gap
between subdwarf helium stars and Wolf-Rayet stars. We propose that these stars could be progenitors
of stripped-envelope supernovae.
A pproximately 70% of massive stars [ini-
tial masses of >8 solar masses (M☉)] in-
teract with a binary companion during
their lifetimes (1, 2). Those binary interac-
tions are expected to strip the hydrogen-
rich envelopes from many massive stars, leaving
an exposed hot and compact helium core. The
resulting stripped stars have sufficiently long
lifetimes to be observed and are expected to be
numerous (3).
Binary-stripped massive stars are expected
to influence multiple astrophysical processes:
They are thought to be the progenitors of most
hydrogen-poor core-collapse supernovae (4–6).
The neutron stars that have been observed in
gravitational wave events are thought to have
undergone two phases of envelope stripping
(7). And the high surface temperatures of
stripped stars make them potential sources of
ionizing photons (8, 9).
Despite their predicted ubiquity, few binary-
stripped helium stars with masses between ∼2
and 8 M☉—which are expected to be produced
by stars with initial masses between ∼8 and
25 M☉—have been found. Many other types of
hydrogen-deficient stars have been observed
(10). These are classified as high-mass Wolf-
Rayet (WR) stars (11), low-mass subdwarfs (12),
extreme helium stars (13), and central stars
of planetary nebulae (14), all of which have
been found in binary systems (15–17). How-
ever, none of those classes occupy the mass
range that has been predicted to produce most
1David A. Dunlap Department of Astronomy and Astrophysics,
University of Toronto, Toronto M5S 3H4, Canada. 2The
Observatories of the Carnegie Institution for Science, Pasadena,
CA 91101, USA. 3Independent researcher, 2314 Leiden,
Netherlands. 4Max-Planck-Institut für Astrophysik, 85741
Garching, Germany. 5Anton Pannekoek Institute for Astronomy,
University of Amsterdam, 1090 GE Amsterdam, Netherlands.
6Dunlap Institute for Astronomy and Astrophysics, University of
Toronto, Toronto M5S 3H4, Canada. 7Steward Observatory,
University of Arizona, Tucson, AZ 85721, USA.
*Corresponding author. Email: maria.drout@utoronto.ca (M.R.D.);
ygoetberg@carnegiescience.edu (Y.G.)
†These authors contributed equally to this work.
stripped-envelope supernovae or neutron star
mergers (7). Only one hot helium star with an
appropriate mass has been reported: the “quasi-
WR” star in the system HD 45166 (18, 19).
If such systems are truly rare, models of
binary evolution would need to be revised.
Alternatively, there could be an observational
bias: The optical flux from intermediate-mass
stripped stars might be hidden by a bright
main sequence (MS) companion star. Although
helium star mass-loss rates are uncertain (20),
they are predicted to exhibit weaker wind fea-
tures than luminous WR stars, so they could
potentially have eluded detection in previous
surveys targeting those features (21).
Ultraviolet photometry
Some stripped helium star binaries might be
detectable by excess ultraviolet (UV) emission
in their spectral energy distributions (22). To
assess this possibility, we calculated synthetic
spectra for a large set of hypothetical binaries
containing a stripped star and an MS star (20).
We find that many of the hypothetical systems
remain obscured by the brightness of the MS
star, but hot intermediate-mass helium stars
paired with MS companions of ≲10 M☉ oc-
cupy a specific region of UV-optical color-
magnitude diagrams (CMDs): blueward of the
MS at intermediate luminosities of −1 mag >
MUVM2 > −4 mag (where MUVM2 is the ab-
solute magnitude in the UVM2 ultraviolet
filter; figs. S4 and S5).
We searched for massive stars with UV mag-
nitudes that fall within the CMD region pre-
dicted by our synthetic spectra. We targeted
stars in the Large Magellanic Cloud (LMC) and
Small Magellanic Cloud (SMC) galaxies, be-
cause they contain a large number of massive
stars at known distances, with low obscura-
tion by dust. We measured UV photometry
using archival images from the Swift Ultra-
violet Survey of the Magellanic Clouds (23).
These images cover ∼3 square degrees of the
SMC and ∼9 square degrees of the LMC in
three UV filters at a resolution of 2.5 arc sec.
To reduce the effects of crowding at that reso-
lution, we used the forward modeling code
THE TRACTOR (24) to perform forced point-
spread function photometry. We adopted the
known locations of stars in the optical Magel-
lanic Cloud Photometric Survey (25, 26), which
has better spatial resolution.
This process determined UV magnitudes for
>500,000 sources in the directions of the LMC
and SMC (20). Figure 1 shows a UV-optical
CMD of all the sources. We adopted distances
of 50 and 61 kpc and visual dust extinctions AV
of 0.38 and 0.22 mag for the LMC and SMC,
respectively (20). LMC and SMC extinction
curves (27) were used to determine the corre-
sponding dust obscuration in the UV. The
CMD contains a dense band (which we ascribe
to the MS) and multiple sources blueward of
the MS, which we consider to be candidate
stripped helium star binaries.
Optical spectroscopy
We selected 25 candidate systems for follow-
up spectroscopy by choosing targets that have
luminosities and colors consistent with our pre-
dictions for binaries containing intermediate-
mass helium stars (20) (indicated in Fig. 1).
The stars are of similar brightness to MS stars
with initial masses of ∼6 to 15 M☉ but—for the
adopted extinction—are located blueward of
the zero-age MS (ZAMS) in nine distinct UV-
optical CMDs (20). They have UV-optical colors
similar to those of WR stars but are intrinsically
fainter. For some systems, the observed colors
and magnitudes approach predictions for iso-
lated helium stars with masses between ∼2 and
8 M☉ (Fig. 1).
We obtained between 1 and 30 optical spectra
for each system using the Magellan Echellette
spectrograph (28) on the 6.5-m Magellan Baade
telescope at Las Campanas Observatory, Chile.
All 16 systems with more than one epoch show
radial velocity variations, consistent with being
binary systems (table S9).
We used kinematics to reject any likely fore-
ground objects. All 25 systems have average
radial velocities consistent with expectations
for stars in the LMC and SMC (20). We com-
bined these with proper motions (29), finding
that 23 systems have three-dimensional mo-
tion that is consistent with known O-type and
B-type massive stars in the LMC and SMC (20)
(O-type stars typically have initial masses of
≳15 M☉, and B-type stars typically have initial
masses of ∼2 to 15 M☉). The remaining two
objects (stars 5 and 6 in table S9) show slight
offsets in proper motion but have data quality
issues in the proper motion catalog. We there-
fore retained them in our sample.
Figure 2A shows examples of the spectra;
the full sample is provided in figs. S16 to S21.
We classify the stars into three broad groups:
Drout et al., Science 382, 1287–1291 (2023)
15 December 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Candidate stripped helium star binaries in UV-optical color-magnitude diagrams. Gray dots
show absolute magnitude photometry in the UVM2 ultraviolet band as a function of the UVM2-V (where V is
the visual band) color for stars in (A) the LMC and (B) the SMC. Numbered circles indicate the 25 stars
we investigated further with optical spectroscopy (table S7), color coded according to their observed spectral
morphologies (see legend). Error bars are 1s. These systems have similar UV-optical colors, but lower
brightnesses than either WR stars (dark-purple diamonds) (48, 49) or the weaker-wind WN3/O3 stars (light-
purple diamonds) (36). The connected black dots indicate models of isolated helium-core burning stripped stars,
which are labeled with the current mass of the stripped star (Mstrip) (22). The thick curved line indicates the
expected position of the ZAMS for O-type (light gray) and B-type (dark gray) stars. All observed data have been
corrected for dust extinction (indicated by the arrows), and all magnitudes are in the AB system.
1) Class 1 (eight stars): These spectra are dom-
inated by absorption lines of He II. In some
cases, lines of N IV and/or N V are visible in
emission or absorption. These three ions are
characteristic of very hot stars (30).
2) Class 2 (eight stars): These spectra show
He II in absorption but also have strong ab-
sorption lines of hydrogen in the blue part of
the spectrum (the short-wavelength Balmer
lines). Six of these spectra also show He I; two
do not, which indicates that the spectrum is
a blend of two stars with different tempera-
tures (20).
3) Class 3 (nine stars): No He II lines are
visible in the optical spectrum. These spectra
are dominated by strong Balmer and He I ab-
sorption lines, closely resembling those of
B-type MS stars.
Interpretation of the spectra
The CMD locations and spectral morphologies
of these 25 stars are consistent with our theo-
retical predictions for binary systems con-
taining hot intermediate-mass helium stars.
Our synthetic spectra of such objects (calcu-
lated above) show spectral features similar to
those of WR stars, but with substantially
weaker emission or absorption lines owing to
their lower luminosities and mass-loss rates
(20). In the set of helium star plus MS star
composite spectral models, we identify the
same three broad spectral classes as in our ob-
servations. Reproducing the absorption line
spectra of the observed sample requires mass-
loss rates for intermediate-mass helium stars
that are at least an order of magnitude lower
than extrapolations of WR mass loss (20),
which is consistent with theoretical predic-
tions (31).
We interpret the progression from class 3 to
class 1 as an increasing contribution from the
helium star to the optical flux of the system
(20). Figure 2B compares the equivalent width
(EW) of the He II l5411 line as a function of
Hh + He II l3835 [where l notation indicates
the wavelengths in angstroms (Å), Hh is the
seventh Balmer line, and + indicates blended
lines] for both our observed sample and the
composite models. The He II l5411 line arises
from the helium star; it is not expected at the
cooler temperatures of B-type MS stars. We
use the Hh + He II l3835 blend to probe the
presence of an MS companion, because strong
short-wavelength Balmer lines are not ex-
pected in hot hydrogen-depleted stars [we
have to use a blended line because our syn-
thetic spectra show no isolated hydrogen lines
in the optical range for these stars (20)]. In this
parameter space, all 25 observed stars overlap
with the predictions from the composite mod-
els, regardless of which helium star mass-loss
rates are adopted (20).
For class 3 stars, we find 3s upper limits on
the EW of He II l5411 of ≲0.2 Å (table S8).
This corresponds to models where the helium
star contributes <20% of the optical flux, and
hence the spectra appear like those of B-type
MS stars. Although such stars can show a UV
excess, our models predict that they should be
close to the ZAMS (as observed for the class 3
stars in Fig. 1). In contrast, the class 1 stars all
have Hh + He II l3835 EWs of <1.2 Å, con-
sistent with models for systems where a helium
star contributes >80% of the optical flux, so
the spectra appear like those of isolated helium
stars. The models only produce such spectra
when the MS companion has a mass of ≲3.5 M☉.
We infer that the class 1 objects either (i) come
from binary systems where the two stars had
very different initial masses; or (ii) have com-
panions that are compact objects (neutron
stars or black holes), not MS stars. Those sys-
tems would also have sufficiently blue colors
to match the observations; in the CMD, they
are close to models of isolated helium stars.
Although dust extinction toward individual
objects is uncertain, the location of these
models in the CMD is similar to that of the
class 1 objects (Fig. 1). The intermediate class 2
stars fall close to models where the helium
star contributes 20 to 80% of the optical flux,
so they have composite optical spectra with
contributions from both stars.
Estimates of stellar properties
To assess the nature of the hot stars in these
systems, we considered EW diagnostics that
distinguish the optical spectra of stripped stars
from MS stars and estimate their surface prop-
erties. We used the 1D nonlocal thermodynamic
equilibrium radiative transfer code CMFGEN (32)
to compute a set of spectral models with a
range of effective temperatures (30 kK ≤ Teff ≤
100 kK, where kK is kilokelvin), surface grav-
ities [4.0 < log(g/cm s−2) < 6.0], and surface
hydrogen mass fractions (XH,surf = 0.01, 0.1,
0.3, and 0.5, which are all depleted below MS
values) (20). We choose a baseline mass-loss
rate of 10(cid:1)9M☉ year−1 because it produces ab-
sorption line spectra. We then tested the im-
pact of this choice on our results by varying
the assumed mass-loss rate by two orders of
magnitude. These models cover a broad param-
eter space without making assumptions about
the detailed evolutionary state of each system.
Figure 3 shows these models compared with
both the class 1 stars and O- and B-type MS
model spectra (33, 34) in three parameter
spaces. We focused on the class 1 stars because
we expect them to have minimal contamina-
tion from any companion.
To constrain the effective temperatures of
these stars, Fig. 3A shows the EW of He II
l5411 as a function of the EW of He I l5876,
which provides a temperature diagnostic due to
variations in the the helium ionization balance
Drout et al., Science 382, 1287–1291 (2023)
15 December 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
A
B
Fig. 2. Optical spectra with three spectral morphologies. (A) Three example observed spectra (colored
lines) classified as class 1, 2, or 3 (see text), offset for display. Spectra of all the other stars in our sample
are shown in figs. S16 to S21. The gray line shows the optical spectrum of an example WR star (WR 152
divided by a factor of 5) for comparison (50); it has similar line transitions as the class 1 stars, but in
emission. Vertical dotted lines indicate locations of spectral lines, which are identified by the labels above.
Gray shaded bands indicate the lines used in (B). (B) EWs of He II l5411 and Hh + He II l3835 for all 25 stars
in our spectroscopic sample (large numbered circles). T-shaped error bars indicate 1s uncertainties on
detected lines; triangle-terminated error bars indicate 3s upper or lower limits (for absorption or emission)
on undetected lines. For comparison, we show synthetic models of single stripped stars (black dots
enclosed in the gray shaded region), single B-type MS stars (light-green squares), and composites of the two
(colored dots). Model equivalent widths were calculated assuming a signal-to-noise ratio of 35, consistent with
the median signal-to-noise of the observed stars (20). Models are colored to indicate the fraction of V-band
flux contributed to the binary by the stripped star (color bar). Shaded and labeled boxes define the three classes
of spectral morphology we identify; observed data points use the same colors as the shading. Star 15 does
not exhibit He II l5411 but does show He II l4686, so we classify it as class 2 (20). The observed sample forms a
sequence that overlaps with the theoretical predictions for stripped helium star binaries.
with temperature. For all but two stars, we
find Teff > 70 kK owing to the lack of detected
He I; those are temperatures typical of WR
stars, higher than the hottest O-type stars (35).
For some objects, the detection of N IV and/
or N V can provide an alternative temperature
estimate, which ranges between ∼70 to 80 kK
and ≳90 kK (fig. S11) (20).
Figure 3B shows an equivalent plot for the
EWs of He II l3835 + Hh and He II l4860 + Hb.
This provides a rough estimate of surface
gravity, because of the decrease in line strength
relative to the observed pseudo-continuum
for short-wavelength Balmer lines as they
broaden as a result of increasing log(g). The
observed sample is consistent with surface
gravities log(g) ≳ 5, higher than is observed
in MS stars.
Figure 3C shows the pure helium blend
He I + He II l4026 as a function of the hydrogen/
helium blend He II + Hd l4100, which probe
the hydrogen and helium surface mass frac-
tions. The observed stars are all consistent
with hydrogen-depleted surfaces, spanning
the location of the model grid from XH,surf =
0.01 (almost hydrogen-free) to XH,surf = 0.3.
We chose these diagnostics to avoid more wind-
sensitive spectral lines. Our tests with differ-
ent assumptions for the mass-loss rate and
wind velocity do not change these results (20).
The properties we estimated for each star
are listed in table S2. These diagnostics indicate
that the class 1 stars are hot, compact, and
hydrogen-poor. Figure 1 shows that their bright-
nesses fall along a sequence, connecting WR
stars and the slightly lower luminosity WN3/
O3 stars (36) to subdwarfs. Figure S13 shows
that this sequence also appears in the strengths
of stellar wind lines in the optical spectra.
Figure 4 compares our derived constraints on
Teff and log(g) with predictions for intermediate-
mass helium stars (22). The observed stars
have surface gravities between those of MS stars
and white dwarfs—consistent with our expecta-
tions for helium stars—and temperatures hotter
than most subdwarf stars (37). Figure 4 also
shows a set of evolutionary tracks (22). The ob-
served stars are consistent with predictions
for the core-helium burning phase of ∼2.5 to
8 M☉ stripped stars, which have progenitors
with initial masses of between ∼9 and 25 M☉.
These ranges are high enough for the stars to
later undergo core collapse (38), so they will
explode as stripped-envelope supernovae (39).
The winds from stars with initial masses of
<25 M☉ are too weak to remove the hydrogen-
rich envelope (40), so binary interaction is
thought to be the primary mechanism for strip-
ping stars in that mass range (see supplementary
text section of the supplementary materials).
Excluding alternative explanations
We considered other possible interpretations
of the stars in our sample (see supplementary
text). Some of the class 3 stars could be ordi-
nary (B-type) MS stars in regions with very
little dust obscuration; in this case, they would
appear to show a UV excess resulting from
an overcorrection of the photometry. How-
ever, the same is not true for the class 1 and
class 2 stars, whose locations on the CMD (Fig.
1) are inconsistent with massive O-type MS
stars (the only type of MS star expected to
have detectable He II) for any extinction values.
The absorption-line spectral morphologies of
the class 1 and class 2 stars are also distinct
from those of WR stars (Fig. 2A).
Other types of stars can reach very high tem-
peratures and similar brightnesses, such as
accreting white dwarfs, central stars of planetary
Drout et al., Science 382, 1287–1291 (2023)
15 December 2023
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
A
B
C
Fig. 3. Diagnostic diagrams used to estimate stellar
properties. Each panel plots EWs of pairs of spectral
lines (or line blends), chosen to constrain (A) effective
temperature, (B) surface gravity, and (C) helium
enrichment. Black data points (with 1s error bars) show
the measured EWs for the eight class 1 stars, which
have helium star–type spectra. Colored dots show
predictions from a grid of model spectra, with proper-
ties indicated by color (see legends; shaded regions
encompass the full range for each value). Model equivalent
widths were calculated assuming a signal-to-noise of
100, consistent with the median signal-to-noise of the
observed class 1 stars (20). Gray shaded regions indicate
MS star models (20). As shown in (A), all the class 1
stars have an effective temperature of >50 kK, so we only
show models with Teff ≥ 50 kK and O-type MS models
in (B) and (C). The class 1 stars are hot, compact,
hydrogen-poor, and do not overlap MS stars.
Fig. 4. Physical proper-
ties of the eight class 1
systems compared
with other types of
stars. Estimated effective
temperatures and surface
gravities of the class 1
stars are shown as
numbered black circles.
Error bars and limits
correspond to the
properties of the models
that the star overlaps
with in Fig. 3 and are
listed in table S2.
Colored ellipses indicate
the regions occupied by
main-sequence stars
(yellow), white dwarfs
(purple), theoretical
predictions for helium
stars fusing helium in the center (the helium main-sequence, blue), subdwarfs (green), and red giants and
supergiants (red). Black lines show evolutionary tracks for stars with ZAMS masses of 5.5, 9.0, and 18.2 M⊙,
starting on the main sequence and extending until the end of central helium burning (20). Single stars
of these masses (dotted lines) would evolve into cool and extended red giants or red supergiants. Binary
stars of these initial masses undergo envelope stripping by means of mass transfer (thick solid lines)
and subsequently evolve to burn helium as hot and compact helium stars with stripped masses of 1.4, 2.7,
and 7.1 M⊙ (thin solid lines). These stripped helium stars have surface properties similar to our observed sample.
nebulae, and very young post-asymptotic giant
branch (post-AGB) stars (10), but those types
all have circumstellar material, which pro-
duces emission lines or an infrared excess
(41, 42), neither of which we observe. Young
post-AGB stars are also expected to be very rare
(see supplementary text). Very fast rotation
could fully mix stars—resulting in hot and com-
pact helium stars—but this is only expected
at higher masses and luminosities (43). Some
hot low-mass objects (such as evolved sub-
dwarfs and white dwarf merger products)
could pollute our sample, but our targeting
of the LMC and SMC means that they would
need to be foreground objects located in the
halo of the Milky Way. By examining the fre-
quency of UV excesses in a control sample,
we predict that there are <1 foreground ob-
jects along the line of sight to the LMC or SMC
with colors, magnitudes, and kinematics sim-
ilar to our spectroscopic sample (see supple-
mentary text).
The properties we infer for the observed stars
differ from previously identified helium-rich stars.
They are much hotter and more compact than
cool helium giants, such as the star n Sgr [Teff ≲
15 kK, log(g) ∼ 2] (44, 45). While the mass of
n Sgr is uncertain (46), it could be in a subse-
quent evolutionary phase; some stripped stars
expand substantially upon completion of core-
helium burning (47). The surface properties
we estimated for the class 1 stars are similar
to those of HD 45166, but their spectra are
distinct, with HD 45166 having a spectrum
dominated by emission lines. This indicates
that the anomalously slow wind speed observed
in HD 45166 might not be common (see sup-
plementary text). Instead, the absorption spectra
of stars in our sample imply low mass-loss rates,
consistent with theoretical predictions (20).
The properties, binary companions, and evo-
lutionary history of the individual systems in
our sample are likely diverse. Nevertheless, we
conclude that they constitute a population of
massive stars stripped through binary interac-
tion. Because only a subset of stripped star bi-
naries are expected to show a UV excess (20),
the population we observe represents only a
small fraction of the predicted intermediate-
mass helium stars. Many other examples could
be hidden by brighter companion stars. With
estimated masses of ∼2 to 8 M☉, the stars we
observed fill a gap in previously identified he-
lium stars, connecting subdwarfs with WR stars.
REFERENCES AND NOTES
1. H. Sana et al., Science 337, 444–446 (2012).
2. M. Moe, R. Di Stefano, Astrophys. J. Suppl. Ser. 230, 15 (2017).
3. Y. Shao, X.-D. Li, Astrophys. J. 908, 67 (2021).
4.
5. N. Smith, W. Li, A. V. Filippenko, R. Chornock, Mon. Not. R.
J. D. Lyman et al., Mon. Not. R. Astron. Soc. 457, 328–350 (2016).
6.
Astron. Soc. 412, 1522–1538 (2011).
J. J. Eldridge, M. Fraser, S. J. Smartt, J. R. Maund,
R. M. Crockett, Mon. Not. R. Astron. Soc. 436, 774–795 (2013).
7. T. M. Tauris et al., Astrophys. J. 846, 170 (2017).
8. E. R. Stanway, J. J. Eldridge, G. D. Becker, Mon. Not. R. Astron.
Soc. 456, 485–499 (2016).
9. Y. Götberg et al., Astron. Astrophys. 634, A134 (2020).
10. C. S. Jeffery, in Hydrogen-Deficient Stars, A. Werner, T. Rauch,
Eds., vol. 391 of Astronomical Society of the Pacific Conference
Series (Astronomical Society of the Pacific, 2008), pp. 3–16.
11. P. A. Crowther, Annu. Rev. Astron. Astrophys. 45, 177–219 (2007).
12. U. Heber, Publ. Astron. Soc. Pac. 128, 082001 (2016).
Drout et al., Science 382, 1287–1291 (2023)
15 December 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
13. C. S. Jeffery, in Hydrogen-Deficient Stars, A. Werner, T. Rauch,
Eds., vol. 391 of Astronomical Society of the Pacific Conference
Series (Astronomical Society of the Pacific, 2008), pp. 53–63.
14. W. A. Weidmann, R. Gamen, Astron. Astrophys. 526, A6 (2011).
15. L. Wang et al., Astron. J. 161, 248 (2021).
16. M. M. Shara et al., Mon. Not. R. Astron. Soc. 464, 2066–2074
(2017).
17. M. H. van Kerkwijk et al., Nature 355, 703–705 (1992).
18. J. H. Groh, A. S. Oliveira, J. E. Steiner, Astron. Astrophys. 485,
245–256 (2008).
19. T. Shenar et al., Science 381, 761–765 (2023).
20. Materials and methods are available as supplementary materials.
21. P. Massey, K. F. Neugent, N. Morrell, D. J. Hillier, Astrophys. J.
788, 83 (2014).
22. Y. Götberg et al., Astron. Astrophys. 615, A78 (2018).
23. L. M. Z. Hagen et al., Mon. Not. R. Astron. Soc. 466,
4540–4557 (2017).
24. D. Lang, D. W. Hogg, D. J. Schlegel, Astron. J. 151, 36 (2016).
25. D. Zaritsky, J. Harris, I. B. Thompson, E. K. Grebel, P. Massey,
Astron. J. 123, 855–872 (2002).
26. D. Zaritsky, J. Harris, I. B. Thompson, E. K. Grebel, Astron. J.
128, 1606–1614 (2004).
27. K. D. Gordon, G. C. Clayton, K. A. Misselt, A. U. Landolt,
M. J. Wolff, Astrophys. J. 594, 279–293 (2003).
28. J. L. Marshall et al., Proc. SPIE 7014, 701454 (2008).
29. Gaia Collaboration, Astron. Astrophys. 674, A1 (2023).
30. R. O. Gray, C. J. Corbally, Stellar Spectral Classification,
Princeton Series in Astrophysics (Princeton Univ. Press, 2009).
31. J. S. Vink, Astron. Astrophys. 607, L8 (2017).
32. D. J. Hillier, D. L. Miller, Astrophys. J. 496, 407–427 (1998).
33. T. Lanz, I. Hubeny, Astrophys. J. Suppl. Ser. 146, 417–441
(2003).
34. T. Lanz, I. Hubeny, Astrophys. J. Suppl. Ser. 169, 83–104 (2007).
35. N. R. Walborn et al., Astron. J. 123, 2754–2771 (2002).
36. K. F. Neugent, P. Massey, D. J. Hillier, N. Morrell, Astrophys. J.
841, 20 (2017).
37. J. S. Drilling, C. S. Jeffery, U. Heber, S. Moehler, R. Napiwotzki,
Astron. Astrophys. 551, A31 (2013).
38. T. Ertl, S. E. Woosley, T. Sukhbold, H. T. Janka, Astrophys. J.
890, 51 (2020).
39. S.-C. Yoon, L. Dessart, A. Clocchiatti, Astrophys. J. 840, 10 (2017).
40. E. R. Beasor, B. Davies, N. Smith, Astrophys. J. 922, 55 (2021).
41. H. Van Winckel, Annu. Rev. Astron. Astrophys. 41, 391–427
(2003).
42. W. R. Hamann, M. Peña, G. Gräfener, M. T. Ruiz, Astron.
Astrophys. 409, 969–982 (2003).
43. I. Brott et al., Astron. Astrophys. 530, A115 (2011).
44. D. Schoenberner, J. S. Drilling, Astrophys. J. 268, 225–227 (1983).
45. T. Kipper, V. G. Klochkova, Balt. Astron. 21, 219–240 (2012).
46. A. Gilkis, T. Shenar, Mon. Not. R. Astron. Soc. 518, 3541–3555
(2023).
47. E. Laplace, Y. Götberg, S. E. de Mink, S. Justham, R. Farmer,
Astron. Astrophys. 637, A6 (2020).
48. J. Breysacher, M. Azzopardi, G. Testor, Astron. Astrophys.
Suppl. Ser. 137, 117–145 (1999).
49. P. Massey, A. S. Duffy, Astrophys. J. 550, 713–723 (2001).
50. W.-R. Hamann, L. Koesterke, U. Wessolowski, Astron. Astrophys.
Suppl. Ser. 113, 459–471 (1995).
51. M. R. Drout et al., An observed population of intermediate-
mass helium stars that have been stripped in binaries –
theoretical, computational and observational data,
Zenodo (2023); https://doi.org/10.5281/zenodo.7600727.
52. Y. Götberg et al., Stellar properties of observed stars stripped
in binaries in the Magellanic Clouds – spectral models,
Zenodo (2023); https://doi.org/10.5281/zenodo.7976200.
53. M. R. Drout et al., An observed population of intermediate-
mass helium stars that have been stripped in binaries – raw
spectroscopic data, Zenodo (2023); https://doi.org/10.5281/
zenodo.10035849.
ACKN OWLED GMEN TS
We thank K. Auchettl, K. Breivik, C. Burns, A. Carpenter, J. Fuller,
B. Hovis-Afflerbach, A. Ji, C. Johnston, D. Kelson, M. van Kerkwijk,
D. Lang, C. Norman, T. Piro, M. Renzo, A. Roc, P. Senchyna,
S. Torres, and M. Zapartas for fruitful discussions and support,
and we acknowledge feedback from the anonymous referees that
improved this manuscript. This paper includes data gathered
with the 6.5 m Magellan Telescopes located at Las Campanas
Observatory, Chile. We thank J. Mulchaey (Carnegie Observatories
director), L. Infante (Las Campanas Observatory director), and
the entire Las Campanas staff for their hard work and dedication
to keeping the observatory operating through the COVID-19
pandemic, when a large fraction of this data was obtained.
Computing resources used for this work were made possible by
a grant from the Ahmanson Foundation. Funding: M.R.D.
acknowledges support from the NSERC through grant RGPIN-
2019-06186, the Canada Research Chairs Program, the Canadian
Institute for Advanced Research (CIFAR), and the Dunlap
Institute at the University of Toronto. Y.G. was supported by NASA
through the NASA Hubble Fellowship Program grant HST-HF2-
51457.001-A awarded by the Space Telescope Science Institute,
which is operated by the Association of Universities for Research
in Astronomy, Inc., for NASA, under contract NAS5-26555.
A.J.G.O. acknowledges support from the Lachlan Gilchrist
Fellowship Fund. S.E.d.M. acknowledges funding by the
Netherlands Organization for Scientific Research (NWO) as part of
the Vidi research program BinWaves under project number
639.042.728. N.S. was supported by NASA through HST grant GO-
15824. Author contributions: M.R.D., S.E.d.M., and Y.G. designed
the UV-optical color-magnitude diagrams used for candidate
identification. M.R.D. and B.A.L. performed photometry on the
Swift UVOT images and selected candidate targets for follow-up.
M.R.D., Y.G., B.A.L., and A.J.G.O. performed the spectroscopic
observations with the Magellan Baade telescope. M.R.D. reduced
and analyzed the spectroscopic data and performed the kinematic
analysis. Y.G. computed the spectral models with CMFGEN with
input from J.H.G. and designed the equivalent width diagnostics
used to estimate stellar properties. M.R.D. and Y.G. interpreted
the observed sample in the context of the stellar models. M.R.D.
and Y.G. wrote the paper, with input from B.A.L., J.H.G., S.E.d.M.,
A.J.G.O., and N.S. Competing interests: The authors have no
competing interests to declare. Data and materials availability:
Reduced and combined spectra of the 25 stars in our
spectroscopic sample, our stripped helium star models, our
main-sequence spectral models, equivalent width measurements
from the theoretical models, and the index files used for
astrometry are archived at Zenodo (51, 52). The UV and optical
photometry for the observed stars and models are provided in
tables S6, S4, and S7 and are archived in machine-readable form
at Zenodo (51). Raw Magellan Echellette spectra are also archived
at Zenodo (53). License information: Copyright © 2023 the
authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.science.org/about/science-
licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade4970
Materials and Methods
Supplementary Text
Figs. S1 to S21
Tables S1 to S9
References (54–167)
Submitted 30 August 2022; accepted 27 October 2023
10.1126/science.ade4970
Drout et al., Science 382, 1287–1291 (2023)
15 December 2023
5 of 5
|
10.1126_science.ade4985
|
RES EARCH
ROBOTICS
Multilegged matter transport: A framework for
locomotion on noisy landscapes
Baxi Chong1,2, Juntao He3, Daniel Soto3, Tianyu Wang3, Daniel Irvine4,
Grigoriy Blekherman4, Daniel I. Goldman1,2,3*
Whereas the transport of matter by wheeled vehicles or legged robots can be guaranteed in engineered
landscapes such as roads or rails, locomotion prediction in complex environments such as collapsed
buildings or crop fields remains challenging. Inspired by the principles of information transmission,
which allow signals to be reliably transmitted over “noisy” channels, we developed a “matter-transport”
framework that demonstrates that noninertial locomotion can be provably generated over noisy
rugose landscapes (heterogeneities on the scale of locomotor dimensions). Experiments confirm that
sufficient spatial redundancy in the form of serially connected legged robots leads to reliable transport
on such terrain without requiring sensing and control. Further analogies from communication theory
coupled with advances in gaits (coding) and sensor-based feedback control (error detection and
correction) can lead to agile locomotion in complex terradynamic regimes.
T he transport of matter across land is cru-
cial to societies and groups (1, 2, 3), as well
as to individuals during locomotion (4).
Engineered self-propulsion as a means
of terrestrial matter transport has been
studied across scales from enormous multi-
wheeled trains (Fig. 1A) to small few-wheeled
and legged robots (5–7). For large devices, low
dissipation, inertia-dominated locomotion is a
commonly used matter-transport scheme. Spe-
cifically, locomotion in wheeled systems on
smooth surfaces such as tracks and roads will
persist over long distance unless acted on by
dissipative internal or external forces (Fig. 1A).
On natural terrain, dissipation attributable
to external forces can occur through interac-
tions with terrain heterogeneities such as ob-
stacles, gaps, or inclines (8), as well as through
interactions with flowable materials (9). In
dissipation-dominated applications, such as
those encountered in robot movement in cer-
tain agricultural (e.g., crop fields) or confined
and crowded search-and-rescue (e.g., collapsed
buildings) scenarios, a system must contin-
uously and actively generate forces and/or
reduce dissipation. In part, because of our
lack of understanding of the terradynamic
(9, 10) interactions with the environments
listed above, principles by which locomotors
can be designed and controlled to guarantee
reliable and predictable matter transport are
lacking.
One engineering solution (11) to facilitate
matter transport in complex terradynamic
regimes is to use structures such as limbs to
1Interdisciplinary Graduate Program in Quantitative Biosciences,
Georgia Institute of Technology, North Avenue, Atlanta,
GA 30332, USA. 2School of Physics, Georgia Institute of
Technology, 837 State St NW, Atlanta, GA 30332, USA.
3Institute for Robotics and Intelligent Machines, Georgia
Institute of Technology, 801 Atlantic Dr NW, Atlanta,
GA 30332, USA. 4School of Mathematics, Georgia Institute
of Technology, 686 Cherry St NW, Atlanta, GA 30332, USA.
*Corresponding author. Email: daniel.goldman@physics.gatech.edu
periodically make and break contact with the
environment (12). Such dynamics can poten-
tially simplify the thrusting interactions into a
collection of discrete units, which minimize
unexpected interference (4), thus providing
an alternative to wheeled carriers on “noisy”
landscapes. There have been two basic ap-
proaches of limb use in dissipation-dominated
environments. The first relies heavily on sen-
sors (13) to detect and respond to terrain het-
erogeneity in real time (5, 14). This approach is
used for the increasingly agile locomotion in
state-of-the-art legged robots (mostly bipedal
or quadrupedal) (5, 12, 14, 15). However, the
use of sensors and high bandwidth control
can be expensive and restricted to specific
applications.
The second approach is to instill legged ro-
bots with “mechanical intelligence” such that
locomotion can be performed with minimal
environmental awareness. This has been most
effective with devices with more than four legs,
such as hexapods (6) and myriapods (16, 17).
Whereas more limbs help avoid catastrophic
failures (e.g., loss of stability), terrain hetero-
geneity can still cause deficiencies in thrust-
ing interactions, which substantially reduce
locomotor performance (movie S1) (18–20).
This raises the question of how variable num-
bers of limbs and sensors should be arranged
such that one can guarantee that a locomo-
tor can go from point A to point B in a speci-
fied time and across an arbitrarily complex
landscape, and furthermore, how much sens-
ing, feedback, bandwidth, and/or control are
needed.
This question is analogous to that of infor-
mation and signal transmission over noisy
channels as first analyzed by Shannon nearly
a century ago (21). Over a noiseless channel, a
continuous analog signal is, in principle, able
to convey an infinite amount of information
(22). Despite its efficiency, an analog signal can
be distorted by channel noise inherent in all
communication modalities, a property similar
to heterogeneities introduced to smooth sur-
faces in inertia-driven matter transport (Fig. 1A).
To counter channel noise in communication,
Shannon (21) constructed a scheme in which
the central idea was to digitize (encode) infor-
mation into binary bit sequences and “buffer”
(correct) the transmission error through re-
dundancy (Fig. 1C).
With the analogy to information theory, it
is reasonable—at least, conceptually—to an-
ticipate reliable matter transport over noisy
landscapes given sufficient redundancy of
terrestrial interaction. In this work, we show
that this anticipation is correct and leads to an
open-loop framework for matter transport by
which, for a complex terradynamic task, we can
guarantee that sufficiently redundant multi-
legged robots can reliably and predictably self-
transport over a given distance through “buffer
and tolerate” dynamics without the need for
sensing and feedback control or environmen-
tal awareness (Fig. 1B and movie S1) (16).
Development of the matter-transport framework
Our framework proceeds as follows (Fig. 2):
We define a transport task as a physical en-
tity moving to a specific destination D at a
fixed time T. As shown in Fig. 2B (i), this is
analogous to the intended message being trans-
mitted across a noisy channel at a given rate.
Similar to the bit-based digital signal trans-
mission, we focus on a dissipation-dominated
system in which locomotion is driven by thrust-
ing interactions from basic active contacts
(bacs, our analogy to bits), which are discrete
units of active terradynamic interaction. Ex-
amples of bacs include limbs making contact
with the environment (23) or vertical waves of
contact in limbless robots (24). We quantify
the temporal and spatial distribution of bacs
as a binary sequence Xm, in which 1 denotes
contact and 0 denotes noncontact [Fig. 2B (ii)].
As the locomotor implements the desired bac
sequence over a noisy landscape (the analog of
a noisy channel) [Fig. 2B (iii)], the terrain un-
certainty can introduce contact noise to the
actual bac sequence, Ym [Fig. 2B (iv)]. This leads
to a discrepancy between the actual destina-
tion ^D (evaluated at the scheduled time T ) and
the desired destination D [Fig. 2B (v)].
We next discuss our characterization and
quantification of noisy landscapes. A dissipation-
dominated terrain can have different types
of heterogeneities, each with different com-
plex terradynamic effects on bacs (9). Con-
sider a terrain characterized by a height map,
h(x,y). Depending on the scale of the gra-
dient, @h=@x;
@h=@y(cid:2), the terrain heteroge-
½
neity can affect the locomotion in the form
of slopes, walls, or obstacles (Fig. 1B), which
directly impact the thrust-generation pro-
cess in the plane parallel to the terrestrial
Chong et al., Science 380, 509–515 (2023)
5 May 2023
1 of 7
RES EARCH | R E S E A R C H A R T I C L E
A Matter transportation
Source matter
Continuous
Discrete
B Noisy landscapes
Rugosity
Granular media
Obstacles
Wall
Environment awareness
C Signal transmission
Original signal
Analog
Digital
...
...
...
...
Noisy wire
Retransmission
...
...
...
...
Received matter
Delayed
Delivered
Delivered
Recovered signal
Fig. 1. Signal transmission and matter transport. (A) Matter transport
with either continuous or discrete active contacts can be effective on “noise-
free” tracks. Discrete redundant contacts enable effective matter transport over
rugose tracks through redundancy or environmental awareness. (B) Multi-
segmented robophysical locomotors with directionally compliant legs traverse
noisy landscapes: (from left to right) a laboratory model of rugose terrain,
entangled granular media, boulders, and steps. (C) The transmission of
analog and digital signals through noisy wires. A digital signal allows reliable
transmission through a noisy wire through either redundancy or a
retransmission channel.
surface (e.g., a stumble) (25). Parallel thrust
disturbances can be minimized by proper
design of mechanical structures or passively
compliant mechanisms [supplementary ma-
terials (SM), section 1.2] (16, 26). In this study,
we focus on a class of noisy landscapes (ru-
gose terrains) in which the height distribu-
tion, h(x,y), can affect the supporting force
distribution (e.g., missing steps) in the direc-
tion perpendicular to the terrestrial plane
and therefore contaminate the intended bac
sequence X m→Y m
ð
Þ.
With the notion of bacs and contact noise
established, we can now model matter trans-
port as a stochastic process. We first consider
an abstract characterization of thrust genera-
tion if given one pair of legs. We quantify the
instantaneous thrust over a bac, f(t), as the
instantaneous external force required to keep
the locomotor in place at time t∈ 0; t
Þ, where
t is the duration of the bac. An example of
thrust function f(t) is illustrated in Fig. 3A.
The nominal (undisturbed on flat terrain)
∫t
average thrust is fn ¼ 1
0 f tð Þdt.
t
Next, we introduce a coefficient function
which encapsulates the uncertainty in the bac,
½
t
½
ð
f
c(t). The terrain-disturbed thrust can be for-
mulated by ^f ¼ 1
∫t
0c tð Þf tð Þdt. We assume c(t)
∫t
has the property 1
0c tð Þ ¼ 1 so that the sup-
t
porting force balances gravity. Further, we
assume that the initiation of a bac is delayed by
some time, c1, and the duration of a bac is short-
ened to tu : c tð Þ ¼ 0; t ∉ c1; c1 þ tu
g. Spe-
(cid:2)
cifically, we assume c1 to be a random variable
having a uniform distribution of c1 ∼ U 0; t
Þ ,
and the duration of the bac, tu, is assumed to
be a random variable determined by the ter-
rain rugosity. We sample tu from the cumu-
lative distribution function given by G tuð
Þ ¼
1 (cid:3) b
(cid:2) so that there is a fi-
ð
½
nite probability of complete bac loss, p tu ¼ð
0Þ ¼ b , and b < 1 characterizes the contact
noise level and offers an approximation to
the rugosity of the terrain (Fig. 3C). Whenever
c1 þ tu > t , we extend the excessive contact
duration c1 þ tu (cid:3) t
Þ into the next bac [Fig. 2B
ð
(iii)] (SM, section 1.4). For simplicity, we as-
sume that c(t) is otherwise uniformly distributed
(cid:4)
during the bac: c tð Þ ¼ t(cid:3)1
g.
In this way, the terrain-disturbed average
thrust reduces to ^f ¼ sign tuð
u fu , where
fu ¼ ∫c1þtu
f tð Þdt is the thrust disturbance. The
c1
Þtu=t þ b; tu ∈ 0; t
t ∈ c1;c1 þ tu
u t;
Þt(cid:3)1
(cid:1)
(cid:3)
sign function sign(tu) implies that no thrust
(cid:6)
^f ¼ 0
with complete bac
will be generated
loss (tu = 0).
(cid:5)
As demonstrated in previous studies of
dissipation-dominated multilegged locomotion
on flat ground (27, 28), because of the periodic
limb lifting and landing, an effective viscous
(rate-dependent) cycle-averaged thrust-velocity
(the average thrust and velocity over a period,
respectively) relationship emerges in frictional
environments, despite such thrusts being in-
stantaneously independent of velocity. Specif-
ically, the relationship of cycle-averaged robot
locomotion velocity is derived to be linearly
correlated with the cycle-averaged thrust:
^v ¼ g(cid:3)1^f , where g is the effective viscous drag
coefficient (Fig. 3D). In this way, the terrain-
disturbed velocity can be approximated by
^v ¼ g(cid:3)1sign tuð
Þt(cid:3)1
u fu.
Taking the analogy from information theory
in which redundant bits can bound the un-
certainty from channel noise, we hypothesize
that locomotors with redundant bacs can offer
robustness over terrain uncertainty. A straight-
forward scheme to include redundancy is to
decrease the transport rate by allowing more
Chong et al., Science 380, 509–515 (2023)
5 May 2023
2 of 7
RES EARCH | R E S E A R C H A R T I C L E
A(i)
Signal transmission
W
Message
A(ii)
Locomotion
D
Mass/destination
Encoder
Gait design
Xm
Desired bits
Xm
Desired bacs
Noisy channel
Complex terrain
Ym
Received bits
Ym
Actual bacs
B(i)
D
B(ii)
Xm
Bac
No contact
Contact noise
B(iii)
Desired bacs
B(iv)
Ym
L[1]
D
L[3]
R[1]
R[3]
Actual bacs
c1
c1
Time
Time
Decoder
W
Estimated message
Gait realization
D
Actual destination
B(v)
D
D
Fig. 2. Framing the matter-transport problem as a sequence of basic
active contacts (bacs). (A) The correspondence of processes in (i) signal
transmission [adapted from (21)] and (ii) locomotion (matter transport).
(B) (i) A multilegged robot, the matter to be transported to a destination D.
(ii) The desired bac sequence to reach the locomotion destination. (iii) Noisy
landscapes can introduce contact errors such as delaying bacs and shortening
the duration of bacs. We compare the desired bac (which spans a duration t)
and two terrain-contaminated bacs (each begins at c1) with shorter duration
(tu). (iv) A bac sequence contaminated by contact errors leads to a (v) locomotion
destination
^
D smaller than the expected D.
transport time (temporal redundancy). Thus,
we have:
1½ (cid:2) ¼
^vT
1
gT
XT
i¼1
(cid:5) (cid:6) f i
sign ti
u
u
ti
u
ð1Þ
(cid:8)
where ^vT
1½ (cid:2)
is the average terrain-disturbed
velocity over T periods, ti
u and f i
u are the con-
tact and thrust disturbance, respectively, over
the ith period. Here, T represents the order of
temporal redundancy. We expect the vari-
(cid:7)
ance of the average terrain-disturbed velocity,
s2 ^v 1½ (cid:2)
, to decrease as T increases. Further,
T
^vT
1½ (cid:2) converges to a Dirac delta function as T
approaches infinity (proof given in the SM,
proposition 3). Moreover, the expected aver-
age terrain-disturbed velocity, ^v 1½ (cid:2)
, remains
T
constant (by the law of large numbers).
E
D
We next evaluate the effectiveness of this
temporal redundancy scheme. For simplicity,
we assume D is one-dimensional, ^DT
1½ (cid:2) ¼ T ^v1½ (cid:2)
T ,
and we note that there is no reason to expect that
^DT
1½ (cid:2) should converge (to D) as T increases. There-
fore, temporal redundancy can only guarantee
the completion of a matter-transport task, but
the exact transport duration can be indefinite.
Given the inefficiency of temporal redun-
dancy, we develop a framework, analogous to
Shannon’s encoding scheme, to remove inef-
ficient redundancy and compensate it with
“redundancy of the right sort” [(29) p. 164] for
more effective locomotion. In particular, the
appropriate redundancy facilitates the simul-
taneous “communication” (e.g., redistribution)
of bacs in response to contact noise. To devel-
op a specific scenario for legged systems, we
consider redundancy in the form of repeating
serially-connected modules, in which a mod-
ule is defined as a pair of legs. With proper
coordination, the effect of contact noise will be
shared among all bacs instead of acting on an
individual bac. Because such redundancy is
distributed in the spatial domain, we refer to
it as spatial redundancy. Effectively, this spa-
tial redundancy serves as a moving average
filter over the contact-noise profile. For sim-
plicity, we consider a simple module coordi-
nation that the instantaneous thrust f(t) on
each module is identical and invariant to
the number of modules. The average terrain-
disturbed velocity for N serially connected
modules over T periods is:
N½
^vT
(cid:2) ¼
1
gT
XT
i¼1
0
B
B
B
B
B
@
sign
!
XN
j¼1
tij
u
1
C
C
C
C
C
A
XN
j¼1
XN
j¼1
f ij
u
tij
u
ð2Þ
u ¼ 0
u and f ij
where tij
u are disturbances on the jth
module over the ith temporal repetition. In-
tuitively, in the case where there are M com-
plete bac losses in the ith temporal repetition,
(cid:10)
(cid:10)
(cid:9)
(cid:1)
(cid:10), the locomotor with N mod-
(cid:10)
M ¼ j; tij
ules will essentially reduce to the configuration
with N – M modules. In other words, loco-
motors with spatial redundancy N can afford
up to N – 1 complete bac losses without sub-
stantial thrust deficiency, which indicates that
spatial redundancy can also serve to bound
the uncertainty in thrust generation. We show
that for fixed T ≥ 1, ^vT
N½
(cid:2) will also converge to a
(cid:2)
Dirac delta function as spatial redundancy
N approaches infinity (SM, proposition 4).
D
E
The expected average terrain-disturbed veloc-
ity, ^vT
N½
, can be approximated by (1 − bN)Cs,
where Cs is a constant determined by f(t), g, and
b (SM, proposition 4). Therefore, greater spatial
redundancy not only reduces variance, but also
improves the expected average terrain-disturbed
velocity, a feature otherwise not possible with
only temporal redundancy.
N½
For any fixed T, the distribution of actual
destinations ^DT
(cid:2) ¼ T ^vT
N½
(cid:2) will also converge
to a Dirac delta function as N approaches infin-
ity. Given a matter-transport task over desired
distance D at scheduled time T, we consider it
successful if
(cid:11)
(cid:11)
(cid:5)
^D; ^T
(cid:6)
(cid:3) D; Tð
(cid:11)
(cid:11)
Þ
(cid:10)
(cid:10)
¼ ^D
(cid:10)
(cid:10)
(cid:10)
N½
(cid:2)
(cid:10) < e
T (cid:3) D
ð3Þ
where e is the tolerance. In this way, for arbitrary
e > 0 and p0 < 1, there exists a finite N such that
the probability of successful matter transport
(subject to e) is greater than p0 (proof given in
the SM, proposition 5). The “minimal spatial
redundancy” required to achieve the proba-
bility of p0 for tolerance e > 0 is bounded by
, where k is a constant that
Ne;p0
describes the locomotor speed on flat terrain,
and D is the desired destination specified by dis-
tance, for one-dimensional locomotion (proof
given in the SM, propositions 5 and 6).
kDp
logðbÞ
≤ logð1(cid:3)
ffiffiffiffi
p0
Þ
Numerical tests of the framework
We first tested our theoretical predictions by nu-
merical simulations. We chose a saptiotemporal
Chong et al., Science 380, 509–515 (2023)
5 May 2023
3 of 7
RES EARCH | R E S E A R C H A R T I C L E
bac distribution pattern that was based on
limb-stepping patterns of biological centipedes
[and whose efficacy in generating locomotion
in multilegged robots was previously studied
in (27, 30)]. Each bac generates an instanta-
neous thrust given by f(t), which is indepen-
dent of our choice of spatial redundancy (proof
given in the SM, proposition 1). We illustrate
f(t) in Fig. 3B. Assuming b = 0.5, we compared
the distribution of normalized (by the nominal
speed, vopen) terrain-disturbed velocity, v/vopen,
for different combinations of temporal and
spatial redundancy in Fig. 3E. In this ex-
ample, Cs = vopen indicates that the terrain-
disturbed velocity converges to the nominal
velocity (dashed black curve) given sufficient
spatial redundancy.
To illustrate that spatial redundancy allows
the matter-transport process (defined by the
actual distance achieved, ^D, and duration to
achieve this displacement, ^T ) to converge to
the desired transport-distance and duration
pair (D, T), we calculate the 90% confidence
interval (CI) of ^D as a function of T for N = 8
(red curves) and N = 1 (blue curves) (Fig. 3F).
With greater spatial redundancy, the vari-
ance of ^D is substantially reduced, approach-
ing the limit of nominal distance-duration
relationship (dashed black curve) (Fig. 3F).
Further, we numerically calculate the proba-
bility of successful matter transport (evaluated
at T = 1) as a function of spatial redundancy
N subject to different choices of tolerance e
(Fig. 3G). For all choices of e, the success prob-
ability converges to one as N increases. Figure
3G also shows that the marginal benefit of
having more legs decreases as N increases;
the predicted limit is given by the dashed black
curve. Finally, the locomotion performance
can be affected by the terrain rugosity. We
calculated Ne;p0 as a function of b, the con-
tact noise level, and plot this in Fig. 3H. We
compare three combinations of e and p0 and
notice that more rugose terrains require
greater (yet finite) Ne;p0 ; the theoretical
bound is given by the dashed black curve
when p0 = 0.9.
Experimental tests of the framework
Because the model assumes simplified envi-
ronmental interactions, we next tested if our
framework could predict locomotor per-
formance in a physical system. We chose to
work with a well-controlled laboratory multi-
legged robophysical model (movie S1) (design
and control details are provided in the SM,
sections 1.2 and 1.3), similar in design to those
in (27, 30, 31). We measured the average speed
and variance in such robots with different
leg numbers and terrain complexity. Given
that our framework indicates that matter
transport can be achieved without the need
for environmental awareness, we controlled
all robots such that they executed their pre-
A
f [6]
B
1
)
t
(
f
0
0
E(i)
1
N = 1
F
D
C
0
0
E(ii)
T = 1
F
D
C
f [1]
(I)
(II)
C
1
)
u
τ
(
G
F
D
C
b
0
0
f [5]
~
f
t
T=1
T=10
T=60
0.5
1
1.5
N=1
N=10
N=60
0
0.5
1
1.5
E(iii)
N = 8
F
D
C
T=1
T=5
T=10
fn=0.72
~
f
0.8
N=8
14
80
D
40
N
D
1
n
e
p
o
v
/
v
0
0
N=1
F
40
20
T
G
H
0
0
1
0
p
0
0
N0.15, 0.95
30 N0.15, 0.9
N0.2, 0.9
0
0.5
1
v/vopen
1.5
0
0
0.4
b (contact noise level)
0.8
Fig. 3. Numerical simulation to test the matter-transport framework. (A) Illustrations of (left) thrust
generation from bacs and (right) the thrust-velocity relationship. Self-propulsion with (I) nominal contact and
(II) contact errors are compared. (B) The instantaneous thrust f(t) as a function of time, derived from
(28). (C) The cumulative distribution function (CDF) of tu. (D) The thrust-velocity (normalized by the nominal
velocity) relationship. (E) The numerical CDF of terrain-disturbed velocity for robots with different
combinations of temporal (T) and spatial (N) redundancy. (i to iii) Black dashed curves indicate the nominal
CDF. (F) Numerically calculated 90% CI of actual destination
^
D‐T relationship is shown with the black
spatial redundancies: N = 1 (blue) and N = 8 (red). The nominal
dashed curve. (G) The probability of successful scheduled matter transport (p0) as a function of spatial
redundancy evaluated at different tolerance e. We plot the theoretical predicted limit as a black dashed curve.
(H) The relationship between Ne,p0 , the minimal spatial redundancy required to achieve the desired success
probability (p0) subject to tolerance e, and b, contact noise level. The black dashed line indicates the
theoretical predicted limit when p0 = 0.9.
^
D as a function of T. We compare two
programmed stepping patterns (open loop)
and did not sense or respond to features of
the environment. To facilitate comparison
across different spatial redundancies N, our
chosen bac sequence (the same as that in the
numerical tests) has the property that, in the-
ory, all robots share the same thrust function
f(t), the same performance on flat terrain, and
the same thrust-velocity relationship (proof
given in the SM, proposition 2).
We constructed laboratory models of rugose
terrains composed of (10 × 10 cm2) blocks with
variation in heights (Fig. 4A). The block
heights, h(x,y), are randomly distributed (SM,
section 1.1). Such rugose terrains ensure that
limbs will experience thrust deficiency from
Chong et al., Science 380, 509–515 (2023)
5 May 2023
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
stochastic contact (32). The contact error can also
arise from robot motor noise because of actua-
tion delay, insufficient torque, or body compli-
ance. We define the terrain rugosity, Rg, as the
standard deviation of heights normalized by
block side length. We tested the performance
of 3-8 segmented [each segment has two di-
rectionally compliant limbs (31)] robophysical
models on rugose terrains and recorded the
bac duration (tu) on each leg. The distributions
of tu measured from 225 and 309 bacs for ter-
rain with rugosity 0.17 and 0.32, respectively,
are shown in fig. S5. For 0 < tu < t, the mea-
sured cumulative distribution function of tu can
be approximated by a linear function, a feature
in accord with (and therefore, justifying) our as-
sumed tu bac duration distribution in Fig. 3B.
We also recorded the normalized interstep
average velocity (the average velocity over a
step that spans the duration of a bac) of the 12-
legged robot on rugose terrains during multi-
ple steps (v/vopen) and compared this with the
corresponding normalized bac duration (tu/t)
(Fig. 4B). Specifically, in each trial (five trials
on each terrain), we programmed the robot to
run for two periods such that there was at least
one bac generated by each leg (6 legs visual-
izable from side-view camera). The correlation
in these quantities (Fig. 4B) indicates that bac
contamination is an important source of loco-
motor speed variability, which is in accord with
our model assumption. Further, we measured
the distributions of average v/vopen with dif-
ferent combinations of temporal and spatial
redundancy (Fig. 4D). Specifically, we measured
the average displacement (over T cycles; T is
specified in Fig. 4D) of a 6-legged and a 12-legged
robot on each terrain. The measured cumu-
lative distribution functions of average v/vopen
were in qualitative agreement with predictions
from numerical tests (Fig. 3E). The sources for
discrepancies between numerical and experi-
mental tests include over-simplification of ter-
rain characterization and contact error from
robot motor noise.
As discussed earlier, for a given transport
distance D, the average and the variance of
transport duration are both important met-
rics for evaluating matter-transport perfor-
mance. We measured the transport duration
(in units of numbers of periods) as a function
of D for a 6-legged robot and a 14-legged
robot over a rugose terrain (Rg = 0.32). Sim-
ilar to our numerical prediction in Fig. 3F,
experimental results also show that spatial
redundancy reduces the variance (illustrated
by error bars) in transport duration (Fig. 4C).
Thus, we can guarantee predictable and reliable
speed on a noisy terrain during open-loop noni-
nertial matter transport, analogous to that of
the predictability of inertia-driven dynamics
on noiseless (e.g., rails and roads) terrain.
To determine how spatial redundancy af-
fects transport duration, we recorded T[D=60],
the time required for a robot to locomote 60 cm
over rugose terrains, for robots with varying
number of legs on Rg = {0, 0.17, 0.32} (10 trials
in each condition) (Fig. 4E). A hexapod can
eventually self-transport 60 cm, but there is a
large variance in T[D=60]. By contrast, systems
with 16 legs can finish the self-transport task
with short average time and small variance.
Moreover, for systems with sufficiently high spa-
tial redundancy (e.g., N ≥ 5), further increases
in N do not result in substantial changes in
T[D=60], including both the average and the
A(iii)
0.2
F
D
P
0
0
C
120
12
0
Height (cm)
Rg=0.17
0.32
6
Height (cm)
12
A(i)
Rg=0.17
A(ii)
Rg=0.32
10cm
B
0.6
n
e
p
o
v
/
v
0
0.4
0.7
τu/τ
1
D
Rg=0.17
Rg=0.32
E
1
0
1
)
3
=
N
(
F
D
C
)
6
=
N
(
F
D
C
0
0
T=12
T=24
T=1
T=6
T=1
1 0
avg. v/vopen
T=12
T=1
1
l
)
s
e
c
y
c
(
T
60
N = 3
0
40
N = 7
70
100
D, destination (cm)
40
20
]
0
6
=
D
T
[
0
Rg=0.32
Rg=0.17
Rg=0
3
7
N: number of leg pairs
5
6
CM
Fig. 4. Experimental test of the matter transport framework on multilegged robots. (A) Renderings of
rugose terrains with rugosity (i) Rg = 0.17 and (ii) Rg = 0.32. The block-height distributions are shown in (iii).
(B) Bac contamination leads to speed degradation in a 12-legged robot on rugose terrains. Each point
denotes the robot’s normalized interstep averaged velocity v/vopen) and a corresponding contact duration
^
(tu/t) in one bac. Color schemes are identical to those in (A). (C) The empirical transport duration (
T, in units
of gait periods) as a function of destination distance D. We compare two robots with 14 (red) and 6 (blue)
legs (renderings shown as insets). There is a large variance in transport duration for the six-legged robot,
and the variance grows as destination distance increases. The 14-legged robot has a comparably tightly
bounded transport duration. (D) The empirical distribution of average velocity on terrains with (left) Rg = 0.17
and (right) Rg = 0.32 on (Top) the six-legged robot and (Bottom) the 12-legged robot. The empirical
distributions were obtained from 30 trials. (E) For robots with different numbers of leg pairs N, we recorded
T[D= 60], the number of periods required to transport D = 60 cm on terrains with (blue) Rg = 0, (green)
Rg = 0.17, and (black) Rg = 0.32. The error bar was calculated from at least 10 trials. T[D=60] for contact-
modulated gaits are illustrated in the dashed purple rectangle.
Chong et al., Science 380, 509–515 (2023)
5 May 2023
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
variance. This finding is consistent with our
numerical prediction of the marginal benefit
from increases in spatial redundancy after a
large enough N (Fig. 3G).
From the Shannon scheme for signal trans-
mission, it is reasonable to anticipate im-
proved performance with more elaborate coding
schemes, which we define in matter transport
as designing the terradynamic interaction pro-
file of bacs (e.g., the instantaneous thrust func-
tion f(t) and thrust-velocity relationship). One
straightforward method to modulate the bac-
interaction profile is to change the body wave
amplitude (27) or the number of waves on the
body (33), which alters the temporal and spa-
tial distribution of bacs and the associated body
postures. To illustrate the potential of appro-
priate coding, we show one example of gait
design modulation. Specifically, we imposed
a head-to-tail vertical travelling wave along
the body with twice the spatial frequency as the
horizontal wave. This had the effect that the
duration of bacs (t) was actively and system-
atically shortened (contact modulation; SM,
section 1.5). We tested the performance of
contact-modulated (CM) multilegged robots
over our rugose terrains and observed im-
proved locomotion robustness over terrain
rugosity (indicated by smaller error bars), al-
though with some reduction of the nominal ve-
locity vopen for locomotion on flat ground (Fig.
4E, dashed purple rectangle). Furthermore, with
sufficient spatial redundancy (N = 6) as well
as contact modulation, our multilegged robot
was capable of traversing diverse laboratory
(obstacles, inclines, and walls) environments
(Fig. 1B and movie S1) and field-like environ-
ments (granular media, pebbles, and rock piles)
with completely open-loop operations.
Discussion
One value of our framework lies in its cod-
ification of the benefits of redundancy, which
lead to locomotor robustness over environ-
mental contact errors without requiring sensing.
This contrasts with the prevailing paradigm
of contact-error prevention in the conven-
tional sensor-based closed-loop controls that
take advantage of visual, tactile, or joint-torque
information from the environment to change
the robot dynamics (5, 14). In this way, the
complexity of matter transport can be trans-
ferred from the real-time feedback-based control
(e.g., dealing with the flow of sensor informa-
tion) to preprogrammed gait design. Thus,
our framework could simplify matter transport
tasks such as search-and-rescue (34), extrater-
restrial exploration (35), or even microrobotics
(36), in which robot deployments are often
preferred but are challenging because of un-
predictable terradynamic interactions and
unreliable sensors.
However, sensory feedback can clearly be of
value when a robot becomes “stuck,” when terra-
dynamic interactions vary substantially (e.g.,
moving from low to high rugosity terrain) or
when large spatial redundancy is undesirable.
In such cases, the robot must understand its
state (through proprioception or exterocep-
tion) (13) and change dynamics accordingly.
An analogy from information theory, error-
detection coding, facilitates the augmentation
of our framework with sensory-based control: in
so-called error-detection coding, the presence
of a “reverse channel” can facilitate the retrans-
mission of signals, thereby improving signal
transmission accuracy (21, 29). The introduction
of feedback can increase the capacity of a noisy
channel and can therefore decrease the coding
complexity (37). We posit that a sensor-based
framework in locomotion for a fixed number
of bacs shares a mechanism similar to error-
detection coding. If performance improvement
with increased redundant bacs becomes mar-
ginal, then sensors (e.g., monitoring foot con-
tact) that detect bac contamination could
facilitate rapid environmental adaptation (e.g.,
whole-body gait adaptation or local leg place-
ment adaptation) and further improve loco-
motion performance. Thus, a combination of
redundancy-based and sensor-based mecha-
nisms can offer unique advantages over chal-
lenging terrains, a feature similar to hybrid
automatic repeat-request method (hybrid ARQ)
in signal transmission (38).
In addition to the importance of locomo-
tion in artificial locomotors, we posit that our
matter-transport framework can give insights
into aspects of neuromechanical and mor-
phological evolution (39) from a physics of
living systems perspective. That is, animals
ranging from those which generate propul-
sion through a single bac-pair (i.e., bipeds)
(40, 41), to those which use many bacs (i.e.,
myriapods) (42), are capable of traversing
complex natural terrains. The importance of
environmental awareness and whole-body
coordination is hypothesized to diminish as
the number of bacs (redundancy) increases
(42–44). Thus, in biological terrestrial loco-
motors, there appears to be a shift toward
either advanced neuromechanical control with
reduced body appendages, or redundant body
appendages with simplified neuromechanical
control. Integration of our framework with
advances in biological experimentation could
yield insights into the benefits and trade-offs
of diverse control architectures.
RE FERENCES AND NOTES
1.
2.
J. Aguilar et al., Science 361, 672–677 (2018).
J. Mindell, H. Rutter, S. Watkins, Encyclopaedia of Environmental
Health (Elsevier Masson, 2011), pp. 578–589.
3. H. Kreutzmann, Mod. Asian Stud. 25, 711–736 (1991).
4. R. M. Alexander, Principles of animal locomotion (Princeton
Univ. Press, 2003).
7. U. Saranli, M. Buehler, D. E. Koditschek, Int. J. Robot. Res. 20,
616–631 (2001).
8. M. G. Bekker, Off-the-road locomotion: Research and development
in terramechanics (Univ. of Michigan Press, 1960).
9. C. Li, T. Zhang, D. I. Goldman, Science 339, 1408–1412
(2013).
10. C. Li et al., Bioinspir. Biomim. 10, 046003 (2015).
11. W. Weber, E. Weber, Mechanik der menschlichen Gehwerkzeuge
(Verlag der Dietrichschen Buchhandlung, Göttingen, Germany,
1836).
12. J. Lee, J. Hwangbo, L. Wellhausen, V. Koltun, M. Hutter,
Sci. Robot. 5, eabc5986 (2020).
13. J. Borenstein, H. R. Everett, L. Feng, D. Wehe, J. Robot. Syst.
14, 231–249 (1997).
14. M. Hutter et al., 2016 IEEE/RSJ international conference
on intelligent robots and systems (IROS) (IEEE, 2016),
pp. 38–44.
15. E. R. Westervelt, J. W. Grizzle, C. Chevallereau, J. H. Choi,
B. Morris, Feedback control of dynamic bipedal robot
locomotion (CRC press, 2018).
16. Y. Ozkan-Aydin, D. I. Goldman, Science Robotics 6, eabf1628
(2021).
17. S. Aoi et al., Sci. Rep. 6, 30199 (2016).
18. E. Moore, D. Campbell, F. Grimminger, M. Buehler, Proceedings
2002 IEEE International Conference on Robotics and
Automation (Cat. No. 02CH37292), Washington, DC, 11 to
15 May, 2002 (IEEE, 2002), vol. 3, pp. 2222–2227.
19. C. Ordonez, J. Shill, A. Johnson, J. Clark, E. Collins, Proc SPIE
Int Soc Opt Eng 8741, 240–251 (2013).
20. X. A. Wu, T. M. Huh, R. Mukherjee, M. Cutkosky, IEEE Robot.
Autom. Lett. 1, 1125–1132 (2016).
21. C. E. Shannon, Bell Syst. Tech. J. 27, 379–423 (1948).
22. A. Feinstein, IEEE Trans. Inf. Theory 4, 2– 22 (1954).
23. R. J. Full, D. E. Koditschek, J. Exp. Biol. 202, 3325–3332
(1999).
24. H. C. Astley et al., Proc. Natl. Acad. Sci. U.S.A. 112, 6200–6205
(2015).
25. H. Forssberg, J. Neurophysiol. 42, 936–953 (1979).
26. P. E. Schiebel et al., Robophysical modeling of bilaterally
activated and soft limbless locomotors, Biomimetic and
Biohybrid Systems: 9th International Conference, Living
Machines, Freiburg, Germany, 28 to 30 July, 2020 (Springer
International Publishing, 2020), pp. 300–311.
27. B. Chong et al., Proc. Natl. Acad. Sci. U.S.A. 120, e2213698120
(2023).
28. D. Zhao, B. Bittner, G. Clifton, N. Gravish, S. Revzen, Proc. Natl.
Acad. Sci. U.S.A. 119, e2113222119 (2022).
29. J. R. Pierce, An Introduction to Information Theory: Symbols,
Signals and Noise (Courier Corporation, 2012).
30. B. Chong et al., Bioinspir. Biomim. 17, 046015 (2022).
31. Y. Ozkan-Aydin, B. Chong, E. Aydin, D. I. Goldman, 2020 3rd
IEEE International Conference on Soft Robotics (RoboSoft),
New Haven, CT, USA, 15 May to 15 July, 2020 (IEEE, 2020),
pp. 156–163.
32. A. Jacoff, A. Downs, A. Virts, E. Messina, Proceedings of the 8th
Workshop on Performance Metrics for Intelligent Systems,
Gaithersburg, MD, USA, 19 to 21 August, 2008 (NIST, 2008),
pp. 29–34.
33. B. Chong et al., Int. J. Robot. Res. 40, 1547–1562
(2021).
34. J. Whitman, N. Zevallos, M. Travers, H. Choset, 2018 IEEE
International Symposium on Safety, Security, and Rescue
rRobotics (SSRR), Philadelphia, PA, USA, 6 to 8 August, 2018
(IEEE, 2018), pp. 1–6.
35. N. D. Naclerio et al., Sci. Robot. 6, eabe2922 (2021).
36. J. J. Abbott, Z. Nagy, F. Beyeler, B. J. Nelson, IEEE Robot.
Autom. Mag. 14, 92–103 (2007).
37. S. Tatikonda, S. Mitter, IEEE Trans. Inf. Theory 55, 323–349
(2008).
38. C. Lott, O. Milenkovic, E. Soljanin, 2007 IEEE Information Theory
Workshop on Information Theory for Wireless Networks,
Bergen, Norway, 1 to 7 July, 2007 (IEEE, 2007), pp. 1–5.
39. B. Chong, T. Wang, E. Erickson, P. J. Bergmann, D. I. Goldman,
Proc. Natl. Acad. Sci. U.S.A. 119, e2118456119 (2022).
40. D. S. Marigold, Exerc. Sport Sci. Rev. 36, 145–151
(2008).
41. T. Krasovsky, A. Lamontagne, A. G. Feldman, M. F. Levin,
Gait Posture 39, 378–385 (2014).
42. K. Diaz, E. Erickson, B. Chong, D. Soto, D. I. Goldman,
5. P. Holmes, R. J. Full, D. Koditschek, J. Guckenheimer,
J. Exp. Biol. 226, jeb244688 (2023).
SIAM Rev. 48, 207–304 (2006).
6. M. Raibert, K. Blankespoor, G. Nelson, R. Playter, IFAC
Proceedings Volumes 41, 10822 (2008).
43. S. Sponberg, R. J. Full, J. Exp. Biol. 211, 433–446 (2008).
44. B. Bläsing, H. Cruse, J. Comp. Physiol. A Neuroethol. Sens.
Neural Behav. Physiol. 190, 173–183 (2004).
Chong et al., Science 380, 509–515 (2023)
5 May 2023
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
45. D. Soto, B. Chong, Code for rugose terrain generation, Zenodo
(2021); https://doi.org/10.5281/zenodo.7121219
ACKN OW LEDG MEN TS
We thank T. Murphey, P. Umbanhowar, G. A. Sartoretti, D. Luo,
T. Berrueta, E. Flores, and X. Sun for helpful discussion. We thank
K. Diaz for proofreading. Funding: The authors received funding
from NSF-Simons Southeast Center for Mathematics and Biology
(Simons Foundation SFARI 594594), Georgia Research Alliance
(GRA.VL22.B12), Army Research Office (ARO) MURI program, Army
Research Office grant W911NF-11-1-0514, and a Dunn Family
Professorship. Author contributions: B.C. designed the study and
model and performed numerical simulations. J.H. collected the raw
data for the robophysical experiments. J.H., T.W., D.S., and B.C. built
the experimental platform. D.I., G.B., and B.C. performed the
mathematical proof. D.I.G. oversaw the study. All authors contributed
to the preparation of the manuscript and were involved in the
interpretation of results. Competing interests: Some of the
subject matter herein may be implicated in one or more pending
patent applications such as PCT Patent Application No. PCT/
US2022/043362, entitled “DEVICES AND SYSTEMS FOR
LOCOMOTING DIVERSE TERRAIN AND METHODS OF USE”, which
claims the benefit of priority to US Provisional App. No. 63/
243,435 filed 09/13/2021, and US Provisional App. No. 63/
318,868 filed 03/11/2022. The authors declare that they have no
other competing interests. Data and materials availability: All
data that support the claims in this manuscript are available
on the Zenodo repository (45). License information: Copyright ©
2023 the authors, some rights reserved; exclusive licensee
American Association for the Advancement of Science. No claim
to original US government works. https://www.science.org/about/
science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade4985
Materials and Methods
Supplementary Text
Figs. S1 to S6
Movie S1
References (46, 47)
Submitted 29 September 2022; resubmitted 7 February 2023
Accepted 3 April 2023
10.1126/science.ade4985
Chong et al., Science 380, 509–515 (2023)
5 May 2023
7 of 7
|
10.1126_science.ade5156
|
RES EARCH
EVOLUTION
Evolutionary transitions from camouflage to
aposematism: Hidden signals play a pivotal role
Karl Loeffler-Henry1†, Changku Kang2,3*†, Thomas N. Sherratt1
The initial evolution of warning signals in unprofitable prey, termed aposematism, is often seen as
a paradox because any new conspicuous mutant would be easier to detect than its cryptic
conspecifics and not readily recognized by naïve predators as defended. One possibility is that
permanent aposematism first evolved through species using hidden warning signals, which are
only exposed to would-be predators on encounter. Here, we present a large-scale analysis of
evolutionary transitions in amphibian antipredation coloration and demonstrate that the
evolutionary transition from camouflage to aposematism is rarely direct but tends to involve an
intermediary stage, namely cryptic species that facultatively reveal conspicuous coloration.
Accounting for this intermediate step can resolve the paradox and thereby advance our
understanding of the evolution of aposematism.
S election to avoid being killed by pred-
ators has contributed to the diversity of
animal color patterns (1). These color
adaptations include crypsis and disrup-
tive coloration to avoid being detected
and/or recognized (2), conspicuous warning
signals in defended species to indicate un-
profitability to would-be predators [an associa-
tion known as aposematism (3)], and mimetic
signals that share or exploit the aposematic
signals of other organisms (4). Although our
understanding of the genetics, development,
perception, and function of these color signals
has substantially progressed in recent years
(5), we still know little about the macroevolu-
tionary patterns of color pattern evolution. Spe-
cifically, large-scale macroevolutionary studies
on animal color defense are surprisingly scarce,
and even these have tended to consider sim-
ple binary classifications of species color, no-
tably whether they are cryptic or conspicuous
(6–8). Although this binary classification cap-
tures some well-known antipredator strategies
of animals (crypsis and potential aposematism
or mimicry), this may not be enough to ex-
plain how diverse color defense strategies have
evolved in nature. For example, some species
are cryptic at rest but have bright color signals
hidden on body surfaces that are only exposed
when signaling to conspecifics, fleeing, or as
part of a defensive posture (9–14). These flex-
ible signaling strategies could represent inter-
mediate stages and therefore might play a
pivotal role in evolutionary processes gener-
ating different antipredator defenses (15).
Amphibians are an excellent group to ex-
plore the evolutionary transitions among dif-
1Department of Biology, Carleton University, Ottawa, Ontario
K1S 5B6, Canada. 2Department of Agricultural Biotechnology,
Seoul National University, Seoul 08826, South Korea.
3Research Institute of Agriculture and Life Sciences, Seoul
National University, Seoul 08826, South Korea.
*Corresponding author. Email: changkukang@snu.ac.kr
†These authors contributed equally to this work.
ferent antipredator strategies. Their phylogeny
is available in nearly all extant taxa (16), and
their color patterns have been strongly shaped
by predators through natural selection (17). A
previous study examining macroevolutionary
patterns of amphibian antipredator adapta-
tions found that rates of speciation, extinc-
tion, and transition vary among species with
different defensive traits (6). Although this
large-scale study revealed important evolu-
tionary pathways from camouflage to aposem-
atism, the inference was based largely on a
simple two-color state classification scheme
(cryptic and conspicuous) of each species’ dor-
sal coloration. However, many amphibian spe-
cies show a more complex set of color patterns,
such as having a cryptic dorsum yet conspic-
uous patches on normally hidden body parts
(18–21). These hidden color signals are tax-
onomically widespread in the animal kingdom
yet have seldom been considered in macro-
evolutionary studies (13, 14).
The hidden color signals of amphibians
tend to occur in one of two different forms:
conspicuous color present on (i) the whole
venter (lower surface), such as those in the
genus Bombina, or (ii) part of the concealed
body surface, such as ventral shanks or hind-
limbs commonly found in the family Hylidae.
These hidden signals are often exposed through
behavioral displays [e.g., via an unken reflex
toward approaching predators, or foot flag-
ging for intraspecific signaling (14, 22–24)].
In addition, amphibians may sometimes use
flashing signals during escape: These colors
would be visible only when the prey is mobile
and may mislead a predator into assuming
that the prey’s flash color is its resting color
and in so doing hinder subsequent search (25).
Hidden conspicuous color signals may have
evolved from typical aposematic signals to off-
set the costs of conspicuousness while sta-
tionary. Alternatively, they may serve as an
intermediate state from camouflage to apo-
sematism because this strategy can gain the
advantages of both by only signaling when
discovered (15, 26, 27).
Here, using discrete character evolution
models, we investigate the role of hidden con-
spicuous color signals during the evolutionary
transitions between camouflage and aposema-
tism. Specifically, we examine the transitions
between different antipredator strategies based
on a five-category color classification scheme,
accounting for crypsis, conspicuousness, two
different types of hidden signals [PV; partially
conspicuous venter: cryptic dorsum with con-
spicuous color present as small patches on
normally hidden body parts and FV; fully
conspicuous venter: cryptic dorsum with con-
spicuous colors that fully cover the venter
(28)], and polymorphism (Poly; defined here
as a species having both cryptic and conspic-
uous forms regardless of whether they are
regional variants or coexist in the same pop-
ulation; see Fig. 1 for example species and
materials and methods for details). We iden-
tified two classes of hidden signals because
(i) the classes are distinguishable morpholog-
ically and (ii) their putative functions differ in
that FV coloration is likely to solely function as
an aposematic signal to attacking or approach-
ing predators, whereas PV coloration may also
serve as a flashing signal or territorial display
(14). We analyzed two datasets: color (1106
species with color information available) and
color+chem (315 species with both color and
chemical defense information available).
Results and discussion
Main evolutionary transitions in the color model
We tested nine different models: three models
that allowed transitions between almost all
states (with two exceptions, see below) at
equal or differing rates [All rates different
(ARD), Symmetric (SYM), Equal rates (ER)
models], and six models that restricted certain
transitional pathways (Intermediate, FV inter-
mediate, Stepwise, PV/FV secondary, FV cost-
offset PV secondary, Cost-offset) (table S1 and
fig. 1A). Transitions between the FV and Poly
states and between the PV and Poly states
were not allowed (see materials and methods
for details). An intermediate model that did
not allow the direct transition of species be-
tween the cryptic and conspicuous states was
the best-supported (lowest Akaike informa-
tion criterion) model (Fig. 1A and tables S1 and
S2 for descriptions and full results), while the
estimated transition rates were qualitatively
equivalent among the three best-supported
models [the Intermediate, ARD, and a modi-
fied Intermediate model in which the PV state
cannot evolve directly to the conspicuous state
but only through the FV state (FV interme-
diate); fig. S1]. The parameter estimates from
fitting the Intermediate model revealed sev-
eral major patterns of evolutionary pathways
Loeffler-Henry et al., Science 379, 1136–1140 (2023)
17 March 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
(Fig. 1B). First, although species in the cryptic
state can evolve directly toward all other states
with the exception of the conspicuous state
(which is precluded in the best-supported mod-
el), the cryptic state is stable (i.e., the transition
rates from all other states to the cryptic state
are at least three times higher than the rates
away from it) and the most likely basal ances-
tral state (support probability = 69%; Fig. 2).
Second, the PV state is most strongly asso-
ciated with the cryptic state: Species with
the PV color mainly evolve from cryptic color-
ation and tend to transit back to it. However,
although the transition rate is low, the PV
state is also the most likely state leading to
the FV state, which could facilitate the sub-
sequent transition toward the conspicuous
state. Third, species in the conspicuous state
largely evolve from the FV or polymorphic
state, and other routes are less likely. How-
ever, because species in the polymorphic state
evolve almost exclusively from the conspicu-
ous state, the major pathway to the initial
evolution of conspicuous coloration is likely
through the FV state. Fourth, the polymorphic
state appears unstable and transits rapidly to
the cryptic or conspicuous state.
Main evolutionary transitions in the
color+chem model
In the color+chem model, the evolutionary
pathways among the states are substantially
more complex than the color-only models (Fig.
1C). As with the color model, the cryptic state
was found to be the most likely basal ancestral
color state in the subset of species considered
in the color+chem model (support probabil-
ity = 60%; Fig. 3). Although there was slightly
stronger support for the hypothesis that the
basal defensive state of frogs and salaman-
ders involved chemical defense, the data were
more equivocal (Fig. 3; 58% support proba-
bility for chemical defense versus 42% for no
chemical defense). Chemical defense frequent-
ly coincides with conspicuous coloration: In-
deed, the majority of species classified as either
conspicuous, FV color, or polymorphic have
chemical defense (more than 90% in all three
groups; table S3). However, chemical defense
also occurs in less-conspicuous species as well
(65% in species with PV color, 51% in the cryp-
tic species; table S3). This is not surprising, as
chemical defense is known to provide survival
advantages to both conspicuous and cryptic
species through aposematic signaling and taste
rejection, respectively (29–31). On average, the
transition rates toward the acquisition of de-
fense are higher than the transition rates away
from this defensive state, but the only color
state that is more likely to lose than acquire
chemical defense is crypsis (Fig. 1C). This is
consistent with previous findings that the ac-
quisition of alkaloid sequestration is favored
over losing it in poison frogs (32). One expla-
Fig. 1. Diagrammatical representations of the models that we compared and the parameter estimates
of the best-supported models. No arrows in the fitted transition models (A) indicate that the transition rates
were constrained to be zero. See table S1 for the descriptions of the models. Panels (B) and (C) show the
estimated transition rates among different states from the best-supported models using color dataset [(B);
species with color information available; N = 1106] and color+chem dataset [(C); species with both color and
chemical defense information available; N = 315 species]. Species that were classified as uncertain color (N = 47)
were excluded. Arrow thickness reflects the estimated transition rates, which are also given by the values
listed next to each arrow (those transitions that are possible but not depicted have estimated rates less than
0.00001). The radius of circles is proportional to the log-transformed number of species in each state.
Gray arrows indicate that the strength of evidence is weak because the estimated transition rates were
inconsistent between different functions (“fitMk” and “corHMM”) (28), most likely because they were
estimated from very few changes in the tree or they were estimated from only one extant species
(undefended Poly). Cry: cryptic; PV (partially conspicuous venter): cryptic dorsum with conspicuous color
present as small patches on normally hidden body parts; FV (fully conspicuous venter): cryptic dorsum
with conspicuous colors fully covered on the venter; Con: conspicuous; Poly: a species showing multiple
distinct morphs that include both cryptic and conspicuous forms. ARD: All rates different model; SYM:
symmetric model; ER: equal rates model (see table S1 for details). The photos show sample species from
each color category that we used. From left to right, Theloderma corticale (Dan Rosenberg), Hyla andersonii
(Troy Hibbitts), Paramesotriton hongkongensis (Dan Rosenberg), Dendrobates tinctorius (Michael Gäbler).
nation for the loss of chemical defense in cryp-
tic amphibians is that they experience less risk
of detection by predators and therefore less
selective pressure for the maintenance of
postdetection defense (33). Considering that
chemical defense may pose a cost to its bearer
(34, 35), this may make deterrent toxins less
favored in some cryptic lineages. Another non–
mutually exclusive reason for the differential
in acquisition or loss of chemical defenses in
the different types of signaler may be that
species with any form of conspicuous sig-
nals can potentially benefit from “go slow”
signaling that makes predators cautiously
sample the prey to determine the presence
of chemical defense (36). By contrast, preda-
tors may sample cryptic prey without caution,
possibly making selection for chemical de-
fense to reduce injury after capture weaker
in cryptic species.
After accounting for chemical defense, the
evolutionary transition from crypsis to apo-
sematism (chemically defended conspicuous
state) is not simple but is instead composed
Loeffler-Henry et al., Science 379, 1136–1140 (2023)
17 March 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
of multiple pathways that involve intermediate
states (Fig. 1C; see fig. S2 for the visual com-
parisons of how accounting for chemical de-
fenses modifies the transition rates among the
color states from the color model results).
Whereas the transition from the undefended
conspicuous state to the defended conspicu-
ous state is high, the transition rates toward
the undefended conspicuous state from any
other states are either zero or very low. This is
reasonable because any mutations that make
an individual conspicuous without having
chemical defense would be detrimental and
as such would be selected against unless other
defensive strategies, such as Batesian mimicry,
are involved (37, 38). Thus, a more stable route
to aposematism is via the chemically defended
FV state. There are multiple pathways to
evolve the chemically defended FV state, but
the main routes appear through the defended
PV state or undefended FV state, which also
mainly evolve from the undefended PV state.
These observations collectively suggest that at
least the FV state (but often involving the PV
state) is likely to be involved in transitions
from camouflage to aposematism. Once apo-
sematism has evolved, it goes back to neither
the PV nor FV state but instead either evolves
back to the cryptic state directly or becomes
cryptic/conspicuous-mixed polymorphic.
Hidden signals and their implications for
amphibian color evolution
We hypothesized that the PV signals poten-
tially function as a secondary defense (flash
displays or postdetection warning) or are used
for intraspecific signaling in normally cryptic
species (14, 25); thus, the PV signals are mainly
associated with cryptic coloration. Our results
of both color and color+chem models support
this view in that the PV state is primarily as-
sociated with the cryptic state with a stronger
tendency to go back to the cryptic state (Fig. 1,
B and C). Also, the PV state is the most likely
state that can lead to the evolution of the FV
state, which is a major precursory state toward
conspicuous coloration. The presence of PV
signals implies that a species has managed to
express bright (e.g., carotenoid or pteridine
pigments based) colors potentially acquired
through diet and/or manufactured de novo
and presumably evolved for antipredator or
Fig. 2. Ancestral state estimation of each color state (N = 1106 species) in frogs and salamanders.
Pie charts at each node show the probabilities of ancestral states. The ancestral state of frogs and
salamanders is likely to be cryptic coloration. The hidden color signals (PV and FV) are widespread and have
evolved multiple times in different lineages. PV: cryptic dorsum with conspicuous color present as small
patches on normally hidden body parts; FV: cryptic dorsum with conspicuous colors fully covered on the
venter. See table S11 for photo credits.
conspecific signaling (14, 39, 40). This could
further facilitate the expression of conspicuous
colors on other parts of the body, resulting in
the evolution of the FV states.
In both color and color+chem models, spe-
cies having FV color evolve from either the
cryptic or PV color, but not from the conspic-
uous color (Fig. 1, B and C). Thus, the con-
jecture that the FV color evolves from the
conspicuous color to offset the cost of being
continuously conspicuous is not supported (15).
Instead, the FV state is the most likely inter-
mediate stage that is required for the transi-
tion from crypsis to aposematism. About 91%
of species with FV color have chemical defense
(table S3), suggesting that their ventral warn-
ing coloration is likely an honest signal of their
defense, rather than a bluff. Theoretically, the
FV state can have a selective advantage over
the conspicuous state when a species has no
chemical defense: Having a conspicuous dor-
sum without defense should be highly detri-
mental to individuals, leaving less opportunity
to evolve chemical defense subsequently. How-
ever, because the FV strategy does not involve
the loss of crypsis, this strategy may be able to
persist until the evolution of chemical defense
follows. Indeed, the results of the color+chem
model suggest that the nonchemically de-
fended conspicuous state rarely evolves from
any other nonchemically defended states, but
the FV state could evolve from the PV state in
the absence of chemical defense (Fig. 1C).
Transitional patterns of the cryptic/
conspicuous-mixed polymorphic state
Only 2.3% of all species in our dataset were
considered to exhibit both cryptic and con-
spicuous morphs, i.e., were classified as poly-
morphic. Despite its relative rarity, our results
suggest that this cryptic/conspicuous-mixed
polymorphism has evolved multiple times in
different lineages independently (Fig. 2). Most
of these polymorphic species have chemical
defense (10 out of 11 species) and have evolved
mainly from the conspicuous states (Fig. 1, B
and C). Both the color and color+chem models
suggest that once the cryptic/conspicuous-
mixed polymorphic state is reached, it tends to
rapidly evolve toward either the cryptic or con-
spicuous state with a stronger tendency toward
the cryptic state (Fig. 1, B and C). The low
number of species in the mixed-polymorphic
state may reflect this evolutionary instability.
Conclusion
Our study highlights the importance of hidden
color signals for the evolutionary processes
that generate diverse antipredator colora-
tion in amphibians. Our results suggest that
(i) species with hidden color signals, especially
those with conspicuous colors that cover the
whole venter, represent a key stage in the evo-
lution of aposematic species from cryptic
Loeffler-Henry et al., Science 379, 1136–1140 (2023)
17 March 2023
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Ancestral state estima-
tion of each combination
of color and chemical defense
in frogs and salamanders
(N = 315 species). Pie charts
at each node show the proba-
bilities of ancestral states. The
ancestral state of frogs and
salamanders is likely to be
cryptic coloration, but the
evidence of whether the basal
state was chemically defended
or not was equivocal. Most
states have evolved multiple
times in different lineages.
Transitions from cryptic to
aposematic (chemically
defended conspicuous) states
have usually occurred via
intermediate states.
species; (ii) cryptic/conspicuous-mixed poly-
morphism plays a pivotal role in transitions
from aposematism back to crypsis; and (iii) the
transition rates toward acquisition of chem-
ical defense are higher than those to its loss
in most color states, with the exception of
the cryptic state. A number of complementary
models confirmed the robustness of these
conclusions (see materials and methods for
details and figs. S3 to S9 and tables S4 to S10
for the results).
Biologists have long wondered how rare
conspicuous mutants of a cryptic defended
species can spread in a population when they
have a higher predation risk before predators
learn to avoid them (41, 42). The fact that apo-
sematic species appear to seldom derive directly
from cryptic species confirms the long-standing
intuition that the evolution of aposematism
from crypsis is challenging. Here we show
macroevolutionary evidence for an impor-
tant yet unrecognized solution in which the
problem is effectively side-stepped: Rather
than evolve from cryptic species, aposematic
mutants appear to derive largely from species
with hidden signals. One intuitive way that
this might work is if would-be predators that
are already exposed to a species with hidden
warning signals continue to treat permanently
aposematic mutants of this species with cau-
tion. The initial evolution of hidden color sig-
nals might be facilitated by predatory pressure
that promotes the evolution of secondary
defenses such as flash or warning displays
(15, 24, 25), or via sexual selection that results
in the expression of conspicuous signals in
body parts that are only visible during behav-
ioral displays (14). These complex transitional
routes from crypsis to aposematism would not
be revealed if traditional dichotomous clas-
sifications of animal antipredator coloration
(either cryptic or conspicuous) are applied (see
fig. S10 for the results when the binary clas-
sification was used). Thus, macroevolutionary
studies on animal coloration should take into
account these underappreciated hidden sig-
nals, which are both common and widespread
across the animal kingdom (13, 43, 44), to ad-
vance our understanding of the evolution of
antipredator defenses. Indeed, many animal
taxa such as snakes, fishes, and a variety of
arthropods (see fig. S12 for example groups)
include species that are cryptic, are aposematic,
and have hidden conspicuous signals. We
therefore encourage follow-up studies in other
taxa to evaluate the generality of the stepping-
stone hypothesis as a route to aposematism.
RE FERENCES AND NOTES
1. H. B. Cott, Adaptive Coloration in Animals (Methuen, 1940).
2. M. Stevens, S. Merilaita, Animal Camouflage: Mechanisms and
Function (Cambridge Univ. Press, 2011).
3. E. B. Poulton, The Colours of Animals: Their Meaning and Use,
Especially Considered in the Case of Insects (D. Appleton,
1890).
4. G. D. Ruxton, A. L. William, T. N. Sherratt, M. P. Speed,
Avoiding Attack: The Evolutionary Ecology of Crypsis, Warning
Signals, and Mimicry (Oxford Univ. Press, ed. 2, 2018).
I. C. Cuthill et al., Science 357, eaan0221 (2017).
5.
6. K. Arbuckle, M. P. Speed, Proc. Natl. Acad. Sci. U.S.A. 112,
13597–13602 (2015).
7. B. B. Blaimer, J. R. Mawdsley, S. G. Brady, Evolution 72,
1874–1889 (2018).
8. K. Przeczek, C. Mueller, S. M. Vamosi, Integr. Zool. 3, 149–156
(2008).
9. M. Vidal-García, J. C. O’Hanlon, G. J. Svenson, K. D. L. Umbers,
Proc. Biol. Sci. 287, 20201016 (2020).
10. M. Edmunds, Zool. J. Linn. Soc. 51, 1–32 (1972).
11. T. Caro, T. N. Sherratt, M. Stevens, Evol. Ecol. 30, 797–809
(2016).
12. W. Hödl, A. Amézquita, in Anuran Communication, M. J. Ryan,
Ed. (Smithsonian Institution Press, 2001), pp. 121–141.
13. K. Loeffler-Henry, C. Kang, T. N. Sherratt, Am. Nat. 194, 28–37
14.
(2019).
I. Starnberger, D. Preininger, W. Hödl, J. Comp. Physiol. A
Neuroethol. Sens. Neural Behav. Physiol. 200, 777–787 (2014).
15. K. D. L. Umbers et al., Biol. Lett. 13, 20160936 (2017).
16. W. Jetz, R. A. Pyron, Nat. Ecol. Evol. 2, 850–858 (2018).
17. B. Rojas, Biol. Rev. Camb. Philos. Soc. 92, 1059–1080 (2017).
18. J. Bajger, Herpetologica 36, 133–137 (1980).
19. A. Rudh, A. Qvarnström, Semin. Cell Dev. Biol. 24, 553–561
(2013).
20. M. Martins, J. Herpetol. 23, 305–307 (1989).
21. L. F. Toledo, C. F. B. Haddad, Int. J. Zool. 2009, 910892 (2009).
22. L. F. Toledo, I. Sazima, C. F. B. Haddad, Ethol. Ecol. Evol. 23,
1–25 (2011).
23. D. P. Ferraro, M. O. Pereyra, P. E. Topa, J. Faivovich, Zool. J.
Linn. Soc. 193, 388–412 (2020).
24. R. B. Ferreira et al., Behav. Ecol. Sociobiol. 73, 69 (2019).
25. K. Loeffler-Henry, C. Kang, Y. Yip, T. Caro, T. N. Sherratt,
Behav. Ecol. 29, 528–533 (2018).
26. T. Guilford, Am. Nat. 131, S7–S21 (1988).
27. M. P. Speed, G. D. Ruxton, Proc. Biol. Sci. 272, 431–438(2005).
28. Materials and methods are available as supplementary materials.
29. J. Mappes, N. Marples, J. A. Endler, Trends Ecol. Evol. 20,
598–603 (2005).
30. Y. Yoshimura, E. Kasuya, PLOS ONE 8, e81280 (2013).
31. J. Skelhorn, C. Rowe, Behav. Ecol. Sociobiol. 60, 550–555
(2006).
32. J. D. Carvajal-Castro, F. Vargas-Salinas, S. Casas-Cardona,
B. Rojas, J. C. Santos, Sci. Rep. 11, 19047 (2021).
33. L. Wang, G. D. Ruxton, S. J. Cornell, M. P. Speed, M. Broom,
J. Theor. Biol. 473, 9–19 (2019).
34. M. D. Bowers, in Insect Chemical Ecology: An Evolutionary
Approach, B. D. Roitberg, M. B. Isman, Eds. (Chapman and Hall
New York, 1992), pp. 216–244.
35. R. A. Blennerhassett, K. Bell-Anderson, R. Shine, G. P. Brown,
Proc. Biol. Sci. 286, 20190867 (2019).
36. T. Guilford, J. Theor. Biol. 170, 311–316 (1994).
37. C. R. Darst, M. E. Cummings, Nature 440, 208–211 (2006).
38. L. Lindström, R. V. Alatalo, A. Lyytinen, J. Mappes, Proc. Natl.
Acad. Sci. U.S.A. 98, 9181–9184 (2001).
39. K. D. L. Umbers, A. J. Silla, J. A. Bailey, A. K. Shaw,
P. G. Byrne, Biol. J. Linn. Soc. Lond. 119, 436–444 (2016).
40. T. Goodwin, The Biochemistry of the Carotenoids, Volume I,
Plants (Springer Science & Business Media, 2012).
41. R. A. Fisher, The Genetical Theory of Natural Selection:
A Complete Variorum Edition (Oxford Univ. Press, 1930).
42. R. V. Alatalo, J. Mappes, Nature 382, 708–710 (1996).
Loeffler-Henry et al., Science 379, 1136–1140 (2023)
17 March 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
43. T. Caro, H. Raees, T. Stankowich, Behav. Ecol. Sociobiol. 74, 44
(2020).
44. K. Loeffler-Henry, C. Kang, T. N. Sherratt, Proc. Biol. Sci. 288,
20210866 (2021).
45. K. Loeffler-Henry, C. Kang, T. N. Sherratt, Amphibian color
paper data and analysis code. figshare. Dataset.
https://doi.org/10.6084/m9.figshare.19890313 (2023).
ACKN OW LEDG MEN TS
We are grateful to three anonymous reviewers who provided a
plethora of insightful feedback during the peer review process for
this manuscript. We thank E. Kerr and J. Butler from the Canadian
Herpetological Society for assistance in classifying each species.
Funding: C.K. is supported by the National Research Foundation of
Korea (grant no. NRF-2019R1C1C1002466) and the New Faculty
Startup From Seoul National University. T.N.S. and K.L-H. are
supported by the Natural Sciences and Engineering Research
Council of Canada (NSERC, Discovery Grant to T.N.S.). Author
contributions: C.K. conceived the study. K.L-H. conducted the
image collecting and classifications. C.K. conducted the
phylogenetic analysis with input from K.L-H. and T.N.S. All authors
contributed to discussions and interpretations of the data. All
authors contributed to formulating the initial draft of the
manuscript and subsequent revisioning. Competing interests: The
authors declare no competing interests. Data and material
availability: All data and analysis codes have been deposited at
Figshare (45). License information: Copyright © 2023 the
authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.sciencemag.org/about/
science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade5156
Materials and Methods
Figs. S1 to S12
Tables S1 to S12
References (46–70)
Submitted 23 August 2022; accepted 21 February 2023
10.1126/science.ade5156
Loeffler-Henry et al., Science 379, 1136–1140 (2023)
17 March 2023
5 of 5
|
10.1126_science.ade5050
|
RES EARCH
QUANTUM INFORMATION
An atomic-scale multi-qubit platform
Yu Wang1,2†, Yi Chen1,2,3,4†, Hong T. Bui1,5†, Christoph Wolf1,2, Masahiro Haze1,6,
Cristina Mier1,7, Jinkyung Kim1,5, Deung-Jang Choi1,7,8,9, Christopher P. Lutz10,
Yujeong Bae1,5*, Soo-hyon Phark1,2*, Andreas J. Heinrich1,5*
Individual electron spins in solids are promising candidates for quantum science and technology, where
bottom-up assembly of a quantum device with atomically precise couplings has long been envisioned.
Here, we realized atom-by-atom construction, coherent operations, and readout of coupled electron-spin
qubits using a scanning tunneling microscope. To enable the coherent control of “remote” qubits that
are outside of the tunnel junction, we complemented each electron spin with a local magnetic field
gradient from a nearby single-atom magnet. Readout was achieved by using a sensor qubit in the tunnel
junction and implementing pulsed double electron spin resonance. Fast single-, two-, and three-qubit
operations were thereby demonstrated in an all-electrical fashion. Our angstrom-scale qubit platform
may enable quantum functionalities using electron spin arrays built atom by atom on a surface.
C onstructing and coherently controlling
nanoscale qubit systems lies at the very
heart of quantum-coherent nanoscience
(1). This length scale requires the use of
fundamental quantum properties of atoms,
such as the spin of electrons, which naturally
occurs in many solid-state environments and
allows high-fidelity operations and readout by
electromagnetic means (2). Despite decades of
effort, however, it remains a formidable task to
realize an atomic-scale quantum architecture
in which multiple electron spin qubits can
be precisely assembled, controllably coupled,
and coherently operated. Electron spin qubits
created in dopants in semiconductors and color
centers in insulators, for example, can be well
controlled individually (3, 4) but are difficult
to couple together into a circuit. An attractive
alternative approach is to use atomic-level fabri-
cation with a scanning tunneling microscope
(STM), in which atom manipulation (5) or se-
lective desorption (6) can lead to designed
quantum spin architectures (7, 8). As a first step
toward in situ operation of atomic quantum
devices, a radiofrequency (RF) voltage was
used to coherently control a single electron
spin in the STM tunnel junction (9, 10) in a
so-called electron spin resonance (ESR)–STM
setup (11–14). However, achieving the potential
of this platform requires multiple addressable
and detectable qubits that lie outside of the
subnanometer tunnel junction region.
Building a multi-qubit platform atom by atom
Figure 1A illustrates our strategy to construct
such an atomic-scale multi-qubit platform. A
“sensor” qubit consisting of a spin-1/2 hydro-
genated Ti atom on a bilayer MgO film (14, 15)
was positioned in the tunnel junction, driven,
and detected by a magnetic tip (16, 17). Outside
of the tunnel junction, addressable “remote”
qubits were created by positioning a spin-1/2
Ti atom (14, 15) ~0.6 nm away from a single
Fe atom (18) using atom manipulation [see
section 1 of (19)]. An Fe atom can be seen as a
single-atom magnet in this work because its
spin relaxation time far exceeds the time scale
of qubit operations and its resonance frequen-
cies are far off the applied RF range (11, 18, 20)
[see section 2 of (19)]. The main purpose of Fe
atoms here is to substitute for the tip in pro-
viding a static local magnetic field for remote
Ti spins (those not directly under the tip apex)
to allow their electrical driving. Under a tip-
supplied RF electric field, Ti spins are believed
Fig. 1. Bottom-up construction of multiple coupled electron spin qubits.
(A) Schematic: A sensor spin qubit (Ti, blue) is placed under the apex of a
spin-polarized STM tip for readout. Remote qubits are constructed at
precise separations to the sensor qubit by atom manipulation. Each remote
qubit is composed of a spin-1/2 Ti atom (red) and a single-atom magnet (Fe)
(green), where Fe’s magnetic field gradient, in combination with the RF
electric field between the tip and the sample, coherently drives remote qubits.
(B and C) Constant-current STM images showing atom-by-atom construction
of a multi-qubit structure composed of two (B) and three (C) qubits (image
size: 5.0 × 5.0 nm). Inset: atomic registry of the structure in (C). Structure in (B)
has the same configuration but without remote qubit 2. (D and E) Continuous-wave
ESR spectra measured with the tip positioned on the sensor qubit in the
two-qubit (D) and three-qubit (E) structure. The quantum states of the sensor
qubit are labeled by the first kets. Each ESR transition of the sensor qubit
distinguishes a quantum state of the remote qubits (the second kets).
The measured spin-polarized tunnel-current signal DI is the difference between
averaged signals in lock-in A and B subcycles (fig. S10), which reflects the
change of the sensor spin polarization due to applied coherent control pulses
[see section 4 of (19)]. Imaging conditions in (B) and (C): sample bias voltage
VDC = 100 mV, time-averaged tunnel current IDC = 10 pA. ESR conditions
in (D) and (E): VDC = 50 mV, IDC = 20 pA, zero-to-peak RF voltage VRF = 30 mV.
The sample was kept at 0.4 K during measurements.
Wang et al., Science 382, 87–92 (2023)
6 October 2023
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
to undergo relative oscillations inside Fe’s mag-
netic field gradient, thus experiencing an effec-
tive resonant driving magnetic field (21, 22), in
analogy to micromagnet driving in quantum
dots (23) [more details of the driving mecha-
nism can be found in section 2 of (19) and in
(24)]. By contrast, ESR of the sensor qubit, as
well as previous ESR-STM spectroscopy (12–14),
relies on the tip's magnetic field gradient
(21, 22). Initialization of the qubits was per-
formed thermally by cooling the sample to
0.4 K and applying an external magnetic field,
producing a thermal spin state having a pre-
dominant population in the spin ground state
(~90% under a 0.7 T field).
A representative structure composed of two
remote qubits and one sensor qubit was con-
structed atom by atom using atom manipu-
lation [see section 1 of (19)], as shown in Fig. 1,
B and C. The qubit-qubit couplings, including
dipolar and exchange spin-spin couplings, are
sensitive to their atomic separations down to
the angstrom scale (14, 15) (figs. S1 and S2),
consistent with results in other material sys-
tems (25). The STM-based atomically precise
construction scheme thus allowed us to engi-
neer the resonant frequencies and couplings
among all of the spins, an essential step for
addressing and detecting multiple qubits indi-
vidually [see section 3 of (19) and fig. S2].
Detection of the remote qubits was achieved
through ESR spectroscopy of the sensor qubit,
the ESR transition frequency of which depends
on the quantum states of other qubits (Fig. 1, D
and E). The ESR frequencies of the qubits were
designed to be sufficiently separated compared
with the qubit-qubit couplings so that the
multi-qubit states could be well described by
Zeeman product states, for which we used
the first (second) brackets to denote the quan-
tum states of the sensor (remote) qubit. For
example, in a structure composed of a sensor
qubit and a remote qubit (Fig. 1D), the two ESR
transitions of the sensor qubit at frequencies f1
(corresponding to transition |0i|0i ↔ |1i|0i)
and f2 (|0i|1i ↔ |1i|1i) detect the populations
of |0i and |1i states of the remote qubit,
respectively. The difference between f1 and
f2 is determined by the Ti-Ti coupling [see
1Center for Quantum Nanoscience, Institute for Basic
Science (IBS), Seoul 03760, Korea. 2Ewha Womans
University, Seoul 03760, Korea. 3International Center for
Quantum Materials, School of Physics, Peking University,
Beijing 100871, China. 4Collaborative Innovation Center of
Quantum Matter, Beijing 100871, China. 5Department of
Physics, Ewha Womans University, Seoul 03760, Korea.
6The Institute for Solid State Physics, University of Tokyo,
Kashiwa 277-8581, Japan. 7Centro de Física de Materiales
CFM/MPC (CSIC-UPV/EHU), 20018 Donostia-San Sebastián,
Spain. 8Donostia International Physics Center (DIPC), 20018
Donostia-San Sebastián, Spain. 9Ikerbasque, Basque
Foundation for Science, 48013 Bilbao, Spain. 10IBM Research
Division, Almaden Research Center, San Jose, CA 95120, USA.
*Corresponding author. Email: bae.yujeong@qns.science (Y.B.);
phark@qns.science (S.P.); heinrich.andreas@qns.science (A.J.H.)
†These authors contributed equally to this work.
section 3 of (19)]. Multiple remote qubits
can be simultaneously sensed in a similar
fashion (Fig. 1E).
Single-qubit operation and readout of a
remote qubit
To demonstrate this qubit control and readout
scheme, we began by measuring a remote
qubit’s ESR spectrum with the tip positioned
above the sensor qubit, as sketched in Fig. 2A.
The energy levels and ESR transitions of this
two-qubit structure are illustrated in Fig. 2B,
and its STM image is shown in Fig. 1B. To
individually address the sensor and remote
qubits, we used two RF sources to apply two
consecutive RF voltage pulses to the STM tip.
Fig. 2. Coherent control of a single remote qubit. (A) Schematic of the measurement scheme. Two RF
tones, RF-S and RF-R, were applied to the STM tip for the coherent control of the sensor and remote qubit,
respectively. The atomic structure used in this figure is shown in Fig. 1B. (B) Energy diagram and ESR
transitions of the two-qubit system. (C) Typical pulse sequence composed of a control pulse on the remote
qubit (red) followed by a sensing pulse on the sensor qubit (blue). The control pulse can induce conditional
or unconditional qubit rotations depending on its frequency content. (D) ESR spectra of the remote qubit
measured with the tip positioned on the sensor qubit. The two curves correspond to different sensing
frequencies (top, fS = f1; bottom, fS = f2). Insets: pulse sequence and ESR transitions involved in each
spectrum. (E) Rabi oscillations of the remote qubit performed by simultaneously driving fR = f3 and f4 to
induce unconditional qubit rotations. The sensing pulse was subsequently applied at fS = f1. Red curve is a
fit to an exponentially decaying sinusoid. A second y axis on the right shows the population of state |0i, which
is determined by temperature and transition frequency at tR = 0 (here 90%) (3). (F) Two-axis control of
the remote qubit shown by sweeping the relative phase f between two consecutive p/2 pulses (fR = f3 and f4;
fS = f1). The data have been vertically shifted by 300 fA to allow easier comparison to (E). f = 180 degrees
does not correspond to DI = 0 here because of the RF rectification effect [see section 4 of (19) and fig. S10].
Solid curve shows a cosine fit. The axes of the two-step qubit rotation are illustrated by red arrows on
the Bloch sphere in the rotating frame. ESR conditions in (D): VDC = 50 mV, IDC = 20 pA, VS = VR = 30 mV,
tR = tS = 200 ns; and in (E) and (F): VDC = 50 mV, IDC = 20 pA, VS = 50 mV, VR = 120 mV, tS = 200 ns.
The sample was kept at 0.4 K during measurements.
Wang et al., Science 382, 87–92 (2023)
6 October 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Two-qubit operations in the coupled-qubit structure shown in Fig. 1B.
(A) Controlled rotation of the remote qubit, conditional on the sensor qubit
state being |0i, obtained by driving the transition |0i|0i ↔ |0i|1i at fR = f3. Red
points show coherent oscillations of the remote qubit detected through a sensing
pulse at fS = f1, as shown schematically on the left. Red curve is a fit to an
exponentially decaying sinusoid. The CNOT operation time is ~13 ns (dotted line).
(B) Rate of controlled rotations, Ω/2p, of the remote qubit. The rate increased
linearly with the RF voltage VR (upper) but remained unchanged as the tunnel
current IDC (lower) was varied. Error bars are comparable with the size of
symbols. Corresponding CNOT times are plotted on a second y axis on the right.
Solid curves are linear fits. Inset illustrates the change of tip-sensor separations.
(C and D) Controlled rotations of the sensor qubit without (C) and with (D) a
CNOT operation at fR = f3 on the remote qubit. Blue (cyan) points measured
with fS = f1 (f2) show sensor Rabi oscillations (solid lines are exponentially
decaying sinusoidal fits) only when the remote qubit has a significant population
in state |0i (|1i). The CNOT operation transferred the predominant population
from state |0i|0i to state |0i|1i, thus causing the opposite trends of the
oscillations between (C) and (D). Blue curves are shifted vertically by 50 fA
for clarity. ESR conditions in (A): VDC = 50 mV, IDC = 10 pA, VS = 30mV, VR =
120 mV, tS = 200 ns; in (C): VDC = 20 mV, IDC = 7.5 pA, VS = 100 mV; and in (D):
VDC = 20 mV, IDC = 7.5 pA, VS = VR = 100 mV. The sample was kept at 0.4 K
during measurements.
A control pulse was applied at frequency fR
to control the remote qubit, followed by a sens-
ing pulse at frequency fS acting on the sensor
qubit (Fig. 2, B and C). A sensing pulse applied
at fS = f1 (f2) results in an ESR signal that
depends on the population of state |0i (|1i) of
the remote qubit. To obtain the ESR spectrum
of the remote qubit, we swept the frequency
fR of the control pulse across the remote qubit
resonances while keeping the sensing pulse
fixed at a resonance frequency of the sensor
qubit. When fR matched an ESR transition
of the remote qubit (e.g., |0i|0i ↔ |0i|1i), the
joint two-qubit state populations were altered,
resulting in a detectable increase (decrease)
of the sensor’s ESR signal at f1 (f2) [see section 4
of (19) and Fig. 2D]. This measurement is
conceptually similar to ensemble double-
electron spin resonance spectroscopy and
allowed us to directly obtain the resonance
frequencies of the remote qubit (f3 and f4;
Fig. 2D) even though no tunnel current passed
through it.
Single-qubit control can be performed in our
platform by coherently driving a qubit inde-
pendent of other qubits’ quantum states. This
was achieved by exciting all ESR transitions of
a remote qubit with the same driving strength
(26), as illustrated in the insets of Fig. 2, E and
F. We measured Rabi oscillations of a remote
qubit by varying the pulse duration tR of the
control pulse and subsequently sensing at f1
(Fig. 2E).
Two-axis control (27) of the remote qubit
can be demonstrated by varying the relative
phase f of two consecutive p/2 pulses (Fig. 2F).
On the Bloch sphere in the rotating frame, the
first p/2 pulse rotates the qubit state from the
z axis to the y axis, and the second performs a
p/2 rotation around an axis in the x–y plane
at an angle of f to the x axis (Fig. 2F, bottom
inset). The resulting signal was well described
by the expected cos(f) dependence (27). Two-
axis control allows arbitrary single-qubit oper-
ations (28).
Wang et al., Science 382, 87–92 (2023)
6 October 2023
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Three-qubit operations in a multi-qubit
atomic structure. (A) Schematic of the control
scheme of a multi-qubit structure (Fig. 1C)
composed of two remote qubits and a sensor qubit.
(B) Energy diagram and ESR transitions used for the
measurement in (C). (C) Controlled-controlled
operation of remote qubit 1 performed by driving the
transition |0i|00i ↔ |0i|10i [(A), red pulse; (B),
red arrow]. Four different frequencies of the sensor
qubit were used for sensing the four remote qubit
states [(A), blue pulse; (B), blue arrows, where
the detected states of the remote qubits are labeled
by red kets in (C)]. The two oscillating curves in
(C) (orange and magenta) correspond to detection
of the remote qubits’ |00i or |10i states upon the
controlled-controlled operation at |0i|00i ↔ |0i|
10i. The CCNOT operation time is ~20 ns (dotted line).
The other two qubit states were not driven
so they showed no oscillations. Green, orange,
and blue curves are vertically shifted for clarity.
(D) Strategy to perform single-, two-, and three-
qubit operations in our platform. Single-qubit
operations are performed by nonselectively exciting
all ESR transitions of a qubit with the same strength.
Two-qubit controlled operations were performed by
selectively exciting ESR transitions that correspond to a
specific quantum state of one control qubit. Three-
qubit controlled-controlled operations are performed
by selectively exciting ESR transitions that correspond
to a specific quantum state of multiple control qubits.
ESR conditions in (C): VDC = 50 mV, IDC = 20 pA,
VS = 50 mV, VR = 80 mV, tS = 200 ns. The sample
was kept at 0.4 K during measurements.
Fig. 5. Relaxation and coherence
properties of remote qubits.
The two-qubit structure shown
in Fig. 1B was used in these
measurements. (A) Relaxation time
T1 of the remote qubit measured
by an inversion recovery scheme.
The exponential fit (solid line)
yields a relaxation time T1 of 166 ±
14 ns for the remote qubit. Inset:
Labels of ESR transitions used
in this figure (same as Fig. 2B).
(B) Ramsey measurements of the
remote qubit with fS = f1 and fR =
f3 + 30 MHz yielding a coherence
time T2* = 86 ± 13 ns. Inset: T2*
shows a dependence on the tunnel
current IDC. (C) Spin-echo measurements of the remote qubit measured with fS = f2 and fR = f3. The exponential fit (solid line) yields a coherence time of T2
Inset: T2
IDC = 10 pA, VS = 60 mV, VR = 120 mV, tS = 200 ns. The sample was kept at 0.4 K during measurements.
Echo = 300 ± 54 ns.
Echo shows no dependence on IDC. ESR conditions in (A): VDC = 50 mV, IDC = 20 pA, VS = 50 mV, VR = 120 mV, tS = 200 ns; and in (B) and (C): VDC = 50 mV,
Two-qubit controlled operation
Controlled NOT (CNOT) operations are straight-
forward to perform in our platform by selectively
exciting ESR transitions that correspond to a
specific quantum state of a control qubit. In
the case of a two-qubit structure, we could per-
form a CNOT operation of the remote qubit
by driving a single ESR transition such as |0i|
0i ↔ |0i|1i at fR = f3 (Fig. 3A), which changes
the state of the remote qubit if and only if the
sensor qubit is in state |0i. Here, the quantum
state of the remote qubit was subsequently
Wang et al., Science 382, 87–92 (2023)
6 October 2023
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
measured by applying a sensing pulse at f1
(Fig. 3A). Sensing at f2 (instead of at f1) showed
oscillations of the opposite sign because this
effectively detected the population of state
|1i (instead of state |0i) of the remote qubit
(fig. S4) (26). The oscillations in Fig. 3A show a
CNOT operation time of ~13 ns for the remote
qubit. The gate operation rate increased in
proportion to the RF amplitude (Fig. 3B, top,
and fig. S3A), with an upper limit set by RF
heating effects. The fast operation of remote
qubits is believed to result from the strong
magnetic field gradient of the neighboring Fe
atom (24). As a consequence of Fe-dominated
driving, the CNOT rate is independent of how
far the tip is to the remote qubits, as reflected
by the tunnel current IDC (Fig. 3B, bottom, and
fig. S3B) (24).
To evaluate the effect of the CNOT operation
on the remote qubit, we performed controlled
rotations of the sensor qubit without and with
this CNOT operation (Fig. 3, C and D, respectively).
Because the initial thermal population of the
two-qubit system was predominantly in state
|0i|0i, a controlled rotation of the sensor qubit
at f1 starting from the initial state |0i|0i (i.e.,
without a CNOT operation) showed clear oscil-
lations because a pulse at f1 excites the tran-
sition |0i|0i ↔ |1i|0i. By contrast, measurement
at transition f2 showed no detectable oscillations
(Fig. 3C) because remote qubit state |1i was
nearly unoccupied. These coherent oscillations
were reversed after a CNOT operation of the
remote qubit at f3 (Fig. 3D), which transferred
the predominant population from state |0i|0i
to state |0i|1i, hence becoming accessible to a
sensing pulse at fS = f2 (i.e., |0i|1i ↔ |1i|1i) but
not f1. We further verified that a CNOT oper-
ation of the remote qubit at f4 (instead of f3) did
not strongly affect the oscillations observed on
the sensor qubit (fig. S6), highlighting the se-
lective, controlled nature of CNOT operations.
Multi-qubit controlled operation
Although single- and two-qubit operations are
sufficient to generate arbitrary quantum cir-
cuits, multi-qubit operations can significantly
reduce the execution time and mitigate accu-
mulated operation errors (29). As long as the
qubit-qubit couplings can be spectroscopically
resolved, our platform allows fast, native qubit
operations having multiple control qubits (by
selectively exciting certain ESR transitions, as
shown in Fig. 4D). To demonstrate this ability,
we constructed a three-qubit structure composed
of two remote qubits (referred to as RQ1 and
RQ2; see fig. S2 for details) and a sensor qubit
(as sketched in Fig. 4A and imaged in Fig. 1C).
We selectively drove RQ1’s |0i|00i ↔ |0i|10i
transition (Fig. 4B, red arrow), corresponding to
a rotation of RQ1 if and only if the sensor qubit
and RQ2 were both in state |0i (Fig. 4A). When
varying the pulse duration tR, this controlled-
controlled operation caused oscillating popu-
lations between states |00i and |10i of the two
remote qubits, whereas populations of states
|01i and |11i were left unchanged (Fig. 4C). Here,
the changes of populations were determined
from the intensities of the four ESR transitions
of the sensor qubit (Fig. 4A, blue pulse, and
Fig. 4B, blue arrows). Similarly, the controlled-
controlled operation |0i|01i ↔ |0i|11i caused
oscillating populations between states |01i
and |11i of the two remote qubits, albeit with
reduced amplitudes because of the reduced
initial thermal populations in these states (fig.
S7E). From these measurements, we obtained
a controlled-controlled NOT (CCNOT) opera-
tion time as short as 20 ns. Single-, two-, and
three-qubit gates in our platform have com-
parably fast operation times because they only
differ in the frequency content of the RF pulses
(30) (Fig. 4D).
Characterizing the coherence of remote qubits
In STM-based approaches, tunneling electrons
have posed severe limitations on the energy
relaxation time T1 and coherence time T2 of
the spins (18, 31). This limitation is overcome
in our scheme because tunneling electrons pass
only through the sensor qubit and do not flow
through the remote qubits (Fig. 1A). To charac-
terize the remote qubits, we implemented an
inversion recovery measurement by first apply-
ing a p pulse on the remote qubit to invert its
population. After a delay time t, we applied a
sensing pulse to the sensor qubit to measure
the state of the remote qubit and obtained
an energy relaxation time T1 = 166 ± 14 ns (14).
A pronounced improvement can be seen in
the quantum coherence of remote qubits, which
is already visible in the higher quality of the
remote qubit’s Rabi oscillations (Fig. 3A) com-
pared with the sensor qubit (Fig. 3, C and D).
Figure 5B shows the Ramsey signal measured
at a tunnel current of IDC = 10 pA, from which
we extracted a coherence time T2* = 86 ± 13 ns.
We found that T2*, the coherence time subject
to inhomogeneous broadening, depended on
the tunnel current and thus on the tip height
(Fig. 5B, inset, and fig. S8A), suggesting that
the proximity of the STM tip to the remote
qubits influenced their quantum coherence time
despite the absence of tunnel current through
them. This decoherence effect was likely caused
by slow thermal or field fluctuations, the lat-
ter possibly arising from slight, uncontrolled
tip motions (31). To cancel the effect of this
inhomogeneous broadening, we performed a
spin-echo measurement and observed that the
measured T2
Echo showed negligible dependence
on the tunnel current (Fig. 5C, inset, and fig.
S8B). The measured coherence time T2
Echo =
300 ± 54 ns (Fig. 5C) approaches the theo-
retical limit of 2T1. These measurements high-
light that the quantum coherence of remote
qubits after spin-echo filtering is limited by
energy relaxation events, in contrast to other
solid-state qubits (3, 27), where T1 ≫ T2
usually applies.
Echo
Discussion and outlook
Our work introduces an atom-by-atom con-
structed qubit platform using electron spins
on surfaces where cryogenic initialization,
universal multi-qubit operations, and multi-
qubit detection have been demonstrated. Two
types of future developments will improve the
performance of this qubit platform. The first
seeks to obtain more gate operations within
the spin decoherence time. Our fast gate oper-
ation of ~20 ns implies that a significant im-
provement can be attained with a moderate
increase of T2, e.g., through the use of thicker
insulators (18), a detection mechanism that
avoids magnetic tips (32), the depletion of sur-
rounding nuclear spins (~5% of atoms in MgO
have nuclear spins), and the use of quantum gates
that are protected by dynamical decoupling-
schemes (33). For reference, electron spins in
bulk insulators (34) and semiconductors (35)
have been demonstrated with coherence times
of ~1 ms.
The second development seeks to increase
the number of controllable qubits. Given that the
RF electric field from the tip can have a radiation
range of tens of nanometers (corresponding
to the effective tip diameter) (36), scaling up
the number of remote qubits should not be
limited by driving but rather by detection. A
direct extension of the work shown here would
lead to up to four remote qubits being placed
around a sensor qubit for direct read-out. Fur-
ther increase will rely on remote qubits at next-
nearest neighbor sites, where a SWAP gate (28)
could be used to transfer the quantum infor-
mation (37). Finally, the computational Hilbert
space can be further expanded by using mo-
lecular qubits to construct three-dimensional
architectures (38) or by using qudit systems
with multiple electron and nuclear spin states
(39, 40). Such extensions rely on the specific
benefit of our surface-based approach, in which
a myriad of available spin species (17, 41–43)
and geometries (44, 45) can be precisely assem-
bled using scanning probe techniques.
REFERENCES AND NOTES
1. A. J. Heinrich et al., Nat. Nanotechnol. 16, 1318–1329 (2021).
2. D. D. Awschalom, R. Hanson, J. Wrachtrup, B. B. Zhou,
Nat. Photonics 12, 516–527 (2018).
3. M. T. Ma¸ Dzik et al., Nat. Commun. 12, 181 (2021).
4. P. Neumann et al., Nat. Phys. 6, 249–253 (2010).
5. D. M. Eigler, E. K. Schweizer, Nature 344, 524–526 (1990).
6. M. Kiczynski et al., Nature 606, 694–699 (2022).
7. H. J. Lee, W. Ho, M. Persson, Phys. Rev. Lett. 92, 186802 (2004).
8. E. Liebhaber et al., Nat. Commun. 13, 2160 (2022).
9. K. Yang et al., Science 366, 509–512 (2019).
10. P. Willke et al., ACS Nano 15, 17959–17965 (2021).
11. S. Baumann et al., Science 350, 417–420 (2015).
12. T. S. Seifert et al., Sci. Adv. 6, eabc5511 (2020).
13. M. Steinbrecher et al., Phys. Rev. B 103, 155405 (2021).
14. L. M. Veldman et al., Science 372, 964–968 (2021).
15. Y. Bae et al., Sci. Adv. 4, eaau4159 (2018).
16. R. Wiesendanger, Rev. Mod. Phys. 81, 1495–1550 (2009).
17. Y. Chen, Y. Bae, A. J. Heinrich, Adv. Mater. 35, e2107534 (2023).
Wang et al., Science 382, 87–92 (2023)
6 October 2023
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
18. W. Paul et al., Nat. Phys. 13, 403–407 (2016).
19. See the supplementary materials.
20. P. Willke et al., Nano Lett. 19, 8201–8206 (2019).
21. J. L. Lado, A. Ferrón, J. Fernández-Rossier, Phys. Rev. B 96,
205420 (2017).
22. K. Yang et al., Phys. Rev. Lett. 122, 227203 (2019).
23. Y. Tokura, W. G. van der Wiel, T. Obata, S. Tarucha, Phys. Rev. Lett.
96, 047202 (2006).
24. S.-H. Phark et al., Adv. Scie2302033 (2023).
25. B. Koiller, X. Hu, S. Das Sarma, Phys. Rev. Lett. 88, 027903
(2002).
26. W. Huang et al., Nature 569, 532–536 (2019).
27. E. Kawakami et al., Nat. Nanotechnol. 9, 666–670 (2014).
28. G. Wolfowicz, J. J. L. Morton, eMagRes 5, 1515–1528 (2016).
29. M. D. Reed et al., Nature 482, 382–385 (2012).
30. L. M. Vandersypen et al., Nature 414, 883–887 (2001).
31. P. Willke et al., Sci. Adv. 4, eaaq1543 (2018).
32. L. Sellies, R. Spachtholz, P. Scheuerer, J. Repp, Atomic-scale
access to long spin coherence in single molecules with scanning
force microscopy. arXiv:2212.12244 [cond-mat.mes-hall] (2022).
33. D. Suter, G. A. Álvarez, Rev. Mod. Phys. 88, 041001 (2016).
34. E. D. Herbschleb et al., Nat. Commun. 10, 3766 (2019).
35. J. T. Muhonen et al., Nat. Nanotechnol. 9, 986–991 (2014).
36. Z. Chu, L. Zheng, K. Lai, Annu. Rev. Mater. Res. 50, 105–130
(2020).
37. Y. P. Kandel et al., Nature 573, 553–557 (2019).
38. X. Zhang et al., Nat. Chem. 14, 59–65 (2022).
39. K. Yang et al., Nat. Nanotechnol. 13, 1120–1125 (2018).
40. S. Thiele et al., Science 344, 1135–1138 (2014).
41. G. Czap et al., Science 364, 670–673 (2019).
42. S. Kovarik et al., Nano Lett. 22, 4176–4181 (2022).
43. R. Kawaguchi et al., Nano Lett. 23, 213–219 (2023).
44. A. A. Khajetoorians, D. Wegner, A. F. Otte, I. Swart, Nat. Rev. Phys. 1,
703–715 (2019).
45. K. Yang et al., Nat. Commun. 12, 993 (2021).
46. Y. Wang, Data and codes for: An atomic-scale multi-qubit platform,
Zenodo (2023); https://doi.org/10.5281/zenodo.7984789.
ACKN OWLED GMEN TS
We thank A. Ardavan, M. Ternes, J. F. Rossier, D. Loss, F. Donati, and
F. Cho for fruitful discussions. Funding: This work was supported
by the Institute for Basic Science (grant IBS-R027-D1). C.P.L.
acknowledges support from the Office of Naval Research (grant
N00014-21-1-2467). M.H. acknowledges support from the University
of Tokyo Global Activity Support Program for Young Researchers
(FY2020). D.-J.C. and C.M. acknowledge support from the Spanish
State Research Agency (grants RTI2018-097895-B-C44, Excelencia
EUR2020-112116, and PID2021-127917NB-I00) funded by MCIN/AEI/
10.13039/501100011033 and by “ERDF: A way of making Europe”
and Eusko Jaurlaritza (project PIBA 2020_1_0017). Y.C. acknowledges
support from the National Natural Science Foundation of China
(grant 12250001). Author contributions: A.J.H., S.P., and C.P.L.
conceived the project. Y.W., Y.C., H.T.B., M.H., C.M., J.K., D.-J.C., Y.B.,
and S.P. performed the measurements and analyzed the data. C.W.
performed simulations to optimize the measurement schemes. All
authors discussed and prepared the manuscript together. Competing
interests: The authors declare no competing interests. Data and
materials availability: All data that support the plots within the paper
and other findings of this study are present in the main and
supplementary figures. Raw data and Matlab codes are available from
Zenodo (46). License information: Copyright © 2023 the authors,
some rights reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government works.
https://www.science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade5050
Materials and Methods
Supplementary Text
Figs. S1 to S10
References (47, 48)
Submitted 20 August 2022; resubmitted 4 March 2023
Accepted 30 August 2023
10.1126/science.ade5050
Wang et al., Science 382, 87–92 (2023)
6 October 2023
6 of 6
|
10.1126_science.ade7471
|
RES EARCH
CHEMICAL DYNAMICS
Stereodynamical control of the H + HD → H2 + D
reaction through HD reagent alignment
Yufeng Wang1†, Jiayu Huang1†, Wei Wang1,2, Tianyu Du1,2, Yurun Xie1,3, Yuxin Ma1,2, Chunlei Xiao1,4*,
Zhaojun Zhang1*, Dong H. Zhang1,3,4*, Xueming Yang1,3,4*
Prealigning nonpolar reacting molecules leads to large stereodynamical effects because of their
weak steering interaction en route to the reaction barrier. However, experimental limitations
in preparing aligned molecules efficiently have hindered the investigation of steric effects in
bimolecular reactions involving hydrogen. Here, we report a high-resolution crossed-beam
study of the reaction H + HD(v = 1, j = 2) → H2(v′, j′) + D at collision energies of 0.50, 1.20, and
2.07 electron volts in which the vibrationally excited hydrogen deuteride (HD) molecules were
prepared in two collision configurations, with their bond preferentially aligned parallel and
perpendicular to the relative velocity of collision partners. Notable stereodynamical effects in
differential cross sections were observed. Quantum dynamics calculations revealed that strong
constructive interference in the perpendicular configuration plays an important role in the
stereodynamical effects observed.
T he fundamental goal for chemical reac-
tion dynamics is to provide a detailed
and quantitative understanding of the
chemical reaction process and to pro-
vide new tools to control the outcome
of a chemical event beyond the traditional
ways, such as adding suitable catalysts and
changing the temperature or pressure of a
reaction mixture. One efficient way to con-
trol chemical reactions is to deposit some
energy in the reaction coordinate of the re-
actant to make a desired molecular bond
more easily cleaved (1–4). Numerous dynam-
ical studies have been carried out to realize
such an idea through vibrational excitation
of reagent molecules, leading to the discovery
and deep understanding of bond-selective or
mode-specific chemistry (5–7). In addition to
vibrational control, it is well established that
the mutual orientation of the colliding part-
ners also has a big effect on the chemical
reaction outcome. Hence, by controlling col-
liding molecular orientation, it is possible to
either promote or hinder the yield of products
into specific final states or scattering angles
(8–10).
For many years, steric control has been per-
formed for inelastic and reactive systems main-
ly involving polar molecules (11–13). Many
methods have been developed for aligning
1State Key Laboratory of Molecular Reaction Dynamics,
Dalian Institute of Chemical Physics, Chinese Academy of
Sciences, Dalian, Liaoning 116023, China. 2School of
Chemical Sciences, University of Chinese Academy of
Sciences, Beijing 100049, China. 3Department of
Chemistry and Shenzhen Key Laboratory of Energy
Chemistry, Southern University of Science and
Technology, Shenzhen 518055, China. 4Hefei National
Laboratory, Hefei 230088, China.
*Corresponding author. Email: chunleixiao@dicp.ac.cn (C.X.);
zhangzhj@dicp.ac.cn (Z.Z.); zhangdh@dicp.ac.cn (D.H.Z.);
xmyang@dicp.ac.cn (X.Y.)
†These authors contributed equally to this work.
or orienting molecules in scattering experi-
ments, including optical pumping (14), hexa-
pole state selection (15), and brute force
orientation (16). An elegant theoretical frame-
work for the characterization of steric effects
has been developed by Aldegunde et al. and
Jambrina et al. (17, 18). Recently, Heid et al.
investigated end-on and side-on collisions
of Ar with oriented NO and demonstrated
that the collision outcome could be controlled
by varying the bond axis orientation (19).
Wang et al. and Pan et al. carried out a series
of experiments to probe the steric effect on
the differential cross sections (DCSs) in the
Cl + CHD3 reaction (20–22). A strong steric
effect was observed, which suggests that re-
orientation effects of CHD3 en route to the re-
action barrier are not strong in this system
owing to the essentially nonpolar nature of
CHD3 (23).
Clearly, aligning nonpolar reacting mole-
cules can have large steric effects because of
their weak steering interaction en route to the
reaction barrier. H2 is undoubtedly the best
candidate for the purpose because it is both
the most widely studied molecule in dynam-
ical experiments and the most tractable theo-
retically (24–27). However, until recently, it
has been difficult to prepare sufficient concen-
trations of H2 in specific quantum states for
scattering experiments (28, 29). The develop-
ment of the Stark-induced adiabatic Raman
passage (SARP) technique not only opened the
door to exciting a large concentration of H2
and its isotopologues in specific quantum
states to study collision dynamics for vibra-
tionally excited H2 molecules, but it also made
it possible to align these molecules for steric
dynamics experiments (30–32). Perreault et al.
observed a strong stereodynamic preference
of angular distributions in the inelastic scat-
tering between aligned HD and D2 molecules
at temperatures down to 1 K (33), which in-
dicates that the weak steering interaction
can suppress the reorientation effects and
expose more pronounced steric effects (34).
They also created a quantum mechanical dou-
ble slit by preparing the rovibrationally excited
D2 molecule in a biaxial state with coherently
coupled bond axis orientations and demon-
strated that they act as the two slits of a double-
slit interferometer manifesting interference
as a strong modulation in the measured an-
gular distribution when inelastic scattering
with a He atom (35, 36). It would be highly
desirable to see whether such striking steric
effects can be observed in the simplest chem-
ical reactions involving H2 molecules and
can be understood at the most fundamental
level.
Experimental demonstration of
stereodynamical control
We carried out a fully quantum state–resolved,
crossed–molecular beam study for the H +
HD → H2 + D reaction, with HD molecules
prepared in two preferentially aligned states
using the stimulated Raman pumping (SRP)
scheme. We found that the DCS of the reac-
tion changed drastically with the direction
of the HD bond axis, which indicated that we
could effectively control the DCS of chemical
reactions.
The experiment was conducted on a modi-
fied crossed-beam apparatus based on the
Rydberg D atom time-of-flight (TOF) detec-
tion technique (37), as described in the sup-
plementary materials. The HD beam was
generated by supersonic expansion through
a pulsed valve cooled by liquid nitrogen. The
H beam was produced by ultraviolet laser
photolysis of HI molecules in a pure HI beam
at the nozzle tip of another pulsed valve. The
HD beam and the H beam were collimated by
skimmers and then entered the scattering
chamber, where they collided at a crossing
angle of 90°. The velocity of the HD beam
was 1250 m/s. The speeds of H beam were
11,230, 17,470, and 22,949 m/s, corresponding
to collision energies of 0.50, 1.20, and 2.07 eV,
respectively. The HD molecules were excited
from (v = 0, j = 0) to (v = 1, j = 2) (where v is the
vibrational quantum number and j is the ro-
tational quantum number) by SRP through
the S(0) transition at the center of the scat-
tering region of the two molecular beams. A
single–longitudinal mode, optical parame-
tric oscillator-amplifier produced the high-
energy Stokes laser, which was the key for
the high SRP excitation efficiency (38). After
the reaction, the D atoms produced from the
reaction were excited to a high-lying Rydberg
state at the crossing region, then flew ~318 mm
before reaching the MCP detector, where they
were field-ionized by an electric field applied
between the MCP and a metal mesh. The
Wang et al., Science 379, 191–195 (2023)
13 January 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
ion signals were then amplified, discrimi-
nated, and recorded in the form of a TOF
spectrum.
Figure 1 shows the schematic of the prepa-
ration of vibrationally excited HD in two
different collision geometries, similar to the
scheme used in (33). Linearly polarized pump
and Stokes lasers with parallel directions of
polarization were used, so the HD molecules
were excited to the (v = 1, j = 2, m = 0) state
with quantization axis along the laser polar-
ization direction. Because the HD bond axis
in the (v = 1, j = 2, m = 0) state was pref-
erentially aligned parallel to the laser polar-
ization direction, we were able to control the
direction of the HD bond axis in scattering by
changing the direction of the laser polariza-
tion. By setting the polarization direction of
the pump and Stokes lasers parallel or per-
pendicular to the relative velocity of colliding
partners, the bond axis of HD was preferen-
tially aligned in parallel or perpendicular to
the relative velocity. We named these two col-
lision configurations parallel and perpendicu-
lar, respectively.
Figure 2, A to C, presents TOF spectra of the
H + HD(v = 1, j = 2) → H2 + D reaction in the
sideways direction with HD(v = 1, j = 2) pre-
pared in parallel and perpendicular configu-
rations at collision energies of 0.50, 1.20, and
2.07 eV, respectively, measured on the scat-
tering plane. Many sharp peaks were observed
in the TOF spectra. Based on the conservation
of momentum and energy, they could be as-
signed to various rovibrational states of H2
products. It was obvious that TOF spectra
obtained with parallel and perpendicular con-
figurations were quite different.
By measuring TOF spectra at different scat-
tering angles on the scattering plane, DCSs of
the reaction on the plane were obtained. It
should be noted that in the perpendicular
configuration, the alignment of the HD molec-
ular bond on the x axis breaks the scattering
symmetry about the z axis. As a result, the
measured DCS on the scattering plane differs
from the conventional DCS, whose integration
over the scattering angle q gives the integral
cross section. Figure 3, A and C, shows DCSs
of H + HD(v = 1, j = 2) → H2 + D obtained at
the collision energy of 0.50 eV for parallel
and perpendicular configurations, respec-
tively. Even casual inspection reveals that
the DCSs for these two configurations were
different. For the parallel configuration, the
H2 products were predominantly backward
scattered, with some small peaks for the
products with low translational energy in the
sideways direction. For the perpendicular con-
figuration, the DCS showed pronounced side-
ways scattered peaks, in particular for the
products with low translational energy. Evi-
dent differences between the two DCSs indi-
cated the existence of strong stereodynamical
effects in this reaction.
The difference in the DCSs for these two
configurations at the collision energy of 1.20 eV
was even more obvious. For the parallel con-
figuration, the H2 products remained predo-
minantly backward scattered, whereas the
sideway peaks became higher and some for-
ward components showed up (Fig. 3E). In
strong contrast, the angular distribution for
the perpendicular configuration was dominated
by sideways peaks (Fig. 3G) with backward-
scattered amplitude suppressed substantially,
underscoring strong stereodynamical effects
in the reaction.
With a further increase of collision energy
to 2.07 eV, the angular distributions for both
the parallel and perpendicular configurations
looked quite similar—dominated by sideways
Parallel
Perpendicular
x
HD
H
x
HD
Stokes
y
pump
z(k)
Stokes
y
pump
H
z(k)
x
α
x
y
y
Bond axis
β
z(k)
k'
θ
z(k)
Fig. 1. The schematic of two collision geometries prepared by SRP. The scattering frame is defined
so that the z axis is parallel to the reactant relative velocity, and xz is the detection plane (or scattering
plane). The alignment of the molecular bond axis was controlled using the polarized pump and Stokes
laser pulses, as indicated by the green and red double arrows, respectively. The preferential direction
of the HD molecular bond axis in the scattering frame is specified by the polar and azimuthal angle
(b, a). In theory, the direction of k′ is described by the polar and azimuthal angle (q, f). In the present
experiment, f = 0.
peaks with more or less the same scattering
angles and the same translational energies
(Fig. 3, I and K). However, the relative inten-
sities for large-angle scattering and for for-
ward scattering were much stronger for the
parallel configuration.
Figure 3 also shows the change of relative
reactivity for these two configurations. At low
collision energies, the parallel configuration
that leads to end-on collisions is predominant
because of a narrow cone of acceptance and
small impact parameters. As the collision en-
ergy increases, the side-on configurations be-
come increasingly prevalent and sideways or
forward scattering takes over with the broad-
ening of the acceptance cone and the increas-
ing of the impact parameters.
Quantum dynamical simulation of
stereodynamical effect
To understand the strong stereodynamical ef-
fects in the reaction, we carried out nonadia-
batic time-dependent wave packet calculations
on the diabatic potential energy surface we
constructed for this reaction (39). Details of
the theoretical calculation can be found in
A
100
v'=0
v'=1
Parallel
Perpendicular
)
s
t
n
u
o
C
(
l
a
n
g
S
i
B
)
s
t
n
u
o
C
(
l
a
n
g
S
i
50
0
300
200
100
0
C
450
)
s
t
n
u
o
C
(
l
a
n
g
S
i
300
150
0
35
40
45
50
55
Time of flight ( s)
60
v'=0
v'=1
v'=2
EC=0.50 eV
65
70
75
EC=1.20 eV
25
30
35
40
45
50
55
60
Time of flight ( s)
v'=0
v'=1
v'=2
v'=3
v'=4
EC=2.07 eV
20
25
30
35
40
45
50
55
Time of flight ( s)
Fig. 2. TOF spectra of the D atom product from
the H + HD(v = 1, j = 2) → H2(v , j ) + D reaction.
(A to C) They were obtained at three collision
energies: 0.50 eV (A), 1.20 eV (B), and 2.07 eV (C)
in parallel (blue solid line) and perpendicular (red
solid line) configurations in the sideways direction,
at laboratory angles of 25° (A), 28° (B), and 29° (C),
respectively. The sharp peaks can be assigned to
various rovibrational states of the H2 product, as
indicated in the figure.
Wang et al., Science 379, 191–195 (2023)
13 January 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
EXP
THEORY
A
C
E
G
I
K
H
Parallel/3
EC=0.50 eV
Perpendicular
Parallel × 2
EC=1.20 eV
Perpendicular
Parallel × 3
EC=2.07 eV
Perpendicular
B
HD
D
F
H
J
L
Fig. 3. Three-dimensional scattering H2 product
contour plots on the scattering plane. (A to
L) Product contour plots are shown in the center-
of-mass frame for experimental measurements
(left column) and quantum dynamics calculations
(right column) at collision energies of 0.50 eV
[(A) to (D)], 1.20 eV [(E) to (H)], and 2.07 eV [(I)
to (L)], with parallel [(A), (B), (E), (F), (I), and (J)]
and perpendicular [(C), (D), (G), (H), (K), and (L)]
configurations, respectively.
the supplementary materials. For the paral-
lel configuration, the state of HD prepared
by SRP is v ¼ 1; j ¼ 2; m ¼ 0
i, and for the
perpendicular configuration, the state of HD
prepared is
j
i
v ¼ 1; j ¼ 2; mx ¼ 0
j
r
i
ffiffiffi
3
v ¼ 1; j ¼ 2; m ¼ þ2
j
8
1
j
2
r
v ¼ 1; j ¼ 2; m ¼ 0
ffiffiffi
3
v ¼ 1; j ¼ 2; m ¼ (cid:2)2
i
j
8
i
¼
(cid:2)
þ
ð1Þ
where m denotes the projection of angular
momentum along the quantization axis z at
the direction of relative velocity, and mx de-
notes the component along the x axis. The
DCS for the parallel configuration for an out-
going channel with specific quantum state
(v′j′m′) has a cylindrical symmetry with re-
ð
f m ¼ þ2
Þ (cid:2)
1
2
ð
f m ¼ 0
Þ
1.2
A
0.50 eV
spect to the z axis and can be evaluated
easily as
ð
ds v ¼ 1; j ¼ 2; mz ¼ 0 → v′j′m′
dW
Þ
¼ f v ¼ 1; j ¼ 2; m ¼ 0 → v′j′m′
ð
j
j2
j2 ≡ f0j
Þ
ð2Þ
ð
where f v ¼ 1; j ¼ 2; m ¼ 0 → v′j′m′
Þ repre-
sents the state-to-state scattering amplitude
within a solid angle dW along the direction
(q, f) defined with respect to the quantiza-
tion z axis, which is abbreviated as f0 with
the under script 0 denoting m = 0 and other
index omitted. For the perpendicular configu-
ration, the DCS results from the interference
of the scattering amplitudes associated with
the three input channels as follows
ð
ds v ¼ 1; j ¼ 2; mx ¼ 0 → v′j′m′
dW
Þ
¼
ffiffiffi
3
8
(cid:3)
r
(cid:3)
(cid:3)
(cid:3)
(cid:3)
ffiffiffi
r
3
8
(cid:3)
(cid:3)
2
(cid:3)
(cid:3)
Þ
¼
3
8
(cid:4)
j2 þ Re
þ
ð
f m ¼ (cid:2)2
j
fþ2
j2 þ
1
4
f0j
j2
þ
j
f(cid:2)2
3
8
r
ffiffiffi
3
8
(cid:2)
fþ2 f0
(cid:3) (cid:2)
3
fþ2 f(cid:2)2
4
r
ffiffiffi
3
8
f(cid:2)2 f0
(cid:3)
(cid:3)
(cid:5)
ð3Þ
In Fig. 3, the theoretical DCSs at each col-
lision energy and for the internuclear axis
preparations are also shown. The excellent
agreement between experiment and theory
demonstrates the high accuracy of the quan-
tum calculations.
Influence of quantum interference on
stereodynamical effect
Given that the quantum dynamical simulation
was capable of accurately reproducing the ob-
served DCSs, we were confident of using
theory to determine the physical origin of the
strong stereodynamical effects. At the collision
energy of 0.50 eV, the DCS for the parallel
configuration was determined by single input
channel with m = 0 and manifested a pre-
dominated backward feature, as shown in
Fig. 4A, as a result of head-on collision dy-
namics. By contrast, the DCSs for m = ±2,
which are the main input channels for the
perpendicular configuration, peak at q = 100°,
indicative of peripheral dynamics with large
impact-parameter collisions as can been seen
from the opacity functions shown in fig. S12
and the dependence of DCS on the total
angular momentum shown in figs. S13 to S15.
The direct combination of the m = 0 and m =
±2 DCSs with ¼ and ¾ weights without the
interference term shown in Eq. 3 gave rise to
an essentially straight line with a small bump
at q = 90°. However, the actual DCS for per-
pendicular configuration showed an evident
peak around q = 90°, with a height consid-
erably higher than the direct combination
result, apparently as a result of the construc-
tive interference between the m = 0 and m =
±2 channels. Therefore, the pronounced side-
ways peaks for the perpendicular configuration
shown in Fig. 3C came from the constructive
interference between the m = 0 and m = ±2
channels. It is worthwhile to point out that
the interference term at the forward and back-
ward directions is zero, as explained in the
supplementary materials.
Figure 4B shows the angular distribution
at the collision energy of 1.2 eV. Although the
DCS for the parallel configuration was still
backward dominated, it extended all the way
up to the forward direction with a substantial
amplitude. The DCS for m = ±2 channels
)
.
u
.
a
(
S
C
D
0.8
0.4
0.0
0.9
B
Parallel(m=0)
m=2
Perp. w/o mix term
Perpendicular
1.20 eV
C
2.07 eV
)
.
u
.
a
(
S
C
D
)
.
u
.
a
(
S
C
D
0.6
0.3
0.0
1.0
0.5
0.0
0
30
60
90
120
150
180
Fig. 4. Calculated DCSs on the scattering plane.
(A to C) All final states of the H2 product from the
H + HD(v = 1, j = 2) → H2 + D reaction are included
at collision energies of 0.50 eV (A), 1.20 eV (B),
and 2.07 eV (C), respectively. The results of
m = −2 are equal to that of m = 2, so only m = 2
is shown. a.u., arbitrary units.
Wang et al., Science 379, 191–195 (2023)
13 January 2023
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
resembled that at the collision energy of
0.5 eV as a broad peak, but the peak po-
sition moved forward to q = 70°. The direct
combination of the m = 0 and m = ±2 DCSs
without the interference terms resulted in
a rather uniform distribution with a small
but broad peak at q = 70°. In strong con-
trast, the actual DCS had a pronounced peak
also at q = 70° with the amplitude doubled
as compared with that without the interfer-
ence terms because of the constructive in-
terference between the m = 0 and m = ±2
channels—obviously higher than the backward
scattering.
With the further increase of the collision
energy to 2.07 eV, the DCS for m = ±2 chan-
nels resembled those at low collision energies
as a broad peak but with the peak position
moving forward further to q = 60°. The DCS
for the parallel configuration changed sub-
stantially and became rather uniform, with a
broad peak at q = 60° and another narrower
peak in the forward direction, apparently as
a result of a broader cone of acceptance and
larger impact parameters. The direct combi-
nation of the m = 0 and m = ±2 DCSs without
the interference terms looked close to m =
±2 with one peak in the forward direction
and another at q = 60° with the same heights.
The interference between the m = 0 and m =
±2 channels substantially increased the peak
intensity at q = 60° and doubled the peak
height, but it had no effect on the forward
peak intensity, making the relative intensity
of the forward-scattering peak considerably
suppressed.
To verify this strong interference behavior,
we show in Fig. 5, A and B, the comparison
between experimental and theoretical DCSs
for the product H2(v′ = 0, j′ = 1) and H2(v′ = 1,
j = 3) states, respectively, at the collision en-
ergy of 0.50 eV. As seen, the theory agreed
with experiment well on the DCSs for both
product states. The DCS for the H2(v′ = 0,
j′ = 1) state exhibited two clear peaks at q =
125° and 180°, respectively, whereas the direct
combination without the interference terms
only showed a tiny bump at q = 120°. The
DCS for the H2(v′ = 1, j = 3) state exhibited
two peaks with a pronounced one at q = 90°;
by contrast, the direct combination without
the interference terms only had one broad
peak. The interference effects were obvious
for the perpendicular configuration.
The concept of intrinsic polarization-
dependent differential cross sections (PDDCSs)
has been widely used to study the stereo-
dynamical effects in chemical reactions (17).
Because the quantum dynamical simulation
reproduces the observed DCSs very well, as
shown in Fig. 3 (as well as the total DCSs, as
shown in figs. S6 to S8), we can use the theo-
retical PDDCSs to analyze the observed stereo-
dynamical effects. As shown in figs. S9 to S11,
H2(v'=0, j'=1)
from the parallel configuration, manifesting
strong stereodynamical effects.
Exp perpendicular
Theory perp. w/o mix term
Theory perp.
H2(v'=1, j'=3)
A
B
.
)
.
u
a
(
S
C
D
0.04
0.03
0.02
0.01
0.00
0.04
.
)
.
u
a
(
S
C
D
0.02
0.00
0
30
60
90
120
150
180
Fig. 5. Comparisons between experiment and
theory on product state–resolved DCSs. (A and
B) Comparisons are made on the scattering plane
at the collision energy of 0.50 eV for the H2 product
in the (v′ = 0, j′ = 1) (A) and (v′ = 1, j′ = 3) (B) states.
In this experiment, TOF spectra at different
laboratory angles were acquired by scanning the
laboratory angle back and forth 36 times. By
analyzing the summed TOF signals for the above
H2 product states at each laboratory angle in these
scans, the error bars of one standard deviation
(1s) in the experimental DCS shown in this figure
were estimated to be ~10%. Analysis of experi-
mental data and evaluation of error bars are also
discussed in the supplementary materials.
qð Þ and S 4ð Þ
0
the S 2ð Þ
qð Þ moments are respon-
0
sible for the effects in the parallel configu-
ration as found in many theoretical studies
(12, 17, 19), whereas the feature of stereo-
dynamical effects in the sideways direction
in the perpendicular configuration are main-
ly originated from the S 2ð Þ
qð Þ
T2
moments.
qð Þ and S 4ð Þ
T4
Therefore, the m = 0 and m = ±2 channels
for the HD(v = 1, j = 2) state had different
angular distributions, with one backward do-
minated and the other peaked in a mostly
sideways direction. For the perpendicular con-
figuration, the angular distribution on the scat-
tering plane is determined by Eq. 3, with the
interference term between the m = 0 and m =
±2 channels. Notably, strong constructive in-
terference occurred in the sideways direction,
substantially enhancing the peak height in the
sideways direction particularly at the collision
energies of 1.20 and 2.07 eV. As a result, the
measured angular distributions for the perpen-
dicular configuration were markedly different
REFERENCES AND NOTES
1. R. N. Zare, Science 279, 1875–1879 (1998).
2. F. F. Crim, Acc. Chem. Res. 32, 877–884 (1999).
3. F. F. Crim, Proc. Natl. Acad. Sci. U.S.A. 105, 12654–12661
(2008).
4. P. M. Hundt, B. Jiang, M. E. van Reijzen, H. Guo, R. D. Beck,
Science 344, 504–507 (2014).
5. B. R. Strazisar, C. Lin, H. F. Davis, Science 290, 958–961
(2000).
6. H. Guo, B. Jiang, Acc. Chem. Res. 47, 3679–3685
(2014).
7. G. Czakó, J. M. Bowman, Science 334, 343–346 (2011).
8. K. Liu, Annu. Rev. Phys. Chem. 67, 91–111 (2016).
9. B. L. Yoder, R. Bisson, R. D. Beck, Science 329, 553–556
(2010).
10. R. B. Bernstein, D. R. Herschbach, R. D. Levine, J. Phys. Chem.
91, 5365–5377 (1987).
11. D. H. Parker, R. B. Bernstein, Annu. Rev. Phys. Chem. 40,
561–595 (1989).
12. F. J. Aoiz et al., Phys. Chem. Chem. Phys. 17, 30210–30228
(2015).
13. H. J. Loesch, Annu. Rev. Phys. Chem. 46, 555–594
(1995).
14. Z. Karny, R. C. Estler, R. N. Zare, J. Chem. Phys. 69, 5199–5201
(1978).
15. D. H. Parker, H. Jalink, S. Stolte, J. Phys. Chem. 91, 5427–5437
(1987).
16. H. J. Loesch, A. Remscheid, J. Chem. Phys. 93, 4779–4790
(1990).
17. J. Aldegunde et al., J. Phys. Chem. A 109, 6200–6217
(2005).
18. P. G. Jambrina, M. Morita, J. F. E. Croft, F. J. Aoiz,
N. Balakrishnan, J. Phys. Chem. Lett. 13, 4064–4072
(2022).
19. C. G. Heid, V. Walpole, M. Brouard, P. G. Jambrina, F. J. Aoiz,
Nat. Chem. 11, 662–668 (2019).
20. F. Wang, J. S. Lin, K. Liu, Science 331, 900–903 (2011).
21. F. Wang, K. Liu, T. P. Rakitzis, Nat. Chem. 4, 636–641
(2012).
22. H. Pan, F. Wang, G. Czakó, K. Liu, Nat. Chem. 9, 1175–1180
(2017).
23. K. Liu, J. Chem. Phys. 142, 080901 (2015).
24. G. C. Schatz, A. Kuppermann, J. Chem. Phys. 59, 964–965
(1973).
25. J. Z. H. Zhang, S.-I. Chu, W. H. Miller, J. Chem. Phys. 88,
6233–6239 (1988).
26. J. Jankunas et al., Proc. Natl. Acad. Sci. U.S.A. 111, 15–20
(2014).
27. Y. Xie et al., Science 368, 767–771 (2020).
28. T. Wang et al., Science 342, 1499–1502 (2013).
29. T. Yang et al., Science 347, 60–63 (2015).
30. T. Wang, T. Yang, C. Xiao, D. Dai, X. Yang, J. Phys. Chem. Lett.
4, 368–371 (2013).
31. N. Mukherjee, R. N. Zare, J. Chem. Phys. 135, 024201
(2011).
32. N. Mukherjee, W. Dong, R. N. Zare, J. Chem. Phys. 140, 074201
(2014).
33. W. E. Perreault, N. Mukherjee, R. N. Zare, Science 358,
356–359 (2017).
34. W. E. Perreault, N. Mukherjee, R. N. Zare, Nat. Chem. 10,
561–567 (2018).
35. H. Zhou, W. E. Perreault, N. Mukherjee, R. N. Zare, Science
374, 960–964 (2021).
36. X. Wang, X. Yang, Science 374, 938–939 (2021).
37. M. Qiu et al., Rev. Sci. Instrum. 76, 083107 (2005).
38. Y. Wang et al., Rev. Sci. Instrum. 91, 053001 (2020).
39. D. Yuan et al., Science 362, 1289–1293 (2018).
40. Y. Wang et al., Stereodynamical Control of the H+HD→H2+D
Reaction through HD Reagent Alignment, dataset, Dryad
(2022); https://doi.org/10.5061/dryad.x0k6djhpk.
AC KNOWLED GME NTS
We thank K. Liu for the helpful discussion on the stereodynamical
experiment. Funding: This work was supported by the National
Natural Science Foundation of China (grant nos. 22288201,
22173097, 41827801, and 22103084), the Chinese Academy of
Sciences (grant no. GJJSTD20220001), the Innovation Program for
Wang et al., Science 379, 191–195 (2023)
13 January 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
Quantum Science and Technology (grant no. 2021ZD0303300),
the Guangdong Science and Technology Program (grant nos.
2019ZT08L455 and 2019JC01X091), and the Shenzhen Science
and Technology Program (grant no. ZDSYS20200421111001787).
Author contributions: Y.W., W.W., T.D., Y.X., Y.M., C.X., and
X.Y. performed the crossed-beam experiments and data analysis.
J.H., Z.Z., and D.H.Z. performed the quantum dynamics
calculations and data analysis. C.X., Z.Z., D.H.Z., and X.Y. designed
the research and wrote the manuscript. Competing interests:
The authors declare no competing interests. Data and materials
availability: All data needed to evaluate the conclusions in this
paper are present in the paper or the supplementary materials.
All data presented in this paper are deposited at Dryad (40).
License information: Copyright © 2023 the authors, some rights
reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government
works. https://www.science.org/about/science-licenses-journal-
article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade7471
Materials and Methods
Figs. S1 to S15
Table S1
References (41–46)
Submitted 12 September 2022; accepted 14 December 2022
10.1126/science.ade7471
Wang et al., Science 379, 191–195 (2023)
13 January 2023
5 of 5
|
10.1126_science.ade7651
|
RES EARCH
QUANTUM SIMULATION
A superconducting quantum simulator based
on a photonic-bandgap metamaterial
Xueyue Zhang1,2†, Eunjong Kim1,2†, Daniel K. Mark3, Soonwon Choi3, Oskar Painter1,2,4*
Synthesizing many-body quantum systems with various ranges of interactions facilitates the study of
quantum chaotic dynamics. Such extended interaction range can be enabled by using nonlocal degrees
of freedom such as photonic modes in an otherwise locally connected structure. Here, we present
a superconducting quantum simulator in which qubits are connected through an extensible photonic-
bandgap metamaterial, thus realizing a one-dimensional Bose-Hubbard model with tunable hopping
range and on-site interaction. Using individual site control and readout, we characterize the statistics of
measurement outcomes from many-body quench dynamics, which enables in situ Hamiltonian learning.
Further, the outcome statistics reveal the effect of increased hopping range, showing the predicted
crossover from integrability to ergodicity. Our work enables the study of emergent randomness from
chaotic many-body evolution and, more broadly, expands the accessible Hamiltonians for quantum
simulation using superconducting circuits.
R ealizing a scalable architecture for quan-
tum computation and simulation is a
central goal in the field of quantum in-
formation science. Although architec-
tures with nearest-neighbor (NN) coupling
between quantum particles on a lattice are
prevalent, quantum systems with long-range
interactions can realize a richer set of com-
putational tasks and physical phenomena
(1–4). For instance, in the case of gate-based
quantum computation, coupling beyond the
NN level enables nonlocal gate operations be-
tween qubits, which can reduce the overhead
of quantum algorithms and lift the restrictions
on code rate and distance of local-interaction–
based quantum error-correcting codes (5, 6).
In the case of analog quantum simulation, the
inclusion of long-range interactions can alter
the behavior of otherwise integrable many-
body systems (7, 8), resulting in quantum chaotic
dynamics, which is at the root of such topics
as quantum thermalization (9) and quantum
information scrambling (10). Furthermore,
control over the range of lattice connectivity
grants access to different physical regimes
and the crossover between them, such as in
many-body quantum phase transitions (11–13)
and the hydrodynamics of nonequilibrium
quantum states (14).
For engineered quantum systems consist-
ing of interacting quantum particles on a lat-
tice, it is often challenging to scale to larger
lattice sizes while maintaining a high degree
of lattice connectivity and single-site con-
1Thomas J. Watson, Sr., Laboratory of Applied Physics and
Kavli Nanoscience Institute, California Institute of Technology,
Pasadena, CA 91125, USA. 2Institute for Quantum Information
and Matter, California Institute of Technology, Pasadena,
CA 91125, USA. 3Center for Theoretical Physics, Massachusetts
Institute of Technology, Cambridge, MA 02139, USA. 4AWS
Center for Quantum Computing, Pasadena, CA 91125, USA.
*Corresponding author. Email: opainter@caltech.edu
†These authors contributed equally to this work.
trol. One common approach, developed for
trapped-ion and neutral-atom systems, is to
use resonant modes of either vibrational (1)
or optical (15) cavities as a quantum bus for
mediating interactions between the internal
states of atoms across the lattice. Similar schemes
have been adopted in superconducting quantum
circuits, realizing systems as large as 20 qubits
with all-to-all coupling via a common microwave
cavity (16). Increasing the number of lattice
sites in this case, however, leads to either par-
asitic coupling arising from dense placement
of sites in a fixed-volume cavity or frequency-
crowding effects stemming from the increased
spectral density of cavity modes when in-
creasing the cavity size (17).
An alternative approach for connecting
quantum particles on a lattice is to construct
a quantum bus from an intrinsically extensible
structure, such as a waveguide. Along this di-
rection, engineered photonic-bandgap wave-
guides have been proposed as a quantum bus
that simultaneously protects quantum par-
ticles from radiative damping through the
waveguide while allowing for extended-range
lattice connectivity (18). The waveguide-bus
concept has been investigated in the context of
many-body simulation with cold atoms cou-
pled to engineered nanophotonic waveguides
(18, 19), and recent experiments have explored
qubit-photon bound states in superconducting
quantum circuits with microwave photonic-
bandgap waveguides (20–24). However, the
realization of a scalable many-body quantum
simulator, with single-site quantum-particle
control and a high level of lattice connectivity,
has remained an open challenge.
We demonstrate a scalable many-body quan-
tum simulator consisting of a one-dimensional
(1D) lattice of superconducting transmon qubits
coupled to a common metamaterial wave-
guide. This system provides both tunable-range
connectivity between qubits and full single-
site control and state measurement of individ-
ual qubits. The waveguide acts both as a bus
for mediating exponentially decaying long-
range interactions between qubits and as a
Purcell filter enabling multiplexed, rapid read-
out of the qubit states with high fidelity. This
system realizes an extended version of the
Bose-Hubbard model with tunable hopping
range and on-site interaction. Using our abil-
ity to efficiently collect measurement outcomes
from many-body quench dynamics—enabled
by the fast experimental repetition rate of our
system—we perform direct analysis of out-
come statistics to learn Hamiltonian parame-
ters in situ and study the effect of hopping
range on the evolution of randomness across
the system. Specifically, we observe a distri-
bution of outcome bit-string probabilities re-
flecting the ergodic nature of the Hamiltonian
with long-range hopping. This result exper-
imentally confirms the expectation from quan-
tum chaos for interacting many-particle systems,
highlighting the connection between ergodic
unitary dynamics and its effective statistical de-
scription in terms of random matrix theory (25).
Metamaterial-based quantum simulator
The backbone of the many-body quantum sim-
ulator in this work is a metamaterial waveguide
formed from a chain of lumped-element inductor-
capacitor (LC) microwave resonators. The wave-
guide can be described by a generic model
(Fig. 1A) of a 1D cavity array with NN coupling
t (26, 27). The corresponding dispersion re-
lation (Fig. 1B) is given by wk ¼ wc þ 2tcos kdð
Þ,
exhibiting a passband centered around the
cavity frequency wc with a bandwidth of 4t,
where k is the wave vector and d is the lattice
constant of the array. The bandgap at frequen-
cies below we;(cid:2) ¼ wc (cid:2) 2t (above we;þ ¼ wc þ
2t ) is denoted as the lower (upper) bandgap,
abbreviated as LBG (UBG). Inside the band-
gaps, the off-resonant coupling between a bare
quantum emitter and the waveguide modes
gives rise to an emitter-photon bound state
(28) whose photonic tail is localized around
the emitter. Localization follows a spatial pro-
j=x in the LBG or UBG (27), where
file ð∓1Þ
Dx is the displacement in the number of unit
cells from the emitter and x is the localization
length controlled by the detuning D between
the band-edge frequency and the transition
frequency of the bound state. The overlap of
two bound states results in photon-mediated
coupling with a range covering multiple unit
cells, i.e., long-range coupling, exhibiting a
greater strength and a more extended range
x at a smaller detuning Dj
j (Fig. 1C).
Dxe(cid:2) Dxj
The metamaterial waveguide consists of a
42–unit-cell array of capacitively coupled lumped-
element microwave resonators (Fig. 1, D and
E) and is equipped at both ends with engi-
neered tapering sections, designed to reduce
the impedance mismatch to external 50-ohm
Zhang et al., Science 379, 278–283 (2023)
20 January 2023
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
A
B
c
t
e,−
e,+
g
g
. . .
. . .
F
D
Ct
Cg
Cg
. . .
. . .
)
B
d
(
n
o
i
s
s
i
y
c
n
e
u
q
e
r
F
0
-25
-50
-75
-100
Fig. 1. Metamaterial-
based quantum simula-
tor. (A) Schematic
showing a 1D array of
coupled cavities with
nearest-neighbor coupling
t. Each cavity is occupied
by a quantum emitter
(orange ball) with coupling
g to the cavity. (B) Dis-
persion relation of the
coupled cavity array in (A)
with a passband between
we;(cid:3) centered at wc
(bandwidth of 4t). The
LBG (UBG) below (above)
the passband is shaded in
green (purple). (C) Top
(Bottom): Cartoon of two
emitter-photon bound
states at small (large)
detuning Dj
j, indicated by
dark (light) orange arrows
in (B), exhibiting an
extended (restricted) spatial range and large (small) photonic component in the
bound states. (D) Electrical circuit realization of (A) with capacitively coupled
LC resonators and transmon qubits corresponding to the cavity array and
the quantum emitters, respectively. The coupling capacitors are color coded in
accordance with (A). (E) Optical micrograph (false colored) of the fabricated
quantum simulator with 42 metamaterial resonators (lattice constant
d ¼ 292 mm) colored blue connected to input-output ports (red) via tapering
m
s
n
a
r
T
4.0
R10
E
UBG
LBG
d
0
Wavevector k
C
. . .
. . .
. . .
. . .
R2
R4
R6
R8
R10
Q2
Q4
Q6
Q8
Q10
4
t
/d
R9
R2
R1
R7
R6
R5
R4
R3
R8
5.0
6.0
7.0
Frequency (GHz)
Q1
8.0
Q3
Q5
Q7
Q9
R1
R3
R5
R7
R9
300 μm
50 μm
d
1 mm
sections (purple). Ten qubits (Qi, colored orange), controlled by individual charge
drive lines (pink) and flux bias lines (dark blue), and their readout resonators
(Ri, colored green) couple to the 10 inner unit cells of the metamaterial waveguide
with a zoomed-in view in the left inset. Detailed view of the coupling region is shown
in the right inset. Two auxiliary qubits (yellow) are not used in this experiment.
(F) Transmission spectrum through the metamaterial waveguide (red curve) with
black arrows indicating the 10 resonances of the readout resonators Ri.
input-output ports at frequencies lying within
the passband of the waveguide (24, 29). Each
of the middle 10 metamaterial resonators (unit
cells labeled by i = 1 to 10) couples to a trans-
mon qubit (30), which serves as the quantum
emitter. Individual addressing of each qubit
is achieved by excitation (XY control) from a
charge drive line and frequency tuning (Z con-
trol) from a flux bias line. Dispersive qubit read-
out is enabled by capacitively coupling each
qubit Q i to a compact readout resonator Ri,
which itself is then coupled to the metamate-
rial resonator of the same unit cell. The entire
metamaterial and transmon qubit system (the
device) is fabricated by using evaporated thin-
film aluminum on a high-resistivity silicon
substrate, with fabrication procedures detailed
in (21, 29). Further details of the device modeling
and the experimental setup used to measure
and test the device are discussed in supple-
mentary text I and II, respectively.
This realization of the device enables qubit
readout using the passband of the metamate-
rial waveguide with built-in protection against
Purcell decay channels (supplementary text III).
The transmission spectrum through the wave-
guide (Fig. 1F) shows a passband ranging from
we;(cid:2)=2p ≈ 5:01 GHz to we;þ=2p ≈ 7:08 GHz
with ripples smaller than 8dB near the center.
The extinction ratio of the transmission be-
tween that measured in the passband and that
measured in the bandgaps is greater than
65 dB, with a sharp transition in the trans-
mission occurring within 100MHz of the band
edges. In the middle of the passband, reso-
nances associated with the readout resonators
are observed between 5.574 and 6.328 GHz.
The average decay rate of 10 readout resona-
tors is kRi
=2p ¼ 11:8 MHz, enabling fast, high-
fidelity multiplexed readout while maintaining
a low level of readout cross-talk. For details of
readout methods and characterization, refer
to supplementary text IV.
Bose-Hubbard model with long-range hopping
The spatially extended bound-state excitations,
formed between transmon-qubit excitations
and waveguide photons of the metamaterial-
waveguide bus, creates a lattice of interacting
microwave photons (31). This quantum system
is described by an extended version of the 1D
Bose-Hubbard model with tunable long-range
hopping and on-site interaction. Specifically,
each bound state formed from qubit Q i, inher-
iting the level structure of an anharmonic os-
cillator from a transmon qubit (30), serves as a
bosonic site with local site energy Di ¼ w01;i and
the on-site interaction Ui ¼ w12;i (cid:2) w01;i. Here,
w01;i and w12;i are the transition frequencies
of the bound state on site Q i from its ground
state j0〉 to the first excited state j1〉 and that
from the first to the second excited state j2〉,
respectively. In addition, the long-range hop-
ping Ji;j is enabled by the overlap between a
pair of qubit-photon bound states on sites Q i
and Q j . The Hamiltonian of this model that
captures the basic processes mentioned above
can be written as
^H =ℏ ¼
†^bj þ
^bi
X
i
Ui
2
^ni ^ni (cid:2) 1
ð
Þ
X
i;j
þ
Ji;j
X
Di ^ni
ð1Þ
i
† (^bi) is the creation (annihilation)
where ^bi
operator and ^ni ≡ ^bi
†^bi is the number operator
on site Qi. The parameters of the Hamiltonian
realized in this simulator can be learned through
experiments enabled by the precise, single-
site–level control over qubits.
j
We measure the on-site interactionUi (Fig. 2A)
by performing spectroscopy of w12 after ini-
tializing Q i in its first excited state j1〉. From
j decreases as w01
within either bandgap, Ui
approaches the closest band edge owing to
dressing from the passband modes of the
metamaterial (22)—i.e., the Lamb shift. In the
UBG, a wide tuning range of Ui
j is achievable
from the strong hybridization between the
j1〉–j2〉 transition and the band-edge modes at
we;þÞ=2p < 300 MHz. The magnitude
w01(cid:2)ð
(cid:2)
(cid:2) is measured from vacuum
of hopping Ji;j
Rabi oscillations between sites Q i and Q j by
(cid:2)
(cid:2)
j
Zhang et al., Science 379, 278–283 (2023)
20 January 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
A
250
200
150
100
50
0
102
101
1
)
z
H
M
(
2
/
|
U
i
|
B
)
z
H
M
(
2
j
,
i
J
1
1
i
,
i
J
j
,
i
0.1J
10-1
6
C
5
4
3
2
1
4.4
Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Q10
···
D
E
Initialization
zinit = m1 m2
m1
( )
X
Q1
m10
i
j
1
2
3
4
5
6
7
8
9
i
1
2
3
4
5
6
7
8
9
1
1
i
,
i
J
j
0.1J
,
i
2
4
i
6
j
8
2
4
8
6
j
i
4.6
4.8
5.0
7.2
7.4
7.6
Frequency (GHz)
Qi
Qj
Q10
1.0
F
d
F
0.8
0.6
0.4
0.2
0.0
Interaction
mi
( )
X
mj
( )
X
m10
( )
X
+–+–
++++
Optimized
0.2
Readout
z = n1 n2
n10
···
···
+–+–
···
++++
n1
ni
nj
n10
4.0
3.5
G
)
z
H
M
(
2
4
3
2
j
,
i
J
1
1
5
i
9
0
2
4
8
6
i-j
d
F
0.4
0.3
0.2
3.0
4.0
3.5
J7,8 2 (MHz)
4.5
0.4
0.6
(μs)
0.8
1.0
j
j
(cid:2)
(cid:2)
j between sites. Four
j versus frequency w01
Fig. 2. Hamiltonian learning. (A) On-site interaction Ui
(cid:2)
(cid:2)
with measured values indicated by colored circles. (B) Hopping amplitude Ji;j
versus frequency with experimental data shown as markers (error bars indicate a
standard deviation). Colors represent the distance i (cid:2) j
gray-scale arrows specify frequencies in Fig. 3. (C) Localization length x extracted
by fitting the exponential decay of measured hopping rate [results of polynomial
fitting of datapoints in (B)] as a function of distance i (cid:2) j
different frequencies. The darkness of a marker matches a fitting curve in the
inset at the same frequency. In (A) to (C), green (purple) shading on the left
(right) corresponds to LBG (UBG), and theory curves (solid) are obtained from
numerical calculations using an identical circuit model. (D) Left (Right): Cartoon
illustrating hopping from a site inside the LBG (UBG) with positive or negative
sign represented in red or blue and the amplitude represented by opacity, with
alternating (all positive) signs denoted as þ (cid:2) þ(cid:2) (þ þ þþ). (E) Pulse
sequence for many-body evolution. Ten sites, Q1 to Q10, are initialized (X gate) in a
bit-string zinit at their idle frequencies, then tuned to resonance for time t during
j (insets) at a few
j
the interaction stage, followed by a site-resolved single-shot readout at their idle
frequencies to obtain a final bit-string z. (F) Many-body fidelity estimator Fd at
w01=2p ¼ 4:72 GHz versus evolution time t. The Fd curves assume Ji;j’s from two-
qubit measurement indicated by the dashed line in (B) with alternating signs
(orange) and all positive signs (green), and from the numerical optimization
(blue). The shading corresponds to a standard deviation for 40 randomly
chosen zinit’s in the five-excitation sector. Inset: Fd at t ¼ 0:6 ms versus J7;8
with optimized parameters on the remaining Ji;j’s, where the brown and the black
dashed lines indicate J7;8 extracted from (B) and the numerical optimization,
(cid:2)
(cid:2) interpolated from two-qubit measurements
respectively. (G) Comparison of Ji;j
(colored circles) and from numerical optimization (blue triangles with error bars
for 68% confidence interval), corresponding to the orange and the blue curves in
(cid:2)
(cid:2)’s are shown (details in the
(F), respectively. All nearest-neighbor hopping Ji;iþ1
j > 1 only the average values are
inset), whereas for larger hopping distances i (cid:2) j
indicated. The navy stars represent the maximum Fd at t ¼ 0:6 ms in (F) and the
optimized J7;8 in the insets of (F) and (G).
(cid:2)
(cid:2)
(cid:2)
(cid:2)
j
(cid:2)
(cid:2)
(cid:2)
(cid:2)
j, Ji;j
initializing one site with a p-pulse and tuning
w01 of both sites on resonance for a duration
τ with fast flux pulses. For a fixed distance
(cid:2)
(cid:2) increases with a decreasing Dj
j,
i (cid:2) j
j
resulting from larger photonic components
of the bound states (Fig. 2B). Compared to
(cid:2)
(cid:2) at the
the LBG, the UBG exhibits larger Ji;j
same Dj
j, owing to a stronger coupling g of the
bare qubits to the metamaterial at higher fre-
quencies and the breakdown of the tight-binding
cavity array model in the circuit realization
(Fig. 1D; also see supplementary text I). At a
(cid:2)
(cid:2) decreases exponentially as a
specific w01, Ji;j
function of distance i (cid:2) j
j , resembling the
j
profile of the photonic tail in a qubit-photon
bound state (Fig. 2C). From fitting the expo-
nential decay curve we extract the localization
length x, which ranges from x ¼ 1:4 to 4:2, with
the largest localization length occurring at
the smallest achievable band-edge detunings.
For even smaller detunings, the eigenstate of
(cid:2)
(cid:2)
the two interacting bound states merges into
the passband and becomes radiative to the
waveguide.
Many-body Hamiltonian learning
Beyond the above single- and two-qubit mea-
surements, we perform in situ many-body
characterization of Hamiltonian parameters
(32, 33), which are otherwise hard to access.
For example, the sign of the hopping term Ji;j
inherits the spatial profile of the photonic
component of the bound states. In the case of
the bound states in the UBG, the sign of the
hopping terms are all uniform (positive), where-
as for bound states in the LBG, the hopping
terms alternate sign as the distance between
lattice sites increases by one (Fig. 2D). This is
due to the photonic component of the bound
state behaving as a defect mode inside the
bandgap, exhibiting a spatial profile resem-
bling the wave vector at the nearest band edge
(k = 0 at the upper band-edge and k ¼ p=d at
the lower band edge). Although insignificant
in measurements involving only two lattice
sites, the sign of the hopping terms does alter
the many-body dynamics of the system. Here,
we use a many-body fidelity estimator Fd pro-
posed in (33) to reveal this information. This
fidelity estimator, which closely tracks the true
many-body fidelity, is obtained for ergodic
quench evolution of simple initial states (sup-
plementary text VI).
We follow the sequence described in Fig. 2E
to perform the many-body quench evolution.
The sequence consists of preparing a set of five
randomly chosen sites in their first excited
state, followed by using flux pulses to alignw01 of
all 10 sites for time t, and then finally per-
forming site-resolved single-shot measure-
ment on all lattice sites to obtain a 10-bit string
z ¼ n1n2⋯n10. The many-body fidelity estima-
tor Fd is calculated by comparing bit-string
Zhang et al., Science 379, 278–283 (2023)
20 January 2023
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
A
B
i
,
i
1
+
6
4
8
10
d
a
r
Integrable
4.50 GHz
4.55 GHz
4.72 GHz
4.80 GHz
Fig. 3. Two-particle quantum
walk with increasing hopping
range. (A) Evolution of the popu-
lation h^nii on sites Q1 to Q10
as a function of normalized evolu-
tion time Ji;iþ1 t. The system is
initialized in zinit ¼ 0000110000
and the evolution occurs at
w01=2p ¼ 4:50, 4.55, 4.72, and
4.80 GHz with the longest evolu-
tion times of 904, 781, 430,
and 200 ns from left to right. (B) The second moment m
2 as a function of normalized evolution time Ji;iþ1 t.
Results calculated from the data in (A) are shown as solid curves with gray scales corresponding to frames
in (A) and arrows in Fig. 2B. Result from numerical simulation of the integrable Hamiltonian is shown as
the dotted curve, and me
2 for a generic ergodic system is indicated by the red dashed line.
0.1
µ2
Q10Q1
Q10Q1
Q10Q1
Site
Site
Site
Site
Q10
µ2
Q1
2
1
0
J
0
1
e
statistics of repeated measurements with nu-
merical simulation of the evolution assuming
a set of Hamiltonian parameters in Eq. 1. The
maximum Fd is achieved at the parameter val-
ues closest to the Hamiltonian realized in the
experiment. The fast repetition rate of this ex-
periment enables us to perform a large number
of measurements (1.6 × 105 in total), reducing
statistical error and increasing sensitivity to
small Hamiltonian parameter variations (see
supplementary text V for details of qubit con-
trol, pulse sequence, and repetition rate).
We compare Fd at w01=2p ¼ 4:72 GHz using
three different parameter sets for Ji;j in Fig.
2F: a first set with amplitudes derived from the
two-qubit experiments in Fig. 2B assuming al-
ternating signs (Fig. 2D, left); a second set
with the same amplitudes as the first but all
positive signs (Fig. 2D, right); and a third
set of optimized parameter values that maxi-
mize Fd . The optimized hopping terms are
restricted to be real-valued, with independent
Ji;iþ1 for each i = 1 to 9 and Ji;j for each dis-
j > 1 (all qubit pairs of the same
tance i (cid:2) j
j
distance having the same Ji;j). An alternating
sign of Ji;j with distance is favored, yielding
a higher many-body fidelity compared to hop-
ping terms with all positive signs. This is fur-
ther evidenced by the alternating signs of the
resulting optimized parameter set. Although
we find small differences between the set of
optimized hopping amplitudes and those
from the two-qubit experiments with alter-
nating signs (Fig. 2G), Fd of the optimized
parameter set is markedly better. The sensitivity
of Fd to the hopping terms is highlighted in
the inset of Fig. 2F, where the variation of the
fidelity versus J7;8 is shown. For details of the
Fd calculation and parameter optimization,
refer to supplementary text VI.
Ergodic many-body dynamics with
long-range hopping
We now use the platform to study the effect of
long-range hopping on the many-body dynam-
ics. Specifically, the ergodicity of the 1D Bose-
j ≫ 1)
Hubbard model in the hardcore limit ( U =J
j
depends on the range of hopping, which ex-
hibits integrable behavior with NN hopping,
and chaotic behavior with long-range hopping.
We study this crossover with various hopping
ranges and investigate the resulting dynamics
using both conventional one- and two-site cor-
relators, and the statistics of the global bit-
strings resulting from qubit-state measurement
outcomes across the lattice. This latter tech-
nique is particularly useful in identifying uni-
versal signatures of ergodicity and the effect of
decoherence at long evolution times. The cross-
over between integrable and ergodic dynamics
can be qualitatively visualized by a two-particle
quantum walk (34–36) with initial excitations
on sites Q 5 and Q 6 using the sequence shown in
Fig. 2E. The measured quantum walk at a few
different w01’s indicated by arrows in Fig. 2B is
shown (Fig. 3A) as a function of normalized
evolution time Ji;iþ1t, where Ji;iþ1 is the av-
erage NN hopping rate (the corresponding
numerical simulations are provided in sup-
plementary text VII, showing that the quan-
tum walk patterns are not visibly affected
by decoherence). The excitation wave packets
smear over the system when w01 is close to the
band-edge frequency. More quantitatively,
this trend can be probed by computing the
probability pz of measuring a certain bit-string
z in the two-excitation sector at evolution
time t. For a generic ergodic Hamiltonian,
≡
the second moment m
z (25), which re-
2
flects the probability fluctuations, converges
to me
Þ after initial evolution (33)
owing to the chaotic nature of its quantum dy-
namics (D ¼ 45 is the dimension of the two-
excitation Hilbert space). No such convergence
is expected in an integrable Hamiltonian owing
to revivals associated with ballistic propaga-
tion of wave packets. As an example, we show in
Fig. 3B the results from the spin-1=2 XY model
obtained from modifying the Hamiltonian in
Eq. 1 by keeping only NN hopping terms in the
hardcore limit. When w01 is closer to the band
edge, the measured second moment deviates
from the simulated integrable result and con-
verges to me
2 at an earlier normalized evolution
2 ¼ 2= D þ 1
ð
zp2
X
j
time Ji;iþ1τ consistent with the breaking of
integrability due to the extended hopping
range. With U =J
j > 36 for all the measure-
ments illustrated in Fig. 3, finite on-site in-
teractions of the Bose-Hubbard model play a
negligible role in the breaking of integrabil-
ity (supplementary text VIII).
To further probe this ergodic nature of
Hamiltonian with long-range hopping, we
use the experimental evolution at w01=2p ¼
4:72 GHz as an example. At a short time (t ¼
16 ns), the excitations remain in their initial
sites. This is visualized for a quantum walk
with initial excitations on sites Q 5 and Q 6 in
the left panel of Fig. 4C (evolution of popula-
tion h^ni〉) and in the bottom left panel of Fig.
4D (two-site correlator h^ni ^nj〉). The histogram
Þ of experimentally measured bit-string
P pzð
probabilities pzf g at this early evolution stage
(Fig. 4B, left) shows a distribution with a tail
of large pz values, giving a large m2 (Experiment
curve in Fig. 4A). This is associated with an
insufficient scrambling of the initially local-
ized quantum information. At an intermediate
time (t ¼ 360 ns ), the excitations are more
spread out over the entire 1D lattice (Fig. 4D,
middle left panel), forming a “speckle” pattern
with site-to-site fluctuation that is associated
with quantum interference. The quantitative
signatures of this speckle pattern manifest
Þ following the Porter-
in the histogram P pzð
Thomas (PT) distribution (37) (Fig. 4B, mid-
dle) and in the second moment m2 settling to
the ergodic value me
2. The PT distribution re-
sults from the randomness in the distribution
of wavefunction magnitudes, which is pre-
dicted by Berry’s conjecture (38) stating that
the single-particle eigenstates of a chaotic sys-
tem behave like random superpositions of
plane waves. Similarly, in the many-body set-
tings, the distribution of wave function mag-
nitudes across basis states also follows the PT
distribution. Our observation is the first ex-
perimental verification of this many-body ver-
sion of Berry’s conjecture in a Bose-Hubbard
system, whose extension in the thermodynamic
limit provides the modern theory of quantum
thermalization such as eigenstate thermaliza-
tion hypothesis (39, 40). This draws a connec-
tion between quantum many-body chaos and
random matrix theory, leading to a deeper
understanding of the randomness in many-
body dynamics (32). Distinct from the ran-
domness inherent in random circuits (25, 41),
the randomness in our case originates from the
ergodicity of the time-independent Hamiltonian.
In contrast to the experimental results, theo-
retical calculations at the same evolution time
using the integrable Hamiltonian show aggre-
gated excitations on a few sites (Fig. 4E, middle)
and the resulting larger value of m
2 (Integrable
theory curve in Fig. 4A). This comparison high-
lights the effect of long-range hopping in prob-
ing the ergodic regime.
Zhang et al., Science 379, 278–283 (2023)
20 January 2023
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
A
1
2
µ
0.1
e
µ2
Experiment
Theory
Integrable theory
B
)
z
p
(
P
102
100
10-2
0.0 0.2 0.4 0.6 0.8
pz
0.0
0.05
pz
D
0.1
0.15
0.0
0.05
0.1
0.15
pz
C
s
μ
1.0
0.8
0.6
0.4
0.2
0.0
0.01
0.1
(μs)
1
Q1
Site
Q10 Q1
Site
Q10 Q1
Site
Q10
j
Q
e
t
i
S
j
Q
e
t
i
S
j
Q
e
t
i
S
1.0
0.8
0.6
0.4
0.2
0.0
0
1
Q
1
Q
0
1
Q
1
Q
0
1
Q
1
Q
Q1
Site Qi
Q10
Q1
Site Qi
Q10
E
j
Q
e
t
i
S
j
Q
e
t
i
S
j
Q
e
t
i
S
0
1
Q
1
Q
0
1
Q
1
Q
0
1
Q
1
Q
0.06
0.04
0.02
0.00
0.10
0.05
0.00
0.75
0.50
0.25
0.00
0.15
0.10
0.05
0.15
0.10
0.05
0.6
0.4
0.2
0.0
Q1
Site Qi
Q10
2 as a function of evolution time t in our system from
Fig. 4. Ergodic many-body dynamics with long-range hopping at 4.72 GHz.
(A) Second moment m
the experiment (orange) and the theory with the optimized parameter set
in Fig. 2G (blue), compared to theoretical predictions of the integrable model
(green). The shading on each curve corresponds to a standard deviation of the
mean second moment for 20 randomly chosen initial bit-strings zinit in the
two-particle sector, and the red dashed line represents the ergodic value me
2.
(B) Density histogram P pzð
probabilities pzf g with the 20 initializations zinit’s at evolution times t ¼ 16 ns,
Þ of the distribution of experimental bit-string
360 ns, and 5.4 ms from left to right [indicated by the dotted lines in (A)]. The
solid lines show the PT distribution, and the dashed line in the right plot shows
the value pz ¼ 1=D of a classical uniform distribution. (C) Evolution of the
population h^nii on sites Q1 to Q10 as a function of time t with zinit ¼ 0000110000
in the cases of experiment, theory, and integrable theory from left to right. The
white dashed lines at the bottom (in the middle) indicate t ¼ 16 ns (360 ns).
(D and E) Two-site correlator h^ni
t ¼ 16 ns, 360 ns, and 5.4 ms from bottom to top in the cases of experiment
[left column of (D)], theory [right column of (D)], and integrable theory (E).
^nji with zinit ¼ 0000110000 at evolution times
In addition, we study the impact of decoher-
ence by juxtaposing the measurement results
and the decoherence-free theoretical calcula-
tion using the optimal learned Hamiltonian
with long-range hopping. Before the evolution
time of t ≈ 1 ms, the two cases agree in the sec-
ond moment m
2 (Experiment and Theory curves
in Fig. 4A), the quantum walk population (Fig.
4C, left and middle), and the two-site correlator
(Fig. 4D, middle panels), suggesting that these
results are not affected by decoherence. After a
long evolution time (t ¼ 5:4 ms, larger than the
2;i ¼ 1:16 ms),
averaged Ramsey coherence time T (cid:4)
the second moment of the two cases deviates
from one another, and the experimental speckle
pattern begins to wash out compared to the
theoretical modeling (Fig. 4D, top panels). An-
other probe of the decoherence is the histogram
Þ of the measured bit-string probabilities
P pzð
(Fig. 4B, right). Here, the histogram deviates from
the PT distribution, narrows substantially, and
approaches a uniform distribution correspond-
ing to a completely decohered, maximally mixed
state. Additional numerical simulations of m2
and P pzð
Þ for ergodic and integrable systems
can be found in supplementary text VIII.
Conclusion and outlook
Our many-body quantum simulator is based
on a 1D lattice of transmon qubits connected
together using a superconducting metamate-
rial, which exhibits photonic bandgaps that
protect qubit-photon bound states from de-
cay and a transmission passband used for
high-fidelity multiplexed qubit-state readout.
Furthermore, the metamaterial plays the role
of a scalable photonic bus to mediate tunable
long-range coupling between qubit-photon
bound states. This system of interacting bound
states realizes a Bose-Hubbard model with long-
range hopping. We characterize the system
using conventional single- and two-qubit mea-
surements along with a sample-efficient many-
body Hamiltonian learning protocol. Lastly,
we study the many-body quench dynamics of
the system versus the range of the lattice hop-
ping, revealing the ergodic nature of the ex-
tended Bose-Hubbard model, as distinct from
its NN-coupling counterpart. The major chal-
lenge in probing long-time quantum evolu-
tion in our experiment is the short Ramsey
2;i ¼ 1:16 ms, limited by flux-
coherence time T (cid:4)
noise–induced dephasing. Incorporating a
single refocusing pulse has been shown to
increase the coherence time to T2E;i ¼ 5:64 ms
at the single-qubit level (supplementary text
II). The extended quantum evolution times
enabled by further dynamical decoupling,
combined with the tunable-range coupling
investigated in this work, provide valuable
opportunities to explore nonequilibrium dy-
namics with or without coupling to an envi-
ronment and quantum phases of matter in
the presence of frustration.
REFERENCES AND NOTES
1. K. Mølmer, A. Sørensen, Phys. Rev. Lett. 82, 1835–1838 (1999).
2. V. D. Vaidya et al., Phys. Rev. X 8, 011002 (2018).
3. A. Periwal et al., Nature 600, 630–635 (2021).
4. D. Bluvstein et al., Nature 604, 451–456 (2022).
5. S. Bravyi, D. Poulin, B. Terhal, Phys. Rev. Lett. 104, 050503
(2010).
6. N. Delfosse, M. E. Beverland, M. A. Tremblay, arXiv:2109.14599
[quant-ph] (2021).
7. P. Jurcevic et al., Nature 511, 202–205 (2014).
8. P. Richerme et al., Nature 511, 198–201 (2014).
9. R. Nandkishore, D. A. Huse, Annu. Rev. Condens. Matter Phys.
6, 15–38 (2015).
10. M. K. Joshi et al., Phys. Rev. Lett. 124, 240505 (2020).
11. R. Islam et al., Science 340, 583–587 (2013).
12. R. Landig et al., Nature 532, 476–479 (2016).
13. S. Ebadi et al., Nature 595, 227–232 (2021).
14. M. K. Joshi et al., Science 376, 720–724 (2022).
15. S.-B. Zheng, G.-C. Guo, Phys. Rev. Lett. 85, 2392–2395 (2000).
16. Q. Guo et al., Nat. Phys. 17, 234–239 (2021).
17. Y. P. Zhong et al., Nat. Phys. 15, 741–744 (2019).
18. J. S. Douglas et al., Nat. Photonics 9, 326–331 (2015).
19. D. E. Chang, J. S. Douglas, A. González-Tudela, C.-L. Hung,
H. J. Kimble, Rev. Mod. Phys. 90, 031002 (2018).
20. Y. Liu, A. A. Houck, Nat. Phys. 13, 48–52 (2017).
21. M. Mirhosseini et al., Nat. Commun. 9, 3706 (2018).
22. N. M. Sundaresan, R. Lundgren, G. Zhu, A. V. Gorshkov,
A. A. Houck, Phys. Rev. X 9, 011021 (2019).
23. M. Scigliuzzo et al., Phys. Rev. X 12, 031036 (2022).
24. E. Kim et al., Phys. Rev. X 11, 011015 (2021).
25. S. Boixo et al., Nat. Phys. 14, 595–600 (2018).
26. M. J. Hartmann, F. G. S. L. Brandão, M. B. Plenio, Nat. Phys. 2,
849–855 (2006).
27. G. Calajó, F. Ciccarello, D. Chang, P. Rabl, Phys. Rev. A 93,
033833 (2016).
28. S. John, J. Wang, Phys. Rev. Lett. 64, 2418–2421 (1990).
29. V. S. Ferreira et al., Phys. Rev. X 11, 041043 (2021).
30. J. Koch et al., Phys. Rev. A 76, 042319 (2007).
31. A. A. Houck, H. E. Türeci, J. Koch, Nat. Phys. 8, 292–299
(2012).
32. J. Choi et al., arXiv:2103.03535 [quant-ph] (2021).
Zhang et al., Science 379, 278–283 (2023)
20 January 2023
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
33. D. K. Mark, J. Choi, A. L. Shaw, M. Endres, S. Choi, arXiv:2205.
12211 [quant-ph] (2022).
34. A. Peruzzo et al., Science 329, 1500–1503 (2010).
35. P. M. Preiss et al., Science 347, 1229–1233 (2015).
36. Z. Yan et al., Science 364, 753–756 (2019).
37. C. E. Porter, R. G. Thomas, Phys. Rev. 104, 483–491 (1956).
38. M. V. Berry, J. Phys. Math. Gen. 10, 2083–2091 (1977).
39. J. M. Deutsch, Phys. Rev. A 43, 2046–2049 (1991).
40. M. Srednicki, Phys. Rev. E Stat. Phys. Plasmas Fluids Relat.
Interdiscip. Topics 50, 888–901 (1994).
41. F. Arute et al., Nature 574, 505–510 (2019).
42. Data and Simulation Code for, A superconducting quantum
simulator based on a photonic-bandgap metamaterial, version 1,
Zenodo (2022); https://doi.org/10.5281/zenodo.7502375.
ACKN OW LEDG MEN TS
The authors thank A. Gorshkov, A. González-Tudela, D. Chang,
O. Motrunich, R. Ma, F. Brandão, G. Refael, S. Meesala, V. Ferreira,
G. Kim, A. Butler, and Z. Zheng for helpful discussions. We
appreciate MIT Lincoln Laboratories for the provision of traveling-
wave parametric amplifiers used for both spectroscopic and time-
domain measurements in this work, and the AWS Center
for Quantum Computing for the Eccosorb filters installed in the
cryogenic setup for infrared filtering. We also thank the Quantum
Machines team for technical support and discussions on the
Quantum Orchestration Platform. Funding: This work was
supported by the AFOSR Quantum Photonic Matter MURI (grant
FA9550-16-1-0323), the DOE-BES Quantum Information Science
Program (grant DE-SC0020152), the Institute for Quantum
Information and Matter, an NSF Physics Frontiers Center (grant
PHY-1125565) with support of the Gordon and Betty Moore
Foundation, the Kavli Nanoscience Institute at Caltech, and the
AWS Center for Quantum Computing. D.K.M. acknowledges
support from the NSF QLCI program (2016245) and the DOE
Quantum Systems Accelerator Center (contract no. 7568717).
Author contributions: X.Z., E.K., D.K.M., S.C., and O.P. came
up with the concept. X.Z. and E.K. planned the experiment,
performed the device design and fabrication, and performed the
measurements. X.Z., E.K., D.K.M., and S.C. analyzed the data.
O.P. supervised the project. All authors contributed to the writing
of the manuscript. Competing interests: The authors declare
no competing interests. Data and materials availability:
Experimental data shown in the main text and supplementary
materials, as well as the simulation code, are available in Zenodo
(42). License information: Copyright © 2023 the authors,
some rights reserved; exclusive licensee American Association for
the Advancement of Science. No claim to original US government
works. https://www.sciencemag.org/about/science-licenses-
journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade7651
Materials and Methods
Supplementary Text
Figs. S1 to S14
Tables S1 and S2
References (43–88)
Submitted 6 September 2022; accepted 16 December 2022
10.1126/science.ade7651
Zhang et al., Science 379, 278–283 (2023)
20 January 2023
6 of 6
|
10.1126_science.ade7147
|
RES EARCH
CHEMISTRY
Modification of ground-state chemical reactivity via
light–matter coherence in infrared cavities
Wonmi Ahn1, Johan F. Triana2, Felipe Recabal2, Felipe Herrera2,3*, Blake S. Simpkins4*
Reaction-rate modifications for chemical processes due to strong coupling between reactant molecular
vibrations and the cavity vacuum have been reported; however, no currently accepted mechanisms
explain these observations. In this work, reaction-rate constants were extracted from evolving cavity
transmission spectra, revealing resonant suppression of the intracavity reaction rate for alcoholysis
of phenyl isocyanate with cyclohexanol. We observed up to an 80% suppression of the rate by tuning
cavity modes to be resonant with the reactant isocyanate (NCO) stretch, the product carbonyl (CO)
stretch, and cooperative reactant-solvent modes (CH). These results were interpreted using an open
quantum system model that predicted resonant modifications of the vibrational distribution of reactants
from canonical statistics as a result of light–matter quantum coherences, suggesting links to explore
between chemistry and quantum science.
C ontrolling chemical reactions with electro-
magnetic fields is a long-standing goal in
chemistry and physics (1, 2). Femtosecond
laser pulses can transiently excite vibra-
tional modes of reactant molecules to
selectively promote breaking or forming of
chemical bonds (3–5). However, fast energy
redistribution in polyatomic molecules se-
verely limits this approach, despite efforts
to overcome this obstacle using laser pulse
shaping (6, 7).
Chemical control without lasers has been
recently demonstrated using cavities (8–12).
In this approach, hybrid light–matter polar-
iton states arise from strong interactions of
dipole-allowed molecular transitions with the
cavity vacuum at optical (13) and infrared fre-
quencies (14–16). Experiments show inhibition
of excited-state processes such as photoisome-
rization (17) and photobleaching (18) in visible
cavities and also modification of bond forma-
tion and cleavage rates in infrared cavities as
a result of vibrational strong coupling (VSC)
(8–10, 19). VSC is characterized by collective
molecular response in transmission (20, 21),
a spatial dependence of the interaction that
follows the mode profile (22, 23), and revers-
ible modulation of the system using ultrafast
lasers (24, 25) or electrochemistry (26, 27).
Achieving coupling-induced selective chem-
istry would enable chemical catalysis by design,
but challenges to its reproducibility (28, 29)
and lack of mechanistic explanation have stifled
progress. Here, we report robust experimental
evidence of cavity-modified chemistry and de-
scribe a theory consistent with measurements.
1UNAM — National Nanotechnology Research Center and
Institute of Materials Science and Nanotechnology, Bilkent
University, Ankara, Turkey. 2Department of Physics, Universidad
de Santiago de Chile, Santiago, Chile. 3Millennium Institute for
Research in Optics (MIRO), Concepción, Chile. 4Chemistry
Division, US Naval Research Laboratory, Washington, DC, USA.
*Corresponding author. Email: blake.simpkins@nrl.navy.mil (B.S.S.);
felipe.herrera.u@usach.cl (F.H.)
We studied the alcoholysis of phenyl isocyanate
(PHI) with cyclohexanol (CHol) in tetrahydro-
furan (THF) to give urethane [cyclohexyl car-
banilate (CC)]. The reaction is exothermal (30),
has a low activation energy (31), and reso-
nant cavity modes can be tuned to reactant,
product, or solvent vibrational modes. We mea-
sured a strong cavity-tuning dependence of
the reaction kinetics, with rate constants re-
duced by 30 to 80%, and developed a quantum
model that qualitatively agrees with observa-
tions and provides mechanistic understanding
for intracavity reaction kinetics. Our theory
proposes that the intracavity reactivity de-
pends on stationary light–matter coherences,
and we discuss the importance of energy dis-
order in preserving coherence over chemical
time scales.
Results and discussion
The alcoholysis of isocyanates is well understood
(31–33) and proceeds through concerted nu-
cleophilic addition at the NC bond in isocya-
nate (31, 33, 34). The geometry of the PHI-CHol
complex (Fig. 1A) involves an NHO hydrogen
bond that evolves into a cyclic NHOC structure
in the transition state (3). Cleavage of the HO
bond results in ring opening and exothermic
formation of urethane (DHrxn ≈ −20.5 kcal/mol)
with activation energy of 6.7 kcal/mol (2343 cm−1)
in THF. The second-order rate constant at room
temperature is k0 = 0.59 × 10−5 M−1·s−1 (3). Back
reactions are negligible.
We injected the reactant solution into a
thin-layer cell bounded by transparent CaF2
windows, for out-of-cavity control measure-
ments, or by Au-coated CaF2 windows, for
cavity-coupled measurements [Fig. 1B; addi-
tional details in section 1 of the supplementary
materials (SM)]. Figure 1C shows two sets of
transmission spectra. For control measure-
ments (upper curves, red), reactant bands de-
creased (NCO stretch of PHI at ~2260 cm−1
and OH band of CHol at ~3470 cm−1), and
product bands grew (CO at 1730 cm−1 and NH
at 3293 cm−1; see detailed spectra in figs. S1 and
S2). The NCO band absorption was converted
to reactant concentration through direct pro-
portionality (see procedure in section 1 of the
SM and fig. S3 for extinction coefficient cali-
bration), then inverted and plotted against time
(Fig. 1D), yielding a line whose slope equaled
the second-order rate constant (35). The aver-
age of six such measurements gave a control
rate constant k0 = (2.34 ± 0.2) × 10−5 M−1·s−1
(datasets in fig. S4). This rate was higher than
in previous reports (32), which involved lower
reactant concentrations. The rates measured under
our conditions were consistent in independent
measurements performed in a period of 24
months, with all reactions (control and cavity-
coupled) carried out using the same initial
reactant concentrations.
Typical transmission spectra for a cavity-
coupled sample are shown in Fig. 1C (lower
curves, blue). These example data exhibited
multiple resonant peaks, with one coupled to
the NCO band of PHI (2260 cm−1), giving a
splitting at normal incidence of 112 cm−1 (cav-
ity Q ~ 100, cavity linewidth k ≈ 38 cm−1; see
table S2). These evolving transmission spectra
were fit to a function that accounted for the
absorbance of the intracavity medium (Fig.
1E), which, again, was directly proportional to
reactant concentration. We inverted and plot-
ted this data (Fig. 1F) to extract a rate constant
k = (1.48 ± 0.2) × 10−5 M−1·s−1 for this sample,
~37% lower than uncoupled controls. Collec-
tion of the entire cavity dispersion allowed
identification and fitting of spectra showing
strong interaction with the mode of interest
regardless of tuning at normal incidence
[see the model in section 3.3 of the SM and
(14, 21, 23, 25, 27)].
Reaction rates were extracted for different
cavities. The resulting “action spectrum” (9) is
shown in Fig. 2A. The initial (blue) and final
(orange) transmission spectra of the control
solution are shown to identify relevant vibra-
tional modes. There was a strong dependence
of the reaction rate on the cavity mode tuning,
with rate suppression due to VSC on reactant
(NCO), product (CO), and cooperative reactant-
solvent (CH) modes (full dataset in fig. S8 and
table S1). Cavity-induced suppression spanned
30 to 80%, relative to uncoupled controls.
The largest suppression was found for cavities
tuned to the NCO reactant mode, with a fre-
quency dependence that closely followed the
shape of the NCO absorption band. The rate
constants for far-detuned cavities (squares)
were close to the out-of-cavity rates (k/k0 ~ 0.91;
see table S1). Figure 2B shows representative
inverse concentration plots and linear fits
for cavities tuned to the reactant NCO and
reactant-solvent CH bands, highlighting the
lower slopes (rates), relative to out-of-cavity
controls. Our mechanistic discussion below
Ahn et al., Science 380, 1165–1168 (2023)
16 June 2023
1 of 4
RES EARCH | R E S E A R C H A R T I C L E
A
C
)
%
(
5
i
i
n
o
s
s
m
s
n
a
r
t
y
t
i
v
a
C
0
E
)
.
i
m
r
o
n
(
n
o
s
s
m
s
n
a
r
T
i
θ
100
80
60
40
20
0
ν
ν
ν
ν
ν
l
S
o
u
t
i
o
n
t
r
a
n
s
m
s
s
o
n
i
i
4000
3500
3000
2500
Frequency (cm-1)
(
%
)
2000
1500
data
fit
2350
2300
2250
Frequency (cm-1)
2200
B
D
)
1
-
M
(
]
I
H
P
[
e
s
r
e
v
n
I
F
3.0
2.5
2.0
1.5
1.0
0.5
)
1
-
M
(
]
I
H
P
[
e
s
r
e
v
n
I
1.4
1.3
1.2
1.1
0
0
data
fit
20
40
Time (103 s)
Time (103 sec)
60
80
data for "191127"
10
5
20
Time (103 s)
Time (103 s)
15
100
25
Fig. 1. Description of urethane monomer formation and reaction monitoring. (A) The reactants
phenyl isocyanate (PHI) and cyclohexanol (CHol) were combined in tetrahydrofuran (THF) to form cyclohexyl
carbanilate (CC). (B) Solution was contained between two CaF2 windows that were either transparent
(for control measurements) or coated with Au/SiO2 (for cavity-coupled experiments). (C) Time-dependent
Fourier transform infrared transmission spectra for out-of-cavity control measurements (red hues) showed
reactant absorptions, nNCO of PHI at 2260 cm−1 and nOH of CHol at 3470 cm−1, diminished as the reaction
proceeded, while product features, nCO at 1730 cm−1 and nNH at 3293 cm−1, increased. The nNCO absorption
was converted to PHI concentration, inverted, and plotted versus time to extract the second-order reaction
rate constant as shown in (D). The blue curves in (C) correspond to a time series of cavity-coupled
transmission spectra showing strong coupling between the cavity and NCO vibrational mode of the PHI reactant.
These spectra were fit, as shown in (E), to yield time-dependent PHI concentration, which was inverted,
plotted, and fit to yield the reaction rate constant under cavity-coupled conditions. One typical cavity-coupled
dataset is shown in (F).
focuses on the NCO band because it plays a
prominent role in the reaction, however, mod-
ification of one mode can influence others
(intramolecular vibrational relaxation, Fermi
coupling, etc.). Further, we note that the product-
coupled cavity also supported a higher-order
mode that weakly coupled to the reactant OH
mode (~3500 cm−1), however, we have only
highlighted modes under strong coupling. The
role of weak coupling in chemical reactivity
has yet to be fully understood.
Our mechanistic description of VSC-modified
reactivity started by modeling the vibrational
structure of the PHI molecule in the frequency
region of the NCO band, which includes a
fundamental NCO stretch, n6, and a Fermi
resonance between n6 and a combination of
low-frequency CH bending modes (analysis in
section 4 of the SM). The NCO fundamental at
2260 cm−1 was inhomogeneously broadened
(full width at half maximum ≈ 47 cm−1; see
fig. S13). We modeled an ensemble of N reac-
tant NCO vibrations under VSC at 300 K, con-
sidering vibrational relaxation, cavity decay,
thermalization, and many-body correlations,
using an open quantum system approach (see
sections 5 and 6 of the SM). Field-dependent
dipole self-energy terms (36) were not included.
The theory suggested that although the coupled
vibration-cavity system was at thermal equi-
librium with its environment, as confirmed by
experiments (37), when tracing out the pho-
tonic degrees of freedom, the stationary vibra-
tional population of reactants could deviate
T
r
a
n
s
m
s
s
o
n
(
i
i
%
)
A
)
1
-
s
1
-
M
5
-
0
1
(
k
2.4
2.0
1.6
1.2
0.8
0.4
B
)
d
e
t
f
i
h
s
,
1
-
M
(
]
I
H
P
[
e
s
r
e
v
n
I
0.4
0.3
0.2
0.1
0.0
0
100
50
0
ν
ν
ν
3200 3000 2800 2600 2400 2200 2000 1800
Frequency (cm-1)
5
10
Time (103 sec)
15
20
Fig. 2. Cavity-modified chemical reactivity.
The action spectrum presented in (A) shows the
extracted reaction rate constants (orange symbols)
as a function of cavity tuning (i.e., Fabry-Perot mode
position at normal incidence). The gray horizontal
band represents the average out-of-cavity control
rate with its width equal to the standard deviation
of six measurements. Blue and orange dashed
curves correspond, respectively, to initial and final
transmission spectra. Reaction suppression was
observed when the cavity was tuned to prominent
vibrational modes. Several example linear fits, from
which reaction rate constants were extracted, are
shown in (B).
from canonical Boltzmann statistics. This
phenomenon has been shown to occur for
other strongly coupled subsystems (38), but its
potential consequences in cavity chemistry
have yet to be fully explored. In this picture,
chemical bonds can break and form through
local two-body processes with the vibrational
level statistics modified by the strongly inter-
acting photonic environment. This insight can
be complemented by other approaches wherein
the cavity photon quadrature is treated as
another classical coordinate that contributes
to a static polaritonic potential energy surface
(39, 40).
In Fig. 3A, we show the deviation of the
most-probable n6 occupation from its cavity-
free canonical Boltzmann value, Dn6=nth
6 , as
a function of cavity frequency, for an ensemble
of PHI molecules (N = 50) with a Gaussian dis-
tribution of n6 mode frequencies (variance =
s2), coupled to a single cavity mode. Cavity-
modified rate measurements (orange points
reproduced from Fig. 2A) and the PHI trans-
mission spectrum (orange dashed curve) are
also shown. This single-mode theory predicted
a narrow vibrational depopulation feature (blue
Ahn et al., Science 380, 1165–1168 (2023)
16 June 2023
2 of 4
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Comparison of cavity-
modified reactivity with
theoretical prediction.
(A) The NCO region of the action
spectrum of Fig. 2 is reproduced
here. Single-mode theory pre-
dicted a resonant depopulation
of the n6 mode (blue points)
that qualitatively followed the
experimental data well. Example
cavity-induced population redis-
tribution is shown in (B). A
multimode single-molecule
treatment, solid black curve
in (A), yielded a resonant
depopulation effect that was
considerably stronger in magni-
tude (note the 9× scaling applied
to the single-mode analysis)
but was broader and blue-shifted
relative to the experimental
results. Although the cavity-
induced effect was predicted
to diminish with increased oscilla-
tor number, N, the scaling
power strongly depended on
molecular disorder (C). Molecular
disorder, s, was defined as a
Gaussian broadening of the
Lorentzian linewidth.
A
B
ω
circles, Dn6=nth
< 0) which closely followed
6
the transmission lineshape and was most prom-
inent at the frequency of the combination band
(2280 cm−1), in qualitative agreement with the
measured action spectrum. Figure 3B shows
that at large detuning, the vibrational occupa-
tion of n6 was symmetrically centered at the
canonical Boltzmann average, but near-resonant
cavities gave a skewed distribution of Dn6=nth
6 ,
whose most-probable value corresponded to
net vibrational depopulation (additional histo-
grams in fig. S14). The formal connection be-
tween Dn6=nth
6 and the reaction rate has yet to
be developed (see ansatz in section 6 in the SM).
The single-mode many-particle analysis qual-
itatively agreed with the experimental data and
improved our understanding of cavity-suppressed
reactivity. However, the predicted values of
Dn6=nth
6 were relatively small. This result was
due to the many-particle model not account-
ing for the full dispersion of the cavity field.
Treating large N and a continuum of cavity
modes is prohibitive, but we could gain insight
×
C
ω
σ
σ
σ
σ
N
larger than the single cavity mode approach,
suggesting that the entire photon spectrum con-
tributed to cavity chemistry phenomena.
Finally, we addressed the N-scaling of the
vibrational depopulation effect. Because the
total vibration-cavity system was in thermal
equilibrium (37), deviation of the photonic
occupation from a canonical distribution by
dn would correspond to a redistribution of
vibrational occupation per molecule of −dn/N.
Therefore, for typical values of N ~ 106 in
Fabry-Perot cavities, the population redistrib-
ution on individual molecules should be neg-
ligible. However, in Fig. 3C we showed that
this many-body dilution behavior did not hold
in general for ensembles with frequency dis-
order, by plotting Dn6=nth
6 as a function of N
for different values of the disorder width s (see
also fig. S15). We found that inhomogeneous
broadening could protect resonant popula-
tion redistribution from the homogeneous 1/N
scaling, possibly due to partial delocalization
of molecular states (41, 42).
X
by treating a single PHI molecule in a multi-
mode Fabry-Perot cavity with a quasi-continuous
spectrum wk, where k denotes the in-plane
wave number (details in section 6.2 of the SM).
We showed that Dn6=nth
6 in this case is pro-
E
D
(cid:3)
(cid:2)
†
^b6
Im ^ak
g6;k=g6
, where
portional to
E
D
†
^b6
^ak
is the stationary light–matter coher-
ence between the vibrational mode and the kth
cavity mode, g6,k is the Rabi frequency, and g6 is
the homogeneous vibrational linewidth. Solv-
ing for the steady-state coherence gave
ss
ss
k
Dn6=nth
6
¼
X
P wkð
k
(cid:5)
(cid:4)
Þ e(cid:2)Dk;6=kBT (cid:2) 1
ð1Þ
where Dk;6 ¼ wk (cid:2) wn6
is the detuning of n6
from the kth mode, and P(wk) is a normalized
(cid:3)
(cid:2)
2
distribution function scaling with g6;k=Dk;6
(cid:6)
(cid:6)
(cid:6) ≫ g6. Figure 3A (solid black curve)
(cid:6)
for Dk;6
shows that the frequency response from Eq. 1 was
much broader and blue-shifted relative to ex-
periments. However, integrating over the cavity
dispersion gave values of Dn6=nth
6 considerably
Ahn et al., Science 380, 1165–1168 (2023)
16 June 2023
3 of 4
RES EARCH | R E S E A R C H A R T I C L E
Conclusions
In this work, we suppressed a ground-state
addition reaction through strong coupling be-
tween molecular vibrational modes and cavity
vacuum fields. This suppression reached 80%,
and we found a strong cavity frequency de-
pendence that closely followed the reactant
infrared absorption spectrum. The strongest
effect was for cavities resonant with an NCO
mode that participated in the transition state
of the reaction. We described the mechanism
quantum mechanically as the emergence of
stationary noncanonical vibrational populations
as a result of strong vibration–cavity coupling
but noted that the composite vibration-cavity
polaritonic state remained in a Boltzmann
thermal state. Deviations of the vibrational
occupations from canonical statistics were due
to stationary light–matter coherences that de-
pended on the details of the cavity spectrum
and dispersion. For molecular ensembles, we
showed evidence that inhomogeneous spec-
tral broadening could protect the light–matter
coherences that influenced vibrational reac-
tivity, suggesting fundamental links between
chemistry and quantum science that have yet
to be fully developed.
10. F. J. Garcia-Vidal, C. Ciuti, T. W. Ebbesen, Science 373,
39. T. E. Li, J. E. Subotnik, A. Nitzan, Proc. Natl. Acad. Sci. U.S.A.
eabd0336 (2021).
117, 18324–18331 (2020).
11. B. S. Simpkins, A. D. Dunkelberger, J. C. Owrutsky, J. Phys. Chem. C
40. C. Schäfer, J. Flick, E. Ronca, P. Narang, A. Rubio, Nat.
125, 19081–19087 (2021).
12. F. Herrera, J. Owrutsky, J. Chem. Phys. 152, 100902 (2020).
13. G. Khitrova, H. M. Gibbs, M. Kira, S. W. Koch, A. Scherer,
Nat. Phys. 2, 81–90 (2006).
14. J. P. Long, B. S. Simpkins, ACS Photonics 2, 130–136 (2015).
15. A. Shalabney et al., Nat. Commun. 6, 5981 (2015).
16. J. Del Pino, J. Feist, F. J. Garcia-Vidal, New J. Phys. 17, 053040
(2015).
Commun. 13, 7817 (2022).
41. D. Wellnitz, G. Pupillo, J. Schachenmayer, Commun. Phys. 5,
120 (2022).
42. M. Du, J. Yuen-Zhou, Phys. Rev. Lett. 128, 096001 (2022).
43. J. F. Triana, F. Recabal, F. Herrera, B. S. Simpkins, W. Ahn,
Data: Modification of ground state chemical reactivity via light-
matter coherence in infrared cavities, version 2, Zenodo
(2023); https://doi.org/10.5281/zenodo.7915873.
17. J. A. Hutchison, T. Schwartz, C. Genet, E. Devaux,
T. W. Ebbesen, Angew. Chem. Int. Ed. 51, 1592–1596 (2012).
AC KNOWLED GME NTS
18. B. Munkhbat, M. Wersäll, D. G. Baranov, T. J. Antosiewicz,
T. Shegai, Sci. Adv. 4, eaas9552 (2018).
19. J. Lather, P. Bhatt, A. Thomas, T. W. Ebbesen, J. George,
Angew. Chem. Int. Ed. 58, 10635–10638 (2019).
20. B. S. Simpkins et al., ACS Photonics 2, 1460–1467 (2015).
21. C. Weisbuch, M. Nishioka, A. Ishikawa, Y. Arakawa, Phys. Rev. Lett.
69, 3314–3317 (1992).
22. W. Ahn, I. Vurgaftman, A. D. Dunkelberger, J. C. Owrutsky,
B. S. Simpkins, ACS Photonics 5, 158–166 (2018).
23. S. Wang et al., J. Phys. Chem. Lett. 5, 1433–1439 (2014).
24. A. D. Dunkelberger et al., ACS Photonics 6, 2719–2725 (2019).
25. R. Houdré et al., Phys. Rev. B 52, 7810–7813 (1995).
26. J. J. Pietron, K. P. Fears, J. C. Owrutsky, B. S. Simpkins, ACS
Photonics 7, 165–173 (2020).
27. W. Ahn, B. S. Simpkins, APL Photonics 5, 076107 (2020).
28. M. V. Imperatore, J. B. Asbury, N. C. Giebink, J. Chem. Phys.
154, 191103 (2021).
29. G. D. Wiesehan, W. Xiong, J. Chem. Phys. 155, 241103 (2021).
30. E. Delebecq, J.-P. Pascault, B. Boutevin, F. Ganachaud,
Chem. Rev. 113, 80–118 (2013).
31. F. Kössl, M. Lisaj, V. Kozich, K. Heyne, O. Kühn, Chem. Phys. Lett.
621, 41–45 (2015).
32. G. Raspoet, M. T. Nguyen, M. McGarraghy, A. F. Hegarty, J. Org. Chem.
We thank J. Owrutsky, A. Dunkelberger, I. Vurgaftman, and
J. Schachenmayer for discussions. Funding: This work was
supported by US Naval Research Laboratory Nanoscience Institute,
grant WU 1J03 (W.A. and B.S.S.); ANID Fondecyt Regular grant
1221420 (F.H.); ANID Fondecyt Doctorado grant 21221970
(F.R.); ANID Fondecyt Iniciación grant 11230679 (J.T.); Millennium
Science Initiative Program grant ICN17_012 (F.H. and J.T.); and
Programa de Cooperación Científica ECOS-ANID ECOS grant
200028 (F.H.). Author contributions: Conceptualization: B.S.S.
and F.H. Methodology: B.S.S., F.H., and W.A. Investigation: W.A.,
J.F.T., F.R., F.H., and B.S.S. Visualization: B.S.S., J.F.T., and
F.R. Funding acquisition: B.S.S. and F.H. Writing – original draft:
W.A. and B.S.S. Writing – review & editing: B.S.S. and F.H.
Competing interests: None declared. Data and materials
availability: All data needed to support the conclusions of the
main text and supplementary materials have been uploaded to
Zenodo (43). License information: Copyright © 2023 the authors,
some rights reserved; exclusive licensee American Association
for the Advancement of Science. No claim to original US
government works. https://www.science.org/about/science-
licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade7147
Materials and Methods
Supplementary Text
Figs. S1 to S19
Tables S1 to S4
References (44–77)
Submitted 2 September 2022; resubmitted 7 March 2023
Accepted 12 May 2023
10.1126/science.ade7147
RE FE RENCES AND N OT ES
63, 6878–6885 (1998).
1. A. H. Zewail, Phys. Today 33, 27–33 (1980).
2. W. S. Warren, H. Rabitz, M. Dahleh, Science 259, 1581–1589 (1993).
3. T. Stensitzki et al., Nat. Chem. 10, 126–131 (2018).
4. M. Delor et al., Science 346, 1492–1495 (2014).
5. Z. Lin et al., J. Am. Chem. Soc. 131, 18060–18062 (2009).
6. A. Assion et al., Science 282, 919–922 (1998).
7. M. P. A. Branderhorst et al., Science 320, 638–643 (2008).
8. A. Thomas et al., Angew. Chem. Int. Ed. 55, 11462–11466 (2016).
9. A. Thomas et al., Science 363, 615–619 (2019).
33. A. A. Caraculacu, S. Coseri, Prog. Polym. Sci. 26, 799–851
(2001).
34. M. Çoban, F. A. S. Konuklar, Comput. Theor. Chem. 963,
168–175 (2011).
35. K. J. Laidler, Chemical Kinetics (Pearson Education Inc., 1987).
36. C. Schäfer, M. Ruggenthaler, A. Rubio, Phys. Rev. A 98, 043801
(2018).
37. M. Seidel et al., ACS Photonics 6, 1823–1825 (2019).
38. D. Xu, J. Cao, Front. Phys. (Beijing) 11, 110308 (2016).
Ahn et al., Science 380, 1165–1168 (2023)
16 June 2023
4 of 4
|
10.1126_science.ade5337
|
tion applied—much faster than the timescale
over which the data and spectator qubits de-
correlate. This second requirement has limited
the experimental implementation of such pro-
tocols because a substantial number of mea-
surements are required to reliably estimate
the effects of a dynamic noise environment.
Furthermore, the spectator qubit readouts must
be performed mid-circuit without perturbing
the data qubits.
In this work, we overcome these challenges
and demonstrate real-time correction of cor-
related phase errors using a dual-species array
of individually trapped neutral atoms. The
protocol is outlined in Fig. 1A. Data qubits
(rubidium atoms) and spectator qubits (cesium
atoms) are laser-cooled into optical tweezer
arrays (26). During logic operations on the
data qubits, mid-circuit readouts (MCRs) on
the array of ~60 spectator qubits enable single-
shot estimation of globally correlated phase
errors. The readout results are processed in
real time and used to infer the noise-induced
phase accrued by the ~60 data qubits. Crucially,
owing to the cross-talk–free operation of the
two species, these readouts do not disturb the
coherence of the data qubits. We leverage a
classical control architecture to perform in-
sequence feedforward, such that correlated
errors on the data qubits are mitigated within
the execution of the quantum circuit. Finally,
we show that the spectator qubits can be
RES EARCH
QUANTUM INFORMATION
Mid-circuit correction of correlated phase errors
using an array of spectator qubits
K. Singh1†, C. E. Bradley2†, S. Anand2†, V. Ramesh3, R. White3, H. Bernien2*
Scaling up invariably error-prone quantum processors is a formidable challenge. Although quantum error
correction ultimately promises fault-tolerant operation, the required qubit overhead and error thresholds are
daunting. In a complementary proposal, colocated, auxiliary “spectator” qubits act as in situ probes of
noise and enable real-time, coherent corrections of data qubit errors. We used an array of cesium spectator
qubits to correct correlated phase errors on an array of rubidium data qubits. By combining in-sequence
readout, data processing, and feedforward operations, these correlated errors were suppressed within
the execution of the quantum circuit. The protocol is broadly applicable to quantum information platforms
and establishes key tools for scaling neutral-atom quantum processors: mid-circuit readout of atom
arrays, real-time processing and feedforward, and coherent mid-circuit reloading of atomic qubits.
susceptible to the same noise sources. Specta-
tor qubits act as in situ probes of that noise,
such that measurement and feedforward can
be used to coherently protect the data qubits
during the execution of a quantum algorithm
(23–25). Notably, under two key conditions,
spectator protocols are agnostic to the spec-
trum and correlation time of the noise source.
First, the noise-induced dynamics must be
correlated between the spectator and data
qubits. Second, an estimate of those dynam-
ics must be made by reading out the spectator
qubits—and a subsequent feedforward opera-
R ealizing large-scale programmable quan-
tum systems that can overcome inevita-
ble noise sources is a central challenge
for modern physics (1, 2). Environmental
noise and experimental parameter drift
necessitate strategies to reduce their impact
and overcome resulting qubit errors. Although
quantum error correction will ultimately be
required, achieving the necessary qubit opera-
tion fidelities is an outstanding challenge for
present quantum computing platforms (3–9).
Moreover, the effectiveness of error-correction
codes is reduced by correlated errors (10, 11),
which may naturally occur when the qubits
are in close spatial proximity or are controlled
by shared hardware (12–16).
To address these challenges, a number of
techniques have been developed to mitigate
the effects of noise, such as composite pulses
(17), optimal control (18), dynamical decou-
pling (17, 19), Hamiltonian learning (20), and
machine learning–based control engineering
(21). These techniques have found great suc-
cess, but they are typically tailored to specific
noise models or require careful calibration and
thus face challenges when used in realistic, fluc-
tuating environments. For example, dynamical
decoupling generates a filter function that miti-
gates a particular spectrum of noise, with pass-
bands remaining that are not suppressed (22).
Additionally, it is only effective if the correla-
tion time of the noise is long with respect to the
interpulse delay.
Recent theoretical work has proposed a com-
plementary technique based on “spectator”
qubits—additional qubits that are colocated
with the computational “data” qubits and are
1Intelligence Community Postdoctoral Research Fellowship
Program, Pritzker School of Molecular Engineering,
University of Chicago, Chicago, IL 60637, USA. 2Pritzker
School of Molecular Engineering, University of Chicago,
Chicago, IL 60637, USA. 3Department of Physics, University
of Chicago, Chicago, IL 60637, USA.
*Corresponding author. Email: bernien@uchicago.edu
†These authors contributed equally to this work.
Fig. 1. Spectator qubit protocol with a dual-species atom array. (A) Feedforward loop for real-time correction
of correlated phase errors between data qubits (Rb atoms, blue) and spectator qubits (Cs atoms, yellow). A mid-circuit,
single-shot phase estimation on the spectators is used to infer the noise-induced phase accrued by the data qubits.
This information enables a real-time correction on the data qubits before the final readout, which suppresses
dephasing. Subsequently, spectator qubits lost during readout can be replenished while maintaining data qubit
coherence. (B) Example fluorescence image of the dual-species atom array. Scale bar indicates ~10 mm. (C) Microwave
Rabi oscillations of the data and spectator qubits. Dashed lines are fits to exponentially decaying sinusoids.
Singh et al., Science 380, 1265–1269 (2023)
26 June 2023
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
replenished within the data qubit coherence
time, an essential step toward repeated mea-
surements and the continuous operation of
atom-based quantum processors.
MCR of spectator qubits
Our experiment is performed on arrays of 10-
by-10 and 11-by-11 sites for the spectator and
data qubits, respectively (Fig. 1B), which are
stochastically loaded with an average loading
fraction of ~55%. The experimental apparatus
has been upgraded from our previous work (26)
to incorporate qubit initialization, manipulation,
and readout, along with classical hardware to
implement real-time processing and feedfor-
ward. Here, the qubits are encoded into long-
lived hyperfine states ( jF ¼ 1; mF ¼ 0i :¼
j0i
Rb for Rb;
jF ¼ 3; mF ¼ 0i :¼ j0i
Cs and jF ¼ 4; mF ¼
0i :¼ j1i
Cs for Cs, where F is the total angular
momentum and m
F is the magnetic quantum
number). Microwave driving of the data and
spectator qubits after optical pumping into
j1i
Cs reveals coherent Rabi oscillations
(Fig. 1C).
Rb and jF ¼ 2; mF ¼ 0i :¼ j1i
Rb andj1i
An essential ingredient for the spectator
protocol is to perform MCR of the spectator
qubits without inducing additional data qubit
decoherence. This is challenging in single-
species atom arrays because all atoms are
resonant with the excitation laser and the
measured qubits scatter light, which can de-
cohere the data qubits through reabsorption.
To overcome this, several ideas have been pro-
posed and demonstrated, including coherently
transporting qubits into readout cavities (27)
or using additional shelving states to hide
atoms from excitations from the readout light,
as demonstrated for trapped ions (4). How-
ever, realizing cross-talk–free imaging in large
atom arrays has remained an outstanding
challenge. A key motivation behind the dual-
species approach is that the different atomic
species have distinct optical transitions, and
measurements on one species are not expected
to influence the other (26, 28).
In a first experiment, we characterized the
spectator qubit MCR and measured its impact
on the data qubit coherence. The quantum
circuit is shown in Fig. 2A. During an XY8
decoupling sequence on the data qubits, an
XY4 sequence was performed on the spectators.
The spectator qubits were measured within
the XY8 sequence by selectively removing all
atoms in the j1i
Cs state by a resonant laser pulse
and then fluorescence imaging for 15 ms. The
coherences of the data and spectator qubits
as a function of their individual decoupling
times are shown in Fig. 2, B and E, respec-
tively. Although the camera exposure time is
fixed, the imaging light is applied for a var-
iable time, 5t (of a total of 16t), to determine
its effect on the data qubits. Crucially, the data
qubit coherence time is unaltered by the MCR
Fig. 2. MCR of atomic qubits. (A) Pulse diagram depicting MCR of atomic qubits. Lowercase and uppercase
letters indicate p/2 and p pulses, respectively, along that axis of rotation. The readout light is left on for the remaining
duration of the sequence after MCR. (B) Measurement of spectator qubit dynamics while preserving data qubit
coherence. During an XY8 decoupling sequence on the data qubits (red diamonds), we performed an XY4
decoupling sequence and subsequent projective measurement on the spectator qubits. The data qubit coherence
q
(
) is unchanged in the absence of MCR (blue circles). Dashed lines are fits (29). The inset shows
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
hsxi2 þ hsyi2
coherence measurements for early (square) and late (triangle) evolution times. (C) Example fluorescence histogram
of a spectator qubit. Solid lines are fits to a bimodal Poisson distribution. (D) Cumulative histogram of the
discrimination infidelities of the spectator qubits during MCR (29). eCDF, empirical cumulative distribution function.
(E) Coherence of spectator qubits. The measured spectator coherence time is TXY4
= 136(7) ms.
2
2;MCR = 0.68(1) s, T XY8
[fitted T XY8
2;No MCR = 0.65(2) s].
The large detuning of the imaging light leads
to negligibly low spontaneous scattering rates
of ~10−7 Hz. Moreover, spontaneous Raman
scattering events that change these m
F states
are further suppressed by a factor of 0.009
owing to destructive interference of the off-
resonant transition amplitudes (12). The the-
oretical T
time from this decay channel is
thus ~108 s, resulting in a data qubit bit-flip
rate from readout cross-talk of ~10−11 during
the 15-ms MCR. This readout duration was
chosen to balance the requirements for achiev-
ing a high discrimination fidelity while mini-
mizing the time for a feedforward operation
(29). The discrimination fidelity of the specta-
tor qubit states (Fig. 2D) is extracted from a
1
bimodal fit to the fluorescence histogram of
each spectator qubit, as exemplified in Fig. 2C.
Across the spectator array, we find a mean
fidelity of 0.989(5), showing that the spectator
qubit states are well resolved by MCR.
Spectator protocol and correction
of phase errors
The preservation of data qubit coherence dur-
ing spectator readout opens the possibility for
feedforward operations within a quantum cir-
cuit. Under simultaneous evolution, noise chan-
nels can induce correlated phase errors between
the data and spectator qubits. Importantly, the
large number of spectator qubits allows single-
shot estimation of the acquired phase from one
simultaneous MCR. The phase accrued by the
Singh et al., Science 380, 1265–1269 (2023)
26 June 2023
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Mid-circuit correction of correlated phase errors. (A) Noise channels
induce correlated phase errors (red arrows) between the two sets of qubits.
Measurement of the spectators along the y axis enables single-shot phase
estimation, from which the phase accrued by the data qubits can be inferred
and corrected in real time (green arrow). (B) Gate sequence. The data and
spectator qubits are synchronously decoupled and acquire correlated errors
owing to magnetic field noise dBz. The spectator qubit decoupling sequence is
truncated, with the remaining time assigned for MCR and feedforward. CPU,
central processing unit; QPU, quantum processing unit. (C) Example coherence
AC
= 36.2 Hz and 10.7 mG RMS. Dashed lines are fits, from which
measurement of the data qubits at the end of the sequence, with the feed-
forward turned on (green squares) and off (blue triangles). Field noise is
applied at f
we extract hsxi = 0.53(1) and 0.02(2) for feedforward on and off, respectively.
(D) Data qubit hsxi as a function of the RMS noise strength at fAC. The shaded
green region indicates the correctable range (see text). (E) Data qubit hsxi
as a function of the noise frequency at 10.7 mG RMS. The shaded gray region
indicates an absolute gain in the measured coherence. For (D) and (E), solid
lines are the results of numerical simulations (see text).
data qubits can then be inferred and corrected
in real time, as illustrated in Fig. 3A.
To demonstrate this capability, we injected
global magnetic field noise with amplitudes
and frequencies comparable to those typically
found in laboratory environments. The phase
of the noise was random in each experimen-
tal repetition, without shot-to-shot temporal
correlations. We focused on monochromatic
noise for ease of synthesis and interpretation
of protocol performance but note that our
scheme is generally agnostic to the noise spec-
trum. The pulse sequence for the experiment
is shown in Fig. 3B. The data and spectator
qubits underwent synchronous dynamical de-
coupling and acquired correlated errors from
the common noise. Although the filter func-
tion of the Carr-Purcell-Meiboom-Gill (CPMG)–
type dynamical decoupling sequence partially
mitigates such noise, certain frequencies still
couple into the sequence, occurring at odd-
harmonics of f
= 1/(4t) = 36.2 Hz, where 2t
AC
is the time between p-pulses (22). The specta-
tors sample this noise for three-quarters of the
total evolution time of the data qubits, with the
remainder of the time assigned for MCR and
feedforward. To achieve fast camera process-
ing and feedback, we used a camera-linked
classical control architecture for in-sequence
processing of the fluorescence images, which
in turn triggers an arbitrary-waveform gener-
ator to perform real-time updates of the phase
of the final data qubit p/2 pulse (29). The
phase update of this final p/2 pulse is equiv-
alent to a z-axis qubit rotation, which is used
to correct the noise-induced phase error on
the data qubits.
To estimate the phase acquired by the spec-
tators, FS, MCR was performed along an axis
orthogonal to the state preparation axis. Ac-
cordingly, the collective expectation value of
the array can be inverted to give an estimate,
F0
Þ, where C is a scaling fac-
S
tor that describes the amplitude of the signal
ð
¼ arcsin hsyi=C
in the absence of injected noise (29). F0
S is
uniquely defined when the accrued phase lies
within [−p/2, p/2], beyond which the protocol
breaks down. The estimated noise-induced
phase accrued by the data qubits is given by
F0
S , where g = 4/3 is the ratio of the
D
sensing times and b = 1.35 is the ratio of the
second-order Zeeman shifts of the clock states
(29). With this knowledge, a real-time correc-
tion can be applied.
¼ gbF0
AC
We first probed the case for which the noise
is maximally coupled, at f
[10.7 mG root
mean square (RMS)]. Without the spectator
protocol, the random phase of the noise leads
to complete dephasing of the data qubits.
Notably, the feedforward corrects the noise-
induced phase in each experimental repeti-
tion, resulting in a recovery of the data qubit
coherence (Fig. 3C). The coherence as a func-
tion of the noise amplitude is shown in Fig. 3D.
In stark contrast to the rapid decay observed
in the absence of feedforward, the spectator
Singh et al., Science 380, 1265–1269 (2023)
26 June 2023
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Reloading of spectator qubits while maintaining data qubit
coherence. (A) Reloading spectators using a pulsed MOT while decoupling the
data qubits. The data qubit coherence time is TXY4
= 0.42(3) s with the pulsed
2
MOT and TXY4
= 0.45(1) s without it. Spectators are reloaded on a timescale of
150(50) ms (time required to reach 1 − 1/e of asymptote), saturating at a
2
loading fraction of 0.49. (B) Reloading spectators using PGC during data
qubit decoupling. The data qubit coherence time is TXY8
PGC light and TXY8
= 0.65(2) s without it. Reloading occurs on a faster timescale
of 90(30) ms compared with (A), saturating at a fraction of 0.32. Dashed lines
are fits (29).
= 0.64(5) s with the
2
2
protocol robustly preserves coherence for field
strengths less than 11 mG. Beyond this value,
the accrued phases on the spectator qubits can
exceed ±p/2, where the protocol can no longer
unambiguously detect phase errors.
AC
AC
Next, we studied the dependence on the noise
frequency for an RMS noise strength of 10.7 mG
(Fig. 3E). For a range of frequencies close to
f
, real-time correction results in an absolute
gain in the measured signal, shielding the data
qubits from otherwise deleterious decoherence.
A pair of small additional features occur near
in the “feedforward-on” spectrum, arising
f
from the finite spectator readout time, which
leads to decorrelation between the data and
spectator qubits. Reducing the fraction of time
used for MCR would suppress these effects.
Outside this region, feedforward causes a slight
reduction in the measured coherence, which
results from imperfect phase estimation. For
both the amplitude and frequency sweep, the
salient features of the data are well described by
simple simulations of the experiment with no
free parameters aside from a global amplitude
rescaling (Fig. 3, D and E). These simulations
are based on the assumption of monochromatic
noise that solely perturbs the frequencies of the
qubits (29). At stronger noise strengths, a slight
discrepancy occurs, which likely arises from a
breakdown of these assumptions.
Alongside our numerical simulations, analytic
expressions can be derived for the error due to
quantum projection noise (QPN) in the phase-
estimation step. In the absence of any correlated
dephasing, QPN-induced feedforward errors
modulate the data qubit expectation values
hsxi by f ≈ 1 (cid:2) g2b2
2NC2 (29). For our experimental
parameters (C = 0.46, N = 61, where N is the
average number of loaded spectator qubits),
we find f ≈ 0.88, which is in good agreement
with the numerical simulations.
In the context of quantum information pro-
cessing, it is interesting to consider the re-
quirements to reach f ≫ 0:99 . Without any
change in g or b, f = 0.99 could be achieved for
N = 165 and C = 1. At present, the value of C is
limited primarily by uncorrelated dephasing of
the spectator qubits, which is caused by thermal
motion in the optical tweezers and tweezer-
induced T
processes. Thermal motion can be
reduced by additional cooling schemes, and T
can be improved by increased detuning of the
optical tweezers.
1
1
Beyond optimizing for g ≈ 1, f can be fur-
ther improved by reducing b, at the cost of a
reduced range of correctable data qubit er-
rors, FD;max ¼ Tgbp=2. This could be achieved
with alternative spectator qubit states, such
as magnetic field–sensitive states.
Although in this work we focus on mag-
netic field noise, the protocol can also mitigate
common-mode control errors. For instance, by
cotrapping the data and spectator qubits using
the same laser system (such as a far-detuned
1064-nm laser), phase errors induced by in-
tensity fluctuations of the trapping-laser light
could be corrected.
Coherent reloading of spectator qubits
In these experiments, fluorescence-based de-
tection of the spectators involves selectively
removing those in the j1i
Cs state before imag-
ing. Therefore, performing repetitive MCRs
will continuously deplete the array. Although
low-loss readout techniques exist (30, 31), fi-
nite losses always remain from both the read-
out itself and the trapping lifetime. Therefore,
continuous operation of atom-based quantum
processors will require reload and reset op-
erations that overcome these erasure errors
(32, 33). In this work, we explored two meth-
ods for reloading spectators while maintaining
coherent data qubits. These build on our stan-
dard procedure, in which a two-dimensional
magneto-optical trap (MOT) generates a beam
of atoms that is laser-cooled into the tweezer
array via a three-dimensional MOT.
2
2
The first reloading approach uses a strobo-
scopic MOT that is applied synchronously with
an XY4 sequence on the data qubits to de-
couple them from the magnetic field gradient
(Fig. 4A). Without the gradient, this decou-
pling sequence gives T XY4
= 0.45(1) s. With it,
we find T XY4
= 0.42(3) s, but the functional
form is modified (29). The spectator array
is reloaded on a much shorter timescale of
150(50) ms, defined as the time taken to reach
1 − 1/e of the asymptotic loading fraction.
The pulsed MOT saturates at a loading frac-
tion of 0.49, which is comparable to that achieved
with the standard procedure. Residual de-
phasing from the field gradient can be over-
come by using low-inductance coils with faster
switching times and by performing decou-
pling pulses using a Raman laser system, which
would enable Rabi frequencies in the mega-
hertz range.
In the second approach, we used polarization-
gradient cooling (PGC) to load spectators di-
rectly from the atomic beam without a field
gradient (Fig. 4B). This both increases the load-
ing speed and allows an arbitrary choice of
decoupling parameters: Here, we used a single
cycle of XY8. In this reloading paradigm, the
Singh et al., Science 380, 1265–1269 (2023)
26 June 2023
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
2
data qubit coherence time [T XY8
= 0.64(5) s]
is unchanged from the values presented in
Fig. 2, and the spectator qubit array is re-
loaded on a timescale of 90(30) ms. The frac-
tion of total reloaded spectators is lower than
that in the previous method, saturating at
0.32. We hypothesize that this is limited by the
2-mm-diameter cooling beams. Incorporating
larger cooling beams will likely increase the
loading fraction for both approaches and would
enable reloading times of a few tens of milli-
seconds (34). Coherence times on the order of
seconds can be achieved by using further de-
tuned trapping light and a larger number of
decoupling pulses (9).
Discussion and outlook
A central challenge for all quantum architec-
tures is to increase system sizes while maintain-
ing low physical error rates. Our demonstration
of the use of spectator qubits to measure and
correct correlated phase noise is a broadly ap-
plicable strategy that can be used to reduce
error rates in quantum computing platforms.
Furthermore, spectator protocols could be
used in conjunction with standard quantum
error correction strategies to protect against
correlated errors as well as increase the fi-
delity of operations beyond the fault-tolerance
threshold. An attractive feature of this protocol
is that it does not necessitate interactions (two-
qubit gates) or individual spectator qubit con-
trol, which reduces hardware complexity. The
use of spectator qubits for noise measurements
may provide opportunities in quantum sens-
ing and metrology (22, 35, 36) and for im-
proving clock coherence within a single device
through differential spectroscopy between the
data and spectator qubits (37). Whereas in
this work we focused on global noise, arrays
of spectator qubits may also enable the de-
tection of spatially varying noise fields that
can be suppressed by local qubit addressing
(24). Careful engineering of the spectator qubits
and their control sequences may improve pro-
tocol performance. For example, spectator qubits
could be encoded in states with enhanced or
reduced noise sensitivity to increase the phase
resolution or the range of tolerable noise (25).
This can be achieved by using nonzero m
F states
or by entangling the spectator qubits (22).
The methods demonstrated in this work
constitute a set of quantum-control techniques
that are essential for atom-array quantum pro-
cessors, including MCR, feedforward opera-
tions, and the reloading of auxiliary qubits
while maintaining quantum data. Combining
these capabilities with programmable intra-
species (9, 38) and interspecies Rydberg gates
will enable auxiliary qubit–assisted measure-
ments as required for quantum error correction
(32, 33, 39) and efficient preparation of long-
range entangled states (40). These same ca-
pabilities also enable the exploration of complex
dynamical quantum behavior under continuous
observation, including measurement-induced
phase transitions (41).
Note added in proof: After the consideration
of this manuscript, a preprint paper was pub-
lished that describes MCR of a neutral atom
by shelving in hyperfine states (42).
RE FERENCES AND NOTES
1. E. T. Campbell, B. M. Terhal, C. Vuillot, Nature 549, 172–179 (2017).
2. B. M. Terhal, Rev. Mod. Phys. 87, 307–346 (2015).
3. C. Ryan-Anderson et al., Phys. Rev. X 11, 041058 (2021).
4. L. Postler et al., Nature 605, 675–680 (2022).
5. Y. Zhao et al., Phys. Rev. Lett. 129, 030501 (2022).
6. S. Krinner et al., Nature 605, 669–674 (2022).
7. R. Acharya et al., Nature 614, 676–681 (2023).
8. M. H. Abobeih et al., Nature 606, 884–889 (2022).
9. D. Bluvstein et al., Nature 604, 451–456 (2022).
10. J. Preskill, Quantum Inf. Comput. 13, 181–194 (2013).
11. A. G. Fowler, J. M. Martinis, Phys. Rev. A 89, 032316 (2014).
12. S. Kuhr et al., Phys. Rev. A 72, 023406 (2005).
13. T. Monz et al., Phys. Rev. Lett. 106, 130506 (2011).
14. C. E. Bradley et al., Phys. Rev. X 9, 031045 (2019).
15. J. M. Boter et al., Phys. Rev. B 101, 235133 (2020).
16. C. D. Wilen et al., Nature 594, 369–373 (2021).
17. L. M. Vandersypen, I. L. Chuang, Rev. Mod. Phys. 76,
1037–1069 (2005).
18. H. A. Rabitz, M. M. Hsieh, C. M. Rosenthal, Science 303,
1998–2001 (2004).
19. T. Gullion, D. B. Baker, M. S. Conradi, J. Magn. Reson. 89,
479–484 (1990).
20. M. D. Shulman et al., Nat. Commun. 5, 5156 (2014).
21. S. Mavadia, V. Frey, J. Sastrawan, S. Dona, M. J. Biercuk,
Nat. Commun. 8, 14106 (2017).
22. C. L. Degen, F. Reinhard, P. Cappellaro, Rev. Mod. Phys. 89,
035002 (2017).
23. S. Majumder, L. Andreta de Castro, K. R. Brown, NPJ Quantum
Inf. 6, 19 (2020).
24. R. S. Gupta, L. C. Govia, M. J. Biercuk, Phys. Rev. A 102,
042611 (2020).
25. H. Song, A. Chantasri, B. Tonekaboni, H. M. Wiseman,
Phys. Rev. A 107, L030601 (2023).
26. K. Singh, S. Anand, A. Pocklington, J. T. Kemp, H. Bernien,
Phys. Rev. X 12, 011040 (2022).
30. A. Fuhrmanek, R. Bourgain, Y. R. P. Sortais, A. Browaeys,
Phys. Rev. Lett. 106, 133003 (2011).
31. M. J. Gibbons, C. D. Hamley, C.-Y. Shih, M. S. Chapman,
Phys. Rev. Lett. 106, 133002 (2011).
32. I. Cong et al., Phys. Rev. X 12, 021049 (2022).
33. Y. Wu, S. Kolkowitz, S. Puri, J. D. Thompson, Nat. Commun. 13,
4657 (2022).
34. A. Steane, M. Chowdhury, C. Foot, J. Opt. Soc. Am. B 9, 2142
(1992).
35. I. S. Madjarov et al., Phys. Rev. X 9, 041052 (2019).
36. M. A. Norcia et al., Science 366, 93–97 (2019).
37. M. E. Kim et al., Nat. Phys. 19, 25–29 (2023).
38. T. M. Graham et al., Nature 604, 457–462 (2022).
39. M. Morgado, S. Whitlock, AVS Quantum Sci. 3, 023501 (2021).
40. R. Verresen, N. Tantivasadakarn, A. Vishwanath, Efficiently
preparing Schrödinger’s cat, fractons and non-Abelian
topological order in quantum devices. arXiv:2112.03061
[quant-ph] (2021).
41. Y. Li, X. Chen, M. P. A. Fisher, Phys. Rev. B 98, 205136
(2018).
42. T. M. Graham et al., Mid-circuit measurements on a neutral
atom quantum processor. arXiv:2303.10051 [quant-ph]
(2023).
43. K. Singh et al., Data and plotting code for “Mid-circuit correction
of correlated phase errors using an array of spectator qubits.”
Zenodo (2023); https://doi.org/10.5281/zenodo.7612718.
AC KNOWLED GME NTS
We thank H. Levine, J. Covey, and A. Clerk for fruitful discussions
and critical reading of the manuscript. Funding: We acknowledge
funding from the Office for Naval Research (N00014-20-1-2510),
the Air Force Office of Scientific Research (FA9550-21-1-0209),
the NSF Quantum Leap Challenge Institutes (QLCI) for Hybrid
Quantum Architectures and Networks (NSF award 2016136), and
the Sloan Foundation. This material is based upon work supported
by the US Department of Energy, Office of Science, National
Quantum Information Science Research Centers, and was
supported by an appointment to the Intelligence Community
Postdoctoral Research Fellowship Program at the Pritzker School
of Molecular Engineering administered by the Oak Ridge Institute
for Science and Education through an interagency agreement
between the US Department of Energy and the Office of the
Director of National Intelligence. Author contributions: K.S., C.E.B.,
S.A., V.R., R.W., and H.B. contributed to the experiments. K.S., C.E.B.,
S.A., and H.B. contributed to the analysis of the results and
preparation of the manuscript. Competing interests: The work
described in this article is included in a patent application filed with
the US Patent and Trade Office by the University of Chicago.
Data and materials availability: The data and code supporting
this study are available at Zenodo (43). License information:
Copyright © 2023 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science.
No claim to original US government works. https://www.science.org/
about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade5337
Materials and Methods
Supplementary Text
Figs. S1 to S5
Tables S1 and S2
References (44–48)
27. E. Deist et al., Phys. Rev. Lett. 129, 203602 (2022).
28. J. T. Zhang et al., Quantum Sci. Technol. 7, 035006 (2022).
29. See supplementary materials.
Submitted 24 August 2022; accepted 15 May 2023
Published online 25 May 2023
10.1126/science.ade5337
Singh et al., Science 380, 1265–1269 (2023)
26 June 2023
5 of 5
|
10.1126_science.ade6360
|
RES EARCH
MPOX
Structure of monkeypox virus DNA
polymerase holoenzyme
Qi Peng1†, Yufeng Xie2†, Lu Kuai1†, Han Wang1,3, Jianxun Qi1, George F. Gao1,2,4,5,6*, Yi Shi1,5,6*
The World Health Organization declared mpox (or monkeypox) a public health emergency of
international concern in July 2022, and prophylactic and therapeutic measures are in urgent need.
The monkeypox virus (MPXV) has its own DNA polymerase F8, together with the processive cofactors
A22 and E4, constituting the polymerase holoenzyme for genome replication. Here, we determined
the holoenzyme structure in complex with DNA using cryo–electron microscopy at the global resolution
of ~2.8 angstroms. The holoenzyme possesses an architecture that suggests a “forward sliding clamp”
processivity mechanism for viral DNA replication. MPXV polymerase has a DNA binding mode similar
to that of other B-family DNA polymerases from different species. These findings reveal the mechanism
of the MPXV genome replication and may guide the development of anti-poxvirus drugs.
A s of 2 December 2022, over 82,000 hu-
man mpox (or monkeypox) cases have
been laboratory confirmed in 110 coun-
tries worldwide (https://www.cdc.gov/).
Most infection cases have been reported
in Europe and other non-endemic countries,
including China (1), and these cases were
mostly found in homosexual young men (2).
Human-to-human transmission usually occurs
through close contact with lesions, respiratory
droplets, body fluids, and contaminated mate-
rials, such as bedding (3).
Although the monkeypox virus (MPXV) was
first isolated from a monkey in Denmark in
1958, its natural host was thought to be rodent
(3, 4). Since the first human mpox case was
identified in the Democratic Republic of the
Congo (5), it has been endemic to several cen-
tral and western African countries (6, 7).
Sporadic infection cases have been reported
outside Africa, including England, the United
States, Singapore, and Israel, and are mainly
associated with travelers from endemic coun-
tries, nosocomial infections, or direct contact
with imported rodents infected with MPXV
(4, 8, 9). Phylogenetic analysis has revealed
that MPXV can be classified into two genetic
clades: the West Africa clade and the more
pathogenic Congo Basin clade (10, 11). The 2022
outbreak of MPXV belongs to the West Africa
clade and most likely has a single origin that
has not been identified (12).
1CAS Key Laboratory of Pathogen Microbiology and
Immunology, Institute of Microbiology, Chinese Academy of
Sciences, Beijing 100101, China. 2Department of Basic
Medical Sciences, School of Medicine, Tsinghua University,
Beijing 100084, China. 3College of Future Technology, Peking
University, Beijing 100871, China. 4Savaid Medical School,
University of Chinese Academy of Sciences, Beijing 100049,
China. 5Center for Influenza Research and Early-warning
(CASCIRE), CAS-TWAS Center of Excellence for Emerging
Infectious Disease (CEEID), Chinese Academy of Sciences,
Beijing 100101, China. 6Research Unit of Adaptive Evolution
and Control of Emerging Viruses, Chinese Academy of
Medical Sciences, Beijing 100052, China.
*Corresponding author. Email: gaof@im.ac.cn (G.F.G.); shiyi@
im.ac.cn (Y.S.)
†These authors contributed equally to this work.
MPXV is a large double-stranded DNA virus
that replicates exclusively in the cytoplasm of
the infected cells. It belongs to the Orthopox-
virus genus of the Poxviridae family, which
also includes the variola virus that causes
smallpox and has killed millions of humans
in recorded history. Similar to the vaccinia
virus (VACV), the prototype of poxviruses,
MPXV may enter host cells by either fusion
with the plasma membrane or endocytosis,
and at least 16 proteins in the virion mem-
brane are involved in the entry process (13).
After entry, the virus initiates early gene tran-
scription events, and viral DNA synthesis oc-
curs at perinuclear sites called viral factories
(14, 15). The MPXV replicative holoenzyme
consists of catalytic polymerase F8 (equivalent
to E9 in VACV), a heterodimeric processivity
factor consisting of A22 (equivalent to A20 in
VACV) and uracil-DNA glycosylase E4 (equiv-
alent to D4 in VACV).
Previous genetic, biochemical, and structural
studies on the VACV E9-A20-D4 core repli-
cation machinery have advanced our under-
standing of poxvirus DNA replication. VACV
E9 was recognized as a member of the
B-family DNA polymerase, and structural
analysis has revealed the canonical features
of DNA polymerases and five poxvirus-specific
insertions (16). The E9 polymerase alone does
not have processive DNA synthesis activity
unless it is bound to its heterodimeric cofactor
A20/D4 (17–20). Although poxvirus DNA poly-
merase shares many features with other B-family
polymerases, the processivity factor is dis-
tinctive. In VACV, A20 serves as an essential
bridge to link E9 and D4 together and shares
no homology with viral proteins beyond pox-
virus. The N-terminal domain of A20 binds to
D4 (21–23), and its C-terminal domain binds
to one insertion in the palm domain of E9
(24). Given that the DNA replication machin-
ery is extremely conserved for orthopoxviruses,
with a sequence identity of more than 97%
between VACV and MPXV, the results obtained
for VACV could also be applied to MPXV. How-
ever, we are still awaiting a reliable high-
resolution structure of the replicating state of
the Orthopoxvirus polymerase holoenzyme,
and the mode of operation of the processiv-
ity factor needs to be elucidated.
Results
Biochemical characterization of the purified
polymerase proteins
We coexpressed MPXV F8 polymerase and the
A22-E4 heterodimer using the baculovirus ex-
pression system and purified the homogeneous
F8-A22-E4 heterotrimer protein for enzymatic
and structural studies (fig. S1). When a 38-
nucleotide (nt) template DNA was used with a
24-nt primer DNA, the wild-type holoenzyme
heterotrimer displayed weak primer extension
activity in the reaction buffer with deoxynu-
cleotide triphosphate (dNTP) (fig. S1). More-
over, exonuclease activity was confirmed using
an enzymatic assay without dNTP substrates.
The holoenzyme could completely degrade
the primer DNA in an adenosine 5´-triphosphate
(ATP)–independent manner (fig. S1). We also
prepared an exonuclease-deficient F8 mutant
protein and found that the F8-mutant-A22-E4
holoenzyme did not cleave the primer-template
DNA, thereby demonstrating a much stronger
product band than the wild-type holoenzyme
protein, and the polymerization product could
be efficiently inhibited by heparin (fig. S1).
Overall architecture of F8-A22-E4
polymerase holoenzyme
To capture the replicating conformation of the
MPXV F8-A22-E4 polymerase holoenzyme, we
incubated the 3′-H modified primer-template
DNA and the exonuclease-deficient polymer-
ase holoenzyme in the reaction buffer with
deoxythymidine triphosphate (dTTP) substrate.
We then prepared cryo–electron microscopy
(cryo-EM) samples using a graphene grid to
avoid preferential orientation observed with
ordinary grids. The holoenzyme–DNA com-
plex was resolved to ~2.8 Å (figs. S2 and S3).
The EM map shows the key structural features
of all proteins and DNA elements (fig. S4).
Although the density of the 5′-end template
was weak, we traced the main chains using the
unsharpened EM density map to demonstrate
the template entry channel (see below).
The structure of the holoenzyme–DNA com-
plex contains one F8, one A22, one E4, and the
primer-template DNA, as well as an incoming
dTTP substrate (Fig. 1). F8, A22, and E4 form
pairwise interactions with each other (Fig. 1).
The F8 structure can be traced for 1004 res-
idues, except for the last two residues, and
the classical N-terminal domain (NTD), exo-
nuclease domain (Exo), palm domain, fingers
domain, and thumb domain were observed in
a closed conformation (Fig. 1). Five “poxvirus-
specific” insertion regions in MPXV F8 can also
Peng et al., Science 379, 100–105 (2023)
6 January 2023
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
be observed, like those inserts seen in the VACV
E9 structure (16) (Fig. 1 and fig. S5, A and B).
The 218-residue MPXV E4 resembles the
VACV D4 structure (25) (Fig. 1A and fig. S5, C
and D). The 426-residue MPXV A22 structure
can be divided into three domains: the A22
NTD, middle domain (Mid), and C-terminal
domain (CTD) (Fig. 1A and fig. S5E). When
we aligned the Mid of A22 on the online Dali
server, we found that it shows high structural
similarities with the African swine fever virus
(ASFV) DNA ligase and the bacteriophage
T4 DNA ligase (26, 27). The Mid can be fur-
ther divided into two subdomains: an adeny-
lation domain (OD), which mostly resembles
the OD of ASFV DNA ligase (fig. S5F), and an
OB-fold domain (OB), which mostly resem-
bles the OB of Thermus filiformis DNA ligase
(fig. S5G). Further enzymatic assay showed
that the MPXV polymerase holoenzyme did
not possess ordinary ligase activity similar to
that of T4 ligase (fig. S6, A and B). Compared
with the structures of adenylation domains
from the T4 and ASFV ligases, the putative
ligase active site of the A22 Mid is replaced
by hydrophobic and negatively charged resi-
dues, which may prevent the binding of ATP
(fig. S6C). As the Mid of A22 lacks the essen-
tial DNA binding domain, A22 may comprise
a degenerative ligase domain acting simply as
a flexible linker.
The MPXV DNA polymerase holoenzyme is
stabilized by pairwise interactions between the
F8, A22, and E4 subunits (fig. S7). A22 acts as a
bridge to bind E4 and F8 via the A22 NTD and
CTD, respectively. The interactions are almost
identical to those of their VACV counterparts
(fig. S8) (21–24). A previous study proposed a
VACV polymerase holoenzyme model with an
elongated shape of the A20-D4 cofactor, lead-
ing to a ~150-Å distance between the E9 poly-
merase active site and the D4 DNA binding site
(28). However, in our replicating MPXV holo-
enzyme structure, the A22-E4 cofactor folds back,
and E4 directly interacts with the Exo domain of
F8 at two sites, one where Trp36 and Arg39 of E4
form hydrogen bonds and hydrophobic inter-
actions with Phe179 and Leu278 from F8 Exo (fig.
S7F), and another where Asn165 of E4 forms a
hydrogen bond with Asn303 from F8 Exo (fig. S7G).
Primer-template DNA recognition by the
polymerase complex
The structure of MPXV polymerase holoenzyme–
DNA complex contains 22-nt DNA in the tem-
plate strand, 14-nt DNA in the primer strand,
and the incoming dTTP, as well as a magne-
sium ion that may serve as catalytic ion near
the active site (Fig. 2, A and B). The double-
stranded primer-template DNA binds in a
groove formed between the palm and thumb
domains of F8, and the single-stranded 5′ ex-
tension of the template strand probably passes
through a channel formed by the NTD and
Fig. 1. Overall structure of the replicating MPXV DNA polymerase holoenzyme. (A) Schematic diagrams of
the domain architecture of MPXV DNA polymerase F8 and processivity factors (A22 and E4). The F8 can be
divided into five domains: NTD, blue; Exo, magenta; palm, green; fingers, yellow; thumb, cyan. Compared to
other B-family polymerases, F8 contains five inserted elements in which the largest one was named as insert2
(purple), and the other four small inserts are indicated as rectangles. A22 is colored by domains: NTD, deep
pink; Mid, pink; CTD, salmon. E4, orange; template strand, gray; primer strand, red. (B and C) Atomic model and
cryo-EM density map of the replicating MPXV DNA polymerase holoenzyme. The structures were colored by
domains, as depicted in (A).
Exo of F8 and the E4 subunit in an orientation
perpendicular to the DNA duplex (Fig. 2, C
and D, and fig. S9). The template DNA has 12
unpaired nucleotides at the 5′ end, but only
two of them are well-ordered with a defined
base structure. For the remaining 10 unpaired
bases, we can only trace partial phosphate-
ribose backbones (C6 to T11) because of the
weak EM density (Fig. 2, C and D).
Upon primer-template DNA binding, F8 poly-
merase undergoes conformational changes, a
common feature of the B-family DNA poly-
merases. Comparison of the MPXV F8 in this
holoenzyme–DNA complex structure with the
Peng et al., Science 379, 100–105 (2023)
6 January 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. The interactions
between DNA and MPXV
polymerase. (A) Sharpened
EM densities and
atomic models of dsDNA.
(B) Enlarged view of the
sharpened EM densities and
atomic models of dsDNA in
the active site of polymerase.
(C) The cut-off view of
unsharpened EM density
map for MPXV polymerase
holoenzyme in complex
with the DNA, revealing the
consecutive density of
template DNA. (D) Enlarged
view of the unsharpened
densities and supposed
atomic models of the
template DNA in the
template entry channel.
The region (C6 to T11)
was built using this unsharp-
ened map. The remaining
5′-terminal five bases of
template DNA were invisible,
which reflects an inherent
flexible conformation of
the 5′-terminal unpaired
region of the template
strand. (E and F) The
primary interfaces between
F8 and DNA. The F8 mainly
interacted with the minor
groove of primer-template
DNA, with only a few
contacts to the major groove
contributed by the residues
of exonuclease domain.
The primer-template DNA
is shown in surface repre-
sentation calculated from the
atomic model, and F8 is
shown in cartoon
representation.
VACV apo E9 structure, which has high se-
quence identity, shows that the fingers domain
rotates toward the palm domain by ~17° in the
replicating state (fig. S9). This rotation drags
the positively charged Arg634 and Lys661 of the
fingers domain closer to the active site, where
they can interact with the triphosphate group
of incoming dNTP. The rotated fingers do-
main interacts with the Exo, and this interac-
tion further stabilizes the closed conformation
of the fingers domain. Moreover, the thumb
domain also makes a distinct rotation to wrap
around the primer-template DNA duplex on
its minor groove side (fig. S9). The DNA du-
plex is accommodated in a positively charged
groove of the thumb domain, as observed in
other B-family DNA polymerases (fig. S10).
The modeled double-stranded DNA helix
is formed by 14 base pairs from the primer-
template DNA and maintains a B-form con-
formation (Fig. 2A and fig. S11). Extensive
Peng et al., Science 379, 100–105 (2023)
6 January 2023
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
protein-DNA interactions are observed be-
tween the primer-template DNA and F8, with
a total of 47 residues from F8 directly partici-
pating in DNA binding (within 4.0 Å distance,
29 residues to the template strand, and 18 res-
idues to the primer strand) (Fig. 2, fig. S11,
and table S2). Most of the key residues in-
volved in primer-template DNA binding are
highly conserved among different B-family
DNA polymerases. Protein-DNA interactions
mainly involve the phosphodiester backbone
of the DNA, with many interactions directly
with the phosphate groups (fig. S11A). Inter-
actions with the template-strand phosphates
are largely hydrogen bonds to the main or side
chains of 13 residues from the thumb, palm,
fingers domains, Exo, and NTD of F8 (fig. S11,
A to C), whereas the primer strand is bound by
both electrostatic interactions and hydrogen
bonds with the nine residues from the thumb,
palm domains, and Exo of F8 (fig. S11, A, D,
and E). There is little contact of F8 with the
base pairs of primer-template DNA, except
for one hydrogen bond interaction between
R832 of the thumb domain and the base of
T22 from the primer strand, which may be
important for stabilizing the B-form confor-
mation of the DNA duplex (fig. S11, A and D).
This is consistent with the fact that the enzy-
matic activity of F8 does not rely on a specific
sequence during the elongation step. In addi-
tion, residue N675 of the fingers domain forms
a hydrogen bond with the base of unpaired
A12 from the template strand, and this inter-
action may be responsible for the kinking of
the single-stranded 5′ extension of the tem-
plate strand near the active site.
Interactions between the polymerase and the
incoming nucleotide
Next to the 3′ terminus of the primer strand is
the incoming dTTP, which binds to the active
site of the polymerase in a manner analogous
to that observed in the structures of other DNA
polymerase complexes (Fig. 3A) (29). The incom-
ing dTTP is accommodated in a groove formed
by residues from the palm and fingers domains
(Fig. 3B). The two highly conserved aspartate
residues, D549 (in motif A) and D753 (in mo-
tif C), together with the triphosphate tail of
dTTP, coordinate one divalent metal ion (as-
sumed as magnesium, which has been added to
the reaction buffer) (Fig. 3C). The triphosphate
tail also interacts with the main chains of
Y550, S552, and L553 from motif A, and the side
chains of two positively charged R634 and K661
from the fingers domain (Fig. 3C).
The ribose of dTTP stacks on top of the
phenyl ring of Y554 from motif A, in a manner
similar to that previously observed with Y416
in the ternary complex structure of RB69 poly-
merase (30) (Fig. 3C). There would be a steric
clash between the 2’OH of ribonucleotides and
Y554, hence providing a “steric gating” effect
Fig. 3. Recognition of the incoming dTTP. (A) Cut-off view of the F8 protein, which is shown in surface
representation to reveal the inner active site. The F8 is colored by domains as in Fig. 1; template strand,
gray; primer strand, red. (B) The binding pocket of the incoming dTTP. It is formed by the fingers domain,
palm domain, and upper base pair. The incoming dTTP is shown as a ball-and-stick model, and the residues
of F8 and the upper base pair are shown in surface representation. (C) Interaction details between F8
and the incoming dTTP. The key residues are shown as sticks and colored as corresponding domains. The
magnesium ion is depicted as a black sphere. Single-letter abbreviations for the amino acid residues are as
follows: A, Ala; C, Cys; D, Asp; E, Glu; F, Phe; G, Gly; H, His; I, Ile; K, Lys; L, Leu; N, Asn; P, Pro; R, Arg; S, Ser;
T, Thr; V, Val; W, Trp; and Y, Tyr.
to select dNTP as the substrate. This similar
“steric gating” effect has also been described
in other DNA polymerases and was first pro-
posed in HIV-1 reverse transcriptase (31–33).
Operation mode of processivity cofactor
For B-family DNA polymerases, proliferating
cell nuclear antigen (PCNA) or PCNA-like pro-
teins are required for high processivity. How-
ever, for poxviruses, including MPXV and
VACV, no homologous PCNA-like proteins
have been identified in the viral genome. In-
stead, the poxvirus-specific A22-E4 hetero-
dimer is responsible for the high processivity
of DNA replication.
A primer-extension assay using a 60-nt tem-
plate DNA in the presence of heparin, which
can trap the dissociated DNA polymerase from
the primer-template DNA to guarantee a single-
turnover reaction, showed that the F8 polymer-
ase alone dissociated from the primer-template
DNA after incorporating less than 14 nt, whereas
the F8-A22-E4 holoenzyme was able to gener-
ate full-length 60-nt products with few abort-
ive ones (Fig. 4A). We then demonstrated that
the addition of the A22-E4 heterodimer con-
ferred processivity to F8 in a concentration-
dependent manner (Fig. 4A). When the molar
ratio of A22-E4 to F8 was 1:1, corresponding to
the stoichiometry of the polymerase holoenzyme,
the amount of the full-length product was al-
most the same as that of the product generated
by the preassembled F8-A22-E4 polymerase
holoenzyme (Fig. 4A). This indicates that the
isolated A22-E4 and F8 can be efficiently as-
sembled into functional holoenzymes to per-
form processive DNA synthesis. Moreover, we
performed alanine scanning of critical resi-
dues responsible for the interaction between
E4 and F8 and found that R39A and N165A
substitutions of E4 showed minor effect, W36A
reduced the synthesis of full-length products,
whereas the W36A/R39A and W36A/R39A/
N165A substitutions abolished the synthesis
of full-length products (Fig. 4B). These results
further confirmed the important function
of the A22-E4 heterodimer in DNA replica-
tion processivity in a pure enzymatic reaction
system.
As described above, the E4 cofactor interacts
with the Exo of F8 polymerase, and together
with the NTD of F8, they form a closed-ring
channel to encircle the single-stranded template
DNA (Fig. 4C). By contrast, in the yeast DNA
polymerase complex (29), a representative of
the other B-family DNA polymerases (fig. S12),
Peng et al., Science 379, 100–105 (2023)
6 January 2023
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. The distinctive mechanism of A22-E4 heterodimer promoting pro-
cessivity of MPXV DNA polymerase. (A) Enzymatic assay of A22-E4
heterodimer protein improving the DNA replication processivity of exonuclease-
deficient F8 under single-turnover conditions. The F8 alone was shown to be
distributive and failed to generate full-length products, whereas the addition of
the A22-E4 complex promoted the yield of full-length DNA in a concentration-
dependent manner. (B) Alanine scanning of critical residues of E4 responsible for the
interaction with F8 was performed to examine the effects on the DNA replication
processivity. The A22-E4 W36A mutant reduced the processivity activity, whereas the
W36A/R39A and W36A/R39A/N165A mutants abolished the processivity activity. The
F8 used in this assay was exonuclease deficient. (C and D) Comparison of the
structures of the MPXV and yeast polymerases (PDB ID: 7KC0) in complex with their
own processivity factors and DNA. Polymerases and processivity factors are shown in
surface representation and colored by domains, as in Fig. 1, whereas the primer-
template DNA strands are shown as a cartoon (template, gray; primer, red). The
processivity of yeast polymerase was strengthened by trimeric PCNA to clamp the
primer-template dsDNA. While in the structure of MPXV polymerase holoenzyme,
A22-E4 heterodimer does not interact with dsDNA. Instead, E4 located on the
template entry channel combined with NTD and Exo domains of F8 to form
a forward clamp structure that would prevent the template strand disassociating
from the polymerase complex during DNA replication. (E and F) Two binding
modes of processivity factors with polymerases. The processivity factors bound
with template in poxvirus function as a “forward sliding clamp” (E) or dsDNA
products in eukaryotes as a “backward sliding clamp” (F).
the Exo and NTD form an open semicircular
channel to accommodate the single-stranded
template DNA, and the trimeric PCNA ring
encircles the template-product DNA duplex
(Fig. 4D). This architectural difference be-
tween MPXV and yeast polymerase complexes
is responsible for the different processivity
mechanisms during DNA replication events.
The MPXV DNA polymerase holoenzyme guar-
antees its high DNA replication processivity
by encircling the single-stranded template
DNA, and we propose that it functions as a
“forward sliding clamp” (Fig. 4E); whereas the
other B-family DNA polymerase complexes
possess continuous DNA replication capacity
by encircling the double-stranded template-
product DNA helix that can be recognized as
a “backward sliding clamp” (Fig. 4F).
Discussion
The interaction between E4 and F8 could gen-
erate a ring channel that encircles the single-
stranded template DNA, which is proposed
to be important for high DNA replication
processivity. This processivity mechanism is
different from that of other B-family DNA
polymerases that utilize PCNA or PCNA-like
proteins to encircle the product-template DNA
duplex (29, 34, 35). The configuration of encir-
cling the single-stranded template DNA prob-
ably allows the MPXV polymerase complex
to perform continuous DNA replication by
preventing template DNA disassociation from
Peng et al., Science 379, 100–105 (2023)
6 January 2023
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
the polymerase holoenzyme. The efficiency
of polymerase holoenzyme assembly in the cel-
lular environment requires further study, and
moreover, it should be investigated whether
the F8-A22-E4 polymerase holoenzyme has
other active conformations (28).
Previous studies have implicated the distinc-
tive role of the polymerase holoenzyme in DNA
recombination, which involves both 3′-5′ exo-
nuclease and DNA-joining activities (36–38).
Although we identified DNA ligase-like do-
mains for the Mid of the poxvirus-specific
cofactor A22, the F8-A22-E4 polymerase holo-
enzyme did not have canonical ligase activity
in our assay. However, we cannot rule out the
possibility that A22 possesses ligase activity
in a different state from the current holoenzyme
structure or in an ATP-independent manner.
Orthopoxviruses encode a DNA ligase that is
not essential for virus replication but affects
virulence and sensitivity to DNA-damaging
agents (39). Further functional dissection of
the middle part of A22 and its cooperation
with viral and host ligases is needed.
Poxvirus replication processes use different
replication models, including self-priming,
primer-dependent, and recombination models
(15). Poxviruses, including MPXV, have linear
double-stranded DNA genomes, and the ter-
mini of the two DNA strands are connected to
form a continuous polynucleotide chain (40).
A rolling cycle mechanism has been proposed
to replicate DNA in the form of unbranched
head-to-tail concatemers, which would be re-
solved by a Holliday junction resolvase to pro-
duce unit genomes (41). The proposed “forward
sliding clamp” mode can help interpret the
self-priming model, as the processivity cofac-
tor would facilitate robust continuous replica-
tion along the single-stranded template DNA
unwound by the primase-helicase. Because
there is also a possible presence of Okazaki
fragments for Orthopoxvirus DNA replication
(42), other replication models should also be
studied. Moreover, the working mechanism of
an intact replisome, including the F8-A22-E4
holoenzyme and other replication proteins,
will be a fascinating area for future studies.
33. G. Gao, M. Orlova, M. M. Georgiadis, W. A. Hendrickson,
S. P. Goff, Proc. Natl. Acad. Sci. U.S.A. 94, 407–411 (1997).
34. Y. Shamoo, T. A. Steitz, Cell 99, 155–166 (1999).
35. C. Lancey et al., Nat. Commun. 11, 1109 (2020).
36. D. B. Gammon, D. H. Evans, J. Virol. 83, 4236–4250
RE FERENCES AND NOTES
1. H. Zhao et al., China CDC Wkly. 4, 853–854 (2022).
2.
J. G. Rizk, G. Lippi, B. M. Henry, D. N. Forthal, Y. Rizk, Drugs
82, 957–963 (2022).
J. Kaler, A. Hussain, G. Flores, S. Kheiri, D. Desrosiers, Cureus
14, e26531 (2022).
3.
4. D. Focosi, F. Novazzi, A. Baj, F. Maggi, Rev. Med. Virol. 32, e2392
(2022).
5. D. B. Di Giulio, P. B. Eckburg, Lancet Infect. Dis. 4, 15–25
(2004).
6. E. M. Bunge et al., PLOS Negl. Trop. Dis. 16, e0010141
(2022).
7. W. Tan, G. F. Gao, China CDC Wkly. 4, 847–848 (2022).
8. N. Erez et al., Emerg. Infect. Dis. 25, 980–983 (2019).
9. A. Vaughan et al., Euro Surveill. 23, 1800509 (2018).
10. N. Chen et al., Virology 340, 46–63 (2005).
11. A. M. McCollum, I. K. Damon, Clin. Infect. Dis. 58, 260–267
(2014).
12. J. Isidro et al., Nat. Med. 28, 1569–1572 (2022).
13. B. Moss, Semin. Cell Dev. Biol. 60, 89–96 (2016).
14. M. D. Greseth, P. Traktman, Annu. Rev. Virol. 9, 239–259
(2022).
15. B. Moss, Cold Spring Harb. Perspect. Biol. 5, a010199
(2013).
16. N. Tarbouriech et al., Nat. Commun. 8, 1455 (2017).
17. E. S. Stanitsa, L. Arps, P. Traktman, J. Biol. Chem. 281,
3439–3451 (2006).
18. A. M. Druck Shudofsky, J. E. Silverman, D. Chattopadhyay,
R. P. Ricciardi, J. Virol. 84, 12325–12335 (2010).
19. M. W. Czarnecki, P. Traktman, Virus Res. 234, 193–206
(2017).
20. K. A. Boyle, E. S. Stanitsa, M. D. Greseth, J. K. Lindgren,
P. Traktman, J. Biol. Chem. 286, 24702–24713 (2011).
21. C. Contesto-Richefeu et al., Acta Crystallogr. F Struct. Biol. Commun.
72, 687–691 (2016).
22. W. P. Burmeister et al., J. Biol. Chem. 290, 17923–17934
(2015).
23. C. Contesto-Richefeu et al., PLOS Pathog. 10, e1003978
(2014).
24. B. Bersch, N. Tarbouriech, W. P. Burmeister, F. Iseni, J. Mol. Biol.
433, 167009 (2021).
25. N. Schormann et al., BMC Struct. Biol. 7, 45 (2007).
26. K. Shi et al., Nucleic Acids Res. 46, 10474–10488 (2018).
27. Y. Chen et al., Nat. Commun. 10, 387 (2019).
28. C. Sèle et al., J. Virol. 87, 1679–1689 (2013).
29. F. Zheng, R. E. Georgescu, H. Li, M. E. O’Donnell, Proc. Natl.
Acad. Sci. U.S.A. 117, 30344–30353 (2020).
30. M. C. Franklin, J. Wang, T. A. Steitz, Cell 105, 657–667 (2001).
31. A. P. Silverman, S. J. Garforth, V. R. Prasad, E. T. Kool,
Biochemistry 47, 4800–4807 (2008).
32. G. J. Klarmann et al., Biochemistry 46, 2118–2126 (2007).
(2009).
37. M. D. Hamilton, A. A. Nuara, D. B. Gammon, R. M. Buller,
D. H. Evans, Nucleic Acids Res. 35, 143–151 (2007).
38. M. D. Hamilton, D. H. Evans, Nucleic Acids Res. 33, 2259–2268
(2005).
39. S. M. Kerr et al., EMBO J. 10, 4343–4350 (1991).
40. O. Faye et al., Lancet Infect. Dis. 18, 246 (2018).
41. N. Sklenovská, in Animal-Origin Viral Zoonoses, Y. S. Malik,
R. K. Singh, K. Dhama, Eds. (Springer Singapore, Singapore,
2020), pp. 39-68.
42. M. Esteban, J. A. Holowczak, Virology 78, 57–75 (1977).
AC KNOWLED GME NTS
We thank X.L., Y.M., and X.B. at the cryo-EM Center, Shanxi Academy of
Advanced Research and Innovation, for their technical support of the
cryo-EM data collection. Funding: National Key Research and
Development Program of China grant 2021YFC2300700 (Y.S.);
Strategic Priority Research Program of CAS grant XDB29010000 (Y.S.,
G.F.G.); National Natural Science Foundation of China (NSFC) grants
82241076, 81871658, 32192452, and 32100119 (Y.S., Q.P.); The
Youth Innovation Promotion Association of CAS grant Y201921 (Y.S.).
Author contributions: Conceptualization: Y.S., G.F.G., Q.P.
Investigation: L.K., Q.P., Y.X., J.Q., H.W. Visualization: Q.P., Y.X., L.K.
Funding acquisition: Y.S., G.F.G., Q.P. Supervision: Y.S., G.F.G. Writing –
original draft: Y.S., Q.P., G.F.G. Writing – review and editing: Y.S., Q.P.,
G.F.G. Competing interests: The authors declare that they have no
competing interests. Data and materials availability: Cryo-EM map is
available in the Electron Microscopy Data Bank with code EMD-34731.
Structural model is available in the Protein Data Bank (PDB) with
accession code 8HG1. All materials are available from the authors on
reasonable request with materials transfer agreements (MTAs).
License information: Copyright © 2023 the authors, some rights
reserved; exclusive licensee American Association for the Advancement
of Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse. In the interest
of rapid dissemination of results with immediate public health
relevance, the author will make the Author Accepted Manuscript (AAM)
version available under a CC BY public copyright license.
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade6360
Materials and Methods
Figs. S1 to S12
Tables S1 and S2
References (43–53)
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
Submitted 29 August 2022; accepted 5 December 2022
Published online 15 December 2022
10.1126/science.ade6360
Peng et al., Science 379, 100–105 (2023)
6 January 2023
6 of 6
|
10.1126_science.ade6219
|
RES EARCH
R E S E A R C H A R T I C L E
◥
QUANTUM SIMULATION
Observing the quantum topology of light
Jinfeng Deng1†, Hang Dong1†, Chuanyu Zhang1, Yaozu Wu1, Jiale Yuan1, Xuhao Zhu1, Feitong Jin1,
Hekang Li1, Zhen Wang1,2,3, Han Cai1, Chao Song1*, H. Wang1,2,3*, J. Q. You1, Da-Wei Wang1,3,4*
Topological photonics provides a powerful platform to explore topological physics beyond traditional
electronic materials and shows promising applications in light transport and lasers. Classical degrees of
freedom are routinely used to construct topological light modes in real or synthetic dimensions. Beyond the
classical topology, the inherent quantum nature of light provides a wealth of fundamentally distinct topological
states. Here we implement experiments on topological states of quantized light in a superconducting circuit,
with which one- and two-dimensional Fock-state lattices are constructed. We realize rich topological
physics including topological zero-energy states of the Su-Schrieffer-Heeger model, strain-induced
pseudo-Landau levels, valley Hall effect, and Haldane chiral edge currents. Our study extends the
topological states of light to the quantum regime, bridging topological phases of condensed-matter
physics with circuit quantum electrodynamics, and offers a freedom in controlling the quantum states
of multiple resonators.
T he quantum Hall effect (1) reveals new
phases of matter that are classified by the
topological invariants of energy bands (2).
For two-dimensional electrons in strong
magnetic fields, the chiral edge states
between Landau levels contribute to the quan-
tized Hall conductivity, which is immune to
local defects. This topological effect can also
1Interdisciplinary Center for Quantum Information, State Key
Laboratory of Modern Optical Instrumentation, and Zhejiang
Province Key Laboratory of Quantum Technology and Device,
School of Physics, Zhejiang University, Hangzhou 310027,
China. 2Hangzhou Global Scientific and Technological
Innovation Center, Zhejiang University, Hangzhou 311215,
China. 3Hefei National Laboratory, Hefei 230088, China.
4CAS Center of Excellence in Topological Quantum
Computation, Beijing 100190, China.
*Corresponding author. Email: chaosong@zju.edu.cn (C.S.);
hhwang@zju.edu.cn (H.W.); dwwang@zju.edu.cn (D.-W.W.)
†These authors contributed equally to this work.
exist without Landau levels, such as in the
Haldane model (3), which lays the basis for
topological insulators (4). The optical simula-
tion of quantum Hall edge states (5) opens
a new research area, topological photonics
(6–8), which brings a wealth of applications
in routing and generating electromagnetic
waves, such as backscattering-free waveguides
(9) and topological insulator lasers (10). Clas-
sical degrees of freedom such as frequencies
and orbital angular momenta have been widely
used to synthesize new lattice dimensions to
embed the topological modes (11–13). Such
pure classical topology of light is in stark con-
trast to the topological phases of electrons,
where the quantum wave and fermionic sta-
tistics play a fundamental role. Intriguingly,
new topological states emerging from light
quantization and bosonic statistics have been
predicted beyond classical interpretation (14–17).
Recent development in circuit quantum electro-
dynamics (QED) (18) makes it possible to real-
ize these intrinsic quantum topological states
of light, which provide quantum degrees of
freedom in engineering photonic topology
(14, 19) and offer topological control knobs
in bosonic quantum information processing
(20–23).
Compared with lattices of modes in real or
synthetic dimensions in classical topological
photonics, the topological states of quantized
light are embedded in lattices of Fock states
Πijni〉 with ni being the photon number in the
i th mode. In the Fock-state lattice (FSL), a
mode provides a dimension (14, 19, 20, 24),
in contrast to a site in traditional lattices, in-
cluding those in synthetic dimensions (11–13, 25).
The FSLs exploit the infinite quantum Hilbert
space of light, enabling the construction of
high-dimensional lattices with only a few
cavity modes. To sketch such dimensional
scalability, we use the Jaynes-Cummings (JC)
model (26), which describes the interaction
between a two-level atom with quantized light.
However, here we use multiple quantized light
modes to couple the atom. With two light modes,
the Fock states form one-dimensional (1D) lat-
tices of the Su-Schrieffer-Heeger (SSH) model
(Fig. 1A) (27). By adding just one other mode,
we obtain two-dimensional (2D) strained honey-
comb lattices (Fig. 1B) (28). These lattices are
featured by site-dependent coupling strengths,
which originate from the property of the boso-
nic annihilation operator a
(cid:3)
(cid:3)
a ni ¼
p
ffiffiffi
j
n
n (cid:2) 1i
ð1Þ
For the vacuum state,aj0〉 ¼ 0, which leads to
natural edges of FSLs when the photon number
Fig. 1. Fock-state lattices of multimode Jaynes-Cummings models.
(A) Topological transport of the zero-energy state of the SSH FSL with N ¼ 5.
The sublattice sites of s ¼ ↑ (↓) are denoted by squares (circles) and labeled by n1n2,
the photon numbers in R1 and R2. The thicknesses of the lines connecting
neighboring sites are proportional to the coupling strengths t1 (red) and t2 (blue).
The wave function envelopes of four zero-energy states are schematically drawn
with different colors. (B) The valley Hall response and the Haldane chiral edge
state in a 2D FSL with N ¼ 10. An excited qubit is coupled to three resonators
with different photon numbers nj (left). The coupling strengths are proportional
p
ffiffiffiffiffiffiffiffiffiffiffi
nj þ 1
, which introduces competition between resonators to obtain a photon
to
from the qubit. All the Fock states with the same N are coupled by the JC
Hamiltonian to form a honeycomb lattice (right). The inhomogeneous coupling
strengths induce an effective magnetic field in the 2D FSL. The VHE is featured
by the wave packets at the two valleys moving in opposite directions
perpendicular to an applied force (the black arrow). A Lifshitz topological edge
(dashed line) separating the semimetallic and insulator phases locates on the
incircle, which can host the Haldane chiral edge states (yellow wave packet with
the arrow indicating the moving direction).
Deng et al., Science 378, 966–971 (2022)
2 December 2022
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
in one of the cavities reduces to zero. FSLs also
have topological edges that host zero-energy
states, resulting from the competition between
resonators in exchanging photons with the
atom. Such a simple mechanism enables FSLs
to realize several important models in topo-
logical physics, in particular the seminal SSH
and Haldane models, which have been the
focus in various quantum platforms (29–36).
Here we demonstrate adiabatic transport of
topological zero-energy states in 1D SSH FSLs,
where Fock states are topologically transferred
from one cavity to another while maintaining
the quantumness in superposition states. In
2D FSLs, we observe the valley Hall effect (VHE)
(37) and the Haldane chiral edge current (38),
which offer a topological route of engineering
quantum states of multiple resonators.
Leveraging the advantageous integrability
and tunability of the circuit QED platform
(39–42), we design and fabricate a superconduc-
ting circuit to build and engineer the FSLs.
The key elements of the circuit are a central
gmon qubit (43) (Q0) and three resonators (Rj
with j running from 1 to 3), all with tunable
frequencies. Each resonator Rj is coupled to Q0
through an inductive coupler (Cj) (Fig. 2A).
The coupling strengths gj=2p can be contin-
uously tuned by changing the magnetic flux in
Cj. In addition, each resonator Rj is capacitive-
ly coupled to an ancilla qubit Qj for the prep-
aration and readout of the resonator state.
Other characteristics of the resonators and
qubits can be found in the supplementary
materials.
The Hamiltonian of the coupled system of
Rj’s and Q0 can be described by a multimode JC
model (26) in the rotating-wave approximation
Xd
sz þ
ℏwja†
j aj þ
j¼1
H ¼
ℏw0
2
Xd
(cid:4)
ℏgj sþaj þ a†
j
(cid:5)
s(cid:2)
ð2Þ
j¼1
where aj is the annihilation operator of Rj with
the transition frequency wj, sþ ≡ ↑i ↓h j
and
s(cid:2) ≡ ↓i ↑h j
j
are the raising and lowering operators
j
Fig. 2. Adiabatic transport of the topological zero-energy states in the Fock-
state Su-Schrieffer-Heeger model. (A) False-color circuit image of the device of
this experiment. Inset: Symbolized configuration of the key elements, a central gmon
qubit Q0 (green circle) coupled to three resonators (cyan, red and yellow squares
for R1, R2, and R3) via tunable couplers (blue twin circles). (B) Experimental pulse
sequences for the adiabatic transport. We prepare the initial Fock state of R2 by
repeatedly exciting its ancilla qubit Q2 with a p-pulse and tuning it in resonance with
R2 to swap the photons (upper panel). After the initialization, we tune R1, R2, and
Q0 in resonance (middle panel) and modulate C1, C2 to tune the coupling strengths
g1 (cyan line) and g2 (red line) (lower panel). Finally, we measure the joint population
of R1, R2, and Q0. (C) The observed evolution of the zero-energy state wave packet
in the numerical simulation (upper panel) and experiment (lower panel). Obviously Y0i
only occupies the ↓ij
of the resonators and qubit listed in the table S1. All data in this paper, except
that for quantum state tomography, are averaged over five runs of experiments.
(D) The two-mode Wigner function of the resonator state at t ¼ 300 ns in the plane-
cut along axes Re a1ð
Þ, and the fidelity F ¼ 0:735 (see
fig. S6 for pulse sequence of tomography and more data at other times).
sublattice. In numerical simulation, we use the parameters
Þ and Im a1ð
Þ-Re a2ð
Þ-Im a2ð
j
Deng et al., Science 378, 966–971 (2022)
2 December 2022
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. The pseudo-Landau levels in the 2D Fock-state lattice with N = 5. (A) The evolution of the
excited-state population of the qubit Q0. (B) Fast Fourier transform of the Rabi oscillation. The vertical axis is
the frequency component divided by 2. The solid line is the numerical simulation, and the circles are the
experimental data. (C) The eigenstates in the zeroth and positive pseudo-Landau levels of the 2D FSL, with
eigenenergies corresponding to the Fourier peaks. Each point labels an eigenstate characterized by the
chirality C. The degeneracy of the nth Landau level is N (cid:2) n þ 1.
of Q0 with the transition frequency w0, and d is
the number of resonator modes. The Hamil-
tonian conserves the total excitation number
X
ð
nj þ sz þ 1
N ¼
is the
photon number of Rj and sz ¼ ↑i ↑h j (cid:2) ↓i ↓h j
.
Þ=2 , where nj
j
j
j
p
ffiffiffiffiffi
nj
Topological transport
For d ¼ 2 and N excitations, 2N þ 1 states
s; n1; n2i
j
are coupled in a bipartite tight-
binding lattices with the spin states s ¼ ↑; ↓
labeling the two sublattices (Fig. 1A). When
Q0 is resonant with both resonators, all these
2N þ 1 states have the same energy, which is
set as the zero energy. Because the coupling
strengths tj ¼ gj
( j ¼ 1; 2) depend on the
photon numbers, t1 > t2 and t1 < t2 on the
left- and right-hand sides of the FSL, resulting
in two different topological phases of the SSH
model (27). A topological zero-energy state
locates around the lattice sites satisfying
t1 ¼ t2, which is the topological edge of the
SSH model. We write gj ¼ g0lj where g0 is a
fixed nonzero coupling strength and lj’s are
¼ 1.
the tunable parameters satisfying l2
1
The topological zero-energy state can be writ-
ten as a two-mode binomial state (14, 44)
þ l2
2
XN
s
jY0i ¼
n¼0
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
N !
n! N (cid:2) n
Þ!
ð
ln
2
(cid:3)
(cid:3)
(cid:3)
ÞN (cid:2)n ↓; n; N (cid:2) ni
ð
(cid:2)l1
ð3Þ
which only occupies the ↓ij
sublattice. By
adiabatically tuning l2 from 0 to 1, we can
transport the topological zero-energy state
from the right end of the lattice to the left end,
or vice versa.
j
j
ð
ð
Þ
Þ
j, and l2 ¼ sin 2pnt
In the experiment we select R1 and R2 to
construct the SSH FSL, with R3 being far
detuned and effectively decoupled from the
system. In the experimental pulse sequences
(Fig. 2B), we first prepare the initial state
↓; 0; 5i
j
, which is the topological zero-energy
state of the SSH FSL with N ¼ 5 and l1 ¼ 1,
by pumping five photons successively into R2
via Q2 (upper panel of Fig. 2B). Then we tune
R1, R2, and Q0 in resonance at the frequency
wint=2p ≈ 4:81 GHz and sinusoidally modulate
the coupling strengths where g0=2p ≈ 9 MHz,
l1 ¼ cos 2pnt
j with
n ¼ 416 kHz ≪ g0 to satisfy the adiabatic
condition (lower panel of Fig. 2B). Finally, the
wave packet of the zero-energy state in the
FSL is measured (see supplementary mate-
rials), with the data shown in Fig. 2C. The
adiabatic transport of the topological edge state
is witnessed by the oscillation of the photons
between R1 and R2 following Eq. 3 with time-
dependent l1 and l2. The zero-energy state is
topologically protected by the energy gap g0
from other eigenstates of the FSL and main-
tains coherence during the transport (45). To
show this, we further measure the density
matrix of the two resonators by quantum state
tomography (Fig. 2D). The two-mode binomial
state remains a Fock state in the combinational
dark mode of the two resonators, l2a1 (cid:2) l1a2,
and the quantumness of the states is evident
from the negative values of the two-mode
Wigner functions.
p
Valley Hall effect
When d ¼ 3, the Fock states in the subspace
with N excitations form a two-dimensional
honeycomb lattice containing ðN þ 1Þ2 sites
(Fig. 1B). The site-dependent coupling strengths
introduce a strain, which has the effect of a
ffiffiffi
magnetic field (46–48) and results in
-scaling
n
pseudo-Landau levels (14) when Q0 is reso-
nant with all three resonators. We observe
the Landau levels by analyzing the spectra
of the lattice dynamics (49). In the experi-
ment, we prepare the initial state ↓; 0; 5; 0i
and resonantly couple R1, R2, and R3 to Q0 with
coupling strengths gj=2p ≈ 9 MHz. We mea-
sure the evolution of the probability of finding
Q0 in the ↑ij
state and then perform fast Fourier
transform. We obtain peaks approximately
p
p
ffiffiffi
W0 with W0 ¼
gj (Fig. 3C). The
located at
n
degenerate states in the same Landau level are
distinguished by their chiralities
ffiffiffi
3
j
þbþ (cid:2) b†
(cid:2)b(cid:2)
C ¼ b†
X
3
j¼1
where bT ¼
ð
ajexp ∓i2 jp=3
Þ=
ð4Þ
are
p
ffiffiffi
3
the annihilation operators of the two chiral
dark modes that are decoupled from the qubit.
The chirality C plays the role of the lattice mo-
mentum in conventional lattices, and C ¼ N
and C ¼ (cid:2)N correspond to the two corners of
the Brillouin zone, denoted as K and K ′ val-
leys, respectively (14, 50).
A Lifshitz topological edge (51) on the in-
circle separates the FSL into two phases, a
semimetallic phase within the incircle and a
band insulator phase outside of it (see the
dashed line in Fig. 1B). The states in the zeroth
Landau level are confined within the incircle
by an outside band gap of the insulator (14).
In the semimetal, the strain-induced magnetic
field has opposite signs at the K and K ′ valleys
(see supplementary materials). By introduc-
ing a linear potential to mimic the effect of an
electric field to electrons, we can observe the
VHE (Fig. 1B); i.e., the Hall response has oppo-
site signs at the two valleys. To experimentally
demonstrate this effect, we first prepare an
initial state Y0i
in Eq. 3
following the procedure in Fig. 2B. Such an
initial state on the Lifshitz topological edge is
a Gaussian wave function in the zeroth Landau
level (14). Then we bring R3 in resonance with
Q0 and set gj=2p ≈ 9 MHz for j ¼ 1; 2; 3. The
linear potential with horizontal gradient is in-
troduced by slightly shifting the frequencies of
R1 and R2
with l1 ¼ l2 ¼ 1=
ffiffiffi
2
p
j
(cid:6)
(cid:7)
V ¼ ℏδ a†
1 a1 (cid:2) a†
2a2
ð5Þ
where the detuning δ=2p ≈ 1:8 MHz. We then
measure populations on each lattice site and
obtain the average photon numbers in the
three resonators (Fig. 4B). The linear potential
drives photons from R1 and R2 to R3 whereas
the qubit stays in the ground state. To visualize
the evolution of the wave function, we draw the
Deng et al., Science 378, 966–971 (2022)
2 December 2022
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
j
through the
Fig. 4. The valley Hall effect in the 2D Fock-state lattice. (A) The pulse
sequences for controlling the frequencies (upper panel) and coupling
strengths (lower panel). We first prepare the initial state Y0i
topological transport in the SSH FSL. Then we tune R3 and Q0 in resonance
at wint=2p ≈ 4:81 GHz while we detune R1 and R2 to introduce the linear
potential. Meanwhile we set the coupling strengths gj=2p ≈ 9 MHz for j ¼ 1; 2; 3,
and finally we measure the joint populations at different times during the
evolution. (B) The valley Hall evolution of the average photon numbers in
the three resonators for an initial two-mode binomial state Y0i
The squares are experimental data, and the dashed lines are numerical
simulations with the detuning d=2p ¼ 1:80 MHz. (C) The populations in the FSL
for N = 5 at t = 0, 150, 250, 350, and 480 ns. The left, right, and top vertices
with N ¼ 5.
j
correspond to the states with all photons in R1, R2 and R3. The radius of the
blue circle on each site is proportional to its population. The trajectory of
the wave function is perpendicular to the direction of the effective force
(black arrow). (D) The evolution of the average photon numbers in the
three resonators for the coherent initial state Yci ¼ ↓; a; 0; (cid:2)a
i with a ≈ 1:8.
j
j
(cid:4)
†
We detune R1 and R3 to introduce a linear potential V ¼ ħd a
1 a1 (cid:2) a
†
3a3
. In
(cid:5)
the numerical simulation (dashed lines), we set d=2p ¼ 2:35 MHz. (E) The
measured Wigner functions of the three resonator states at time t = 100 and
290 ns (see fig. S3 for the numerical simulation). The phases of the largest
amplitudes of the Wigner functions are labeled on the unit circles, which show
the chirality of the corresponding states in the two valleys.
population distributions in the FSL at five dif-
ferent times (Fig. 4C). The wave function first
moves upward perpendicular to the force di-
rection (black arrow) until being reflected
by the Lifshitz topological edge near the top
vertex, and then moves downward back to the
initial state (up to a phase factor). In particu-
lar, when the wave function is at the center
of the lattice but in different valleys—e.g., at
t ¼ 150 and 350 ns—it moves in opposite di-
rections, which is a signature of the VHE (14).
In contrast to the valley Hall effect in photonic
lattices, where edges are routinely needed for
the experimental implementation (52, 53),
here we coherently transport the quantum
states to the two valleys and directly mea-
sure the valley Hall drift, thanks to the high
tunability, controllability, and readability of
the superconducting circuit. It is noteworthy
that the qubit remains in the ground state
during the evolution, which reflects a funda-
mental difference between classical and quan-
tum predictions (fig. S5).
j
and (cid:2)ai
j
Surprisingly, the VHE can also be observed
with initial classical states such as Yci ¼ ↓;j
a; 0; (cid:2)ai; i.e., R1 and R3 are in the coherent
states aij
andR2 is in the vacuum state.
This state can be expanded as a superposition
of two-mode binomial states with different
total excitation numbers N (14). Owing to the
synchronized dynamics in different subspaces,
the fields in the three resonators remain as
a direct product of coherent states, and the
evolution of the average photon numbers in
the three resonators follows curves similar
to that for an initial binomial state (Fig. 4D
and supplementary materials).
The states in the two valleys are identified
by their chiralities. Because the states of the
three resonators are separable for coherent
initial states, we perform simultaneous quan-
tum state tomography and obtain their Wigner
functions (Fig. 4E). As expected, the phases
are distributed in a counterclockwise (C > 0)
and clockwise (C < 0) manner at t ¼ 100 and
290 ns when the wave function moves to the
K and K ′ valleys, respectively. Therefore, the
VHE in FSLs can be used to coherently trans-
port the wave function between two valleys
and control the chirality of the quantum states
of multiple resonators.
Haldane model
By introducing a Floquet modulation of the
coupling strength, gj tð Þ ¼ g0 þ 2gdsin ndt þ½
2 j (cid:2) 1
Þp=3(cid:3), we synthesize a most important
ð
model in topological physics, the Haldane
model (33, 35, 36, 38). The effective Hamil-
tonian in the second-order perturbation is (see
supplementary materials)
X3
(cid:4)
(cid:5)
a†
j
σ(cid:2) þ h:c:
þ ℏkszC ð6Þ
HH ¼ ℏg0
j¼1
=nd. The second term in Eq. 6
where k ¼ (cid:2)3g2
d
introduces the complex next-nearest-neighbor
hoppings in the FSL (14, 20) and transforms
flat Landau levels to a two-band structure with
gapless chiral edge states, which originate from
Deng et al., Science 378, 966–971 (2022)
2 December 2022
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. Chiral edge currents of the Fock-state Haldane model. (A) The energy bands of the Hamiltonian
HH in Eq. 6 with total excitation number N = 10. The two energy bands are connected by chiral edge
states, with each dot indicating an eigenstate. The populations of a binomial state (with l1 ¼ l2) on the chiral
edge states are proportional to the radii of the shaded circles. (B) The control sequence in realizing the
Haldane Hamiltonian. We prepare an initial state ↓; a; (cid:2)a; 0i
with a ≈ 1 and tune three resonators and
the gmon qubit on resonance at wint=2p ≈ 4:82 GHz, followed by a Floquet modulation of the coupling
strengths gj tð Þ with static amplitude g0=2p ≈ 2:5 MHz, dynamic amplitude gd=2p ≈ 3:25 MHz, and modulation
frequency nd=2p ¼ 40 MHz, such that the effective Haldane coupling strength k=2p ¼ (cid:2)0:79 MHz. (C) Chiral
edge currents shown by the average photon numbers in the three resonators. The total average photon
number in the initial state is 2. The gray triangle shows the boundary of the FSL with N = 2, which is the most
occupied subspace at the initial time. The circles show the experimental data, and the depths of the
colors indicate the evolution time. Dashed lines are numerical simulations in the ideal case, whereas solid
lines are those considering relevant parameter imperfections as described in the supplementary materials.
j
j
the zeroth Landau level (Fig. 5A). In the ex-
periment, we directly excite R1 and R2 to obtain
an initial state ↓; a; (cid:2)a; 0i
(see Fig. 5A for its
distribution in the subspace N ¼ 10). Then we
periodically modulate the coupling strengths
gj tð Þ to realize the Haldane Hamiltonian (see
the control sequence in Fig. 5B). The average
photon numbers are subsequently measured
as a function of time (17), which shows the
chiral motion; i.e., the wave function rotates in
a counterclockwise manner in the FSL (Fig. 5C).
Ideally the wave function shall be on the in-
circle, i.e., the Lifshitz topological edge. In the
experiment, the chiral rotating wave function
moves toward the center of the FSL owing
to the decoherence and nonlinearity of the
resonators, as well as the imperfect controlling
pulses (figs. S8 to S11).
Concluding remarks
In this work, we have demonstrated the co-
herent control of topological zero-energy states
in 1D and 2D FSLs. These states only occupy
the sublattice where the qubit is in the ↓ij
state, and they are protected from other eigen-
states by an energy gap of the vacuum Rabi
frequency. Perturbations with energy smaller
than this gap, such as slow modulation of
coupling strengths and small detunings be-
tween the resonators, are used to coherently
control the zero-energy state to realize topologi-
cal transport and VHE. Floquet modulations
are introduced to realize the Haldane chiral
edge currents. The techniques that we have
developed in this study can also be applied
to control other eigenstates in the FSL, such
as the excited states in higher Landau levels.
Our methods can be generalized to investi-
gate topological states of more complex qubit-
resonator coupled systems, where the number
of resonators determines the dimension of the
FSLs and each state of the qubits labels a sub-
lattice, with richness beyond known topologi-
cal phases in condensed-matter physics. Our
study paves the way for investigating topologi-
cal phases in FSLs and developing new control
methods for quantum state engineering of
bosonic modes.
REFERENCES AND NOTES
1. K. v. Klitzing, G. Dorda, M. Pepper, Phys. Rev. Lett. 45,
494–497 (1980).
2. D. J. Thouless, M. Kohmoto, M. P. Nightingale, M. den Nijs,
Phys. Rev. Lett. 49, 405–408 (1982).
3. F. D. M. Haldane, Phys. Rev. Lett. 61, 2015–2018 (1988).
4. C. L. Kane, E. J. Mele, Phys. Rev. Lett. 95, 226801 (2005).
5. F. D. M. Haldane, S. Raghu, Phys. Rev. Lett. 100, 013904 (2008).
6. L. Lu, J. D. Joannopoulos, M. Soljačić, Nat. Photonics 8,
821–829 (2014).
7. A. B. Khanikaev, G. Shvets, Nat. Photonics 11, 763–773 (2017).
8. T. Ozawa et al., Rev. Mod. Phys. 91, 015006 (2019).
9. Z. Wang, Y. Chong, J. D. Joannopoulos, M. Soljacić, Nature 461,
772–775 (2009).
10. M. A. Bandres et al., Science 359, eaar4005 (2018).
11. L. Yuan, Q. Lin, M. Xiao, S. Fan, Optica 5, 1396 (2018).
12. T. Ozawa, H. M. Price, Nat. Rev. Phys. 1, 349–357 (2019).
13. E. Lustig, M. Segev, Adv. Opt. Photonics 13, 426 (2021).
14. H. Cai, D.-W. Wang, Natl. Sci. Rev. 8, nwaa196 (2020).
15. D. Cheng et al., Phys. Rev. Res. 3, 033069 (2021).
16. J. Mumford, Meissner effect in Fock space. arXiv:2208.01539
[quant-ph] (2022).
17. P. Saugmann, J. Larson, A Fock state lattice approach to
quantum optics. arXiv:2203.13813 [quant-ph] (2022).
I. Carusotto et al., Nat. Phys. 16, 268–279 (2020).
18.
19. J. Yuan et al., Phys. Rev. B 104, 035410 (2021).
20. D.-W. Wang, H. Cai, R.-B. Liu, M. O. Scully, Phys. Rev. Lett. 116,
220502 (2016).
21. Y. Y. Gao et al., Nature 566, 509–512 (2019).
22. L. Hu et al., Nat. Phys. 15, 503–508 (2019).
23. W.-L. Ma et al., Sci. Bull. 66, 1789–1805 (2021).
24. J. Mumford, Phys. Rev. A 106, 033317 (2022).
25. H. M. Price, T. Ozawa, N. Goldman, Phys. Rev. A 95, 023607
(2017).
26. E. Jaynes, F. Cummings, Proc. IEEE 51, 89–109 (1963).
27. W. P. Su, J. R. Schrieffer, A. J. Heeger, Phys. Rev. Lett. 42,
1698–1701 (1979).
28. F. Guinea, M. I. Katsnelson, A. K. Geim, Nat. Phys. 6, 30–33
(2010).
29. E. J. Meier, F. A. An, B. Gadway, Nat. Commun. 7, 13986
(2016).
30. M. Kiczynski et al., Nature 606, 694–699 (2022).
31. O. Gröning et al., Nature 560, 209–213 (2018).
32. S. de Léséleuc et al., Science 365, 775–780 (2019).
33. G. Jotzu et al., Nature 515, 237–240 (2014).
34. P. Roushan et al., Nature 515, 241–244 (2014).
35. J. W. McIver et al., Nat. Phys. 16, 38–41 (2020).
36. W. Zhao et al., Realization of the Haldane Chern insulator in a
moiré lattice. arXiv:2207.02312 [cond-mat.mes-hall] (2022).
37. D. Xiao, W. Yao, Q. Niu, Phys. Rev. Lett. 99, 236809 (2007).
38. M. C. Rechtsman et al., Nature 496, 196–200 (2013).
39. F. Arute et al., Nature 574, 505–510 (2019).
40. Y. Wu et al., Phys. Rev. Lett. 127, 180501 (2021).
41. D.-W. Wang et al., Nat. Phys. 15, 382–386 (2019).
42. W. Liu, W. Feng, W. Ren, D.-W. Wang, H. Wang, Appl. Phys.
Lett. 116, 114001 (2020).
43. P. Roushan et al., Nat. Phys. 13, 146–151 (2017).
44. H.-C. Fu, R. Sasaki, J. Math. Phys. 38, 3968–3987 (1997).
45. J. Yuan, C. Xu, H. Cai, D.-W. Wang, APL Photonics 6, 030803
(2021).
46. M. C. Rechtsman et al., Nat. Photonics 7, 153–158 (2013).
47. O. Jamadi et al., Light Sci. Appl. 9, 144 (2020).
48. M. Bellec, C. Poli, U. Kuhl, F. Mortessagne, H. Schomerus, Light
Sci. Appl. 9, 146 (2020).
49. P. Roushan et al., Science 358, 1175–1179 (2017).
50. A. H. Castro Neto, F. Guinea, N. M. R. Peres, K. S. Novoselov,
A. K. Geim, Rev. Mod. Phys. 81, 109–162 (2009).
51. M. O. Goerbig, Rev. Mod. Phys. 83, 1193–1243 (2011).
52. G. Salerno, T. Ozawa, H. M. Price, I. Carusotto, Phys. Rev. B 95,
245418 (2017).
53. F. Gao et al., Nat. Phys. 14, 140–144 (2018).
54. J. Deng, Observing the quantum topology of light [Data set],
Zenodo (2022). https://doi.org/10.5281/zenodo.7267558.
AC KNOWLED GME NTS
We thank K. Wang, W. Ren, J. Chen, and X. Zhang for technical
support during the experiment. D.-W.W. is indebted to R. B. Liu
Deng et al., Science 378, 966–971 (2022)
2 December 2022
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
and C. Wu for inspiration, insightful discussion, and encouragement.
The device was fabricated at the Micro-Nano Fabrication Center
of Zhejiang University. Funding: We acknowledge the support
of the National Key Research and Development Program of China
(grant no. 2019YFA0308100), National Natural Science Foundation
of China (grant nos. 11934011, 12174342, 92065204, U20A2076,
and 11725419), the Zhejiang Province Key Research and
Development Program (grant no. 2020C01019), the Fundamental
Research Funds for the Zhejiang Provincial Universities (grant
no. 2021XZZX003), and Innovation Program for Quantum
Science and Technology (grant no. 2021ZD0303200). Author
contributions: D.-W.W., C.S., and H. W. designed the experiment;
J.D. and H.D. designed the device and carried out the experiments
supervised by C.S. and H.W.; H.L. fabricated the device
supervised by H.W.; J.D., C.Z., Y.W., and J.Y. performed the
numerical simulations supervised by C.S. and D.-W.W.; J.D., C.S.,
and D.-W.W. wrote the manuscript with comments and inputs
from other authors. All authors contributed to the analysis of data
and the discussions of the results. Competing interests: All
authors declare no competing interests. Data and materials
availability: The data presented in the figures and that support
the other findings of this study are available at Zenodo (54). License
information: Copyright © 2022 the authors, some rights reserved;
exclusive licensee American Association for the Advancement of
Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade6219
Supplementary Text
Figs. S1 to S11
Table S1
References (55–61)
Submitted 28 August 2022; accepted 4 November 2022
10.1126/science.ade6219
Deng et al., Science 378, 966–971 (2022)
2 December 2022
6 of 6
|
10.1126_science.ade8654
|
RES EARCH
APPLIED PHYSICS
Fiber pumps for wearable fluidic systems
Michael Smith1*, Vito Cacucciolo1,2, Herbert Shea1*
Incorporating pressurized fluidic circuits into textiles can enable muscular support, thermoregulation,
and haptic feedback in a convenient wearable form factor. However, conventional rigid pumps,
with their associated noise and vibration, are unsuitable for most wearables. We report fluidic
pumps in the form of stretchable fibers. This allows pressure sources to be integrated directly into
textiles, enabling untethered wearable fluidics. Our pumps consist of continuous helical electrodes
embedded within the walls of thin elastomer tubing and generate pressure silently through
charge-injection electrohydrodynamics. Each meter of fiber generates 100 kilopascals of pressure,
and flow rates approaching 55 milliliters per minute are possible, which is equivalent to a power
density of 15 watts per kilogram. The benefits in design freedom are considerable, which we illustrate
with demonstrations of wearable haptics, mechanically active fabrics, and thermoregulatory textiles.
T he widespread use of textiles and their
close contact with the human body make
functional fabrics an attractive technol-
ogy for wearable devices. When creat-
ing technology that is meant to be worn,
fiber-format components offer several advan-
tages because of their flexibility, scalability,
and compatibility with existing textile produc-
tion techniques (1, 2). Fiber-based actuators
(3), sensors (4), energy storage (5), and en-
ergy harvesting (6) devices demonstrate the
progress in this field and the benefits afforded
by this format.
Lacking, however, are fiber-based fluidic
components. Fluidic actuation is fundamental
to much of soft robotics—a key axis of wear-
able technology (7, 8). Fluidic systems also
enable numerous other processes, including
heat and mass transport, within conventional
medical, industrial, and personal devices. The
lack of active fluidic components that can be
straightforwardly integrated into textiles is
inhibiting the development of countless useful
devices—devices such as soft supportive exo-
skeletons, adaptive medical orthoses, and wear-
able haptics.
Generating meaningful fluidic power (the
product of pressure and flowrate) in a wear-
able and portable manner remains a sub-
stantial challenge. Although several “soft”
pumps made entirely from compliant ma-
terials have recently been reported (9–13),
their low levels of power density hinder un-
tethered wearable applications. Furthermore,
none of these pumps has a form factor that
can be woven into a textile, and none demon-
strate the scalability necessary for human-
sized applications.
We present a fluidic pump in the form of
a fiber (Fig. 1A). The pump is a flexible and
1Soft Transducers Laboratory (LMTS), École Polytechnique
Fédérale de Lausanne, Neuchâtel, Switzerland. 2Department
of Mechanics, Mathematics and Management (DMMM),
Politecnico di Bari, Bari, Italy.
*Corresponding author. Email: michael.smith@epfl.ch (M.S.);
herbert.shea@epfl.ch (H.S.)
stretchable tube, on the order of millimeters
in diameter, that can generate continuous
fluid flow without any moving parts or vi-
bration. The amount of pressure and flow-
rate generated by the fibers is a function of
both their length (Fig. 1E) and diameter
(Fig. 2B).
The fiber pumps are able to pump liquid in a
form factor that is inherently compatible with
wearable devices (Fig. 1F). In generating flow
directly within the tubing itself (Fig. 1G), the
need for an external pump is removed, and
fluidic systems can be greatly simplified. This
affords substantial improvements in design
freedom for wearables. Pumps can be distrib-
uted throughout the volume of the device, re-
ducing losses, improving comfort, and enabling
advanced multipump wearables without the
need for valves and connectors. The distrib-
uted nature of the fiber pumps means that the
weight of the pump can be counted against the
weight of tubing required in a conventional
fluidic circuit. In the case of fiber pumps, the
tubing is the pump, and effectively the pump
comes at no extra weight penalty.
Fiber pump fabrication and performance
The pumps operate through the principle of
charge injection electrohydrodynamics (EHD)
(9, 14–16). Embedded in the pumps’ poly-
urethane tube wall are two continuous helical
electrodes, fabricated from copper wire (80 mm
diameter). Applying a dc electric field up to
8 kV/mm between these electrodes ionizes
a dielectric liquid within the tube, creating
negatively charged ions as liquid molecules
accept electrons. These ions are accelerated
toward the positive electrode, where they dis-
charge. As they move, they set in motion the
surrounding liquid molecules to create a net
fluid flow (Fig. 1C). The asymmetrical spacing
of the helical electrodes ensures that for a
given polarity of applied voltage, there is a net
flow in one direction. Inverting the polarity of
the applied voltage will reverse the direction of
flow (fig. S7A).
Helical configurations of electrodes have
previously been simulated for use in EHD con-
duction pumps (a related but distinct pumping
mechanism) (17). We demonstrate that this
configuration can produce exceptional perform-
ance as an ion-drag EHD pump. We report
pressures exceeding 80 kPa and flow rates ap-
proaching 55 ml/min. Considering the weight
of the pump, this corresponds to power
densities in excess of 15 W/kg, which is sub-
stantially larger than previous attempts at
soft pumps and within a factor of two of
class-leading (rigid) miniature pumps (table
S3). Furthermore, this configuration is up to
10 times more efficient than a conventional
layout of interdigitated electrodes (9); can
generate 30 times the pressure response
rate (Fig. 2F); and results in a stable, repeat-
able device (fig. S6).
The structure of the fiber pumps is created by
using a filament winding fabrication method
(Fig. 1D, fig. S5, and movie S2). This involves
simultaneously twisting thermoplastic poly-
urethane (TPU) filament and copper wire
together around a central mandrel and subse-
quently fusing the filament together with heat.
Removing the mandrel reveals a hollow tube
with helical electrodes embedded within the
walls (18). Crucially, these electrodes remain
exposed along the inner surface of the tube
wall (Fig. 1B, ii), enabling the injection and
collection of charge necessary for EHD pump-
ing. The filament winding method is continu-
ous and allows for a simple pump construction
from only two materials. This greatly reduces
the compatibility issues between pump mate-
rials and EHD liquids that were previously
reported (9).
This same fabrication method can be used
to create different pump geometries. The exact
structure is determined by the close packing of
the TPU filaments around the mandrel, which
is dependent on the number of filaments, their
diameter, and the diameter of the mandrel
(supplementary text and fig. S1). Changing the
mandrel diameter leads to different diameter
pumps (Fig. 2A). Because the structure is heli-
cal, however, the pump diameter cannot be
changed independently; some other aspect
of the structure must also change. For the
pumps shown in Fig. 2, the helix angle has
been kept constant at ~60° by changing the
number of filaments for each diameter (table
S1). Consequently, the helix pitch is different
in each case.
Under these conditions, reducing the pump
diameter increases the maximum pressure but
decreases the maximum flowrate (Fig. 2B, i). A
maximum power density of 15.2 W/kg occurs
at a diameter of 2.3 mm (1.5 mm inner diameter)
(Fig. 2B, ii), although the optimum helix angle
and pitch are likely different for each diameter
(supplementary text and fig. S2). The elec-
trode spacing can be controlled independently
Smith et al., Science 379, 1327–1332 (2023)
31 March 2023
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
B
E
G
Fig. 1. A fiber-format electrohydrodynamic pump. (A) The fiber pump is soft and flexible. (B) (i) The
structure of the fiber pump, showing (ii) the helical copper electrodes embedded within the walls of a
polyurethane tube. (C) A schematic illustration of the pump operating principle. Liquid molecules become
ionized at the negative electrode, accepting an electron. The ions are then accelerated to and discharged at
the positive electrode, generating a net fluid flow. (D) Pumps are fabricated by using a filament-winding
method. (E) The scaling of fiber pump pressure and flow rate with pump length, for an applied electric field of
8 kV/mm. Values are mean ± SD for n = 3 pumps of each length. (F) The fiber pumps sewn into a woolen
textile. (G) The fiber pump in operation (movie S1).
of diameter and is set by the number of TPU
filaments between the copper wires. In all
cases in this study, two filaments were used,
giving a (perpendicular) electrode spacing of
800 mm. We anticipate that the pump struc-
ture can be optimized further, leading to ad-
ditional improvements in performance.
With respect to the applied electric field, the
maximum pressure scales nonlinearly (Fig. 2C),
as is typically seen in charge-injection EHD
pumps (15). Maximum flowrate scales lin-
early with applied field once above a value of
5 ml/min (Fig. 2D), a behavior also seen in
other EHD pumps (15, 19). Peak power density
and efficiency both increase with increasing
pump length (Fig. 2E) and electric field (fig. S7,
E and F). The fiber pumps can pump contin-
uously for up to 6 days before chemical deposits
passivate the electrodes (fig. S7H). This is con-
siderably longer than any other reported EHD
pump, soft or otherwise (9), and can likely be
improved further through careful selection of
liquid and electrode material.
Throughout this work, we used the dielectric
liquid Novec 7100 (3M). This is a nontoxic,
nonflammable methoxy-fluorocarbon with
low global warming potential (20). It is com-
monly used as a solvent and for heat man-
agement but also performs well as an EHD
fluid because of readily ionizable fluorine
groups and a high electrical breakdown field
of 10 kV/mm (20). Other liquids (including
nonfluorinated compounds) may be pumped,
provided that they have low conductivity and
viscosity and are electrochemically stable
(21, 22).
The electric fields used to drive these pumps
equate to voltages on the order of kilovolts.
This is necessary to overcome the energy bar-
rier for ionization of the liquid. However, the
amount of current flowing is low (<110 mA)
(fig. S7I). This results in a low power con-
sumption of ~0.9 W/m. Consequently, the fiber
pumps can be powered with a battery-operated
high-voltage power supply, weighing 35 g [in-
cluding battery (fig. S11)], enabling untethered
operation (movie S1). A typical smartphone
battery would power a 1-m length of pump
continuously for ~15 hours.
Pump flexibility and stretchability
A major attraction of fiber-based technologies
is the compliance offered by this format. The
fiber pumps are flexible about every direction
perpendicular to their axis, offering substan-
tial design freedom for integration into wear-
ables and highly compliant soft actuators.
Therefore, we have characterized how the
fiber pumps behave when deformed. Stretch-
ability was assessed by measuring the pressure-
flowrate characteristics of the pump—known
as the “pump curve”—at different values of
uniaxial strain (Fig. 2G). Even at 15% strain (a
strain level sufficient for wearable applications),
Smith et al., Science 379, 1327–1332 (2023)
31 March 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
A
C
G
D
B
E
H
F
Fig. 2. Characterization of the fiber pumps. Values with error bars
correspond to mean ± SD for n = 3 pumps. (A) Fiber pumps fabricated at
different diameters (d) with an approximately constant helix angle q ~ 60°.
(B) Variation in (i) maximum pressure and flowrate and (ii) power density for
the different pump diameters at a constant helix angle of 60°. l, length;
E, applied electric field. (C) Pressure and (D) flowrate scaling with applied
electric field for various pump lengths, all with an outer diameter of 2 mm.
(E) The power density and efficiency as a function of pump length, for a
2-mm-diameter pump at 8 kV/mm. (F) Pressure and voltage as a function of
time. Maximum stable pressure is reached within 200 ms. The oscillations in the
pressure measurement are due to the elasticity of the pump and connecting
tubing. (G) (i) The pressure-flowrate characteristics of a fiber pump at (ii)
different levels of uniaxial strain. (H) (i) The pressure-flowrate characteristics
of a fiber pump subjected to different radii of curvature (r). Curvature is achieved
by wrapping the pump around circular objects with multiple loops (ii) to
ensure that an equal length of pump was deformed in each case.
pump performance was broadly unchanged.
Beyond 25% strain, the pump would begin to
fail as the electrode structure detached from
the tube wall (fig. S8). This stretchability was
achieved by using conventional materials:
The helical nature of the electrodes means that
a stretchable device can be fabricated with-
out lossy and unreliable stretchable conduc-
tive materials (Fig. 2G, ii).
The pump curves captured as the fibers are
deformed to different radii of curvature are
displayed in Fig. 2H, i. To enable a fair com-
parison between different curvatures, we ad-
justed the number of loops to ensure that an
equal length of pump was deformed in each
case (Fig. 2H, ii). Deforming to a bend radius
of 4 mm—four times the radius of the pump
itself—reduced the maximum pressure by 7%
and maximum flowrate by 12% when com-
pared with an undeformed pump. The slight
reduction can be explained by considering the
change in electrode spacing around the inside
and outside of the deformed pump, as well as
the pressure loss inherent to fluid flow through
a curved tube. The pump will kink if subjected
to a sufficiently low bend radius, typically
<3 mm, although the exact value will depend
on the loading conditions and pump dimen-
sions (23). The kink, however, does not dam-
age the pump structure, and the pump can
continue to operate once the kink is removed
(fig. S8 and movie S9).
Under repeated uniaxial strain, the pumps
can withstand at least 5000 cycles to 10%
strain without change in performance (fig. S9).
Repeated flexion to a bend radius of 3 mm for
5000 cycles has no effect on pump perform-
ance (fig. S10).
Fiber and textile-based actuators with
fiber pumps
The ability of these pumps to generate sub-
stantial flow rates and pressures while main-
taining flexibility permits multiple applications
in textile-based and wearable applications, which
previously required an external conventional
pump. We have demonstrated the fiber pumps
in the context of mechanically active fibers
Smith et al., Science 379, 1327–1332 (2023)
31 March 2023
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Actuators powered using fiber pumps. (A) A fiber pump connected to a 9-cm-long inverse hydraulic
artificial muscle (IHAM). (B) The IHAM consists of helices of inelastic thread embedded within the wall of an
elastic tube. (C) A maximum strain of 15% is achieved from a single fiber. (D) Combining IHAMs and pumps in
parallel can generate higher forces. In this image, three IHAMs and pumps are used to move a 300 g load. (E) The
strain of the three-fiber IHAM as a function of the electric field applied to the fiber pumps. (F) The frequency
response of the three-fiber IHAM moving a 200-g weight. Applied signal is 6.4 kV square wave (equivalent to 8 kV/mm).
We observed a small resonance peak at ~2.4 Hz. (G) The three-fiber IHAM can operate for at least 5000 cycles at
2.4 Hz, with minimal loss in amplitude. (H) A fluidic fabric actuator with integrated fiber pump. In this device, an
elastic tube is sewn between two sheets of inextensible fabric. The ruffling of the fabric allows the actuator to
extend when pressurized. (I) The fiber pump integrated directly into the fabric actuator. (J) The maximum
difference in isometric axial force generated by the actuator as a function of the applied electric field.
and textiles (Fig. 3), as well as thermoregula-
tory wearables for heat management and hap-
tic feedback (Fig. 4).
Actuation can be achieved by using fiber-
like artificial muscles, which are similar in con-
cept to inverse pneumatic artificial muscles
(IPAMs) (24) but operate with a liquid rather
than a gas. These inverse hydraulic artificial
muscles (IHAMs) can be integrated directly
with the fiber pumps (Fig. 3A and movie S8).
The IHAM consists of an elastic tube with a
helix of inelastic thread embedded within its
walls. This thread acts as a radial constraint so
that as the IHAM is pressurized by the fiber
pump, the device does not expand radially
but instead extends along its length (Fig. 3B).
Actuation is fast (Fig. 3C) because of the
low internal volume of the actuator (~ 300 ml)
and the high performance of the fiber pump.
A single fiber exerts a maximum difference
in isometric axial force of ~0.5 N, and ac-
tuators can be used in parallel to increase
force output (Fig. 3D). Response to the ap-
plied voltage is repeatable, with low hyster-
esis (Fig. 3E). The fast time response of the
fiber IHAMs facilitates operation beyond the
quasi-static regime (Fig. 3F), where they can
continue to operate for several thousands of
cycles (Fig. 3G).
Fiber pumps can be integrated directly into
textiles to create active fabrics, as demon-
strated in Fig. 3H. The device shown consists
of a fluidic fabric muscle (25) with a fiber pump
woven directly into the fabric (Fig. 3I). The
operating principle is similar to that of the
IHAMs, except here the radial constraint is
supplied by an inextensible fabric sewn around
the elastic tubing. The result is an active fabric,
capable of changing shape upon electrical com-
mand (movie S7). Operating at an applied field
of 8 kV/mm, maximum strains approaching
40% are possible (fig. S12D), and actuation
strains of 16% were reached with an applied
load of 100 g (Fig. 3H). We achieved a maxi-
mum isometric axial force of 1.7 N, which was
controllable by tuning the pump voltage (Fig.
3J). Actuation speed was reduced because of
the larger internal volume; nonetheless, full
actuation occurred in under 20 s, an adequate
timeframe for wearable applications such as
self-adjusting clothing and intermittent com-
pression devices. The entire device can also be
washed with standard laundry detergent (fig.
S13 and movie S10) without compromising its
performance.
The example actuators shown in Fig. 3 use
the pump’s working fluid directly. It is equally
feasible to instead use this fluid to displace
a second fluid, such as air, to allow exist-
ing pneumatic actuators to be used with-
out modification.
Thermoregulation and thermal haptics with
fiber pumps
We also demonstrated two applications of
fiber pumps in thermally active textiles. Shown
in Fig. 4A is a thermal haptic glove, a class
of device used to generate different temper-
ature sensations to enhance the sense of im-
mersion in virtual reality (26). Typically reliant
upon rigid thermoelectric ceramics, in this
work we demonstrated how a thermal haptic
glove can be realized by using soft and flex-
ible fiber pumps.
Smith et al., Science 379, 1327–1332 (2023)
31 March 2023
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Heat management by use of fiber pumps. (A) A thermal haptic glove with separate, individually
controlled fiber pumps attached to each finger. (B) Thermal images of the glove (i) before and (ii) 2 s after
simultaneously activating all pumps. (C) A demonstration of thermal haptic feedback (i) approaching
the object with no pumps active, (ii) activating the pump of the finger in contact with the object, chilling the
finger and providing a thermal haptic stimulus, subsequently activating (iii) the thumb and (iv) remaining
fingers as they come into contact with the object. (D) A fully wearable and untethered thermoregulatory
garment, with integrated fiber pumps, power supply and fluid reservoir.
The glove consists of a reservoir of chilled
Novec 7100 fluid and five separate, indepen-
dently controlled fiber pumps winding around
the fingers. Activating a pump rapidly flows
cold fluid around the finger (Fig. 4B), chilling
the pump to the fluid temperature within 2 s
and providing a thermal haptic stimulus to
the user (movie S4). Once the pump voltage
is removed and without the constant flow of
chilled fluid from the reservoir, the pump re-
turns to room temperature. In a virtual reality
setting, this allows the temperature and ther-
mal properties of objects to be simulated with
a degree of spatiotemporal resolution: The
physical object that represents the virtual one
can be made to feel cold, and that sensation is
felt only through those fingers in contact with
the object (Fig. 4C). Achieving the same effect
with any other pump would require several
separate pumps or an array of valves, hinder-
ing wearability.
The same principle can be applied to an
entire garment (Fig. 4D). Four sections of
fiber pump transport chilled fluid around the
torso and arms. The pumps are powered with
a battery-operated power supply (fig. S11),
creating an untethered system that can be
used in any environment and during physical
activity (movie S5). As well as enabling un-
tethered operation, the low power consump-
tion of the pumps means that Joule heating
is negligible in comparison with the cooling
rate provided by the chilled fluid.
The concept of thermoregulatory fabric
can be developed further (fig. S14). The pump
fibers themselves can be woven together
to create a fluidic fabric that, when attached to
hot and cold reservoirs of fluid, allows the tem-
perature of the fabric to be regulated (movie
S6). Such a patch of woven pumps can also be
considered as a pressure source, sewn di-
rectly into the textile of a wearable device to
provide high-pressure fluid for actuation.
Concluding remarks
We have demonstrated a fluidic pump in the
form of a flexible fiber that is capable of gen-
erating exceptional pressures and flow rates.
This advancement allows rigid and bulky ex-
ternal pumps to be removed from wearable
fluidic devices—effectively turning the tubing
into the pump. The scalability, simplicity, and
stability of both the operation and fabrica-
tion of these fibers enable applications that
are not possible with conventional pumping
technologies.
REFERENCES AND NOTES
1. G. Loke, W. Yan, T. Khudiyev, G. Noel, Y. Fink, Adv. Mater. 32,
e1904911 (2020).
J. Xiong, J. Chen, P. S. Lee, Adv. Mater. 33, e2002640 (2021).
2.
3. C. S. Haines et al., Science 343, 868–872 (2014).
4. W. Yan et al., Nature 603, 616–623 (2022).
5. T. Khudiyev et al., Mater. Today 52, 80–89 (2022).
6. H. Sun, Y. Zhang, J. Zhang, X. Sun, H. Peng, Nat. Rev. Mater. 2,
17023 (2017).
7. T. Abe, S. Koizumi, H. Nabae, G. Endo, K. Suzumori, in 2018
IEEE International Conference on Soft Robotics (RoboSoft)
(2018), pp. 572–578.
8. O. Kilic Afsar et al., in The 34th Annual ACM Symposium on User
Interface Software and Technology (Association for Computing
Machinery, New York, NY, USA, 2021), pp. 1010–1026.
9. V. Cacucciolo et al., Nature 572, 516–519 (2019).
10. Y. Matia, H. S. An, R. F. Shepherd, N. Lazarus, Proc. Natl. Acad.
Sci. U.S.A. 119, e2203116119 (2022).
11. R. S. Diteesawat, T. Helps, M. Taghavi, J. Rossiter, Sci. Robot.
6, eabc3721 (2021).
12. W. Tang et al., Nat. Commun. 12, 2247 (2021).
13. C. Stergiopulos et al., in ASME 2014 Conference on Smart
Materials, Adaptive Structures and Intelligent Systems
(American Society of Mechanical Engineers Digital Collection,
2014); https://asmedigitalcollection.asme.org/SMASIS/
proceedings/SMASIS2014/46155/V002T04A011/286215.
14. A. Ramos, Ed., Electrokinetics and Electrohydrodynamics in
Microsystems (Springer Vienna, 2011).
15. J. Kim, Y. Yamada, S. Yokota, Int. J. Adv. Manuf. Technol. 106,
627–639 (2020).
16. Z. Mao, T. Iizuka, S. Maeda, Sens. Actuators A Phys. 332, 113168 (2021).
17. N. O’Connor, J. Yagoobi, in 2021 IEEE Industry Applications
Society Annual Meeting (IAS) (2021); https://ieeexplore.ieee.
org/document/9677332.
18. A detailed account of pump fabrication is provided in the
materials and methods section of the supplementary materials
(fig. S5 and movie S2).
19. T. Matsubara, H. H. Huynh, K. Yoshida, J. Kim, Sens. Actuators
A Phys. 295, 317–323 (2019).
20. M. S. EL-Genk, H. Bostanci, Int. J. Heat Mass Transf. 46,
1841–1854 (2003).
21. S. Yokota, Y. Otsubo, K. Edamura, Electro-sensitive movable
fluids, methods of using the same and motors for the
electro-sensitive movable fluids (2000); https://patents.
google.com/patent/US6030544A/en.
22. J. E. Bryan, J. Seyed-Yagoobi, IEEE Trans. Electr. Insul. 26,
647–655 (1991).
23. K. Luo, P. Rothemund, G. M. Whitesides, Z. Suo, J. Mech.
Phys. Solids 131, 230–239 (2019).
24. E. W. Hawkes, D. L. Christensen, A. M. Okamura, in 2016 IEEE
International Conference on Robotics and Automation (ICRA)
(2016), pp. 4022–4029.
25. M. Zhu, T. N. Do, E. Hawkes, Y. Visell, Soft Robot. 7, 179–197 (2020).
26. Y. Huang et al., Mater. Today Phys. 22, 100602 (2022).
27. M. Smith, V. Cacucciolo, H. Shea, Research data supporting
‘Fiber pumps for wearable fluidic systems’. Zenodo (2023);
https://doi.org/10.5281/zenodo.7451722.
Smith et al., Science 379, 1327–1332 (2023)
31 March 2023
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
ACKN OW LEDG MEN TS
The authors thank V. Py and M. Benbedda for assistance with
the high-voltage electronics featured in this work. We thank S. Maeda
and F. Hartmann for helpful discussions. We also thank J. La Scala
for developing the automated fabrication system used in this work.
Thanks also to L. Caillaud and E. Valicka for help in creating the
textile-based demonstrations. Further thanks to E. Valicka and
R. Hennig for assistance with videography and video editing.
Funding: We acknowledge the support of the Swiss National
Science Foundation (grant IZLJZ2_183656) Author contributions:
M.S., V.C., and H.S. formulated the concept; M.S. fabricated the
devices, carried out the experiments, analyzed the data, and
created the figures and videos, with supervision from H.S. and
V.C; M.S. prepared the original draft of the manuscript; M.S., V.C.,
and H.S. reviewed and edited the manuscript. Competing
interests: The authors declare no competing interests. Data and
materials availability: All data that support the claims in this
manuscript are available on the Zenodo repository (27). License
information: Copyright © 2023 the authors, some rights reserved;
exclusive licensee American Association for the Advancement
of Science. No claim to original US government works. https://
www.science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade8654
Materials and Methods
Supplementary Text
Figs. S1 to S14
Tables S1 to S3
Movies S1 to S10
Submitted 23 September 2022; accepted 28 February 2023
10.1126/science.ade8654
Smith et al., Science 379, 1327–1332 (2023)
31 March 2023
6 of 6
|
10.1126_science.ade8450
|
RES EARCH
NANOPHOTONICS
Experimentally realized in situ backpropagation for
deep learning in photonic neural networks
Sunil Pai1*†, Zhanghao Sun1, Tyler W. Hughes2‡, Taewon Park1, Ben Bartlett2†, Ian A. D. Williamson1§,
Momchil Minkov1‡, Maziyar Milanizadeh3, Nathnael Abebe1#, Francesco Morichetti3, Andrea Melloni3,
Shanhui Fan1, Olav Solgaard1, David A. B. Miller1
Integrated photonic neural networks provide a promising platform for energy-efficient, high-throughput
machine learning with extensive scientific and commercial applications. Photonic neural networks efficiently
transform optically encoded inputs using Mach-Zehnder interferometer mesh networks interleaved with
nonlinearities. We experimentally trained a three-layer, four-port silicon photonic neural network with
programmable phase shifters and optical power monitoring to solve classification tasks using “in situ
backpropagation,” a photonic analog of the most popular method to train conventional neural networks. We
measured backpropagated gradients for phase-shifter voltages by interfering forward- and backward-
propagating light and simulated in situ backpropagation for 64-port photonic neural networks trained on MNIST
image recognition given errors. All experiments performed comparably to digital simulations (>94% test
accuracy), and energy scaling analysis indicated a route to scalable machine learning.
N eural networks (NNs) are ubiquitous com-
puting models loosely inspired by the
structure of a biological brain. Such mod-
els are trained on input data to implement
complex signal processing or “inference”
(1, 2), powering various modern technologies
ranging from language translation to self-
driving cars. The required energy for training
and inference to power these technologies has
recently been estimated to double every 5 to
6 months (3), and thus necessitates an energy-
efficient hardware implementation for NNs.
To address this problem, programmable
photonic neural networks (PNNs) have been
proposed as a promising, scalable, and mass-
manufacturable integrated photonic hard-
ware solution (4). A popular implementation
of PNNs consists of silicon photonic meshes,
N (cid:1) N networks of Mach-Zehnder interfer-
ometers (MZIs) and programmable phase
shifters (5–7), which optically accelerate the
most expensive operation in a PNN: unitary
matrix-vector multiplication (MVM). The MVM
y ¼ U x is implemented by simply sending
an input mode vector x (optical phases and
modes in N input waveguides) through the
network implementing U to yield output modes
y (4, 6, 8). This fundamental mathematical op-
eration, based on optical scattering theory,
additionally enables various analog signal pro-
cessing applications beyond machine learning
(4, 9) such as telecommunications (8), quantum
computing (10, 11), and sensing (12).
1Department of Electrical Engineering, Stanford University,
Stanford, CA 94305, USA. 2Department of Applied Physics,
Stanford University, Stanford, CA 94305, USA. 3Dipartimento
di Elettronica, Informazione e Bioingegneria, Politecnico di
Milano, Milan, Italy.
*Corresponding author. Email: spai@psiquantum.com
†Present address: PsiQuantum, Palo Alto, CA, USA.
‡Present address: Flexcompute Inc., Belmont, MA, USA.
§Present address: X Development LLC, Mountain View, CA USA.
#Present address: Google, Mountain View, CA USA.
Recently, “hybrid” PNNs, which interleave
programmable photonic linear optical elements
(e.g., meshes) and digital nonlinear activation
functions (9, 13), have proven to be a low-
latency and energy-efficient solution for NN
inference in circuit sizes of up to N ¼ 64 (14).
Compared to current fully analog PNNs with
electro-optic (EO) nonlinear activations (15, 16),
hybrid PNNs get around the critical problem
of photonic loss and offer more versatility than
multilayer PNNs for between-layer logical oper-
ations that do not favor optics. Such features may
be present in a number of state-of-the-art ma-
chine learning architectures such as recurrent
neural networks (17) and transformers (18, 19).
When fully optimized, the energy efficiency of
PNN inference has been estimated to be up to two
orders of magnitude higher than that of state-
of-the-art digital electronic application-specific
integrated circuits (ASICs) in artificial intelli-
gence (AI) (20). However, despite the success in
PNN-based inference, efficient on-chip training
of PNNs has not been demonstrated owing to
substantially higher experimental complexity
compared to the inference procedure.
In this study, we experimentally demon-
strated a photonic implementation of back-
propagation, the most widely used method
of training NNs (1, 2). [A minimal bulk optical
demonstration has been previously explored
(21).] Backpropagation is generally performed
by propagating error signals backward through
the NNs to determine programmable parame-
ter gradients via the chain rule. In our multi-
layer PNN device, we performed in situ training
on a foundry-manufactured silicon photonic in-
tegrated circuit by sending light-encoded errors
backward through the PNN and measuring
optical interference with the original forward-
going “inference” signal (22). Once trained,
our chip achieved an accuracy similar to that
of digital simulations, adding new capabilities
beyond existing inference or in silico learning
demonstrations (4, 23, 24). We further de-
signed and experimentally validated an analog
(electro-optic) phase-shifter update protocol, a
key improvement over past proposals requiring
more energy-intensive “digital subtraction” (22).
Finally, we systematically analyzed energy and
latency advantages of in situ backpropagation
and its scalability to larger (64 (cid:1) 64) PNN sys-
tems. Our findings ultimately pave the way for
energy-efficient optoelectronic training of neu-
ral networks and optical systems more broadly.
Photonic neural networks
(cid:1)
(cid:3)
h→ ‘ð Þ
We built a hybrid PNN by alternating sequences
of analog programmable unitary MVM op-
erations U ‘ð Þ
[implemented on a custom-
designed silicon photonic triangular mesh (6)]
and digital nonlinear transformations f ‘ð Þ [im-
plemented using autodifferentiation software
(25–27)] where layer ‘ ≤ L (total of L layers). The
PNN was parameterized by programmable
phase shifts h→∈ ½0; 2pÞD
, where D represents
number of PNN phase shifters. Mathemati-
cally, the following “inference” function sequence
transformed input x ¼ x 1ð Þ, proceeding in a
“feedforward” manner to the output z^ :¼ x Lþ1
Þ
(Fig. 1, A to D):
ð
y ‘ð Þ ¼ U ‘ð Þx ‘ð Þ
(cid:4)
ð
x ‘þ1
Þ ¼ f ‘ð Þ y ‘ð Þ
(cid:5)
ð1Þ
ð2Þ
ð
The “cost function” is defined as L x; z
Þ ¼
(cid:3)
(cid:1)
c z^ xð Þ; z
, where c represents the error be-
tween z^ and ground truth label z. Backprop-
agation updates parameters h→
that are on
D-dimensional gradient @L=@h→
evaluated for
“training example” x; z
Þ (or averaged over a
batch of examples).
ð
Each MZI was parametrized by thermo-optic
phase shifters that locally heat the waveguides
using current sourced from a separate control
driver board (Fig. 2, A and B). Phase shifts were
placed at the input (f, voltage Vf) and internal
(q, voltage Vq) arms of all MZIs to control the
propagation pattern of infrared C band (1530 to
1565 nm) light, enabling arbitrary unitary matrix
multiplication. We embedded an arbitrary 4 (cid:1) 4
unitary matrix multiply in a 6 (cid:1) 6 triangular
network of MZIs. This configuration incorpo-
rated two 1 (cid:1) 5 photonic meshes on either end
of the 4 (cid:1) 4 “matrix unit” capable of sending
any input vector x and measuring any output
vector y from Eqs. 1 and 2. These “generator”
and “analyzer” optical input/output (I/O) cir-
cuits (Figs. 1E and 2B and fig. S5) require cal-
ibrated voltage mappingsq Vqð
to control
optical phase (4, 28, 29) (fig. S2).
(cid:1)
Þ; f Vf
(cid:3)
Backpropagation demonstration
Our core result (Fig. 1E) was experimental re-
alization of backpropagation on a photonic
Pai et al., Science 380, 398–404 (2023)
28 April 2023
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. In situ backpropagation concept. (A) Example machine learning
problem: An unlabeled 2D set of points that are formatted to be input into a PNN.
(B) In situ backpropagation training of an L-layer PNN for the forward direction
and (C) the backward direction showing the dependence of gradient updates
for phase shifts on backpropagated errors. (D) An inference task implemented on
the actual chip resulted in good agreement between the chip-labeled points and
the ideal implemented ring classification boundary (resulting from the ideal
model) and a 90% classification accuracy. (E) Our proposed scheme performed
the three steps of in situ (analog) backpropagation, using a 6 (cid:1) 6 mesh
implementing coherent 4 (cid:1) 4 bidirectional unitary matrix-vector products using a
reference arm. The (1) forward, (2) backward, and (3) sum steps of in situ
backpropagation are shown. Arbitrary input setting and complete amplitude and
phase output measurement were enabled in both directions using the reciprocity
and symmetries of the triangular architecture. All powers throughout the
mesh were monitored by an IR camera using the tapped MZI shown in the inset
for each step, allowing for digital subtraction to compute the gradient (22).
These power measurements performed at phase shifts are indicated by green
horizontal bars.
triangular mesh MVM chip using a custom
optical rig and silicon photonic chip (fig. S1)
(22). Our backpropagation-enabled architec-
ture differs in three ways from a typical PNN
photonic mesh (4):
1) We enabled “bidirectional light propa-
gation,” the ability to send and measure light
propagating left to right or right to left through
the circuit (as depicted in Fig. 1E).
2) We implemented “global monitoring” to
measure optical power ph propagating through
any phase shift h in the circuit using 3% grating
taps (shown in the inset of Fig. 1E and Fig. 2,
A and B). In our proof-of-concept setup, we
used an infrared (IR) camera mounted on an
automated stage to image these taps through-
out the chip (fig. S1E).
3) We implemented both amplitude and
phase detection [improving on past approaches
(30)] using a self-configuring programmable
matrix unit layer (28) on both generator and
analyzer subcircuits (Figs. 1E and 2B and fig.
S5), which by symmetry worked for sending
and measuring light that propagated forward
or backward through the mesh.
These improvements on an already versa-
tile hardware platform enabled backpropa-
gation entirely using physical optical power
measurements to obtain cost gradients (22).
As shown in Fig. 1E, backpropagation required
global optical monitoring, and bidirectional
optical I/O was required to switch between
forward- and backward-propagating signals
to experimentally realize in situ backpropagation.
Equipped with these additional elements, our
protocol can be implemented on any feed-
forward photonic circuit (31) with the requi-
site analyzer and generator circuitry (Fig. 1 and
fig. S5).
Here we give a brief summary of the pro-
cedure (further explained in the supplemen-
tary text). The “forward inference” signal x ‘ð Þ
‘ð Þ
and “backward adjoint” signal x
adj are sent
forward and backward, respectively, through
the mesh that implements U ‘ð Þ. The “sum” vec-
Þ(cid:4) is sent forward, and subtract-
tor x ‘ð Þ (cid:3) iðx
ing the forward and backward measurements
from it digitally yields the gradient (22), a
reverse-mode differentiation process that we
call an “optical vector-Jacobian product (VJP).”
‘ð Þ
adj
Analog update
Going beyond an experimental implementation
of a past theoretical proposal (22), we addi-
tionally explored a more energy-efficient fully
analog gradient measurement update for the
final step, avoiding a digital subtraction update.
Instead of global monitoring optical power in
the first two steps and the final “sum” step, we
toggled an adjoint phase z tð Þ, a square wave
modulation with period T that periodically
toggles between “sum” and “difference” set-
tings z ¼ 0 and p corresponding to signal
‘ð Þ
‘ð Þ
T ¼ x ‘ð Þ∓iðx
. The gradient is
inputs x
(cid:3)
(cid:1)
adj
@L=@h ¼ ph;þ (cid:3) ph;(cid:3)
=4, or half the “signed
amplitude” of the AC (mean-subtracted) sig-
nal (supplementary text 2.6 and fig. S6). The
‘ð Þ
T were computed
sum and difference inputs x
digitally (off-chip), requiring O Nð
Þ operations to
compute per input. The sum and difference in-
puts were directly programmed at the generator
to compute phase gradients, and correspond-
ing sum and difference signal power measure-
ments at each phase shifter subtracted in the
analog domain to update phase-shift volt-
ages. One option to efficiently achieve a periodic
Þ(cid:4)
Pai et al., Science 380, 398–404 (2023)
28 April 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Analog gradient experiment and simulation. (A) The photonic mesh
chip was thermally controlled and wirebonded to a custom printed circuit
board (PCB) with fiber array for laser input/output and a camera overhead for
imaging the chip. Zooming in (IR camera image) reveals the core control-and-
measurement unit of the chip, enabling power measurement using 3% grating tap
monitors and a thermal TiN phase shifter nearby. (B) A 5-mW 1560-nm laser
and a calibrated control unit was used for input generation and output detection.
The IR camera over the chip imaged all grating tap monitors necessary for
backpropagation. (C) Analog gradient update might optionally be implemented
by introducing a summing interference circuit [not implemented on the chip in
(B)] between the input and adjoint fields. (D) The adjoint phase was toggled
between z ¼ 0 and p to evaluate the analog gradient measurement @Li=@h
for i ¼ 1 to 4. (E) Gradients measured using the toggle scheme yielded
approximately correct gradients when the implemented mesh was perturbed from
the optimal (target) unitary given 1 rad phase error standard deviation.
(F) Measured normalized gradient error decreased with cost function [distance
between implemented
single-example gradients outperformed digital gradients.
and optimal U ¼ DFT 4ð Þ], and analog batch and
U h→(cid:1) (cid:3)
^
Þ(cid:4)
‘ð Þ
adj
z toggle is to use the summing architecture
in Fig. 2C, which sums x ‘ð Þ and iðx
inter-
ferometrically with a fast modulator that im-
plements z. In an optimized scheme, we would
physically measure the gradient and update
the phase-shift voltage in the analog domain
using a photodiode, differential amplifier (im-
plementing an analog subtraction), and a
“sample-and-hold” update circuit using only a
single toggle (fig. S6, B and C). This scheme,
extended to energy-efficient “batch updates”
incorporating data from multiple training
examples, was tested on a single phase shifter
to demonstrate the logic of this electronic feed-
back scheme (materials and methods, supple-
mentary text 2.6, and fig. S7). Our demonstration
avoided a costly digital-analog and analog-
digital conversion; when fully integrated,
our approach avoids additional digital mem-
ory complexity required to program N 2 ele-
ments, enabling a truly analog backpropagation
scheme.
The local feedback just described updates each
phase shifter h using the measured gradient:
@L
@h
(cid:3)
(cid:1)
¼ I xhxh;adj
(cid:6)
(cid:6)
(cid:6)xhj2 (cid:3) jxh;adjj2
(cid:6)
xh;þj2(cid:3)
ph;þ (cid:3) ph (cid:3) ph;adj
2
¼
¼
¼
2
ph;þ (cid:3) ph;(cid:3)
4
ð3Þ
where the sum field xh;þ ¼ xh (cid:3) ix(cid:4)
h;adj and
the last equality of Eq. 3 indicate the mathe-
matical equivalence of “digital subtraction”
(Fig. 1E) and our proposed “analog subtrac-
tion” scheme (Fig. 2, C and D, and figs. S6
and S7). Pseudocode and the complete back-
propagation protocol are provided in supple-
mentary text 2.5. Digital and analog gradient
update steps can both be implemented in
parallel across all PNN layers once the mea-
surements from forward and backward steps
are determined.
We experimentally estimated the accuracy
of the analog gradient measurement for a
matrix optimization problem (7) by digital
processing of the optical power measurements
(Fig. 2D). We programmed a sequence of in-
puts into the generator unit of our chip and
recorded the square-wave response oscillating
between ph;þ and ph;(cid:3) and separately subtracted
the two measurements to find the gradient with
respect to h.
ru(cid:4)
We implemented in situ backpropagation
in a single photonic mesh layer, optimizing
the cost function defined for output port i via
r j2 or a “batch” cost function L ¼
Lr ¼ 1 (cid:3) ju^ T
X
4
Lr=4 averaged over four inputs (“batch
r¼1
size” M ¼ 4). Here, ur is row r of U, a target
matrix that we chose to be the four-point dis-
crete Fourier transform [DFT(4)], and u^
r is row
r of ^U , the implemented matrix on the device.
For our gradient measurement step, we sent in
the derivative yadj ¼ @Lr=@y ¼ (cid:3)2ðu^ T
er
to measure an adjoint field xadj, where er is
the rth standard basis vector (1 at position m,
0 everywhere else).
ru(cid:4)
rÞ(cid:4)
We evaluated gradient direction error as
1 (cid:3) g (cid:5) g^, comparing normalized measured
( g^ ) and predicted gradients g ¼ @L=@h→(cid:5)
∥@L=@h→∥(cid:3)1 . Both digital and analog gradi-
ents were less accurate near convergence, with
Pai et al., Science 380, 398–404 (2023)
28 April 2023
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. In situ backpropagation experiment. In situ backpropagation training
(34) was performed for two classification tasks solvable by (A) a three-layer hybrid
PNN consisting of absolute-value nonlinearities and a softmax (effectively sigmoid)
decision layer. (B) Three-step digital subtraction gradient update given monitored
waveguide powers and the measured gradient output. (C) For the circle dataset,
the digital and in situ backpropagation training curves show excellent agreement
resulting in (D) model accuracy of 96% test and 93% train (depicted here for
iteration 930, showing the true labels and the learned classification model outcomes)
and (E) histogram of low gradient error. (F) For the moons dataset, our phase
measurements were sufficiently inaccurate owing to hardware error affecting training,
leading to a lower model model accuracy of 94% test and 87% train (green). Using
ground truth phase (red), the device achieved (G) sufficiently high model accuracy
of 98% test and 95% train. (H) The histogram of gradient errors improved
considerably by roughly an order of magnitude using the correct phase measurement.
the errors empirically decreasing quadratically
with cost L (Fig. 2F). The analog batch gra-
dient (trained by averaging all four gradients
to give @L=@h) validated the photonic portion
of the batch scheme (figs. S6B and S7). All gra-
dient errors, regardless of implementation, scaled
similarly with convergence distance; uncali-
brated thermal cross-talk likely resulted in
gradient measurement errors that were compa-
rable to systematic power errors at the taps.
Digital subtraction encountered different losses
and coupling efficiencies in bidirectional tap
gratings, whereas analog gradient measurements
involved subtraction of only forward-going fields
at forward gratings, likely resulting in superior
performance (Fig. 2F). Finally, error in the full
analog subtraction scheme was independent
of batch size for the gradient calculation, and
no significant deviation due to timing jitter or
signal distortion was observed (fig. S7).
Photonic neural network training
To test overall on-chip training, we assessed the
accuracy of in situ backpropagation to train
multilayer PNNs using a digital subtraction
protocol (22) (Fig. 3A and fig. S3) automated
with Python software (32). We trained our
chip to implement L ¼ 3 layers with N ¼ 4
ports to assign labeled noisy synthetic data, gen-
erated using Scikit-Learn (33), in 2D space to a
0 or 1 label based on the data points’ spatial
location (Figs. 1A; 3, E and H; and fig. S4, I and
J). We performed an 80%:20% train–test split
(200 train points, 50 test points) and trained
on only train points to avoid overfitting.
To implement classification, our PNN assigned
a probability to each point being assigned a 0
or 1 on the basis of the following model:
z^ xð Þ ¼ softmax2ð U 3ð Þ
(cid:6)
(cid:6)
(cid:6)
(cid:6)
(cid:6)
(cid:6)U 2ð Þ U 1ð Þx
(cid:6)
(cid:6)jjÞ
ð4Þ
where softmax2 is the standard softmax (nor-
malized sigmoid) function applied to two
quantities: the total power in outputs 1 and 2
and total power in ports 3 and 4. The input
data x was engineered such that any 2D point
had the same total input power as a four-port
vector (materials and methods). Each point
was classified red or blue (0 or 1, respectively) on
the basis of whether the output of Eq. 4 obeyed
the condition z0 > z1 for each input (Fig. 3),
which we optimized using a binary cross-entropy
cost function (materials and methods).
Our chip performed data input, output, and
matrix operations for all PNN layers. At each
layer output, we digitally performed a square-
root operation on output power to implement
absolute-value nonlinearities [off-chip via JAX
and Haiku (26, 27)] and recorded output phases
for the backward pass of in situ backpropagation.
Ideally, PNNs are controlled by separate pho-
tonic meshes of MZIs for each linear layer to
achieve low power consumption. However,
to save on carbon footprint, we reprogrammed
the same chip to perform successive linear
layers because basic operating principles re-
main the same. We used the Adam gradient
update (34) with a learning rate of 0.01 and
performed digital simulations at each step to
fully compare measured and predicted per-
formance. Before on-chip training experi-
ments, we calibrated all phase shifters on the
chip (materials and methods and fig. S2) and
performed forward inference with digitally
pretrained neural network weights to verify
Pai et al., Science 380, 398–404 (2023)
28 April 2023
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. In situ backpropagation simulation. (A) A two-layer PNN was simulated on
MNIST data using a previously explored PNN benchmark incorporating rectangular
photonic meshes (31). (B and C) Marginal training curve statistics (shaded regions
indicate standard deviation error range about the mean) were computed over a
grid search of 72 tap noise, loss, and I/O amplitude and phase errors (materials
and methods). The dominant contributers were (B) tap noise factor stap (2.7%
increase for stap ¼ 0:02 from 3:7T0:7% average error) and (C) phase measurement
error sf (1.9% increase for sf ¼ 0:05 from 4T1% average error).
accurate calibration. We achieved 90% and 98%
device test set accuracy for ring and moons
datasets, respectively (fig. S4, I and J). Because
our photonic and digital implementation agreed
closely in inference accuracy, we performed
network training on-chip while conducting
evaluations off-chip for convenience.
During training of the circle dataset, predicted
and measured powers for grating tap-to-camera
monitor measurements showed excellent agree-
ment across all waveguide segments required
for accurate gradient computation (Fig. 3B,
fig. S3, and movie S1). The training curves in
Fig. 3C indicate that stochastic gradient descent
was a highly noisy training process for both pre-
dicted and measured curves owing to the noisy
synthetic dataset about the boundary and our
choice of single-example training as opposed
to batch training. These large swings appeared
roughly correlated between the simulated and
measured training curves (Fig. 3E), and we suc-
cessfully achieved 93% train and 96% test model
accuracy (Fig. 3D and fig. S4, A to C). We then
trained the moons dataset, applying the same
procedure to achieve 87% train and 94% test
model accuracy (Fig. 3F, green versus red). When
using the predicted phase and measured am-
plitudes, we reduced gradient error by roughly
an order of magnitude on average, resulting in
95% train and 98% test model accuracy (fig. S4,
D to F), which agreed with digital training (Fig.
3, F to H, and movie S2). This improvement
underscores the importance of accurate phase
measurement for improved training efficiency.
Further monitoring errors could be reduced by
increasing signal-to-noise ratio using integrated
avalanche photodiodes (35), noninvasive light
monitoring (36), or phase shifter–based power
monitoring (37).
Simulations and scalability
Given that our experimental results for N ¼ 4
PNNs showed evidence of hardware error af-
fecting training, we assessed the scalability for
N ¼ 64 PNNs on the MNIST handwritten
digit dataset (38) in the presence of error to
better understand the relative contributions
at scale. We implemented a PNN simulation
framework in Simphox (25) using JAX and
Haiku (26, 27) to simulate an in situ back-
propagation training given a grid search of
systematic and noise errors (materials and
methods). After 100 epochs using M ¼ 600
batch size, we achieved a maximum test ac-
curacy of roughly 97:2% in the ideal case and
a performance degradation to roughly 95%
on average (Fig. 4, B and C). Phase and am-
plitude errors arising from photodetector noise
and phase-shift quantization and calibration
errors affected convergence in error the most.
Overall, our MNIST simulation results suggest
that in situ backpropagation is relatively robust
at scale to noise and hardware errors, which
are difficult to eliminate completely in current
analog computing systems.
We also considered the energy and latency
trade-off with accuracy for the optimized ana-
log gradient update scheme assuming current
state-of-the-art electronics cointegrated with
active photonic components (supplementary
text 2.7). Collectively, our simulation results
(Fig. 4) and energy calculation contours (fig.
S8, supported by tables S1 to S6) indicated
minimal performance degradation for MNIST
training simultaneously with threefold improve-
ment in backpropagation energy efficiency.
This assumed 100-fJ floating point operations
for equivalent digital models (39) and tap noise
factor of stap < 0:01 in the regime where optical
power begins to dominate the energy consump-
tion. Errors may be further reduced by improv-
ing avalanche photodiode sensitivity, reducing
optical component loss, or increasing overall
input optical power, a key factor in the energy-
error trade-off (tables S1 to S6). Trade-off of
input power and photodiode noise generally
enforces a hard limit on scalability of photonic
meshes (i.e., number of MZI layers N) because
all photonic components have loss (16, 40).
Discussion and outlook
In this study, we have demonstrated practically
useful photonic machine learning hardware
by physically measuring gradients calculated
through interferometric measurements of in
situ backpropagation (Fig. 1). We concluded
that gradient accuracy played an important
role in reaching optimal results during training
and decreases near convergence (Fig. 2). As a
core application, we trained multilayer PNNs
using our gradient measurements and found
good agreement with digital training simula-
tions despite optical I/O calibration errors and
camera noise at the global monitoring taps
(Fig. 3). Correcting for phase measurement error
yielded training curves highly correlated to digital
predictions, so optical I/O calibration accuracy is
vital. Even though individual updates were ideal-
ly faster to compute, higher error resulted in
effectively longer training times that mitigated
this benefit. To better understand this trade-off,
we explored an optimized regime of our system,
which considered cointegration of complemen-
tary metal-oxide semiconductor (CMOS) elec-
tronics with photonics (fig. S8 and tables S1 to
S6), and found that in the regime of photonic
advantage (e.g., N ¼ 64 at sufficiently large
batch sizes), we could successfully train MNIST
close to digital equivalents (Fig. 4).
Our demonstration (Fig. 3) and energy
calculations (fig. S8) suggest that in situ
backpropagation, a technique widely used
in machine learning for its efficiency, also
efficiently trains hybrid PNNs. Our hybrid
approach optically accelerated the most com-
Þ operations, where-
ð
putationally intensive O N 2
as nonlinearities and their derivatives, which
Þ computations, were implemented
are O Nð
digitally. This is reasonable becauseO Nð
Þ time
is required to modulate and measure optical
inputs and outputs for the overall network,
regardless of hybrid or all-analog operation.
Because optics is ideal for low-latency and low-
energy signal communication, our in situ back-
propagation scheme could improve energy
efficiency in data center machine learning and
neural network accelerators (e.g., graphics
processing units) with optical interconnects,
in which data are already optically encoded.
Such schemes may be compatible with mixed-
signal schemes for accelerators that already
aim to reduce the current communication en-
ergy bottleneck (39, 41) in the race to address
the energy-doubling AI problem (3).
Pai et al., Science 380, 398–404 (2023)
28 April 2023
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
Population-based methods (42), direct feed-
back alignment (43, 44), and perturbative ap-
proaches (16) have some advantages but are
ultimately less efficient for training neural net-
works compared to backpropagation, especially
for hybrid PNNs. Unlike “receiverless” fully ana-
log PNNs (16), hybrid PNNs require optoelec-
tronic (i.e., digital-analog and analog-digital)
conversions for each layer, which can slow down
perturbative training. In contrast to perturbative
approaches, in situ backpropagation calculates
gradients in a modular framework compatible
with larger-scale AI applications.
Although this work primarily dealt with hy-
brid PNNs, our backpropagation scheme could
be compatible with all-analog or receiverless
implementations implementing EO nonli-
nearities on-chip (15, 16, 45). Previous all-analog
PNN implementations have suffered from ex-
ponential loss scaling because the same optical
modes propagated through all L layers (16).
We propose to reduce this scaling from ex-
ponential to linear by instead splitting input
light equally across the layers and modulating
each layer input by EO activations that depend
on other layer output powers, which acts to
“connect” the layers without an explicit optical
connection (fig. S9, A and H). After incorporat-
ing electronic and optical switches, this “dis-
tributed nonlinearity” architecture can operate
as a hybrid PNN platform for training or an
all-analog platform for inference with full vis-
ibility of EO nonlinearity response to aid back-
propagation training (fig. S9, B to G). The
scaling and errors of these schemes, given the
need to accurately model nonlinear activations
for backpropagation, are left to a future work.
Ultimately, these all-analog schemes suffer
from limited versatility to manipulate or transform
data. Depending on the problem or architecture,
“hybridizing” the all-optical PNN with digital
platforms can add some flexibility when conve-
nient at the expense of optoelectronic conversion
energy. For instance, flexibility of large-scale hy-
brid PNN models has been demonstrated via
high ResNet-50 image classification accuracy
using commercially viable photonic meshes (14).
Our experimental demonstration indicates a
route to train such models on backpropagation-
enabled devices that few other training methods
can efficiently produce. In situ backpropagation
can also train “optical transformers” that lever-
age hybrid PNNs for natural language pro-
cessing and computer vision applications (19).
The periodic application of digital activations,
currently infeasible in optics [e.g., layer normal-
ization (19)], enables one-to-one correspondence
of hybrid PNNs and state-of-the-art large-scale
NN models.
Our demonstration is an experimental ana-
log of “inverse design” of photonic devices.
Inverse design implements reverse-mode auto-
differentiation with respect to material relative
permittivity by interfering adjoint and forward
fields. This forms the basis of the original proof
of in situ backpropagation (22) because phases
are trivially related to material relative permit-
tivity changes. This suggests an even broader
application domain for our technique to op-
timizing arbitrary programmable linear optical
devices with no obvious calibration scheme,
including robust designs (e.g., using multiport
directional couplers) and recirculating designs
(46, 47). The analog gradient update exper-
iment in Fig. 2 is relevant to calibration (6)
because minimizing the cost function L max-
imizes device fidelity.
Our results ultimately have wide-ranging
implications for bridging the fields of pho-
tonics and machine learning. Backpropaga-
tion is the most efficient and widely used neural
network training algorithm for machine learn-
ing, and our demonstration of this popular
echnique as a physical implementation presents
promising capabilities of hybrid PNNs to re-
duce carbon footprint and counter the expo-
nentially increasing costs of AI computation.
RE FERENCES AND NOTES
1. S. Linnainmaa, BIT 16, 146–160 (1976).
2. D. E. Rumelhart, G. E. Hinton, R. J. Williams, Nature 323,
3.
533–536 (1986).
J. Sevilla et al., 2022 International Joint Conference on Neural
Networks (IJCNN), Padua, Italy (2022), pp. 1–8.
4. Y. Shen et al., Nat. Photonics 11, 441–446 (2017).
5. M. Reck, A. Zeilinger, H. J. Bernstein, P. Bertani, Phys. Rev.
Lett. 73, 58–61 (1994).
6. D. A. B. Miller, Photon. Res. 1, 1 (2013) [Invited].
7. S. Pai, B. Bartlett, O. Solgaard, D. A. B. Miller, Phys. Rev. Appl.
11, 064044 (2019).
8. A. Annoni et al., Light Sci. Appl. 6, e17110 (2017).
9. W. Bogaerts et al., Nature 586, 207–216 (2020).
10. J. Carolan et al., Science 349, 711–716 (2015).
11. B. Bartlett, S. Fan, Phys. Rev. A 101, 042319 (2020).
12. M. Milanizadeh et al., Light Sci. Appl. 11, 197 (2022).
13. N. C. Harris et al., Optica 5, 1623 (2018).
14. C. Ramey, “Silicon Photonics for Artificial Intelligence
Acceleration (Lightmatter)” in IEEE Hot Chips 32 Symposium
(HCS) (2020), pp. 1–26.
I. A. D. Williamson et al., IEEE J. Sel. Top. Quantum Electron.
26, 1–12 (2020).
15.
16. S. Bandyopadhyay et al., arXiv:2208.01623 [cs.ET] (2022).
17. L. Jing et al., in Proceedings of the 34th International
Conference on Machine Learning, Sydney, Australia (2017),
vol. 70, pp. 1733–1741.
18. A. Vaswani et al., Adv. Neural Inf. Process. Syst. 30, 5998–6008
(2017).
19. M. G. Anderson, S.-Y. Ma, T. Wang, L. G. Wright, P. L. McMahon,
arXiv:2302.10360 [cs.ET] (2023).
20. M. A. Nahmias et al., IEEE J. Sel. Top. Quantum Electron. 26,
1–18 (2020).
21. A. A. Cruz-Cabrera et al., IEEE Trans. Neural Netw. 11, 1450–1457
(2000).
22. T. W. Hughes, M. Minkov, Y. Shi, S. Fan, Optica 5, 864 (2018).
23. L. G. Wright et al., Nature 601, 549–555 (2022).
24. J. Spall, X. Guo, A. I. Lvovsky, Optica 9, 803–811 (2022).
25. S. Pai, simphox: Another inverse design library [Computer
software]; https://github.com/fancompute/simphox (2022).
26. J. Bradbury et al., JAX: composable transformations of
Python+NumPy programs [Computer software];
https://github.com/google/jax (2022).
27. T. Hennigan, T. Cai, T. Norman, I. Babuschkin, Haiku:
Sonnet for JAX, [Computer software]; https://github.com/
deepmind/dm-haiku (2020).
28. D. A. B. Miller, Optica 7, 794 (2020).
29. M. Prabhu et al., Optica 7, 551 (2020).
30. H. Zhang et al., Nat. Commun. 12, 457 (2021).
31. S. Pai et al., IEEE J. Sel. Top. Quantum Electron. 26, 1–13
(2020).
32. S. Pai, solgaardlab/photonicbackprop: Adding some new
33. F. Pedregosa et al., J. Mach. Learn. Res. 12, 2825–2830 (2011).
34. D. P. Kingma, J. L. Ba, “Adam: A Method for Stochastic
Optimization,” International Conference on Learning
Representations, 7 to 9 May 2015, San Diego.
35. J. K. Perin, M. Sharif, J. M. Kahn, J. Lightwave Technol. 34,
5542–5553 (2016).
36. F. Morichetti et al., IEEE J. Sel. Top. Quantum Electron. 20,
292–301 (2014).
37. S. Pai et al., Nanophotonics 12, 985–991 (2023).
38. L. Deng, IEEE Signal Process. Mag. 29, 141–142 (2012).
39. D. A. Miller, J. Lightwave Technol. 35, 346–396 (2017).
40. S. Pai et al., Optica 10.1364/OPTICA.476173 (2023).
41. B. Murmann, IEEE Trans. Very Large Scale Integr.
(VLSI) Syst. 29, 3–13 (2021).
42. H. Zhang et al., ACS Photonics 8, 1662–1672 (2021).
43. A. Nøkland, in Proceedings of the 30th Conference on
Neural Information Processing Systems, D. D. Lee et al.,
Eds. (Curran Associates, 2016), pp. 1045–1053.
44. M. J. Filipovich et al., Optica 9, 1323–1332 (2022).
45. X. Guo, T. D. Barrett, Z. M. Wang, A. I. Lvovsky, Photon. Res. 9,
B71–B80 (2021).
46. D. Pérez et al., Nat. Commun. 8, 636 (2017).
47. R. Tang, R. Tanomura, T. Tanemura, Y. Nakano, ACS Photonics
8, 2074–2080 (2021).
48. S. Pai, Z. Sun, T. Park, phox: Base repository for
simulation and control of photonic devices [Computer
software], https://github.com/solgaardlab/phox/ (2022).
49. S. Pai, N. Abebe, dphox: photonic layout and device design [Computer
software], https://github.com/solgaardlab/dphox (2022).
AC KNOWLED GME NTS
We acknowledge Advanced MicroFoundries (AMF) in Singapore for
help in fabricating and characterizing the photonic circuit for
our demonstration and Silitronics for help in packaging our chip for
our demonstration. Thanks also to P. Broaddus for helping with
wafer dicing; S. Lorenzo for help in fiber splicing the fiber switch
for bidirectional operation; J. Kahn for guidance on avalanche
photodetector noise estimates; N. Pai for advice on electronics,
scalability, and electrical and thermal control packaging; R. Quan
for help in building our all-analog gradient measurement
electronics; and C. Langrock and K. Urbanek for help in building
our movable optical breadboard. Funding: We acknowledge
funding from Air Force Office of Scientific Research (AFOSR)
grants FA9550-17-1-0002 in collaboration with UT Austin and
FA9550-18-1-0186 through which we share a close collaboration
with UC Davis under B. Yoo. Author contributions: S.P. ran all
experiments with input from Z.S., T.W.H., T.P., B.B., I.A.D.W., N.A.,
M. Minkov, O.S., S.F., and D.A.B.M. S.P., T.W.H., M. Minkov, and I.A.D.
W. conceptualized the experimental protocol. S.P., N.A., F.M.,
M. Milanizadeh, and A.M. contributed to the design of the photonic
mesh. S.P. and Z.S. wrote code to control the photonic integrated
circuit active elements and camera detection and electronic circuit
for analog gradient measurement. T.P. designed the custom PCB with
input from S.P. S.P. wrote the manuscript with input from all
coauthors. All coauthors contributed to discussions of the protocol
and results. Competing interests: S.P., Z.S., T.W.H., I.A.D.W.,
M. Minkov, S.F., O.S., and D.A.B.M. have filed a patent for the analog
backpropagation update protocol discussed in this work with
provisional application no. 63/323743. D.M. holds two related
patents on the SVD architecture: US Patent no. 10,877,287 and
no. 10,534,189. The authors declare no other conflicts of interest.
Data and materials availability: Materials and methods are
available as supplementary materials. All other software and data
for running the simulations and experiments are available through
Zenodo (32) and Github through the Phox framework, including
our experimental code via Phox (48), simulation code via Simphox
(25), and circuit design code via Dphox (49). License information:
Copyright © 2023 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science. No
claim to original US government works. https://www.sciencemag.
org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade8450
Materials and Methods
Supplementary Text
Figs. S1 to S9
Tables S1 to S6
References (50–76)
Movies S1 and S2
analog gradient measurement data (0.0.3), Zenodo (2023);
https://doi.org/10.5281/zenodo.6557413.
Submitted 27 September 2022; accepted 8 March 2023
10.1126/science.ade8450
Pai et al., Science 380, 398–404 (2023)
28 April 2023
6 of 6
|
10.1126_science.ade9097
|
RES EARCH
R E S E A R C H A R T I C L E
◥
COMPUTER SCIENCE
Human-level play in the game of Diplomacy by
combining language models with strategic reasoning
Meta Fundamental AI Research Diplomacy Team (FAIR)†, Anton Bakhtin1‡, Noam Brown1*‡,
Emily Dinan1*‡, Gabriele Farina1, Colin Flaherty1‡, Daniel Fried1,2, Andrew Goff1, Jonathan Gray1‡,
Hengyuan Hu1,3‡, Athul Paul Jacob1,4‡, Mojtaba Komeili1, Karthik Konath1, Minae Kwon1,3,
Adam Lerer1*‡, Mike Lewis1*‡, Alexander H. Miller1‡, Sasha Mitts1, Adithya Renduchintala1‡,
Stephen Roller1, Dirk Rowe1, Weiyan Shi1,5‡, Joe Spisak1, Alexander Wei1,6, David Wu1‡,
Hugh Zhang1,7‡, Markus Zijlstra1
Despite much progress in training artificial intelligence (AI) systems to imitate human language,
building agents that use language to communicate intentionally with humans in interactive environments
remains a major challenge. We introduce Cicero, the first AI agent to achieve human-level performance
in Diplomacy, a strategy game involving both cooperation and competition that emphasizes natural
language negotiation and tactical coordination between seven players. Cicero integrates a language
model with planning and reinforcement learning algorithms by inferring players’ beliefs and intentions
from its conversations and generating dialogue in pursuit of its plans. Across 40 games of an anonymous
online Diplomacy league, Cicero achieved more than double the average score of the human players
and ranked in the top 10% of participants who played more than one game.
A major long-term goal for the field of
artificial intelligence (AI) is to build
agents that can plan, coordinate, and
negotiate with humans in natural lan-
guage. Although much progress has
been made in language models that imitate
human language (1), effective negotiation
agents must go beyond this by understand-
ing the beliefs, goals, and intentions of their
partner; planning joint actions that account
for their partner’s goals; and persuasively and
intentionally communicating these proposals.
We present Cicero, an AI agent that achieved
human-level performance in the strategy game
Diplomacy. In Diplomacy, seven players con-
duct private natural language negotiations to
coordinate their actions to both cooperate and
compete with each other. By contrast, prior
major successes for multi-agent AI have been
in purely adversarial environments such as
chess (2), Go (3), and poker (4), in which com-
munication has no value. For these reasons,
Diplomacy has served as a challenging bench-
mark for multi-agent learning (5–8).
1Meta AI, 1 Hacker Way, Menlo Park, CA, USA. 2Language
Technologies Institute, Carnegie Mellon University,
Pittsburgh, PA, USA. 3Department of Computer Science,
Stanford University, Stanford, CA, USA. 4Computer Science
and Artificial Intelligence Laboratory, Massachusetts
Insititute of Technology, Cambridge, MA, USA. 5Department
of Computer Science, Columbia University, New York, NY,
USA. 6Department of Computer Science, University of
California, Berkeley, Berkeley, CA, USA. 7EconCS Group,
Harvard University, Cambridge, MA, USA.
*Corresponding author. Email: noambrown@meta.com (N.B.);
edinan@meta.com (E.D.); alerer@meta.com (A.L.); mikelewis@
meta.com (M.L.) †FAIR consists of all listed authors. There are no
additional authors or collaborators. ‡These authors contributed
equally to this work.
Cicero couples a controllable dialogue mod-
ule with a strategic reasoning engine. At each
point in the game, Cicero models how the other
players are likely to act on the basis of the game
state and their conversations. It then plans how
the players can coordinate to their mutual ben-
efit and maps these plans into natural language
messages.
We entered Cicero anonymously in 40 games
of Diplomacy in an online league of human
players between 19 August and 13 October 2022.
Over the course of 72 hours of play that in-
volved sending 5277 messages, Cicero ranked
in the top 10% of participants who played more
than one game.
Challenges of human-AI cooperation
in Diplomacy
Almost all prior AI breakthroughs in games have
been in two-player zero-sum (2p0s) settings, in-
cluding chess (2), Go (3), heads-up poker (9, 10),
and StarCraft (11, 12). In finite 2p0s games,
certain reinforcement learning (RL) algorithms
that learn by playing against themselves—a
process known as self-play—will converge to
a policy that is unbeatable in expectation in
balanced games (13). In other words, any finite
2p0s game can be solved through self-play with
sufficient compute and model capacity.
However, in games that involve cooperation,
self-play without human data is no longer
guaranteed to find a policy that performs well
with humans, even with infinite compute and
model capacity, because the self-play agent may
converge to a policy that is incompatible with
human norms and expectations. This effect
can be clearly seen in settings that involve
language, in which prior work found that
self-play produced uninterpretable language
despite achieving high task success for the
agents (14, 15). Even in dialogue-free versions
of Diplomacy, we found that a self-play algo-
rithm that achieved superhuman performance
in 2p0s versions of the game performed poorly
in games with multiple human players owing
to learning a policy inconsistent with the norms
and expectations of potential human allies
(16, 17). Thus, a major challenge in Diplomacy
is to develop a way to harness the potential
benefits of self-play in a way that leads to
human-compatible language and behavior.
The challenge of maintaining human-
interpretable communication is particularly
acute in Diplomacy, in which our agent sent
and received an average of 292 messages per
game (fig. S8). Messages in the game often
involve coordinating precise plans, and any
miscommunication can result in their fail-
ure. Each message an agent sends must be
grounded in (be contextually appropriate and
consistent with) lengthy dialogue histories,
game states—including proposed hypothetical
states—and goals. If messages are inaccurately
grounded, humans may ask the agent to ex-
plain its errors (a challenging task that may
lead to further mistakes) or choose to coop-
erate with others instead. Further, repeated
messaging creates feedback loops, in which
the language model imitates the style of its
own previous messages—for example, send-
ing a short or incoherent message will increase
the likelihood of such messages in the future
(18). Past work on strategic dialogue systems
has avoided these issues by focusing on sim-
pler settings (14, 19–21), which involve only
a single human partner, shorter dialogue his-
tories, and simpler strategies.
Last, Diplomacy is a particularly challeng-
ing domain because success requires building
trust with others in an environment that en-
courages players to not trust anyone. Each
turn’s actions occur simultaneously after non-
binding, private negotiations. To succeed, an
agent must account for the risk that players
may not stay true to their word, or that other
players may themselves doubt the honesty of
the agent. For this reason, an ability to reason
about the beliefs, goals, and intentions of others
and an ability to persuade and build relation-
ships through dialogue are powerful skills in
Diplomacy.
The game of Diplomacy
Diplomacy is a board game in which seven
players compete to control supply centers (SCs)
on a map, by moving their units into them. A
player wins by controlling a majority of SCs.
The game may also end when all remaining
players agree to a draw, or a turn limit is
reached, in which case scores are determined
by the number of SCs each player controls.
Meta Fundamental AI Research Diplomacy Team (FAIR) et al., Science 378, 1067–1074 (2022)
9 December 2022
1 of 8
Output action
are nonsensical, inconsistent with intents, or
strategically poor.
RES EARCH | R E S E A R C H A R T I C L E
Board state & history
Planning
Joint action
Policies
(all players)
Simulator
State value
Future state
Dialogue-free value model (from RL)
Strategic reasoning
Dialogue
AUSTRIA: Hi Italy! Care to work
together on this one? If you
support me into BOH I think we'd
both be able to grow quickly.
ITALY: Could you support me
into BUL in return?
AUSTRIA: ...
Anchor policies
(all players)
Intents
AUSTRIA: VIE BOH, ...
ITALY: TYR S VIE BOH, ...
Dialogue-conditional
action model
Message candidates
Dialogue
model
Filters
(nonsense,
grounding, value)
AUSTRIA: Hi Italy! Care to work
together on this one? If you
support me into BOH I think we'd
both be able to grow quickly.
ITALY: Could you support me
into BUL in return?
AUSTRIA: Sure thing! I have
ordered SER to support GRE
to BUL.
Dialogue history
Message generation
Output message
Fig. 1. Architecture of Cicero. Cicero predicts likely human actions for each player according to the board
state and dialogue, using that as the starting point for a planning algorithm using RL-trained models.
The output of planning is an action for the agent as well as beliefs about other players’ actions, which are
used to select intents for a dialogue model to condition on. Generated message candidates undergo several
filtering steps before a final message is sent.
Each turn, all players engage in private pair-
wise free-form dialogue with the others during
a negotiation period, and then all players simul-
taneously choose an action comprising one
order per unit they control. A unit may sup-
port other units, including those of another
player, which forms the basis for much of the
negotiation in Diplomacy. A detailed descrip-
tion of the rules is provided in the supplemen-
tary materials (SM), materials and methods,
section C.
Overview of Cicero
At a high level, Cicero combines a dialogue
module with a strategic reasoning module,
along with a filtering process to reject low-
quality messages. A diagram of Cicero is pro-
vided in Fig. 1.
Dialogue
Cicero generates dialogue using a pretrained
language model that was further trained on
dialogue data from human games of Diplomacy.
Crucially, in addition to being grounded in
both the dialogue history and game state, the
dialogue model was trained to be controllable
through intents, which we here define to be a
set of planned actions for the agent and its
speaking partner. This was accomplished by
automatically augmenting the human data
with inferred intents and using this informa-
tion as further conditioning during training.
For example, intents showing the agent moving
into the territory Bulgaria (“BUL”) with sup-
port from its speaking partner might yield a
message such as “Could you support me into
BUL in return?” Grounding in intents re-
lieved the dialogue model of most of the re-
sponsibility for learning which actions were
legal and strategically beneficial. In particu-
lar, this control provided an interface be-
tween the dialogue generation and strategic
reasoning.
Strategic reasoning
Cicero uses a strategic reasoning module to
intelligently select intents and actions. This
module runs a planning algorithm that pre-
dicts the policies of all other players on the
basis of the game state and dialogue so far,
accounting for both the strength of different
actions and their likelihood in human games,
and chooses an optimal action for Cicero that
is based on those predictions. Planning relies
on a value and policy function trained through
self-play RL that penalized the agent for de-
viating too far from human behavior, to main-
tain a human-compatible policy. During each
negotiation period, intents are recomputed
every time Cicero sends or receives a message.
At the end of each turn, Cicero plays its most
recently computed intent.
Message filtering
Cicero passes each generated message through
several filters designed to limit messages that
Methods
Data
We obtained a dataset of 125,261 games of
Diplomacy played online at webDiplomacy.net.
Of these, 40,408 games contained dialogue,
with a total of 12,901,662 messages exchanged
between players. Player accounts were de-
identified, and automated redaction of per-
sonally identifiable information (PII) was
performed by webDiplomacy. We refer to this
dataset hereafter as WebDiplomacy.
Intent-controlled dialogue
Cicero generates messages through a neural
generative Diplomacy dialogue model that
was trained to be controllable through a set
of intents.
(cid:1)
(cid:4)
(cid:5)
(cid:3)
Imitation dialogue model
We took R2C2 (22) as our base model—a
2.7 billion–parameter Transformer-based (23)
encoder-decoder model pretrained on text
from the internet by using a BART denoising
objective (24). The base pretrained model
was then further trained on WebDiplomacy
(Methods, Data) through standard maxi-
mum likelihood estimation. Specifically, with
a dataset D ¼ x ið Þ; y ið Þ
, the model was
trained to predict a dialogue message y(i) from
player A to player B at time t, given all of the
following represented as text x(i): dialogue
history (all messages exchanged between
playerA and the six other players up to time t);
game state and action history (current game
state and recent action history); player rating
(rating for A corresponding to Elo rating
computed from games in WebDiplomacy);
game and message metadata (additional info
about game settings and the current message,
such as time since the last message, and current
turn). Additionally, the model conditions on
intents (a set of proposed actions for players A
and B for the current turn and future turns,
representing the intent for message y(i)). Fur-
ther details on the training data, training pro-
cedure, relevant hyperparameters, sampling
procedures, and other inference-time methods
are provided in the SM, section D.1.
During play, we used additional modules
governing when to speak and to whom, which
are described in the SM, section D.4.
Controllable dialogue model
through intents
Standard language modeling approaches would
train our dialogue model only to imitate the
messages from our dataset but not to outper-
form them. To go beyond imitation learning,
we made the dialogue model controllable by
generating messages conditioned on a plan
specified by the strategic reasoning module
Meta Fundamental AI Research Diplomacy Team (FAIR) et al., Science 378, 1067–1074 (2022)
9 December 2022
2 of 8
RES EARCH | R E S E A R C H A R T I C L E
A Intent model training
B Intent annotation
Intent model
Board state & history
Board state & history
ENG:Do you want NTH to support BEL?
FRA: No, BEL is moving to HOL
ENG:NTH S BEL HOL, ...
FRA: BEL HOL, ...
ENG: Do you want NTH to support BEL?
FRA: No, BEL is moving to HOL
ENG: Alright i’ll support you in
Dialogue history ENG-FRA
Only trained on “truthful”
situations where a zero-shot
lie detector says the player
wasn't lying about their orders.
ENG: Alright i’ll support you in
Dialogue history ENG-FRA
Annotated
intents
ENG:NTH S BEL, ...
FRA: BEL H, ...
ENG: Do you want NTH
to support BEL?
FRA: I’ve entered those orders
Artificially injected
agreement
ENG:...
FRA: ...
ENG:...
FRA: ...
Intent model
C Dialogue model training
D Dialogue model inference
Dialogue model
Board state & history
ENG:Do you want NTH to support BEL?
FRA: No, BEL is moving to HOL
ENG:Alright, I’ll
support you in
ENG: ...
ENG:NTH S BEL HOL, ...
FRA: BEL HOL, ...
Board state & history
ENG: Bounce in the English Channel?
FRA: No, I need to move to MAO
to protect against Italy
ENG: ...
LEGEND
ENG:...
FRA:...
Planned moves
(intents)
Dialogue model
ENG:Okay, I’ll move
to North Sea then.
ENG:LON NTH, ...
FRA: BRE MAO, ...
Model inputs
Dialogue history ENG-FRA
Dialogue history ENG-FRA
Planning
Training targets
Fig. 2. Illustration of the training and inference process for intent-controlled
dialogue. Actions are specified as strings of orders for units; for example, “NTH
S BEL - HOL” means that North Sea will support Belgium to Holland. (A) An “intent
model” was trained to predict actions for a pair of players on the basis of their
dialogue. Training data was restricted to a subset in which dialogue is deemed
“truthful” (SM, section D.2.3). (B) Each message in the dialogue training dataset was
annotated with the output of the intent model on the dialogue up to that point,
with an agreement message injected at the end. (C) The dialogue model was trained
to predict each dataset message given the annotated intent for the target message.
(D) During play, intents were supplied by the planning module instead.
(intents), resulting in higher-quality messages.
More specifically, a message is defined to have
intent z if z is the most likely set of actions that
the sender and recipient will take—for both
the current turn and several future turns—if
no further dialogue occurs after the message is
received. To establish this control, we devel-
oped techniques to automatically annotate
every message in the training set with a set of
actions corresponding to the message content.
During training, the dialogue model learned
, where z(i)
the distribution pq y ið Þ x ið Þ; z ið Þ
represents the intent for datapoint [x(i), y(i)];
as a result, at inference, time z provides a point
of control over generation (25). We later de-
scribe the training and inference process, which
is also illustrated in the pipeline in Fig. 2. The
effect of the intents on the generated dialogue
(cid:6)
(cid:6)
(cid:1)
(cid:3)
is demonstrated in Fig. 3; conditioning on
different planned actions results in different
messages.
We considered other notions of intent during
development, such as controlling messages to
focus on specific subsets of actions, third-party
actions, or to have a particular tone. Richer
intents are harder to annotate on human mes-
sages, are harder to select with the planning
module, and create greater risk of taking the
language model out of distribution.
Annotating training messages with intents
When annotating messages in the training data
with corresponding intents, our goal was for the
proposed actions z(i) to closely reflect the con-
tent of a message y(i) so that at training time, the
model learned to exploit the information in z(i).
Naïvely, we could have used the actual ac-
tions played by the sender and recipient at
the end of each turn in the span of the intent.
However, these actions may not reflect the
content of a message if (i) a message is not
honest or (ii) subsequent messages change the
sender’s plans. To resolve (i), we predicted the
most likely action according to a dialogue-
conditional action prediction model trained
on a “truthful” subset of the dataset, in which
we predicted that a player’s dialogue was not
deceptive to others (SM, section D.2). This is
showcased in Fig. 2A; we refer to this model as
the intent model. To resolve (ii), we restricted
the dialogue history that this intent model saw
up to the message in question, which signaled
to the model to predict actions as though the
dialogue had ended at that point in time. We
Meta Fundamental AI Research Diplomacy Team (FAIR) et al., Science 378, 1067–1074 (2022)
9 December 2022
3 of 8
RES EARCH | R E S E A R C H A R T I C L E
Depicted intents:
England convoys an army to Belgium with the
support of France and Germany while taking Norway
in a manner friendly to Russia.
England convoys an army to Norway and also
aggressively moves to the Barents Sea against
Russia.
England convoys an army to Norway and supports
France to gain Belgium, preventing Germany from
taking Belgium.
Example generations from the dialogue model with these actions as the conditioned intent:
ENG FRA Mind supporting Edi - Bel?
ENG FRA What's the plan for Bel?
ENG GER Do you wanna support my convoy to
Bel? With Italy going aggressive France will fall
ENG GER I will be convoying my army to
Norway and moving to Barrents Sea.
and France.
ENG RUS How are you thinking Germany is
gonna open? I may have a shot at Belgium, but I'd
need your help into Den next year.
ENG RUS Hey, I moved to Edi to keep my
options open. Do you have a problem with me
putting an army in Nwy?
ENG FRA Let me know if you want me to
support you to Bel, otherwise I'll probably poke Hol.
ENG GER Looks like you'll get three builds
unless France bounces you! Are you gonna
bounce Russia or not?
ENG RUS Hey, I moved to Edi to keep my
options open. Do you have a problem with me
putting an army in Nwy?
Fig. 3. The effect of intents on Cicero’s dialogue. Pictured are three different possible intents in the same game situation. In each case, we show a message
generated by Cicero (England; pink) to France (blue), Germany (orange) and Russia (purple) conditioned on these intents. Each intent leads to quite different
messages, which are consistent with the intended actions.
additionally added messages to the dialogue
history that suggested a conclusive agreement
between the two parties (Fig. 2B). As a result,
we obtained a high degree of correspondence
between the action annotated as the intent of
a message and the content, achieving a score
of 97% on a small test set designed to measure
this correspondence (compared with 77% for
a simpler baseline) (table S2). Then, the dia-
logue model could be trained in the manner
described in the above section Imitation dia-
logue model and in Fig. 2C (SM, section D.2).
Selecting intents during play
During play, Cicero used the strategic rea-
soning module to select intent actions for the
current turn (Fig. 2D), whereas intent actions
for future turns were generated by means of a
human-imitation model.
Agent intent action for current turn
Cicero conditioned its dialogue on the action
that it intends to play for the current turn.
This choice maximizes Cicero’s honesty and its
ability to coordinate but risks leaking infor-
mation that the recipient could use to exploit
it (for example, telling them which of their
territories Cicero plans to attack) and some-
times led to out-of-distribution intents when
DIALOGUE QUALITY RATINGS (%)
Consistent
with state
Consistent
with plan
High
quality
Perplexity
Language model
+ game state grounding
+ intent grounding (CICERO)
61.90
84.13
87.30
76.19
83.33
92.86
20.64
29.37
37.30
8.02
7.94
7.70
Fig. 4. Controllable dialogue modeling results. We report dialogue quality ratings and perplexity on the
validation set for the Cicero dialogue model and compare them with a baseline without intent grounding and a
baseline without either intent or game-state grounding (“Language model”). Dialogue quality ratings were
calculated according to expert annotation of generated messages in 126 situations; we report the percent of
messages (before filtering) labeled as consistent with the game state, as consistent with the plan for the
next actions, and as particularly high quality. Lower perplexity corresponds to more probability mass on the
ground-truth human messages.
the intended action was hostile, because in
adversarial situations, humans may rarely com-
municate their intent honestly. We describe
approaches for mitigating these risks in the
section Message filtering.
the recipient and/or that they are believed to
be likely to play it given the dialogue. Among
this restricted set, Cicero selected the recip-
ient action with the highest expected value
for itself (SM, section D.2.4).
Recipient intent action for current turn
Dialogue modeling results
Cicero considered the subset of recipient ac-
tions with high likelihood under its beliefs
about their policy. High likelihood requires
that either an action is deemed beneficial for
We compared the performance of our dialogue
model with a baseline without intent ground-
ing and one without intent or game-state
grounding (a “language model”). We report
Meta Fundamental AI Research Diplomacy Team (FAIR) et al., Science 378, 1067–1074 (2022)
9 December 2022
4 of 8
RES EARCH | R E S E A R C H A R T I C L E
England agrees:
England is hostile:
England tries to take advantage of Cicero:
FRA Yes! I will move out of ENG if you
ENG
head back to NAO.
FRA You've been fighting me all game.
ENG
Sorry, I can't trust that you won't stab me.
FRA Yes! I'll leave ENG if you move KIE ->
ENG
MUN and HOL -> BEL.
Cicero predicts England will retreat from ENG to NTH
85% of the time, backs off its own fleet to NAO as
agreed, and begins to move armies away from the
coast.
Cicero does not back off its fleet but rather attacks EDI
with it, and leaves its armies at the coast to defend
against an attack from England, predicting that England
will attack about 90% of the time.
Strategic planning rejects the possibility of vacating KIE
and HOL, because it would make Cicero too vulnerable.
Cicero backs off its fleet to NAO but keeps armies at
the coast to defend.
Fig. 5. The effect of dialogue on Cicero’s strategic planning and intents.
Cicero (France; blue) and England (pink) are entangled in a fight, but it would be
beneficial for both players if they could disengage. Cicero has just messaged England
“Do you want to call this fight off? I can let you focus on Russia and I can focus on
Italy.” Pictured are three ways that England might reply and how Cicero adapts to
each. (Left and middle) Because Cicero’s planning anchors around a dialogue-
conditional policy model, its predictions for other players and accordingly its own
plans are flexible and responsive to negotiation with other players. (Right) Yet Cicero
also avoids blindly trusting what other players propose by rejecting plans that have low
predicted value and run counter to its own interests.
both perplexity on the validation set and dia-
logue quality rating scores, which were cal-
culated on the basis of expert annotation of
messages generated in 126 Diplomacy game
situations. Experts were asked to label whether
a message was (i) consistent with the game
state, (ii) consistent with the agent’s plan, and
(iii) notably high quality, compared with that
of an average human. Results are shown in
Fig. 4, and more details regarding this eval-
uation are provided in the SM, section D.2.3.
Our model outperformed the baselines on
all metrics. The improvement in validation
perplexity demonstrated that the model can
use additional grounding information to
better predict human messages. Expert anno-
tations showed that the grounding informa-
tion provided by the intents and game state
led to higher-quality messages that were
highly consistent with the agent’s intended
action.
the current turn that responded optimally to
the other players’ predicted policies.
Doing this with human players requires
predicting how humans will play. A popular
approach in cooperative games is to model
the other players’ policies through supervised
learning on human data, which is commonly
referred to as behavioral cloning (BC). How-
ever, pure BC is brittle, especially because a
supervised model may learn spurious corre-
lations between dialogue and actions (fig. S6).
To address this problem, Cicero used variants
of piKL (26) to model the policies of players.
piKL is an iterative algorithm that predicts
policies by assuming each player i seeks to
both maximize the expected value of their
policy pi and minimize the Kullback-Leibler
(KL) divergence between pi and the BC policy,
which we call the anchor policy ti. An anchor
strength parameterl∈ 0; ∞½
Þ trades off between
these competing objectives.
receives a reward ui(a) determined by a value
function ui. We discuss the training of this
value function later below.
piKL assumes player i seeks a policy pi that
maximizes the modified utility function
Ui pi; p(cid:2)i
ð
ð
Þ ¼ ui pi; p(cid:2)i
Þ (cid:2) lDKL pi∥ti
ð
Þ
ð1Þ
where p–i represents the policies of all players
other than i, and ui(pi, p–i) is the expected
value of pi given that other players play p–i.
Specifically, let Qt(cid:2)1
and let
aið
(cid:8)
i
(cid:7)
Þ ¼ ui ai; pt(cid:2)i
(cid:2)i
(cid:10)
(cid:9)
Þexp
Þ
Qt(cid:2)1
aið
i
l
pDt
i aið
Þºti aið
ð2Þ
On each iteration t, piKL updates its predic-
tion of the players’ joint policies to be
pt ¼
(cid:11)
t (cid:2) 1
t
(cid:12)
pt(cid:2)1 þ
(cid:11) (cid:12)
1
t
pDt
ð3Þ
Strategic reasoning
piKL: KL-regularized planning
To generate the intents for dialogue and to
choose the final actions to play each turn, Cicero
ran a strategic reasoning module that predicts
other players’ policies (a probability distribu-
tion over actions) for the current turn accord-
ing to the state of the board and the shared
dialogue and then chose a policy for itself for
piKL is an iterative algorithm that predicts
player policies. A complete description of the
algorithm can be found in the SM, section E.1.
piKL treats each turn in Diplomacy as its own
subgame in which each player i simultane-
ously chooses an action ai that results in joint
action a = (a1, ..., an), and then each player i
piKL provably converges to an equilibrium
in the modified utility space (26). When the
anchor strength l is set to a large value, piKL
predicts that player i’s policy will be close to
the anchor policy ti. When l is small, piKL
predicts that player i’s policy will have high
expected value and may deviate substantially
from ti.
Meta Fundamental AI Research Diplomacy Team (FAIR) et al., Science 378, 1067–1074 (2022)
9 December 2022
5 of 8
RES EARCH | R E S E A R C H A R T I C L E
A generalization of piKL referred to as
Distributional Lambda piKL (DiL-piKL) re-
places the single l parameter in piKL with a
probability distribution over l values (SM,
section E.1.3). On each iteration, each player
samples a l value from their distribution. In
practice, we found this led to better perfor-
mance (17).
Dialogue-conditional planning
Because dialogue influences the BC policy (the
anchor policy ti), piKL provides a mechanism
for dialogue to influence policy predictions.
Different possible messages between Cicero
and another player may produce different an-
chor policies (Fig. 5), which ultimately gives
different final predictions about what that
player will do.
Other players may of course be deceptive
about their plans. Cicero does not explicitly
predict whether a message is deceptive or not
but rather relies on piKL to directly predict the
policies of other players on the basis of both
the BC policy (which conditions on the mes-
sage) and on whether deviating from the BC
policy would benefit that player.
Because dialogue in Diplomacy occurs pri-
vately between pairs of players, Cicero must
reason about what information players have
access to when making predictions. For ex-
ample, if Cicero is coordinating an attack with
an ally against an adversary, Cicero’s predic-
tion of the adversary’s policy must account for
the adversary not being aware of the intended
coordination. Cicero accomplished this by pre-
dicting by means of pairwise piKL what every
other player’s policy will be.
Specifically, during strategic planning, for
each player j, Cicero computed an anchor
policy for both itself and player j on the basis
of their shared conversation, the board state,
and the recent action history. Cicero then ran
DiL-piKL for the two players to predict player
j’s policy. On each iteration, Cicero assumed
that the remaining five players would play
according to a policy computed by means of
RL, conditional on the policies of Cicero and
player j. This process gave an independent
prediction of each player’s policy.
Next, Cicero accounted for the players’ pol-
icies not being independent owing to their
ability to correlate their actions through pri-
vate dialogue that Cicero did not observe.
Cicero accomplished this by constructing an
approximate joint policy for all other players
through self-normalized importance sampling:
We sampled N = 1000 joint actions a from the
independent piKL policies of the other players
and reweighted them by the likelihood ratio of
a under the correlated and independent RL
policies, respectively.
ð
ui ai; p(cid:2)i
Last, Cicero chose the action ai that best
responds to the predicted joint policy p–i of the
other players, while still being as consistent
as possible with its dialogue. Specifically,
Þ þ
Cicero chose the action argmaxai
llogti aið
Þ, where ui is the RL value function,
ti(ai) is the probability of the action under the
dialogue-conditional imitation policy, and l =
3 × 10−3. Cicero used a smaller l for regulariz-
ing its best response than for its computation
of other players’ policies; thus, the dialogue
more strongly informed Cicero’s expectations
of how other players would coordinate while
still allowing Cicero more leeway to deviate
when the action that it predicted humans
would most likely choose in its situation was
suboptimal.
Self-play RL for improved value estimation
Applying piKL requires a state value function.
Self-play provides an avenue for training such
Example of coordination - CICERO is AUSTRIA
Example of negotiation - CICERO is FRANCE
ITALY: What are you thinking long term?
Should I go for Turkey or head west
FRANCE: I'll work with you but I need Tunis for now.
a value function but risks becoming incom-
patible with human play (16, 17). To address
this, we used piKL during self-play to keep the
policies human-compatible.
One challenge in doing self-play in Diplomacy
is that players may adapt their actions sub-
stantially on the basis of dialogue with other
players, including coordinating joint actions.
Explicitly simulating conversations would
be extremely expensive in RL. However, a key
insight is that a joint, shared BC policy trained
on the joint action distribution of the human
data already implicitly captures the effects of
dialogue on the action distribution of human
players by modeling that action distribution
directly.
We therefore developed Correlated and
Shared (CoShar) piKL, which allowed for reg-
ularization toward a joint, correlated anchor
policy t shared by all players rather than
toward per-player policies. In this way, we
relied on the joint anchor policy to capture
the correlation between all players’ policies.
Specifically, CoShar piKL differs from standard
piKL in that the probability of joint action a =
(a1, ..., an) in policy pDt becomes
(cid:9)
pDt að Þºt að Þexp
X
i≤n
(cid:10)
Þ
Qt(cid:2)1
aið
i
l
ð4Þ
We found that CoShar piKL retained much
of the correlation present in the joint anchor
policy t while also modeling strong human
play better than imitation alone.
Our resulting self-play algorithm operated
similarly to AlphaZero (27) and ReBeL (28) by
applying planning “in the loop” as the im-
provement operator for RL. In our case, plan-
ning was by use of an approximated version
of CoShar piKL. We generated self-play tra-
jectories in which on each turn we computed
the CoShar piKL policy using a learned state-
value model. We regressed the joint policy
model toward that policy and regressed the
value model toward the expected values of all
players under that policy. We then sampled a
joint action from that policy to generate the
next state in the trajectory. The anchor policy
was fixed throughout training to anchor the
RL near human play (SM, section E.4).
Prior work has shown that neural language
models suffer from contradictions and incon-
sistency as well as a tendency to “hallucinate,”
or generate factually incorrect information (29).
In the complex domain of Diplomacy, dialogue
models exhibit both these problems and other
more subtle mistakes, such as deviations from
the intents used to control the message or
blunders in the strategic content of the mes-
sage. We approached this problem by filtering
generated messages using a series of classifiers
and checks to detect common issues. We outline
AUSTRIA: Yeah, he went to Armenia which is really
great. You can go either way, but if Turkey is
committing to Russia you could always lepanto
AUSTRIA: A lepanto into Turkey is really really strong,
especially since he committed so hard against Russia
TURKEY: Nope, you gotta let me have it
FRANCE:
No, I need it.
FRANCE: You have Serbia and Rome to take.
Message filtering
TURKEY: they're impossible targets
I'm down to go for it. Would
ITALY:
definitely need your help in 02 though
FRANCE: Greece - Ionian Ionian - Tyrr
AUSTRIA: Of course, happy to do that!
TURKEY: hm, you're right
ITALY: Fantastic!
TURKEY: good ideas
FRANCE: Then in fall you take Rome and Austria collapses.
Fig. 6. Successful dialogue examples. Examples of Cicero (left) coordinating and (right) negotiating with
authors of this paper in test games.
Meta Fundamental AI Research Diplomacy Team (FAIR) et al., Science 378, 1067–1074 (2022)
9 December 2022
6 of 8
RES EARCH | R E S E A R C H A R T I C L E
several of these filters here, with additional
details in the SM, section D.3.
Discriminating between human text
and counterfactuals
Much work has used adversarial or counter-
factual examples to improve the robustness of
natural language systems (30, 31). Following
this approach, we generated many kinds of
counterfactual messages that contained mis-
takes that language models are prone to, in-
cluding heuristically corrupted text as well as
model-generated negatives. We trained a suite
of 16 classifiers to discriminate between the
ground-truth human message and different
kinds of counterfactual messages (sometimes
varying the random seed or context informa-
tion available) and used these classifiers in an
ensemble to filter messages. This approach
risked overly filtering complex messages that
contain precise plans and accepting bland
messages, such as “ok,” which are unlikely to
contain mistakes. However, we found that care-
fully designing our ensemble allowed us to
filter most nonsensical messages with mini-
mal impact on message complexity: On a small
evaluation set with 362 expert-annotated ex-
amples, we found that we could detect 83%
of nonsense messages, without substantial im-
pact to message diversity as measured by the
proxy of message length and the number of
references to Diplomacy-specific entities (SM,
section D.3.1).
Intent correspondence
As noted previously, controlling dialogue gen-
eration through intents has the twofold benefit
of improving the strategic value of a message
and reducing discussion of impossible moves
or other hallucinations. However, this control
is imperfect, and the dialogue model may gen-
erate messages that contradict the intents it
conditions on. To address this, we filtered mes-
sages that would reduce the likelihood of the
actions in the intent. Evaluating this method
on a small test set of 1013 expert-annotated
messages, we achieved a recall of 65%, filtering
24% of all messages (SM, section D.3.2).
Value-based filtering
Conditioning on intents can lead to “infor-
mation leakage,” in which the agent reveals
compromising information about its plan to
an adversary (section Selecting intents during
play). To mitigate this, we developed a method
to score potential messages by their estimated
value impact. We computed the piKL policies
for all agents after each candidate message
and filtered those that led to a lower expected
value (EV) for Cicero playing its intended
action. Expert evaluation on a set of 127 dia-
logue scenarios demonstrated that accepted
messages were preferred over filtered messages
62% of the time (P < 0.05) (SM, section D.3.3).
Other filters
We additionally deployed other filters—for ex-
ample, to detect toxic language (SM, section
D.3.4)—and heuristics to curb bad behaviors,
including repetition and off-topic messages
(SM, section D.3.5).
Cicero in anonymous human play
Cicero participated anonymously in 40 games of
Diplomacy in a “blitz” league on webDiplomacy.net
from 19 August to 13 October 2022. This league
played with 5-min negotiation turns; these time
controls allowed games to be completed within
2 hours. Cicero ranked in the top 10% of par-
ticipants who played more than one game and
second out of 19 participants in the league that
played five or more games. Across all 40 games,
Cicero’s mean score was 25.8%, which was more
than double the average score of 12.4% of its
82 opponents. As part of the league, Cicero
participated in an eight-game tournament that
involved 21 participants, six of whom played
at least five games. Participants could play a
maximum of six games, with their rank deter-
mined by the average of their best three games.
Cicero placed first in this tournament.
During games, players were not able to
see the usernames of other players. Although
webDiplomacy notifies users that the website
has participated in AI research and that cer-
tain game modes allow users to play with AI
agents, we evaluated Cicero in games with
humans in which the participants were not
explicitly informed that they were playing
with an AI agent for that particular game.
Cicero’s participation as an AI was revealed to
all players at the conclusion of the research
(SM, section A.4).
Discussion
Cicero successfully combined strategic reason-
ing and dialogue to cooperate and negotiate
with humans on a complex task, achieving
strong human-level performance in the game
of Diplomacy. Furthermore, Cicero passed
as a human player for 40 games of Diplomacy
with 82 distinct players, and no in-game mes-
sages indicated that players believed that
they were playing with an AI agent. One
player mentioned in post-game chat a suspi-
cion that one of Cicero’s accounts might be
a bot, but this did not lead to Cicero being
detected as an AI agent by other players in
the league.
Two examples of coordination and negoti-
ation are shown in Fig. 6. In the coordination
example, we observed Cicero building an al-
liance through discussion of a longer-term
strategy. In the negotiation example, Cicero
successfully changed the other player’s mind by
proposing mutually beneficial moves. Despite
dishonesty being commonplace in Diplomacy,
we were able to achieve human-level perfor-
mance by controlling the agent’s dialogue through
the strategic reasoning module to be largely
honest and helpful to its speaking partners.
Although Cicero is shown to be effective at
cooperating with humans, it occasionally sent
messages that contained grounding errors,
contradicted its plans, or were otherwise stra-
tegically subpar. Although we reduced errors
with a suite of filters, Diplomacy poses an in-
teresting benchmark for studying this prob-
lem. We suspect that these mistakes did not
raise further suspicions that Cicero was an AI
agent because of the time pressure imposed
by the game, as well as because humans occa-
sionally make similar mistakes. As such, formats
of Diplomacy with longer negotiation periods
could provide an even further challenge for
future work because players typically engage
in more detailed and complex negotiation in
these formats.
From a strategic perspective, Cicero reasoned
about dialogue purely in terms of players’ ac-
tions for the current turn. It did not model
how its dialogue might affect the relationship
with other players over the long-term course
of a game. Considering this might allow it to
deploy dialogue more strategically. Further-
more, the expressive power of our intent rep-
resentation limited Cicero’s ability to control
richer affordances of dialogue such as strate-
gically revealing information, asking questions,
or providing explanations for its actions. There
remain many open problems for intentional
use of dialogue, and Diplomacy provides a rich
testbed to explore these connections between
strategy and communication, with the goal
of improving coordination between humans
and agents.
Ethical considerations
We discuss ethical considerations for this re-
search further in the SM, including privacy
considerations for data usage (SM, section A.1),
potential harms resulting from toxic or biased
language generation (SM, section A.2), avenues
for misuse of goal-oriented dialogue technol-
ogy (SM, section A.3), and AI agent disclosure
to human players (SM, section A.4).
REFERENCES AND NOTES
1. T. Brown et al., Adv. Neural Inf. Process. Syst. 33, 1877
(2020).
2. M. Campbell, A. J. Hoane Jr., F. Hsu, Artif. Intell. 134, 57–83
(2002).
3. D. Silver et al., Nature 529, 484–489 (2016).
4. N. Brown, T. Sandholm, Science 365, 885–890 (2019).
5. S. Kraus, D. Lehmann, Diplomat, an agent in a multi agent
environment: An overview, in IEEE International Performance
Computing and Communications Conference (IEEE Computer
Society, 1988), pp. 434–435.
6. D. d. Jonge et al., The challenge of negotiation in the game of
Diplomacy, in International Conference on Agreement
Technologies (Springer, 2018), pp. 100–114.
7. P. Paquette et al., Adv. Neural Inf. Process. Syst. 32,
4474–4485 (2019).
8. A. Dafoe et al., Nature 593, 33–36 (2021).
9. M. Moravčík et al., Science 356, 508–513 (2017).
10. N. Brown, T. Sandholm, Science 359, 418–424 (2018).
11. O. Vinyals et al., Nature 575, 350–354 (2019).
Meta Fundamental AI Research Diplomacy Team (FAIR) et al., Science 378, 1067–1074 (2022)
9 December 2022
7 of 8
RES EARCH | R E S E A R C H A R T I C L E
12. Dota 2 (32) is two-team zero-sum but with unlimited information
sharing between teammates, which makes the game equivalent to
2p0s. Prior work found that self-play from scratch was sufficient
for achieving superhuman performance in multiplayer poker (4),
but this may be due to poker offering few opportunities for players
to cooperate.
13. J. v. Neumann, Math. Ann. 100, 295 (1928).
14. M. Lewis, D. Yarats, Y. Dauphin, D. Parikh, D. Batra, Deal or no
deal? End-to-end learning of negotiation dialogues, in
Proceedings of the 2017 Conference on Empirical Methods
in Natural Language Processing (Association for Computational
Linguistics, Copenhagen, Denmark, 2017), pp. 2443–2453.
15. A. P. Jacob, M. Lewis, J. Andreas, Multitasking inhibits semantic
drift, in Proceedings of the 2021 Conference of the North American
Chapter of the Association for Computational Linguistics: Human
Language Technologies (2021), pp. 5351–5366.
16. A. Bakhtin, D. Wu, A. Lerer, N. Brown, Adv. Neural Inf. Process.
Syst. 32, 34 (2021).
17. A. Bakhtin et al., Mastering the game of no-press Diplomacy via
human-regularized reinforcement learning and planning.
arXiv:2210.05492 [cs.GT] (2022).
22. K. Shuster et al., Language models that seek for knowledge:
Modular search and generation for dialogue and prompt
completionarXiv:2203.13224 (2022).
23. A. Vaswani et al., Attention is all you need, in Advances in
Neural Information Processing Systems, I. Guyon, et al., eds.
(Curran Associates, Inc., 2017), vol. 30.
24. M. Lewis et al., BART: Denoising sequence-to-sequence pre-training
for natural language generation, translation, and comprehension, in
Proceedings of the 58th Annual Meeting of the Association for
Computational Linguistics, ACL 2020, Online, July 5–10,2020,
D. Jurafsky, J. Chai, N. Schluter, J. R. Tetreault, Eds. (Association for
Computational Linguistics, 2020), pp. 7871–7880.
25. N. S. Keskar, B. McCann, L. R. Varshney, C. Xiong, R. Socher,
CTRL: A conditional transformer language model for
controllable generation. arXiv:1909.05858 [cs.CL] (2019).
26. A. P. Jacob et al., Modeling strong and human-like gameplay
with KL-regularized search, in International Conference on
Machine Learning (PMLR, 2022), pp. 9695–9728.
27. D. Silver et al., Science 362, 1140–1144 (2018).
28. N. Brown, A. Bakhtin, A. Lerer, Q. Gong, Adv. Neural Inf.
Process. Syst. 33, 17057 (2020).
18. A. Holtzman, J. Buys, L. Du, M. Forbes, Y. Choi, The curious
29. Z. Ji et al., Survey of hallucination in natural language
case of neural text degeneration, 8th International Conference
on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia,
April 26-30, 2020 (OpenReview.net, 2020).
19. S. Keizer et al., Evaluating persuasion strategies and deep
reinforcement learning methods for negotiation dialogue
agents, in Proceedings of the 15th Conference of the European
Chapter of the Association for Computational Linguistics: Volume
2, Short Papers (Association for Computational Linguistics,
2017), pp. 480–484.
20. T. Hiraoka, G. Neubig, S. Sakti, T. Toda, S. Nakamura,
Construction and analysis of a persuasive dialogue corpus, in
Situated Dialog in Speech-Based Human-Computer Interaction
(Springer, 2016), pp. 125–138.
21. X. Wang et al., Persuasion for good: Towards a personalized
persuasive dialogue system for social good, in Proceedings of
the 57th Annual Meeting of the Association for Computational
Linguistics (Association for Computational Linguistics,
Florence, Italy, 2019), pp. 5635–5649.
generation. arXiv:2202.03629 [cs.CL] (2022).
30. P. Gupta, Y. Tsvetkov, J. P. Bigham, Synthesizing adversarial
negative responses for robust response ranking and evaluation, in
Findings of the Association for Computational Linguistics: ACL/
IJCNLP 2021, Online Event, August 1-6,2021, C. Zong, F. Xia, W. Li,
R. Navigli, Eds. (Association for Computational Linguistics, 2021),
vol. ACL/IJCNLP 2021 of Findings of ACL, pp. 3867–3883.
31. M. Alzantot et al., Generating natural language adversarial
examples, in Proceedings of the 2018 Conference on
Empirical Methods in Natural Language Processing, Brussels,
Belgium, October 31–November 4,2018, E. Riloff, D. Chiang,
J. Hockenmaier, J. Tsujii, eds. (Association for Computational
Linguistics, 2018), pp. 2890–2896.
32. C. Berner et al., Dota 2 with large scale deep reinforcement
learning. arXiv:1912.06680 (2019).
33. FAIR et al., Supplementary data for “Human-level play in the
game of Diplomacy by combining language models with
strategic reasoning”. Zenodo (2022).
34. FAIR et al., Code for “Human-level play in the game of
Diplomacy by combining language models with strategic
reasoning”. GitHub (2022); https://github.com/
facebookresearch/diplomacy_cicero.
AC KNOWLED GME NTS
We thank K. Kuliukas for providing access to the WebDiplomacy
data and for supporting this research; J. Andreas, N. Goyal,
P. Paquette, K. Shuster, and S. Zhang for helpful support and
discussions; and J. Weston for feedback on early drafts of this
paper. Funding: All funding was provided by Meta. Author
contributions: Authors are listed alphabetically in the byline.
A.B., N.B., E.D., G.F., C.F., D.F., J.G., H.H., A.P.J., M.Ko., M.Kw.,
A.L., M.L., A.R., S.R., W.S., A.W., D.W., and H.Z. contributed
to the development of Cicero algorithms, code, and
experiments. A.G., K.K., and M.Z. provided Diplomacy expertise
and data annotation. S.M. and D.R. contributed to data
collection. A.H.M. and J.S. managed the research team. A.B.,
N.B., E.D., G.F., C.F., D.F., J.G., A.P.J., A.L., M.L., A.H.M., A.R.,
W.S., A.W., and D.W. wrote the paper. Competing interests:
None declared. Data and materials availability: The figure
and table data are deposited in Zenodo (33). The codebase is
available at (34). Training data was licensed from
WebDiplomacy.net by Meta AI. License information:
Copyright © 2022 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of
Science. No claim to original US government works.
https://www.science.org/about/science-licenses-journal-
article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade9097
Materials and Methods
Figs. S1 to S10
Tables S1 to S15
References (35–90)
Submitted 15 September 2022; accepted 9 November 2022
10.1126/science.ade9097
Meta Fundamental AI Research Diplomacy Team (FAIR) et al., Science 378, 1067–1074 (2022)
9 December 2022
8 of 8
|
10.1126_science.ade5671
|
RES EARCH
SEXUAL SELECTION
Female preference for rare males is maintained by
indirect selection in Trinidadian guppies
Tomos Potter1*†, Jeff Arendt2†, Ronald D. Bassar3, Beth Watson4, Paul Bentzen4,
Joseph Travis1, David N. Reznick2
When females prefer mates with rare phenotypes, sexual selection can maintain rather than deplete
genetic variation. However, there is no consensus on why this widespread and frequently observed
preference might evolve and persist. We examine the fitness consequences of female preference for rare
male color patterns in a natural population of Trinidadian guppies, using a pedigree that spans 10
generations. We demonstrate (i) a rare male reproductive advantage, (ii) that females that mate with
rare males gain an indirect fitness advantage through the mating success of their sons, and (iii) the
fitness benefit that females accrue through their “sexy sons” evaporates for their grandsons as their
phenotype becomes common. Counter to prevailing theory, we show that female preference can be
maintained through indirect selection.
W hether female preference for rare males
can sustain genetic polymorphisms in
nature has long been controversial (1).
Rarity, as an attractive trait, compli-
cates sexual selection theory for the
evolution of female preference because it in-
troduces negative frequency-dependence. Nega-
tive frequency-dependent selection occurs when
the fitness of a trait decreases as it becomes
more common and increases as it becomes
rarer. In the absence of negative frequency-
dependence, female preference for certain
male traits can be explained if the sons of at-
tractive males are also attractive (2, 3) or if
paternal attractiveness correlates with enhanced
viability in offspring (4, 5). However, when
attractiveness is frequency-dependent, as is
the case when rare male phenotypes have an
advantage, sons can become victims of their
father’s success. Specifically, the progeny of
successful rare males are doomed to become
common and thus unattractive. Although there
is robust theory describing conditions under
which female preference for rare males may
evolve (6), it is unclear whether the costs of
such a preference will outweigh its benefits (7).
Numerous laboratory studies have demon-
strated a rare-male mating advantage in sev-
eral taxa [reviewed in (8, 9)]. However, we
know of only five such studies in nature (10–14).
As is common in laboratory studies, all studies
in nature except one (12) reduced the options
for female choice to just two male morphs
(albeit in some cases noting variation within
morphs) (11, 14). These studies were further
limited to a single mating season, precluding
the detection of long-term fitness consequences
1Department of Biological Science, Florida State University,
Tallahassee, Florida, USA. 2Department of Evolution, Ecology
and Organismal Biology, University of California, Riverside,
California, USA. 3Department of Biological Sciences, Auburn
University, Auburn, Alabama, USA. 4Department of Biology,
Dalhousie University, Halifax, Nova Scotia, Canada.
*Corresponding author. Email: tomos.potter@protonmail.com
†These authors contributed equally to this work.
for females that mate with rare males. As a
result, although these studies have docu-
mented rare-male advantage and demon-
strated its proximate mechanism through
female preference behaviors, the ultimate,
evolutionary explanation of why females pre-
fer to mate with rare males remains unknown.
We show an advantage for rare color pat-
terns in males under natural conditions in the
highly polymorphic Trinidadian guppy (Poecilia
reticulata) and quantify the fitness conse-
quences of this advantage over multiple gene-
rations. Male color patterns in guppies are
reliably transmitted from father to son (15–18).
We identified 27 distinct color patterns in our
study population and have confirmed that all
males within a patriline share the same pat-
tern (Fig. 1) (19). During courtship, males dis-
play their pattern to potential mates (20).
Numerous laboratory studies (15, 20–30) and
one field manipulation (31) have demonstra-
ted female preference for rare or unfamiliar
male patterns. Whether those results apply to
the much wider level of unmanipulated varia-
tion in wild male color patterns is unknown.
We tested these ideas as part of an experi-
mental study of evolution in a natural stream
in Trinidad (19, 32, 33). We used monthly mark-
recapture data to determine the presence and
movement patterns of individuals within the
population over this period. We used a micro-
satellite-based pedigree to determine the rela-
tedness of individuals and the reproductive
success for each individual every month over
this period (19). Our dataset includes monthly
observations of 7173 individuals spanning 10
generations (34). We used generalized linear
mixed effects models (GLMMs) to test the ef-
fects of male pattern rarity and novelty (defined
below) on components of fitness in guppies.
Measuring rarity and novelty
We assigned a “rarity” score monthly to each
male. We calculated the rarity of a focal pat-
tern (ri ) as a function of the total number of
individuals with that pattern (ni ), the total
number of individuals of all patterns (Np), and
the degree of polymorphism, i.e., the number
of patterns (P):
ri ¼ ln
(cid:3)
P
(cid:1)
ni
Np
Weighting the relative frequency of patterns
( ni =NP ) byP allows meaningful comparison across
localities with different degrees of polymor-
phism. Another useful aspect of this approach
is that log-transformation results in rare pat-
terns having negative values, common patterns
having positive values, and patterns that are
neither rare nor common (ni =NP
P= ) having
a value of zero. This makes linearized model
coefficients directly interpretable. To illustrate
our results, we define a “rare” male as one with
a pattern half as frequent as expected given
the total degree of polymorphism [i.e., ri ¼
lnð0:5Þ], and a “common” male as one with a
pattern twice as frequent as expected [i.e.,
ri ¼ lnð2Þ]. These illustrative values fall well
within the observed distribution of male pat-
tern rarity (fig. S1).
¼ 1
A female’s assessment of male rarity will de-
pend upon the males that she regularly en-
counters, which may be a spatial subset of the
total population. The stream habitat is sub-
divided into discrete pools connected by riffles.
Our spatially explicit mark-recapture censuses
allowed us to reconstruct patterns of move-
ment and make inferences about population
structure. To assess the possibility that female
mate preference is shaped by that structure,
we calculated rarity at three spatial scales: the
local level (e.g., the pool or riffle where the fish
was caught that month), the neighborhood,
and the whole population. Neighborhoods
were defined using network analysis of move-
ment of male guppies between pools (19). We
identified four distinct multipool neighbor-
hoods characterized by high movement within
but low movement between. Males moved
around considerably (62% were new arrivals
to pools each month, 17% were new arrivals
to neighborhoods, fig. S2), whereas females
moved around much less (28% in pools, 4%
in neighborhoods, fig. S2). When a female as-
sesses how rare a male is, she is likely to see all
those males that we collected in the pool with
her that month. Although we did not observe
this directly, our neighborhood-level analyses
include males likely to have passed through the
pool in the previous month.
In addition to an advantage to rarity, several
studies have demonstrated the advantage of
novelty in the form of female preference for
unfamiliar males (15, 22, 23, 26, 35), regard-
less of their color pattern. Female guppies may
identify novel males through olfactory cues
(36). As such, males that are new arrivals in a
Potter et al., Science 380, 309–312 (2023)
21 April 2023
1 of 4
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Male color patterns in guppies are reliably transmitted from father to son. Here, we show three generations (father, son, and grandson) for five example
patrilines showing consistency in color pattern within a Y-lineage. Numbers to the left indicate which lineage of 27 are being shown. Some elastomer marks, used
to identify individuals, are clearly visible. For example, the son of lineage 14 has a red mark on the dorsal side of the caudal peduncle.
pool or neighborhood may experience a re-
productive advantage. To test this, we defined
males as “novel” if they were new arrivals to
the pool or neighborhood in which they were
caught that month. By this definition, males
cease to be novel one month after arriving in
a locality.
Results
We found evidence for negative frequency-
dependent selection operating on male color
patterns, resulting in a rare-male advantage
(Fig. 2 and table S1). These effects were sig-
nificant over all three spatial scales over which
male rarity was calculated but were strongest
and weakest at the neighborhood and local levels,
respectively, as determined by Akaike informa-
tion criterion (AIC) scores (table S1). Each month,
males with rarer patterns at the neighborhood
level had 36% more mating partners (GLMM,
n = 6248, P-value = 2.43 × 10−6) and ultimate-
ly sired 38% more offspring that recruited into
the population (GLMM, n = 6248, P-value =
4.75 × 10−6) (Fig. 2). These results, observed
over multiple generations, provide strong evi-
dence that negative frequency-dependent sex-
ual selection is occurring through rare-male
advantage.
Novel males (new arrivals to a pool or
neighborhood, regardless of their color pat-
tern) also had a large reproductive advantage
over residents (Fig. 2 and table S1), with the
effect strongest at the local level. Compared
with residents of equivalent rarity, new arri-
vals to pools had 45% more mating partners
(GLMM, n = 6248, P-value = 2.77 × 10−4) and
sired 50% more offspring each month (GLMM,
n = 6248, P-value = 3.30 × 10−4).
In line with earlier studies (15, 20–30), our
results indicate that female guppies prefer to
mate with rare color-patterned and/or unfam-
iliar males. One potential explanation for this
preference is inbreeding avoidance: rare or
novel males may be less likely to be kin (37).
However, we found no evidence that male
rarity was associated with the relatedness of
j
ð
mating partners (linear model, n = 1259,
P-value = 0.273) (19). Surprisingly, resident
males were more likely to be unrelated to their
partners than novel males who were new ar-
rivals to pools [P unrelated resident ¼ 0:28
Þ
,
P unrelated novel ¼ 0:15
Þ
ð
j
, logistic regression,
n = 1580, P-value = 5.74 × 10−8)]. This could
occur if males are more likely to remain in
pools with unrelated females. Nevertheless,
the higher relatedness of novel (and thus at-
tractive) males and the absence of any as-
sociation of relatedness with rarity contradict
the inbreeding-avoidance hypothesis (table S2).
We found no direct benefit for females’ pre-
ference for rarity or novelty. Mating with rare
males did not result in more recruited off-
spring (table S3 and Fig. 2C, GLMM, n = 2290,
P-value = 0.448), nor did mating with new
arrivals to neighborhoods (P-value = 0.969)
or pools (P-value = 0.397). Our measure of
recruited offspring refers to those that sur-
vived to be large enough to be individually
marked (~2 months old) (19), meaning that
Potter et al., Science 380, 309–312 (2023)
21 April 2023
2 of 4
RES EARCH | R E S E A R C H A R T I C L E
s
r
e
n
t
r
a
p
f
o
r
e
b
m
u
N
0.6
0.4
0.2
0.0
g
n
i
t
a
m
r
e
p
g
n
i
r
p
s
f
f
O
1.5
1.0
0.5
0.0
A
r
c
−1
0
r i
C
r
c
B
r
c
g
n
i
r
p
s
f
f
o
d
e
t
i
u
r
c
e
R
0.6
0.4
0.2
0.0
1
−1
1
0
r i
r
c
g D
3
2
1
0
n
i
t
a
m
r
e
p
g
n
i
r
p
s
f
f
o
−
d
n
a
r
G
N. S.
−1
0
r i
1
−1
0
r i
1
new arrival
resident
Fig. 2. Rare and novel males have higher
reproductive fitness, and females that mate
with rare males have more grand-offspring.
Effects of male pattern rarity (ri) and novelty (new
arrival or resident) on components of fitness.
(A) Number of mating partners per month (for
males); (B) monthly number of offspring
recruited into the population (for males);
(C) number of recruited offspring per mating
(for males and females); (D) number of grand-
offspring that ultimately recruit into the
population from a single mating (for males and
females). Values for “rare” [ri ¼ ln 0:5ð
“common” [ri ¼ ln 2ð Þ] males are indicated
with dotted lines annotated r and c, respectively;
the dashed line indicates ri ¼ 0. Shaded areas
are 95% confidence intervals, N.S. indicates
that the slope is not significant (P > 0.05).
Predictions are based on models where rarity
was calculated at the neighborhood level (A)
and (B) or the population level (C) and (D).
Þ] and
variation in offspring viability is captured in
this metric. This indicates that preference for
rare or novel males is not under direct selec-
tion through mechanisms that enhance off-
spring viability, such as inbreeding avoidance
or so-called “good genes” (4, 5).
What then is the ultimate benefit of mating
with rare or novel males? Although we de-
tected no indirect fitness benefits for females
that mated with novel males [i.e., that were
new arrivals to neighborhoods (GLMM, n =
1951, P-value = 0.439) or pools (P-value =
0.573)], matings with rare males (at the pop-
ulation level) ultimately resulted in 48% more
grand-offspring recruited into the population
than matings with common males (P-value =
3.42 × 10−4). This is a substantial indirect fit-
ness benefit for those females (table S3 and
Fig. 2D).
Females that mated with rare males gained
this indirect fitness advantage through the en-
hanced reproductive success of their sons: a
so-called “sexy sons” effect [in the sense of
100% 69% 38%
i
r
1
0
−1
−2
f1
f2
Generation
f3
Fig. 3. Rare males have rare sons but common
grandsons. In this figure, we track the trajectory
of male pattern rarity (ri, at the population level,
x axis) over three generations (y axis), focusing on
rare males [f1, ri < ln 0:5ð
grandsons (f3). Lines connect male to sons to
grandsons. Percentages describe individuals in each
generation where ri < 0, i.e., that are rarer than
expected.
Þ], their sons (f2), and
Kokko (6)]. The sons of rare males (at the
population level) sired more offspring per
month (table S4, GLMM, n = 9807, P-value =
0.0039), but there was no such effect in
daughters (P-value = 0.575). This occurred
because the sons of rare males were also rare
(albeit less so than their fathers) and thus still
attractive (Fig. 3). This advantage was short-
lived, however; after two generations of rare-
male advantage the grandsons of rare males
became victims of their forefathers’ success
and were more likely to be common (Fig. 3).
Consequently, for males with equivalently rare
fathers, having a rare grandfather reduced re-
productive success (P = 0.0082).
Discussion
Female guppies that mated with rare males
gained no direct fitness advantage in doing
so. Their offspring did not have increased via-
bility due to any genetic advantage of their
attractive fathers, nor were they less inbred.
Instead, females that mated with rare males
derived substantial indirect fitness through
the attractiveness of their sons. This is at odds
with the long-held prediction that such in-
direct selection cannot maintain female pref-
erence (7, 38–40). This prediction is not an
ineluctable consequence of theory. It is based
on assumptions about the genetic variances
and covariances of female preference and
male traits, which imply that indirect selection
must always be overwhelmed by the direct
costs of female choice (7). We suggest that
this may not hold true when the desirable
trait is rarity and physical traits are arbitrary.
Our study shows that female preference can be
maintained by indirect selection when nega-
tive frequency-dependence is operating.
Our findings offer a resolution to the “lek
paradox” (39). To maintain female preferences
through indirect selection, there must be a
sustained supply of genetic variation in male
traits (39, 41). The crux of the lek paradox is
that selection on male traits will erode that
genetic variation, ultimately resulting in the
loss of female preference (39). However, when
females prefer rare males—regardless of male
genotype—negative frequency-dependent se-
lection occurs, ensuring the necessary mainte-
nance of genetic variation.
A notable result is the absence of any de-
tectable fitness benefit, direct or indirect, for
females that mated with novel males. Novel
males (new arrivals to pools or neighborhoods)
had substantially higher reproductive success
than residents, regardless of the rarity of their
color pattern. Our results illustrate that mat-
ing with rare and novel males has distinct fit-
ness consequences for females: Rare males
conferred a single-generation reproductive ad-
vantage to their sons, driving indirect selec-
tion for female preference, whereas novel males
conferred no fitness advantages to their part-
ners, either directly or through their offspring.
Why then do females prefer novel males?
One possibility is that female preferences
for rare and novel males stem from a single,
simple mechanism: habituation to familiar
males, i.e., females preferring males that are
unlike those they have recently encountered
(21, 22, 27). Males with rare color patterns or
that are new arrivals to pools (i.e., novel) are
both likely to fit this criterion. In this scenario,
selection for choosy females is driven by the
indirect fitness advantage gained when they
mate with rare males. By contrast, preference
for novel males emerges as a nonadaptive by-
product of the simple behavioral mechanism
under selection.
Female choosiness is likely also under
frequency-dependent selection (6, 42). Consi-
der what would happen if all females mated
with a single male bearing the rarest color
pattern: the sexy son benefit would be lost
because all male offspring would have the
same pattern, making it common and thus
unattractive. As a result, selection for choosy
females would evaporate. Although we do not
know the mean frequency of choosy females in
our population, theory suggests that it is likely
to be high. Female preference alleles evolve to
higher frequencies when the ability to express
choice is hindered (6). Here, female choice was
hindered by the different movement patterns
of males and females. The optimum scale on
which females should choose rare males is at
the level of the population: Males frequently
change location so the rarity of sons, upon
which the indirect benefits to females depend,
is best predicted by the rarity of fathers at the
population level (tables S3 and S4). However,
females can only assess the rarity of males they
encounter. The more limited movement of
females meant that they chose males that were
Potter et al., Science 380, 309–312 (2023)
21 April 2023
3 of 4
RES EARCH | R E S E A R C H A R T I C L E
rare at the level of the neighborhood (table S1).
This mismatch between the optimum and
realized exercises of choice creates the hin-
drance that could sustain a high frequency of
choosy females.
In conclusion, our results challenge the
theoretical arguments against the role of sexy
sons in sexual selection (7, 38–40) by showing
that this indirect form of selection can sustain
female preference. At the same time, we show
that female preference for rare male pheno-
types resolves the lek paradox. Both results are
a consequence of negative frequency-dependent
selection operating on sexual signals and pref-
erences in guppies. Female preference for rare
males is well documented in a diversity of or-
ganisms (8–14), but detecting indirect selection
in the wild is uncommon because it requires
multigenerational studies. The replication of
such studies in other organisms will test the
generality of our results and determine the
broader importance of sexual selection in main-
taining, rather than depleting, genetic variation
in the wild.
RE FE RENCES AND N OT ES
10. D. A. S. Smith, Science 187, 664–665 (1975).
11. J. Muggleton, Heredity 42, 57–65 (1979).
12. V. M. Salceda, W. W. Anderson, Proc. Natl. Acad. Sci. U.S.A. 85,
9870–9874 (1988).
38. M. Kirkpatrick, N. H. Barton, Proc. Natl. Acad. Sci. U.S.A. 94,
1282–1286 (1997).
39. M. Kirkpatrick, M. J. Ryan, Nature 350, 33–38 (1991).
40. J. S. Kotiaho, M. Puurtinen, Funct. Ecol. 21, 638–644
13. B. S. Masters, B. G. Hicks, L. S. Johnson, L. A. Erb, Proc. Biol. Sci.
(2007).
270, 1393–1397 (2003).
14. D. Estévez et al., Front. Mar. Sci. 7, 614237 (2020).
15. J. A. Farr, Evolution 31, 162–168 (1977).
16. A. E. Houde, Anim. Behav. 36, 888–896 (1988).
17. A. E. Houde, Heredity 69, 229–235 (1992).
18. J. R. Paris et al., Nat. Commun. 13, 1233 (2022).
19. Materials and methods are available as supplementary materials.
20. J. A. Farr, Behaviour 74, 38–90 (1980).
21. K. A. Hughes, L. Du, F. H. Rodd, D. N. Reznick, Anim. Behav. 58,
907–916 (1999).
22. A. L. Eakley, A. E. Houde, Proc. R. Soc. B. 271 (2004).
23. S. R. K. Zajitschek, J. P. Evans, R. Brooks, Behav. Ecol. 17,
911–916 (2006).
24. S. R. K. Zajitschek, R. C. Brooks, Am. Nat. 172, 843–854
(2008).
25. K. J. Hampton, K. A. Hughes, A. E. Houde, Ethology 115,
475–481 (2009).
26. M. M. Mariette, S. R. K. Zajitschek, C. M. Garcia, R. C. Brooks,
J. Evol. Biol. 23, 1772–1782 (2010).
27. M. J. Daniel, L. Koffinas, K. A. Hughes, Proc. R. Soc. B. 286,
20190435 (2019).
28. T. Lucon-Xiccato, A. Bisazza, A. Pilastro, Anim. Behav. 155,
217–224 (2019).
29. J. J. Valvo, F. H. Rodd, K. A. Hughes, Behav. Ecol. 30,
1672–1681 (2019).
30. M. J. Daniel, L. Koffinas, K. A. Hughes, Am. Nat. 196, 414–428
(2020).
31. K. A. Hughes, A. E. Houde, A. C. Price, F. H. Rodd, Nature 503,
108–110 (2013).
41. H. Kokko, I. Booksmythe, M. D. Jennions, J. Evol. Biol. 28,
259–266 (2015).
42. H. Kokko, M. D. Jennions, R. Brooks, Annu. Rev. Ecol. Evol. Syst.
37, 43–66 (2006).
AC KNOWLED GME NTS
We thank the many field managers and interns, past and present,
for their contributions to the Guppy Project and J. Ramlal and
M. Ramlal and their family for hosting us in Trinidad. Funding:
We gratefully acknowledge the continuing support of the National
Science Foundation [DEB-0623632EF, DEB-0808039, and DEB-
1258231 (to D.N.R.), DEB-1556884 (to J.T.), and DEB-2100163
(to R.D.B.)]. Ethics: All animals were collected and handled with
the approval of the University of California Riverside IACUC AUP
(a-20080008). Author contributions: Conceptualization: J.A.,
D.N.R., T.P., J.T.; Methodology: all authors; Investigation: T.P.,
J.A., J.T., D.N.R.; Visualization: T.P.; Funding acquisition: D.N.R.,
J.T., R.D.B.; Project administration: D.N.R., J.T., R.D.B.; Supervision:
D.N.R., J.T.; Writing – original draft: T.P.; Writing – review and
editing: all authors. Competing interests: Authors declare
that they have no competing interests. Data and materials
availability: All data and code required to reproduce the analyses
presented here are hosted at Zenodo (34). License information:
Copyright © 2023 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science. No
claim to original US government works. https://www.sciencemag.
org/about/science-licenses-journal-article-reuse
1. R. C. Lewontin, The Genetic Basis of Evolutionary Change
32. J. Travis et al., in Advances in Ecological Research,
(Columbia Univ. Press, 1974), vol. 560.
2. R. A. Fisher, The Genetical Theory of Natural Selection (The
Clarendon Press, 1930).
3. R. Lande, Proc. Natl. Acad. Sci. U.S.A. 78, 3721–3725 (1981).
4. M. Andersson, Evolution 40, 804–816 (1986).
5. A. Grafen, J. Theor. Biol. 144, 473–516 (1990).
6. H. Kokko, M. D. Jennions, A. Houde, Proc. Biol. Sci. 274,
1317–1324 (2007).
7. E. Cameron, T. Day, L. Rowes, J. Evol. Biol. 16, 1055–1060 (2003).
8. P. Knoppien, Biol. Rev. 60, 81–117 (1985).
9. H. Ajuria Ibarra, T. Reader, J. Zool. 290, 81–95 (2013).
Vol 50: Eco-Evolutionary Dynamics, J. MoyaLarano, J. Rowntree,
G. Woodward, Eds. (Elsevier, 2014), vol. 50, pp. 1–40.
33. D. N. Reznick et al., Am. Nat. 194, 671–692 (2019).
34. T. Potter et al., Data and code for the manuscript: “Female
preference for rare males is maintained by indirect selection in
Trinidadian guppies”(2023), doi: 10.5281/ZENODO.7644306
35. R. E. Graber, M. Senagolage, E. Ross, A. E. Houde, K. A. Hughes,
Ethology 121, 17–25 (2015).
36. A. J. Shohet, P. J. Watt, Behav. Ecol. Sociobiol. 55, 363–369
(2004).
37. A. M. Johnson et al., Ethology 116, 448–457 (2010).
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade5671
Materials and Methods
Figs. S1 to S4
Tables S1 to S5
References (43–55)
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
Submitted 25 August 2022; accepted 22 March 2023
10.1126/science.ade5671
Potter et al., Science 380, 309–312 (2023)
21 April 2023
4 of 4
|
10.1126_science.ade9803
|
RES EARCH
CALCULUS INSTRUCTION
Establishing a new standard of care for calculus
using trials with randomized student allocation
Laird Kramer1,2*, Edgar Fuller1,3,4*, Charity Watson1,3,4, Adam Castillo1,3,4†,
Pablo Duran Oliva1,3,5, Geoff Potvin1,2
Calculus, the study of change in processes and systems, serves as the foundation for many STEM disciplines.
Traditional, lecture-based calculus instruction may present a barrier for students seeking STEM degrees,
limit their access to STEM professions, and block their potential to address society’s challenges. A large-scale
pragmatic trial with randomized student allocation was conducted to compare two calculus instruction
styles: active student engagement (treatment condition) versus traditional, lecture-based instruction (control
condition). A sample of 811 university students were studied across 32 sections taught by 19 instructors
over three semesters at a large, US-based Hispanic-serving institution. Large effect sizes were consistently
measured for student learning outcomes in the treatment condition, which demonstrates a new standard
for calculus instruction and increased opportunities for completion of STEM degrees.
C alculus instruction needs substantial trans-
formation because it is often a barrier to
STEM degree attainment, especially for
traditionally underrepresented groups
(1–3), depriving both individuals and so-
ciety of the potential benefits of their inclu-
sion. National calls for calculus transformation
are numerous (4, 5), because failing calculus
can contribute to a student’s departure from
STEM degree programs. Only ~40% of stu-
dents entering universities with STEM degree
intentions actually graduate with a STEM de-
gree (6). More concerning is that the odds of
female students switching out of a STEM pro-
gram after a calculus course is ~1.5 times higher
than that of comparable male students (3). Fur-
thermore, Hispanic and Black/African American
students had >50% higher failure rates than
white students in calculus (7, 8).
Evidence-based instruction, which is imple-
mented in many STEM disciplines, has reliably
led to profound improvement in student suc-
cess (9–11). However, common approaches to
calculus instruction continue to rely on tradi-
tional, lecture-based practices in which students
are passive learners in the classroom and are
expected to construct their knowledge mostly
outside of the classroom by doing homework or
in recitation sessions (12). Mathematics as a dis-
cipline thus needs to embrace its role in enabling
STEM careers that will lead to prosperity for
both individuals and society at large. “Calculus …
must become a pump and not a filter” for the
1STEM Transformation Institute, Florida International
University, Miami, FL 33199, USA. 2Department of Physics,
Florida International University, Miami, FL 33199, USA.
3Center for Transforming Teaching in Mathematics, Florida
International University, Miami, FL 33199, USA. 4Department
of Mathematics and Statistics, Florida International
University, Miami, FL 33199, USA. 5Department of
Mathematics and Applied Mathematics, Virginia
Commonwealth University, Richmond, VA 23284, USA.
*Corresponding author. Email: Laird.Kramer@fiu.edu (L.K.);
Edgar.Fuller@fiu.edu (E.F.)
†Present address: Department of Mathematics, The University of
Texas at Arlington, Arlington, TX 76019, USA.
STEM pipeline, as noted in 1988 by Robert
White, president of the National Academy of
Engineering (13). Handlesman et al. (14) re-
cently argued that “we must fix the classrooms
where many students from historically excluded
communities are discouraged from pursuing
STEM” and that “… the continued exclusive use
of lectures is malpractice at best, or an act of
discrimination at worst.” Thus, it is imperative
that substantial transformation in calculus in-
struction takes place to promote more equita-
ble learning environments for all students.
We present a large-scale trial featuring ran-
domized student allocation into treatment or
control conditions to rigorously compare an
evidence-based, active student engagement
calculus course against traditional, lecture-based
calculus instruction. The work extends prior cal-
culus research investigations (15–17) by includ-
ing random assignment of students to treatment
and control sections, as well as anonymized
analysis of the identical end-of-semester learn-
ing outcomes. The study uses a pragmatic (18)
design with random allocation of students to
inform on the effectiveness of similar inter-
ventions at higher education institutions, re-
flecting real-world classroom constraints. In
these contexts, blinding of the treatment and
control conditions to both students and faculty
is not possible because blinding is only feasible
when the treatment and control conditions re-
main unknown to the participants during the
period of study (such as in a clinical trial drug
study). As with other public health or sociologi-
cal interventions, enrollment of participants
in this study revealed some aspects of a cohort
structure, but it was still possible to maintain
the essential aspects of random assignment by
following a modified protocol, as was done in
Zwarenstein et al. (18). The treatment condi-
tion integrated a suite of coherent strategies
that have been independently found (19, 20)
to improve student learning; thus, the treat-
ment was a distinct departure from tradition-
al instruction, and it was not logically possible
for the treatment condition to remain hidden
from students or faculty after the treatment be-
gan. Random assignment of faculty to control
or treatment conditions would not be possible
because an individual faculty member’s knowl-
edge, philosophy, and experience with a variety
of classroom strategies and instructional prac-
tices may intersect with the features of the treat-
ment or control conditions. The experimental
protocol thus included a group of instructors
willing to adopt the instructional methods in
the treatment condition. This comprehensive
experimental approach was intended to secure
the strongest possible evidence for critical stake-
holders to sustain the treatment beyond the trial.
The treatment condition used the modeling
practices in calculus (MPC) curriculum and
pedagogy, and the control condition represented
the preexisting, traditional instructional prac-
tices at the study institution. MPC integrates
the practices of mathematicians as a central
design tenet throughout the course. Instructors
facilitate students’ application of mathematical
“habits of mind” (21) that foster deeper under-
standing of calculus concepts, including the
identifying of patterns, hypothesis development
and testing, making connections, and communi-
cating ideas precisely to learn calculus through-
out the course. Class time is devoted to students
working collectively in small groups on prede-
signed notes and learning activities developing
their calculus understanding with minimal lec-
turing. Treatment included learning assistants
(22) who were undergraduate peers integrated
within the instructional team to facilitate student
learning and promote culturally responsive ins-
truction. The curriculum promotes mathemati-
cal practices (sense-making, problem solving,
argumentation, etc.) and established strategies
to optimize student engagement, including coop-
erative learning, argumentation and metacogni-
tion, mathematical fluency, and a culturally
responsive environment (23) (described in the
supplementary materials section 2). The MPC
design builds on the SCALE-UP calculus model
(24) and intentionally embodies well-established
recommendations for calculus instruction, in-
cluding ambitious teaching practices and strate-
gies promoted by national mathematics societies
and national reports (12, 20, 25–28).
The study was performed at Florida Interna-
tional University (FIU) in Miami, Florida, the
fourth-largest public research university in the
United States, with 58,787 students, of which
41,795 are undergraduates as of fall 2019 (29).
FIU is a Hispanic-serving institution, with 64%
of students identifying as Hispanic/Latino/a/.
Moreover, 79% of the students identify as mem-
bers of historically underrepresented racial/
ethnic minority groups, and 57% are women.
The institution’s size provided a unique oppor-
tunity to perform this study because there are
18 to 34 40-student sections of calculus 1 being
Kramer et al., Science 381, 995–998 (2023)
1 September 2023
1 of 4
RES EARCH | R E S E A R C H A R T I C L E
taught each semester and primarily serving
STEM majors. Furthermore, institutional con-
ditions created urgency to transform calculus
because historic pass rates in introductory cal-
culus averaged 55% (range 13 to 88%) over the
six semesters before the project’s pilot.
Research design
A pragmatic trial (30–32) of the MPC approach
was performed during the fall 2018, spring
2019, and fall 2019 semesters to rigorously test
student outcomes. Students were randomly
assigned individually to treatment and control
conditions at the beginning of the semester
after enrolling in sections on the basis of their
scheduling preferences using the institution’s
enrollment system. To accommodate the ran-
domized assignments, each of the experimen-
tal sections doubled in size from the usual 40
seats to 80 seats before enrollment opening.
Instructor names and section sizes were invis-
ible to students throughout the enrollment
phase. Just before each term, the 80-seat sec-
tions were split into two 40-seat sections by
assigning each student at random to either a
treatment or control section.
Once assigned, the treatment sections im-
plemented the MPC approach, whereas the
control sections were unchanged. After assign-
ment, students were free to change/drop/add
course sections up until the regular institutional
drop/add deadline (7 days after classes began).
To account for such changes, enrollments were
monitored, and only students who were ran-
domly assigned to either a treatment or control
section and remained in that section through
the regular, nonpenalty drop/add deadline were
included in the data for the experimental study
reported below. In total, 1019 students were
randomly assigned to either the treatment or
control groups. Of these, 516 students were
assigned to the treatment group, and 417 re-
mained in the section at the drop/add deadline.
At the same time, 503 students were assigned
to the control group and 394 students remained
in the section at the drop/add deadline. Our
study followed the Consolidated Standards
for Reporting Trials (CONSORT) guidelines
(18, 33, 34). The specifics of recruiting, enroll-
ment, assignment, and completion for the trial
are discussed in sections 1.3 and 3.1 of the
supplementary materials. The randomization
process produced comparable groups by mathe-
matical background and demographics; class
sizes were typical for the course (see section 3
of the supplementary materials).
Faculty participating in the study included
seven individuals teaching 16 treatment sections,
along with 12 individuals teaching 16 control
sections. Faculty recruited to teach the treat-
ment sections indicated a willingness to adopt
and implement the MPC approach, replicating
the authentic condition of faculty reforming
their classroom practice under the study design.
To prepare for the new instructional approach,
faculty participated in a 2-day, pre-semester
professional development workshop and were
provided with the MPC curricular materials.
Consistency of the MPC treatment was moni-
tored through weekly preparation meetings in
which the course objectives and pacing were
discussed. In-class monitoring by the project
team was deemed overly intrusive and disrup-
tive to classroom engagement. Control section
faculty were not guided to use any particular
practices and chose their normal instructional
approach, best described as traditional lecture
format with at most limited student engage-
ment. Potential effects of instructor differences
on learning outcomes were investigated and
are presented in section 3 of the supplementary
materials and summarized below.
The student outcome measures reported in-
clude identical end-of-semester learning mea-
sures, as well as course success data (i.e., course
grades). The end-of-semester learning measures
focused on evaluating learning using a set of
identical assessment items (problems) developed
by instructors spanning all calculus sections and
spanning the major learning objectives of a cal-
culus 1 course. The aim was to determine how
well students understood essential elements of,
and exhibited fluency and technical competency
in, calculus at course end. Assessment items
aligned to both local and national standards (35)
were embedded in a cumulative final examina-
Effect Sizes for End of Semester Learning Measures by Group
Other
Sci./Math
Eng./Comp.
Sci.
Biology
Transfer
FTiC
Black
Hispanic
Female
Male
Overall
−0.25
0.00
0.25
0.50
0.75
1.00
1.25
1.50
Cohen's d
Fig. 1. Overall end-of-semester learning mea-
sures effect sizes (Cohen’s d) broken out by
major, race/ethnicity, and gender. Error bars
indicate the 95% CI for effect size for each group.
Effect Sizes for End of Semester Course Outcome by Group
Other
Sci./Math
Eng./Comp.
Sci.
Biology
Transfer
FTiC
Black
Hispanic
Female
Male
Overall
0.00
0.25
0.50
Cohen's d
0.75
1.00
Fig. 2. Overall course success (earned grades of
A, B, or C) effect sizes (Cohen’s d) broken out
by major, race/ethnicity, and gender. Error bars
indicate the 95% CI for effect size for each group.
tion and were administered to all students in
each treatment and control section. To ensure
fidelity and fairness to both treatment and
control sections, control and treatment faculty
collaboratively developed a set of items to be
administered to both conditions in identical for-
mat and wording. This set of identical items
formed roughly two-thirds of the total final exa-
mination content, with the remaining items
added by individual faculty in a separate section
of the examination, allowing them to address
their specific instructional goals. Furthermore,
the examinations and problems were formatted
identically and without course section identifiers
to allow completely anonymized evaluation dur-
ing the subsequent comparative analysis. The
identical items covered core calculus topics in-
cluding evaluating limits, identifying extrema,
curve sketching, related rates, and evaluating
indefinite integrals. For the second and third
semesters, additional items focusing on implicit
differentiation and optimization were added
to the identical set of items (for details, see sec-
tion 3.3 of the supplementary materials). Course
success data (grades) reflect the overall assess-
ment of students as assigned by each section’s
instructor. Course grade policies were estab-
lished by individual instructors following depart-
mental syllabus guidelines and were broadly
consistent across sections and semesters.
Analysis of the end-of-course learning mea-
sures used a rubric for each problem, with five
researchers testing the initial rubric on a subset
of examinations to establish interrater reliability.
The final rubric represented consensus on all
elements and accounted for initial ambiguity
or disagreement. The analysis was performed by
a team of 10 trained evaluators, each of whom
evaluated a completely anonymous set of stu-
dent solutions. An average of two evaluators
reviewed each solution for correctness on a
scale from 0 to 100%. The evaluators were very
consistent and interrater reliability was high
(Cohen’s kappa 0.827 in fall 2018 and 0.797 in
spring 2019) (36, 37). The same rubric was ap-
plied to the fall 2019 data given its high degree
of agreement. Once all problems were evaluated,
the research team deanonymized and sorted the
results by treatment and control sections for
the comparative analysis.
Results
The results indicate significant improvements
in student learning for the MPC group across
all three semesters. Students in the treatment
group showed substantially higher scores on
the identical end-of-semester learning outcomes:
(fall 2018: d = 0.505, P < 0.01; spring 2019: d =
0.748, P < 0.001; fall 2019: d = 0.925, P < 0.001)
compared with the control group. Combining
results from all three semesters of trials (i.e.,
32 sections and 811 total students), the overall
standardized mean difference between treat-
ment and control was d = 0.774 [95% confidence
Kramer et al., Science 381, 995–998 (2023)
1 September 2023
2 of 4
RES EARCH | R E S E A R C H A R T I C L E
interval (CI) = 0.618 to 0.930] at the individual
level, a medium/large effect size (36, 37). An ad-
justed effect size (38) was computed using section-
level cluster properties and the mixed-effects
model structure described below and was found
to be dT = 0.771 (95% CI = 0.468 to 1.073; see sec-
tion 3.4.2.1 of the supplementary materials).
The success of the MPC intervention could
be seen across racial and ethnic groups, majors
and academic pathways, and genders (Fig. 1).
Similar medium/large overall effect sizes were
observed for students in the treatment condi-
tion who identified as Black/African American
(d = 0.882, P < 0.001) or Hispanic/Latino/a/
(d = 0.772, P < 0.001) when directly comparing
the identical learning measures with their coun-
terparts in the control condition. Although all
STEM majors showed significantly improved
learning, there were larger effect sizes for biology
majors in the treatment group (d = 0.925, P <
0.001). Students matriculating onto campus as
both first time in college (FTiC) and transfer
students showed medium/large effect sizes and
most were FTiC. Overall, treatment group stu-
dents show more consistency in applying the
tools of calculus to optimization problems, using
derivatives to sketch graphs of functions, and
in the evaluation of limits and integrals.
Potential biases arising in the random stu-
dent assignment and faculty selections were
investigated for hidden-level effects or confound-
ers to establish limitations of the study (see sec-
tion 3 of the supplementary materials). Random
allocation of students provided equivariance in
the demographics of student populations. Analy-
ses showed that allowing students to drop/add
sections during the open registration period
after the initial assignment did not affect the
measured outcomes. Faculty characteristics were
compared and found to be similar in both back-
ground and prior course student grade distribu-
tions. A mixed-effects model with student fixed
effects and random cluster effects due to section
and instructor levels was fit with tests of fixed
effects computed using Satterthwaite approxi-
mations (see section 3.2.4.1 of the supplemen-
tary materials). The explanatory power of the
model was found to be high (conditional R2 =
0.39, marginal R2 = 0.31). The effect of treat-
ment was statistically significant (t(660) = 5.68,
P < 0.001, semipartial R2 = 0.119) (39), implying
an estimated effect size of Cohen’s f = 0.368
with covariates and cluster-level effects present.
Random effects correlated with 0.085 of outcome
variance (intraclass correlation coefficient = 0.13
with demographic covariates). A sensitivity anal-
ysis (see section 3.2.4.3 of the supplementary
materials) showed that unmeasured confound-
ers would need to be four times more powerful
than any measured covariate, including student
mathematics background, to be responsible for
the observed effect.
Furthermore, students in the MPC treatment
condition had improved course grades. Average
grades were significantly higher by ~0.4 points
(4.0 grade point scale) in MPC sections across
all semesters of the study (P < 0.001, d = 0.295).
This translated to success rates (A, B, or C grades)
averaging 11% higher in MPC sections compared
with traditional sections (P < 0.001, d = 0.251;
Fig. 2). Outcomes were consistent across the
three semesters of the experiment (Fig. 3). More-
over, the MPC sections also had lower course
late drop rates (departure after the regular drop/
add period ends) across all three semesters
(P < 0.05, d = 0.141), suggesting that students
more clearly perceived that they were likely to
succeed in the course.
The trend of improved outcomes in course
success was also observed for demographic
subgroups, as can be seen in Fig. 2. A logistic re-
gression model of success using gender identifi-
cation, FTiC status, and Hispanic identification
as independent variables showed the odds of a
female-identified student in the treatment group
passing the course to be 58% higher than the
odds of a female-identified student in the control
group (b1 = 0.46, P < 0.05). Hispanic students’
odds of passing the course were almost double
that of their counterparts in the control group
(b1 = 0.70, P < 0.001). The likelihood of FTiC
students in the treatment group passing the
course increased by ~85% compared with FTiC
students in the control group (b1 = 0.61, P < 0.01)
(for details, see section 3.4 of the supplementary
materials).
Discussion and conclusions
This pragmatic trial demonstrates that stu-
dent learning outcomes were significantly im-
proved in the treatment condition. Contrary to
previous research (40), this study shows that
when students are expected to engage with
calculus concepts collaboratively using inten-
tional, evidence-based teaching strategies, they
develop a better understanding of calculus con-
cepts and techniques. The benefits of the MPC
curriculum and pedagogy were realized regard-
less of racial/ethnic group, gender, or major/
academic pathway. These trends suggest that
the treatment includes culturally responsive and
equitable strategies. Specifically, the MPC learn-
ing environment is designed to promote learn-
ing communities that provide ongoing support
for learning mathematics through collaborative
engagement and ongoing formative feedback.
This aims to promote inclusion and increases
access for students with different mathematical
backgrounds, cultural identities, and life expe-
riences by allowing them to use their mathe-
matics skills in a supportive, nonthreatening
environment.
The improved learning and course success
for modeling practices in calculus reported in
this study have profound implications for cal-
culus instruction. This study demonstrates the
substantial benefit to students of the MPC ap-
proach designed around established, evidence-
based principles, and should motivate educators
in mathematics and other STEM disciplines to
adopt the same or similar approaches and
conduct similar studies to replicate these find-
ings. Improved student success also leads to
more efficient student progress to graduation
and boosts institutional effectiveness. Applying
this study’s 11% average improvement in pass
rate to all 2000 first-time calculus students at
FIU would translate to 220 additional students
succeeding in calculus annually and reducing
the instructional load by five sections annually.
Extending this strategy to the ~300,000 stu-
dents across the nation taking calculus 1 each
e
t
a
L
d
n
a
,
l
i
a
F
,
)
/
C
B
A
/
(
s
s
e
c
c
u
S
t
n
e
d
u
t
S
f
o
t
n
e
c
r
e
P
r
o
r
r
E
d
t
S
h
t
i
w
p
u
o
r
G
i
t
n
e
m
n
g
s
s
A
d
n
a
m
r
e
T
y
b
p
o
r
D
75
50
25
0
Fall 2018
Spring 2019
Fall 2019
Outcome
Success
Fail
Late Drop
Treatment
Control
Treatment
Control
Treatment
Control
Fig. 3. Final course grade outcomes broken out by term and curriculum, including success (earned
grades of A, B, or C), fail (earned grades of D or F), and late drops (withdrawals or drops after the
institution’s drop/add deadline). Vertical scale is percentage by outcome. Error bars indicate the 95% CI
for the mean percentage of students in each outcome group over all sections in a term.
Kramer et al., Science 381, 995–998 (2023)
1 September 2023
3 of 4
RES EARCH | R E S E A R C H A R T I C L E
year, these results translate to the potential for
an additional 33,000 students passing calculus
each year, saving students an estimated $23.9
million on tuition [based on a three-credit course
at the average public college/university tuition
rate of $242/credit (41, 42)]. Pragmatic trials
provide guidance on what can be achieved by
engaging faculty willing to change their ins-
truction. These results potentially represent a
lower bound on the long-term effects because
faculty likely develop additional expertise
through continued instruction and realize im-
proved outcomes. The measured effect size pro-
vides the rationale to stop the control due to
treatment benefit if one follows medical re-
search protocols (43, 44).
The experimental methodology described here
establishes a new standard of care for calculus
instruction and a high standard of evidence to
bear on understanding the impacts on student
learning. Improved learning of calculus aims to
foster higher success in future STEM courses
and develop the STEM “habits of mind” that stu-
dents take with them into their future careers.
Further, MPC shows the potential to address
the disparities that differentially affect histori-
cally underrepresented groups, thus offering a
mechanism to address Handelsman et al.’s
(14) call to promote the success of historically
excluded communities. We envision a mathe-
matics experience for all students built on this
approach and recommend that active student
engagement must be deployed across all STEM
disciplines to improve our development of fu-
ture STEM professionals from all backgrounds.
RE FE RENCES AND N OT ES
1. U. Treisman, Coll. Math. J. 23, 362–372 (1992).
2. B. B. Alexander, A. C. Burda, S. B. Millar, J. Women Minor. Sci.
3.
Eng. 3, 145–159 (1997).
J. Ellis, B. K. Fosdick, C. Rasmussen, PLOS ONE 11, e0157447
(2016).
4. S. L. Ganter, D. J. Lewis, D. Hughes-Hallett, Science 279,
2019–2025 (1998).
5. S. Olson, D. G. Riordan, Engage to Excel: Producing One Million
Additional College Graduates with Degrees in Science,
Technology, Engineering, and Mathematics. Report to the
President (Executive Office of the President, 2012).
6. E. Seymour, N. M. Hewitt, Talking About Leaving (Westview
Press, 1997).
7. E. Seymour, A.-B. Hunter, Talking About Leaving Revisited: Persistence,
Relocation, and Loss in Undergraduate STEM Education (Springer,
2019); http://link.springer.com/10.1007/978-3-030-25304-2.
8. A. K. Koch, B. Drake, Digging into the Disciplines: The Impact of
Gateway Courses in Accounting, Calculus, and Chemistry on
Student Success (John N. Gardner Institute for Excellence in
Undergraduate Education, 2018).
9. M. J. Graham, J. Frederick, A. Byars-Winston, A. B. Hunter,
J. Handelsman, Science 341, 1455–1456 (2013).
10. M. Stains et al., Science 359, 1468–1470 (2018).
11. C. Henderson et al., Int. J. STEM Educ. 4, 14 (2017).
12. C. Rasmussen et al., J. Res. Math. Educ. 50, 98–111
(2019).
13. L. A. Steen, in Papers Presented at a Colloquium, Washington,
DC, October 28-29, 1987 (ERIC, 1988); MAA Notes Number 8.
14. J. Handelsman et al., Science 376, 1057–1059 (2022).
15. J. Bookman, C. P. Friedman, in Assessment Practices in
Undergraduate Mathematics, B. Gold, S. Keith, W. A. Marion, Eds.
(Mathematical Association of America, 1999), MAA Notes, p. 253.
16. M. Brown, in Calculus: The Dynamics of Change, W. Roberts,
Ed. (Mathematical Association of America, 1996), vol. 39 of
MAA Notes, pp. 52–58.
17. E. Tidmore, Focus on Calculus 7, 5–6 (1994).
18. M. Zwarenstein et al., BMJ 337, a2390–a2390 (2008).
19. S. Freeman et al., Proc. Natl. Acad. Sci. U.S.A. 111, 8410–8415
(2014).
20. M. Abell, L. Braddy, D. Ensley, L. Ludwig, H. Soto-Johnson, MAA
Instructional Practices Guide (Mathematical Association of
America, 2018).
21. K. Saxe et al., A Common Vision for Undergraduate Mathematical
Sciences Programs in 2025 (Mathematical Association of
America, 2015).
22. V. Otero, S. Pollock, N. Finkelstein, Am. J. Phys. 78, 1218–1224
(2010).
23. G. Ladson-Billings, Theory Pract. 34, 159–165 (1995).
24. L. Benson et al., in Proceedings of the 2009 Annual Conference &
Exposition, American Society of Engineering Education (ASEE,
2009); pp. 14.159.1–14.159.7.
25. A. Kezar, A. C. Chambers, J. C. Burkhardt, Higher Education for
the Public Good: Emerging Voices from a National Movement
(Wiley, 2015).
26. D. Bressoud, Math. Teach. 109, 179–185 (2015).
27. G. Sonnert, P. M. Sadler, S. M. Sadler, D. M. Bressoud, Int. J.
Math. Educ. Sci. Technol. 46, 370–387 (2015).
28. P. Zorn, E. Bressoud, W. David, M. Pearson, “Response to the
PCAST Report to the President, Engage to Excel” (MAA, 2012);
https://www.maa.org/sites/default/files/pdf/sciencepolicy/
MAA-ResponsePCAST-E2E.pdf.
29. D. Burrows, “Common data set 2019-2020” (Florida
International University CDS Archive, 2020); https://opir.fiu.
edu/CDS/CDS2019.pdf.
30. C. J. Torgerson, Education 37, 313–321 (2009).
31. M. Roland, D. J. Torgerson, BMJ 316, 285 (1998).
32. B. Styles, C. Torgerson, Educ. Res. 60, 255–264 (2018).
33. K. F. Schulz, D. G. Altman, D. Moher; CONSORT Group, Trials
11, 32 (2010).
34. D. Moher et al., Int. J. Surg. 10, 28–55 (2012).
35. M. A. Tallman, M. P. Carlson, D. M. Bressoud, M. Pearson,
International Journal of Research in Undergraduate
Mathematics Education 2, 105–133 (2016).
36. J. Cohen, Statistical Power Analysis for the Behavioral Sciences
(Lawrence Erlbaum, ed. 2., 1988).
37. J. C. Valentine, H. Cooper, “Effect size substantive
interpretation guidelines: Issues in the interpretation of effect
sizes” (US Department of Education, 2003); https://wmich.
edu/sites/default/files/attachments/u58/2015/Effect_Size_
Substantive_Interpretation_Guidelines.pdf.
38. L. V. Hedges, in The Handbook Of Research Synthesis and
Meta-Analysis, H. Cooper, L. V. Hedges, J. C. Valentine, Eds.
(Russell sage Foundation, 2009); 399–416.
39. B. C. Jaeger, L. J. Edwards, K. Das, P. K. Sen, J. Appl. Stat. 44,
1086–1105 (2017).
40. M. T. Hora, Proc. Natl. Acad. Sci. U.S.A. 111, E3024
(2014).
41. D. M. Bressoud, M. P. Carlson, V. Mesa, C. Rasmussen, Int. J.
Math. Educ. Sci. Technol. 44, 685–698 (2013).
42. T. Snyder, C. de Brey, S. Dillow, “Digest of education statistics,
2018” (NCES, 2019); https://nces.ed.gov/programs/digest/
d19/tables/dt19_330.10.asp.
43. V. M. Montori et al., JAMA 294, 2203–2209
(2005).
44. S. J. Pocock, JAMA 294, 2228–2230 (2005).
45. Data for: L. Kramer et al., Establishing a new standard of care for
calculus using trials with randomized student allocation, Dryad
(2023); https://doi.org/10.5061/dryad.kkwh70s95.
AC KNOWLED GME NTS
We thank the FIU administration and the State of Florida for initial
support that made this project possible. The work would not be
possible without the faculty, undergraduate learning assistants,
and students participating in the study, who have earned our deep
gratitude. We thank K. Rambo-Hernandez for insightful statistical
analysis discussions, and we are indebted to the reviewers and
editors for their insight and guidance. The research for this study
was conducted under an institutional review board protocol
approved by the Florida International University IRB (approval no.
IRB-18-0211-a.m.02). Students participating in the study consented
to participation at the beginning of each semester. Only students
aged 18 or over were included. Any opinions, findings, and
conclusions or recommendations expressed in this material are
those of the author(s) and do not necessarily reflect the views of
the National Science Foundation. Funding: This work was
supported by the National Science Foundation (grant DUE 1832450
to L.K., E.F., G.P., A.C., and C.W.). Author contributions:
Conceptualization: L.K., E.F., G.P., A.C., C.W.; Funding acquisition:
L.K., E.F., G.P.; Investigation: L.K., E.F., G.P., A.C., C.W., P.D.O.;
Methodology: L.K., E.F., G.P., A.C., C.W., P.D.O.; Project administration:
L.K., E.F.; Writing – original draft: L.K., E.F., G.P., A.C., C.W., P.D.O.;
Writing – review and editing: L.K., E.F., G.P., A.C., C.W., P.D.O.
Competing interests: The authors declare no competing interests.
Data and materials availability: All experimental data are available
at Dryad (45). License information: This work is licensed under a
Creative Commons Attribution 4.0 International (CC BY 4.0) license,
which permits unrestricted use, distribution, and reproduction in any
medium provided the original work is properly cited. To view a copy of
this license, visit https://creativecommons.org/licenses/by/4.0/. This
license does not apply to figures/photos/artwork or other content
included in the article that is credited to a third party; obtain authorization
from the rights holder before using such material.
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade9803
Materials and Methods
Figs. S1 to S10
Tables S1 to S17
References (46–109)
MDAR Reproducibility Checklist
Submitted 20 September 2022; accepted 28 July 2023
10.1126/science.ade9803
Kramer et al., Science 381, 995–998 (2023)
1 September 2023
4 of 4
|
10.1126_science.ade9434
|
Corrected 8 August 2023. See full text.
RES EARCH
ENZYME DESIGN
Combinatorial assembly and design of enzymes
R. Lipsh-Sokolik1, O. Khersonsky1, S. P. Schröder2, C. de Boer2†, S.-Y. Hoch1, G. J. Davies3,
H. S. Overkleeft2, S. J. Fleishman1*
The design of structurally diverse enzymes is constrained by long-range interactions that are
necessary for accurate folding. We introduce an atomistic and machine learning strategy for the
combinatorial assembly and design of enzymes (CADENZ) to design fragments that combine with
one another to generate diverse, low-energy structures with stable catalytic constellations. We
applied CADENZ to endoxylanases and used activity-based protein profiling to recover thousands of
structurally diverse enzymes. Functional designs exhibit high active-site preorganization and more
stable and compact packing outside the active site. Implementing these lessons into CADENZ led
to a 10-fold improved hit rate and more than 10,000 recovered enzymes. This design-test-learn
loop can be applied, in principle, to any modular protein family, yielding huge diversity and general
lessons on protein design principles.
I nnovation in many areas of engineering
relies on the combination of preexisting
modular parts (1). For example, in electri-
cal engineering, standard modular parts,
such as transistors or processing units, are
combined to assemble devices (2). Similarly, in
a hypothetical and entirely modular protein,
fragments could be combined to generate
stable, well-folded, and potentially functional
domains (3). However, in practice, protein
domains exhibit a high density of conserved
molecular interactions that are necessary for
accurate native-state folding. Furthermore,
mutations may be epistatic such that they can
only be incorporated against the background
of other mutations, severely limiting options
for fragment combination (4, 5). Recombina-
tion is an important source of protein diversity
in natural and laboratory evolution (6–8) and
the design of de novo backbones (9); how-
ever, because of epistasis, evolution is typically
restricted to recombining fragments from only
a few high-homology proteins (6).
Despite these challenges, immune system
antibodies present a notable example in which
modularity enables extremely rapid and effec-
tive innovation through the combination of a
small set of genetic fragments [V, (D), and J
genes] (10). The result of this process is an
enormous diversity of binding proteins that
can, in principle, counter any pathogen. Nature
has no equivalent strategy to generate structural
and functional diversity in enzymes, but some
protein folds, such as TIM barrels, b propellers,
and repeat proteins, have evolved through the
duplication, recombination, and mutation of
1Department of Biomolecular Sciences, Weizmann Institute
of Science, 7610001 Rehovot, Israel. 2Leiden Institute of
Chemistry, Leiden University, Einsteinweg 55, 2300 RA
Leiden, Netherlands. 3York Structural Biology Laboratory,
Department of Chemistry, The University of York, Heslington,
York YO10 5DD, UK.
*Corresponding author. Email: sarel@weizmann.ac.il
†Present address: DSM Nutritional Products Ltd, Wurmisweg 576,
4303 Kaiseraugst, Switzerland.
modular fragments and are therefore prom-
inent candidates for fragment combination.
Moreover, these folds constitute some of the
most structurally and functionally versatile
enzymes and binding proteins in nature (11).
In this study, we ask whether enzymes could
be generated from combinable fragments.
We develop a method called CADENZ (combi-
natorial assembly and design of enzymes) to
design and select protein fragments that, when
freely combined, give rise to vast repertoires of
low-energy proteins that exhibit high sequence
and structural diversity. Isolating active en-
zymes in such vast protein libraries requires
high-throughput screening methods (12, 13)
but can be readily and accurately achieved by
using activity-based protein profiling (ABPP).
ABPP uses mechanism-based covalent and ir-
reversible inhibitors composed of a chemical
scaffold that emulates structural features of
the target substrate with an enzyme active
site electrophile and a fluorophore or affinity
tag. To exploit ABPP, we focused on glycoside
hydrolase family 10 (GH10) xylanases (Enzyme
Classification: 3.2.1.8) (14–16) as a model sys-
tem and a dedicated GH10 xylanase–specific
activity-based probe (ABP) as the principal en-
zyme activity readout. We found that CADENZ
generated thousands of functional enzymes
with more than 700 diverse backbones. We
then trained a machine learning model to rank
designs on the basis of their structure and en-
ergy features. Applying the learned model, we
designed a second-generation library that dem-
onstrated an order of magnitude increase in the
success rate of obtaining functional enzymes.
Design of modular and combinable
protein fragments
For a protein fold to be a candidate for mod-
ular assembly and design, its secondary-
structure elements should be conserved among
homologs, but loop regions should exhibit di-
verse conformations, including insertions and
deletions (17–19). In such cases, the secondary-
structure elements typically provide robustness,
whereas the loop regions encode functional
differences. The TIM-barrel fold is a prime ex-
ample of such modularity in which eight b/a
segments comprise an inner b barrel sur-
rounded by a helices (20, 21). The catalytic
pocket is located at the top of the barrel with
critical contributions from all b/a loops. Evo-
lutionary analysis indicates that TIM-barrel
proteins arose through dual duplication of an
ancestral b/a–b/a segment, suggesting that
modern TIM-barrel enzymes can be segmented
into four parts (Fig. 1A) (22–24). Nevertheless,
during evolution, each protein accumulated
mutations to adapt their intersegment inter-
actions for specific functional and stability
requirements. Therefore, recombining frag-
ments from existing proteins mostly produces
unstable and dysfunctional proteins that re-
quire further mutational optimization to become
stable and active (25). To address this problem,
the CADENZ design objective is to compute a
spanning set of backbone fragments that pro-
duce folded and active proteins when freely
combined and without requiring further opti-
mization. The primary challenge CADENZ
addresses is designing mutually compatible
(modular) fragments among which epistasis
is minimal.
The first step of CADENZ is the alignment of
homologous but structurally diverse enzymes
(in the case of this study, 81 structures of GH10
xylanases) and fragmenting them along points
that are structurally highly conserved (within
the core b segments) (see Fig. 1B for a vi-
sual guide to the algorithm) (18, 26). Next, the
fragments are designed to increase stability
while holding the active site fixed. All design
calculations take place within a single arbi-
trarily chosen template [Protein Data Bank
(PDB) entry 3W24 (27)] to provide a realistic
structural context and to promote compati-
bility between fragments. In practice, each
fragment replaces the corresponding one in the
template structure, and we used the PROSS
stability-design algorithm (28) to implement
stabilizing mutations within the fragment (8
to 42 mutations in each fragment; up to 28%)
(fig. S1A). In GH10 xylanases, the active site
interacts with the xylan substrate through
more than a dozen residues from all b/a loops
(29, 30), posing a challenge for modular design
(Fig. 1C). To maintain catalytic activity, in all
design calculations the side chains of four key
catalytic amino acids are restrained to their
crystallographically observed conformations.
At the end of this process, we obtained a set of
fragments that are internally stabilized with-
in a common template and designed to sup-
port the catalytically competent constellation
of active site residues.
However, because of epistasis, combining
the designed fragments would likely result in
mostly high-energy structures that are unlikely
Lipsh-Sokolik et al., Science 379, 195–201 (2023)
13 January 2023
1 of 7
RES EARCH | R E S E A R C H A R T I C L E
Corrected 8 August 2023. See full text.
B
A
C
D
Fig. 1. Key steps in the CADENZ workflow. (A) (Top) Cartoon representation
of selected fragments. (Bottom) Segmentation scheme for GH10 xylanases
(color scheme is consistent in all structural figures). (B) The design pipeline.
Step I: Design maximizes internal stability and compatibility with other fragments
and diversifies active site positions that are not directly involved in the
catalytic step. Step II: DNA oligos encoding fragments are freely ligated with
Golden Gate Assembly (32) to generate DNA molecules encoding the full-length
designs. Step III: Designs are sorted with a xylobiose-emulating activity-based
probe (34) that labels the nucleophilic Glu (red lines) of yeast-displayed
functional enzymes. Activity is confirmed on a subset of the selected enzymes in
a plate-based chromogenic assay. Step IV: An activity predictor is trained on
the basis of features that distinguish presumed active and inactive designs.
(C) Four catalytic residues are restrained throughout design calculations
(in sticks, numbering correspond to PDB entry: 3W24). (D) Fragments can
assemble into low- or high-energy structures depending on other fragments.
(Top left) Segments 3 (blue) and 4 (red) are incompatible (overlap marked by
black circle), resulting in extremely high energy (+1,529 REU). The other designs
exhibit low energies (≤ 950 REU).
to fold into their intended conformation or
support the catalytic constellation (Fig. 1D).
To address this problem, we enumerated all
possible combinations of designed fragments
and ranked them by Rosetta all-atom energy
(Fig. 1B, step I). This process yields hundreds
of thousands of distinct structures, most of
which exhibit unfavorable energies, as expected.
To find mutually compatible (modular) frag-
ments, we present a machine learning–based
approach called EpiNNet (Epistasis Neural
Network), which ranks fragments according
to their probability of forming low-energy
full-length structures (Fig. 2A). EpiNNet is
trained to predict whether a combination of
fragments exhibits favorable Rosetta energy
on the basis of its constituent fragments. The
trained network weights are then used to nom-
inate fragments to generate the enzyme library.
For the next design steps, we used the top six
to seven fragments from each segment, which
assembled into 1764 structures.
To add active site diversity and increase the
chances of favorable fragment combination,
we designed several sequence variants for
each of the backbone fragments. We used the
FuncLib design method (31) to generate low-
energy amino acid constellations at positions
in the active site and in the interfaces between
b/a fragments while fixing the conformations
of the key catalytic residues as observed in
experimentally determined structures (Fig. 1C).
We then used EpiNNet again, this time to find
the single-point mutations that are most likely
to form low-energy full-length proteins in
combination with other mutations (fig. S2A).
The CADENZ strategy does not necessarily
select the lowest-energy fragment combina-
tions, but rather mitigates the risk of com-
bining incompatible ones. The consequences
of intersegment epistasis are notable: Whereas
the energies in the fully enumerated set of
designed GH10s can be as high as +2500
Rosetta energy units (REU), after EpiNNet
fragment selection, the energies are < –890 REU
(Fig. 2B). As a reference, we also generated the
distribution of energies obtained by combin-
ing the sequence of the fragments selected by
EpiNNet before any of the design steps. This
reference simulates the recombination of nat-
ural GH10 genes and exhibits a less favorable
energy distribution than the combination of
PROSS-stabilized fragments (> 100 REU dif-
ference, on average), underscoring the impact
Lipsh-Sokolik et al., Science 379, 195–201 (2023)
13 January 2023
2 of 7
RES EARCH | R E S E A R C H A R T I C L E
Corrected 8 August 2023. See full text.
A
B
Stability &
structural
integrity
11
N
C
D
Selected fragments
Random fragments
Natural sequences
Steric compatibility
Tyr-Leu
7 fragments
90o
EpiNNet rejected
fragment
(steric overlap)
EpiNNet selected fragments are compatible with one another
E
Gly-Gln
Glu
Lys
Leu
Arg
Ala
Steric overlap
Met
Steric overlap
Fig. 2. EpiNNet selects fragments that assemble to low-energy structures.
(A) Schematic representation of the EpiNNet architecture. (B) Most
(89%) of the EpiNNet-selected designs exhibit low energy (< -967 REU,
dashed line) (see materials and methods) relative to proteins generated
by assembling randomly selected or natural fragments. (C) EpiNNet removes
incompatible fragments. (Left) All fragments selected for segment 3
(blue) and 4 (red). (Right) Discarded fragment with a b/a loop that is
incompatible with the other fragments. (D and E) Examples of mutations
selected by EpiNNet (taken from the second-generation library).
(D) Segment 1 from PDB entry 3W24 (27) (yellow) faces segment 2 (green).
EpiNNet prioritizes tyrosine (Tyr) over leucine (Leu) which cannot be
accommodated with neighboring fragments. (E) Segment 3 from PDB entry
1VBR (55) (blue) faces segment 4 (red). EpiNNet prioritizes the small Gly
over the large Gln.
of the design process (Fig. 2B and fig. S1B).
Furthermore, EpiNNet alleviates interseg-
ment epistasis by discarding backbone frag-
ments and designed single-point mutations
that are incompatible with neighboring seg-
ments (Fig. 2, C to E). This analysis highlights
the challenge that epistasis poses for effective
fragment combination while underscoring the
strengths of the EpiNNet selection strategy.
Although EpiNNet eliminates more than 60%
of the fragments, the designed library ex-
hibits high diversity and includes a total of
952,000 sequences adopting 1764 different
backbones.
CADENZ generates thousands of structurally
diverse and active enzymes
We used Golden Gate Assembly to combine
designed fragments into full-length genes
(Fig. 1B, step II) (32) and transformed the
library into yeast cells for functional screening
with cell-surface display (Fig. 1B, step III) (see
materials and methods) (33). To probe en-
zyme activity, we incubated the library with
a xylobiose ABP (34), which reacts within the
enzyme active site to form a covalent and ir-
reversible ester linkage with the glutamic acid
A
B
C
Fig. 3. CADENZ generates functional enzymes with high structure and sequence diversity.
(A) Representative model structures of recovered enzymes designed by CADENZ. Regions that vary
among the four designs are highlighted in colors (top). Active-site electrostatic potential surfaces of
the representative designs exhibit marked differences (putative ligand-bound conformation marked
in yellow sticks on the basis of PDB entry 4PUD) (bottom). (B) Distribution of sequence identity
to nearest natural homologs of recovered designs. (C) The number of distinct structures from which
fragments are sourced. Most recovered designs incorporate fragments from four different sources.
Lipsh-Sokolik et al., Science 379, 195–201 (2023)
13 January 2023
3 of 7
RES EARCH | R E S E A R C H A R T I C L E
nucleophile (35). We then used fluorescence-
activated cell sorting (FACS) to collect the
population of yeast cells expressing active
designs (fig. S3A). ABP labeling depends on
the nucleophilicity of the catalytic glutamic
acid (Glu), the ability of the active site catalytic
acid-base residue to enhance the electrophilic-
ity of the ABP epoxide by protonation, and the
integrity of the xylan molecular recognition
elements within the active-site pocket. There-
fore, ABP labeling acts as a sensitive probe for
design accuracy in the active site (which com-
prises elements from all b/a units). Retaining
glycosidase ABPs report on the first steps of
substrate processing, namely ligand binding
to the active site and subsequent nucleophilic
attack. To confirm that selected proteins ex-
hibit the complete catalytic cycle (36), we trans-
formed Escherichia coli cells with DNA from
the sorted population and randomly selected
186 colonies for screening in 96-well plates
with the chromogenic substrate 4-nitrophenyl
b-xylobioside (O-PNPX2) (37). Of the selected
enzymes, 58% processed the substrate (fig.
S3B), indicating that most designs selected
by the ABP exhibited catalytic activity for
this reaction.
We next applied single-molecule real-time
(SMRT) long-read sequencing (38) to the sorted
population. Encouragingly, sequencing showed
that the sorted population included a large
number of structurally diverse designs: specif-
ically, 3114 distinct designs based on 756 dif-
ferent backbones (Fig. 3A), compared with
only 376 GH10 xylanase entries in the UniProt
database (39). The recovered designs exhibited
many insertions and deletions relative to one
another, with sequence lengths varying from
317 to 395 amino acids and 62% sequence
identity to one another on average. In all mod-
els, residues responsible for the catalytic steps
are held in place by construction, but the active-
site pocket exhibits high geometric and elec-
trostatic differences (Fig. 3A, bottom) because
of loop conformation diversity. The designs
exhibit as many as 169 mutations and 48 to
73% sequence identity (Fig. 3B) to their nearest
natural homolog [in the nonredundant (nr)
sequence database (40)], and most designs
source fragments from four different structures
(Fig. 3C).
Recovered designs are compact and
preorganized for activity
The deep sequencing analysis provides a val-
uable dataset for improving enzyme design
methodology. For each design, we computed
85 structure and energy metrics, some relating
to the entire protein, and others restricted to
the active site. We avoided using the designed
mutations or fragment identities as features
for learning so that we might infer general
lessons that apply to other enzymes. We tested
the differences between the presumed active
Corrected 8 August 2023. See full text.
Fig. 4. Energy and structure features discriminate between active and inactive designs.
(A) Representative features included in the activity predictor. The features show a statistically significant
difference between presumed active and inactive designs with an independent two-sample t test, but no
feature is individually an effective discriminator. All features are normalized by protein length. Low values are
favorable. (B) Separation of designs recovered by ABPP versus other designs on the basis of a logistic
regression model. In green, probability distribution for designs assembled by the fragments selected for
the second-generation library. (C and D) Examples of backbone fragments eliminated by the activity
predictor in the second-generation library. Fragment color scheme as in Fig. 1. (C) 2WYS in segment 2
(green) was selected for the first-generation library but discarded in the second because of low atomic
density. The interface with segment 3 (blue) is poorly packed, leaving a gap between the segments (white).
(D) 1UQZ in segment 4 (red) was selected for the first-generation library but discarded in the second
because of unfavorable hydrogen bond energy. Close inspection revealed two mutations introduced during
sequence design, Arg282→Asn and Arg289→Leu [residue numbering refers to PDB entry: 1UQZ (56)],
eliminating hydrogen bonds that are crucial for b/a loop backbone stabilization. Mutations in red.
and inactive sets using an independent two-
sample t test, finding that 63 metrics exhibited
P values less than 10−10. To select the most
meaningful metrics, we visually inspected their
distributions and focused on 10 (Fig. 4A and
fig. S4) that were not significantly correlated
(see materials and methods). We then trained
a logistic regression model based on these
10 metrics to predict whether an enzyme is
active (Fig. 4B and tables S1 to S3).
The 10 dominant predictive metrics relate
to essential aspects of enzyme catalysis. The
most dominant feature is atomic density,
which gauges protein compactness and corre-
lates with stable packing. Another dominant
feature is the compatibility of the amino acid
identity and the local backbone conformation,
a key determinant of protein foldability (41).
By contrast, this feature is disfavored within
the active-site pocket, presumably because
active-site residues are selected for their im-
pact on activity rather than stability. We also
find that hydrogen-bond energies are highly
discriminating, reflecting the prevalence of
Lipsh-Sokolik et al., Science 379, 195–201 (2023)
13 January 2023
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
Corrected 8 August 2023. See full text.
Fig. 5. Activity predictor significantly increases design success rate.
(A) Backbone fragments selected for the second-generation library (colors as
in Fig. 1, low diversity regions in white). (B) Number of sequences in the
population selected by the xylobiose and cellobiose ABPs (blue and purple,
respectively). In the overlap region, designs selected by both ABPs. (C) Protein
sequence length of recovered designs in the second-generation library.
(D) Distribution of sequence identity to nearest natural homolog of recovered
designs in the second-generation library. (E) Number of unaligned regions to
nearest natural homolog of recovered designs in the second-generation library.
(F) Designs recovered in the second-generation library incorporate more
active site mutations than in the first-generation library. (G) Catalytic efficiency
of seven xylanases from the small-set design and two representative natural
ones [right, names correspond to PDB entry. 3W24 (27) is a xylanase from
the thermophilic organism Thermoanaerobacterium saccharolyticum]. The first
number in the design name indicates the number of different proteins from
which the fragments were sourced. (H) Normalized activity with wheat
arabinoxylan and beechwood xylan. Data are the means ± standard deviation
of duplicate measurements.
buried long-range hydrogen bond networks
in large proteins of a complex fold such as TIM
barrels (fig. S5) (42, 43). However, Rosetta sys-
tem energy makes a small contribution to
predicting activity, presumably because all
designs exhibit low energy by construction
(Fig. 2B).
Within the active site, the model assigns al-
most equal importance to atomic density and
van der Waals energy, two features that pro-
mote precise catalytic residue placement but
penalize overly packed constellations, respec-
tively. The resulting dense yet relaxed packing
arrangements are likely to be key in promot-
ing active site preorganization. Focusing on
the four catalytic residues only, the model in-
cludes a feature that penalizes high repulsive
energy, further emphasizing the importance
of a relaxed and preorganized active site. Our
analysis highlights prerequisites of catalytic
activity that were not observed in previous
high-throughput studies of design methods,
which focused on the kinetic stability of de-
signed miniproteins and binders (44, 45). Ad-
ditionally, the design objective function is
substantially different within the active site
versus the remainder of the protein.
Recently, the AlphaFold2 ab initio structure
prediction method (46) has been shown to
discriminate correctly from incorrectly folded
de novo–designed binders (47). However, when
AlphaFold2 was applied to our set, no discern-
able difference was found between presumed
active and inactive designs in either the root-
mean-square deviation (RMSD) between pre-
dicted and designed models, or in the Alpha-
Fold2 confidence scores (pLDDT) (fig. S6). This
result suggests that despite the high mutational
load and the sequence and structure diversity
in the designs, CADENZ generates sequences
with native-like characteristics.
Order-of-magnitude increase in design
success in second-generation library
We next asked whether the lessons we learned
from the first-generation library could improve
design success rate. We used the same set of
combinatorial designs from the first library,
but instead of ranking them on the basis of
Rosetta energies, we ranked them according
to the activity predictor (Fig. 4B and fig. S2B).
We then applied EpiNNet to nominate frag-
ments that are likely to be mutually com-
patible. As in the first library, we designed
several sequence variants for each backbone
fragment, holding sidechain conformations
of the core catalytic residues fixed in all de-
sign calculations. This second-generation li-
brary included three backbone fragments for
each of the four segments and up to 11 se-
quence variants per fragment (for a total of
100 designed fragments), resulting in 334,125
designed full-length xylanases with 81 differ-
ent backbones. To gain insight into the molec-
ular features that are disfavored by the activity
predictor, we analyzed which backbone frag-
ments were chosen in the first library but dis-
carded in the second. We found, for example,
that atomic density (Fig. 4C) and hydrogen-
bond energy (Fig. 4D) were unfavorable in
many discarded fragments.
We synthesized and screened the second-
generation library as we did with the first-
generation library (fig. S7). Notably, sequencing
confirmed 9859 active designs, an order-of-
magnitude increase in the rate of positive hits
compared with that of the first library (Fig. 5A).
In addition to the xylobiose screen, we also
screened the library with an ABP that is based
on cellobiose (Fig. 5B), the disaccharide repeat
moiety in cellulose rather than in xylan (48).
We found 2778 designs that reacted with the
cellobiose ABP but were not sequenced in
the library sorted with the xylan ABP, for a
total of more than 12,637 active designs (3.8%
of the design population). To verify that the
ABP-selected designs exhibited the full cat-
alytic cycle, we used plate-based validation
with O-PNPX2 and cellPNP (to detect cellulase
activity) confirming 85 and 60% of active clones
in the xylobiose- and cellobiose-labeled popu-
lations, respectively (fig. S7, C and D).
Ranking designs on the basis of the activity
predictor resulted in a more focused library
that included 79 of the 81 designed backbones
and sequence lengths ranging from 312 to
347 amino acids (Fig. 5C). Although the activity
predictor was blind to the identities of the
designed fragments and mutations, we were
Lipsh-Sokolik et al., Science 379, 195–201 (2023)
13 January 2023
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
concerned that it might have focused the sec-
ond library on a set of fragments identified in
functional enzymes in the first library. We ana-
lyzed the source of the active designs in the
second-generation library, finding that 75% of
these designs incorporated backbone fragments
that were not encoded in the first library, and
verifying that the learned energy and structure
features generalized to fragments not included
in the training data. Moreover, the active designs
are as divergent from natural GH10 enzymes
as they were in the first library, exhibiting 50
to 73% sequence identity to the most similar
sequences in the nr database (40) (Fig. 5D)
as well as up to 140 mutations and eight un-
aligned regions (Fig. 5E). Furthermore, the
second-generation library incorporates more
active site mutations (Fig. 5F), increasing the
potential for altered substrate specificities. We
also analyzed the distribution of energy and
structure metrics among active and inactive
designs in the second-generation library. The
discrimination that we observed, however,
was lower than in the first-generation library,
suggesting that the specific learning process
we implemented converged.
As an independent test, we applied the
learned activity predictor to select a small set
of individually designed GH10 enzymes (18).
On the basis of the AbDesign strategy (26),
we used Rosetta atomistic modeling to enu-
merate all fragment combinations and design
their sequences as full-length enzymes, fol-
lowed by selection with the activity predictor.
This strategy encodes more stabilizing inter-
segment interactions than when the fragments
are designed independently, and the designs
are therefore more likely to be stable and fold-
able. Thus, in this implementation, the design
and selection process does not favor modular-
ity but rather optimal structure and energy
properties. The activity predictor selected 27
designs for experimental characterization with
up to 143 mutations and 51 to 74% sequence
identity to their nearest natural homologs. Al-
though these designs were generated by a dif-
ferent process than the one used to train the
activity predictor, notably, 25 (93%) of the de-
signs were active in hydrolyzing O-PNPX2
(table S4), compared with less than 50% in a
previous application of AbDesign to GH10
enzymes (18). We further characterized the
kinetics of the seven most promising designs
with various substrates (table S5). Among these
designs, several exhibited catalytic efficiencies
(kcat/KM) comparable to those of natural GH10
xylanases from thermophiles, including against
natural wood and wheat xylan (Fig. 5, G and
H), despite incorporating over 80 mutations
from any known natural protein sequence.
These results are a marked improvement in
the success of backbone design in enzymes and
underscore that the lessons we learned from
high-throughput screening can be applied
Corrected 8 August 2023. See full text.
to generate a diverse and highly active set of
designs, for either high- or low-throughput
screening.
Discussion
Modularity is a prerequisite for innovation in
numerous engineering disciplines, but protein
domains exhibit high epistasis, severely ham-
pering the ability to combine fragments into
stable and active structures. CADENZ ad-
dresses this conflict by designing a spanning
set of low-energy and mutually compatible
protein fragments that can be assembled into
thousands of diverse and functional proteins.
We have also begun to investigate how EpiNNet
can be implemented to design large repertoires
of active site sequence variants (49). Our ap-
proach increases the number and diversity of
functional enzymes that can be interrogated
relative to the natural diversity, providing
an alternative to metagenomic libraries (12).
Current methods for optimizing and diversify-
ing proteins rely on sequence statistics (50, 51)
or cycles of mutation, recombination, and
screening (52). Because of high epistasis, these
methods explore a small fraction of sequence
and structure space, whereas we show in this
system that CADENZ can generate 106 struc-
turally diverse designs of which >10,000 are
recovered on the basis of activity.
Our results also show that ABPP is an ef-
fective strategy for high-throughput isolation
of successful CADENZ designs and could be
extended to other substrates (53), either natu-
ral or engineered. The combined strategy of
CADENZ and ABPP enabled us to implement
effective design-test-learn cycles on many en-
zyme designs that have previously led to
deeper understanding of the design principles
for de novo–designed miniproteins (44, 45).
The rules we learned increased the design suc-
cess rate by an order of magnitude and were
directly transferable to automated small-scale
design. Such functional data from many homol-
ogous yet structurally diverse enzymes may
guide future improvements in macromolecular
energy functions and advance efforts to de-
velop AI-based enzyme design methods.
RE FERENCES AND NOTES
1. M. E. Csete, J. C. Doyle, Science 295, 1664–1669 (2002).
2. Y. Lazebnik, Cancer Cell 2, 179–182 (2002).
3. M. Parter, N. Kashtan, U. Alon, PLOS Comput. Biol. 4,
e1000206 (2008).
4. D. M. Weinreich, R. A. Watson, L. Chao, Evolution 59, 1165–1174
(2005).
5. D. A. Kondrashov, F. A. Kondrashov, Trends Genet. 31, 24–33
(2015).
6. C. A. Voigt, C. Martinez, Z.-G. Wang, S. L. Mayo, F. H. Arnold,
Nat. Struct. Biol. 9, 553–558 (2002).
7. S. Nepomnyachiy, N. Ben-Tal, R. Kolodny, Proc. Natl. Acad.
Sci. U.S.A. 111, 11691–11696 (2014).
8. N. Ferruz et al., J. Mol. Biol. 432, 3898–3914 (2020).
9. T. M. Jacobs et al., Science 352, 687–690 (2016).
10. K. Murphy, C. Weaver, Janeway’s Immunobiology (Garland
Science, 2016).
11. E. Dellus-Gur, A. Toth-Petroczy, M. Elias, D. S. Tawfik, J. Mol.
Biol. 425, 2609–2621 (2013).
12. P.-Y. Colin et al., Nat. Commun. 6, 10008 (2015).
13. P. Kast, M. Asif-Ullah, N. Jiang, D. Hilvert, Proc. Natl. Acad.
Sci. U.S.A. 93, 5043–5048 (1996).
14. S. G. Withers et al., Biochem. Biophys. Res. Commun. 139,
487–494 (1986).
15. E. Drula et al., Nucleic Acids Res. 50 (D1), D571–D577
(2022).
16. CAZypedia Consortium, Glycobiology 28, 3–8 (2018).
17. A. Tóth-Petróczy, D. S. Tawfik, Curr. Opin. Struct. Biol. 26,
131–138 (2014).
18. G. Lapidoth et al., Nat. Commun. 9, 2780 (2018).
19. R. Netzer et al., Nat. Commun. 9, 5286 (2018).
20. N. Nagano, C. A. Orengo, J. M. Thornton, J. Mol. Biol. 321,
741–765 (2002).
21. R. Sterner, B. Höcker, Chem. Rev. 105, 4038–4055
(2005).
22. B. Höcker, J. Claren, R. Sterner, Proc. Natl. Acad. Sci. U.S.A.
101, 16448–16453 (2004).
23. B. Höcker, S. Beismann-Driemeyer, S. Hettwer, A. Lustig,
R. Sterner, Nat. Struct. Biol. 8, 32–36 (2001).
24. M. Richter et al., J. Mol. Biol. 398, 763–773 (2010).
25. D. L. Trudeau, M. A. Smith, F. H. Arnold, Curr. Opin. Chem. Biol.
17, 902–909 (2013).
26. R. Lipsh-Sokolik, D. Listov, S. J. Fleishman, Protein Sci. 30,
151–159 (2021).
27. X. Han et al., Proteins 81, 1256–1265 (2013).
28. A. Goldenzweig et al., Mol. Cell 63, 337–346 (2016).
29. S. R. Andrews et al., J. Biol. Chem. 275, 23027–23033
(2000).
30. B. Henrissat et al., Proc. Natl. Acad. Sci. U.S.A. 92, 7090–7094
(1995).
31. O. Khersonsky et al., Mol. Cell 72, 178–186.e5
(2018).
32. C. Engler, R. Kandzia, S. Marillonnet, PLOS ONE 3, e3647
(2008).
33. G. Chao et al., Nat. Protoc. 1, 755–768 (2006).
34. S. P. Schröder et al., ACS Cent. Sci. 5, 1067–1078
(2019).
35. D. Tull et al., J. Biol. Chem. 266, 15621–15625 (1991).
36. D. Tull, S. G. Withers, Biochemistry 33, 6363–6370
(1994).
37. H. Taguchi, T. Hamasaki, T. Akamatsu, H. Okada, Biosci.
Biotechnol. Biochem. 60, 983–985 (1996).
38. A. Rhoads, K. F. Au, Genomics Proteomics Bioinformatics 13,
278–289 (2015).
39. A. Bateman et al., Nucleic Acids Res. 49, D480–D489
(2021).
40. E. W. Sayers et al., Nucleic Acids Res. 50 (D1), D20–D26
(2022).
41. A. Goldenzweig, S. J. Fleishman, Annu. Rev. Biochem. 87,
105–129 (2018).
42. O. Khersonsky, S. J. Fleishman, BioDesign Research 2022, 1–11
(2022).
43. D. Baran et al., Proc. Natl. Acad. Sci. U.S.A. 114, 10900–10905
(2017).
44. G. J. Rocklin et al., Science 357, 168–175 (2017).
45. L. Cao et al., Nature 605, 551–560 (2022).
46. J. Jumper et al., Nature 596, 583–589 (2021).
47. N. Bennett et al., bioRxiv (2022), p. 2022.06.15.495993.
48. C. de Boer et al., RSC Chem. Biol. 1, 148–155 (2020).
49. J. Y. Weinstein et al., bioRxiv 2022.10.11.511732 [Preprint]
(2022). https://doi.org/10.1101/2022.10.11.511732.
50. W. P. Russ et al., Science 369, 440–445 (2020).
51. N. Ferruz, S. Schmidt, B. Höcker, Nat. Commun. 13, 4348
(2022).
52. S. L. Lovelock et al., Nature 606, 49–58 (2022).
53. L. Wu et al., Curr. Opin. Chem. Biol. 53, 25–36 (2019).
54. R. Lipsh-Sokolik et al., Data for: Combinatorial assembly and
design of enzymes (2022); http://dx.doi.org/10.5281/zenodo.
7382421.
55. T. Ihsanawati et al., Proteins 61, 999–1009 (2005).
AC KNOWL ED GME NTS
We thank N. London, B. Höcker, S. Barber-Zucker, and D. Listov
for discussions and S. Warzsawski and K. Goldin for technical
help. R.L.-S. is supported by a fellowship from the Arianne
de Rothschild Women Doctoral Program. Funding: This work was
funded by the Volkswagen Foundation grant 94747 (S.J.F.), the
Israel Science Foundation grant 1844 (S.J.F.), the European
Research Council through a Consolidator Award grant 815379
(S.J.F.), the Dr. Barry Sherman Institute for Medicinal Chemistry
(S.J.F.), a donation in memory of Sam Switzer (S.J.F.), the
Royal Society for the Ken Murray Research Professorship
Lipsh-Sokolik et al., Science 379, 195–201 (2023)
13 January 2023
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
Corrected 8 August 2023. See full text.
(G.J.D.), the European Research Council grant ERC-2011-AdG-
290836 ‘Chembiosphing’ (H.S.O.), ERC-2020-SyG-951231
‘Carbocentre’ (H.S.O. and G.J.D.), and the Netherlands Organization
for Scientific Research through the NWO TOP grant 2018-
714.018.002 “Endoglycoprobe” (H.S.O.). Author contributions:
Conceptualization: R.L.S., O.K., S.J.F.; Methodology: R.L.S., O.K.,
S.J.F.; Software: R.L.S.; Validation: R.L.S., O.K.; Formal analysis:
R.L.S., S.Y.H.; Investigation: R.L.S., O.K., S.P.S., C.d.B.; Resources:
R.L.S., O.K., S.P.S., C.d.B., H.S.O., S.J.F.; Data Curation: R.L.S.;
Writing: R.L.S., S.J.F., H.S.O., G.J.D.; Visualization: R.L.S., O.K.;
Supervision: S.J.F., H.S.O.; Project administration: S.J.F.; Funding
acquisition: S.J.F., H.S.O., G.J.D. Competing interests: The
authors declare that they have no competing interests. Data and
materials availability: Custom Python scripts, RosettaScripts,
commandlines, Jupyter notebooks, and datasets are available
at Zenodo (54). License information: Copyright © 2023 the
authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.science.org/about/science-
licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade9434
Materials and Methods
Figs. S1 to S7
Tables S1 to S5
References (56–70)
View/request a protocol for this paper from Bio-protocol.
Submitted 19 September 2022; accepted 12 December 2022
10.1126/science.ade9434
Lipsh-Sokolik et al., Science 379, 195–201 (2023)
13 January 2023
7 of 7
|
10.1126_science.adf0435
|
RES EARCH
NEUROSCIENCE
Psychedelics promote neuroplasticity through
the activation of intracellular 5-HT2A receptors
Maxemiliano V. Vargas1,2, Lee E. Dunlap2,3†, Chunyang Dong4†, Samuel J. Carter2,3,
Robert J. Tombari2,3, Shekib A. Jami5, Lindsay P. Cameron1, Seona D. Patel3, Joseph J. Hennessey6,
Hannah N. Saeger2,7, John D. McCorvy6, John A. Gray2,5,8, Lin Tian2,5,9, David E. Olson2,3,5,9*
Decreased dendritic spine density in the cortex is a hallmark of several neuropsychiatric diseases,
and the ability to promote cortical neuron growth has been hypothesized to underlie the rapid
and sustained therapeutic effects of psychedelics. Activation of 5-hydroxytryptamine (serotonin) 2A
receptors (5-HT2ARs) is essential for psychedelic-induced cortical plasticity, but it is currently unclear
why some 5-HT2AR agonists promote neuroplasticity, whereas others do not. We used molecular and
genetic tools to demonstrate that intracellular 5-HT2ARs mediate the plasticity-promoting properties of
psychedelics; these results explain why serotonin does not engage similar plasticity mechanisms. This
work emphasizes the role of location bias in 5-HT2AR signaling, identifies intracellular 5-HT2ARs as a
therapeutic target, and raises the intriguing possibility that serotonin might not be the endogenous
ligand for intracellular 5-HT2ARs in the cortex.
D ysregulation of the cortex has been hy-
pothesized to play an important role in
the pathophysiology of mental illnesses
such as depression and often manifests
as structural changes, including decreased
dendritic arbor complexity and reduced den-
dritic spine density (1–3). Traditional antide-
pressants, such as selective serotonin reuptake
inhibitors (SSRIs), can rescue these deficits
after chronic treatment, although it seems that
their effects may be independent of serotonin
and perhaps involve the activation of tropomyosin
receptor kinase B (TrkB) signaling (4, 5). A class of
therapeutic compounds known as psychoplasto-
gens (6) are differentiated from SSRIs by their
ability to produce both rapid and sustained ef-
fects on structural plasticity and behavior after a
single administration (7). Psychoplastogens in-
clude both ketamine and serotonergic psychedel-
ics, although their primary targets are distinct (7).
Psychedelics are 5-hydroxytryptamine (sero-
tonin) 2A receptor (5-HT2AR) agonists that can
lead to profound changes in perception, cogni-
tion, and mood (8). Recent evidence suggests
that they promote cortical structural and func-
tional neuroplasticity through activation of
1Neuroscience Graduate Program, University of California,
Davis, Davis, CA 95618, USA. 2Institute for Psychedelics and
Neurotherapeutics, University of California, Davis, Davis, CA
95618, USA. 3Department of Chemistry, University of
California, Davis, Davis, CA 95616, USA. 4Biochemistry,
Molecular, Cellular, and Developmental Biology Graduate
Program, University of California, Davis, Davis, CA 95616,
USA. 5Center for Neuroscience, University of California,
Davis, Davis, CA 95618, USA. 6Department of Cell Biology,
Neurobiology, and Anatomy, Medical College of Wisconsin,
Milwaukee, WI 53226, USA. 7Pharmacology and Toxicology
Graduate Program, University of California, Davis, Davis,
CA 95616, USA. 8Department of Neurology, School of
Medicine, University of California, Davis, Sacramento, CA
95817, USA. 9Department of Biochemistry and Molecular
Medicine, School of Medicine, University of California, Davis,
Sacramento, CA 95817, USA.
*Corresponding author. Email: deolson@ucdavis.edu
†These authors contributed equally to this work.
5-HT2ARs (9, 10). The mechanism by which
5-HT2AR activation leads to changes in neu-
ronal growth is still poorly defined, although it
appears to involve TrkB, mechanistic target of
rapamycin (mTOR), and AMPA receptor signal-
ing (11). It is currently unclear why some 5-HT2AR
ligands can promote neuroplasticity and pro-
duce sustained therapeutic behavioral responses
in the absence of hallucinogenic effects, where-
as other 5-HT2AR agonists do not promote
plasticity at all (12–15). Indeed, serotonin itself
does not produce psychedelic-like effects on
neuronal growth when administered to corti-
cal cultures (9). This enigmatic finding can-
not be easily explained by traditional biased
agonism because serotonin is a balanced agonist
of the 5-HT2AR that exhibits high potency and
efficacy for activating both heteromeric guanine
nucleotide–binding protein (G protein) and
b-arrestin pathways (16, 17).
Unlike psychedelics, the physicochemical
properties of serotonin prevent it from enter-
ing cells by passively diffusing across nonpolar
membranes (8). Thus, we reasoned that an-
other form of functional selectivity, known
as location bias, might explain the difference
in cellular signaling elicited by serotonin and
psychedelics (18, 19). Here, we leveraged both
chemical design and genetic manipulation
to test the hypothesis that activation of an
intracellular population of 5-HT2ARs is nec-
essary for 5-HT2AR ligands to induce cortical
structural plasticity and produce antidepressant-
like behavioral responses.
Lipophilicity correlates with psychoplastogenicity
To firmly establish the role of 5-HT2AR activ-
ation in psychedelic-induced spinogenesis, we
administered5-methoxy-N,N-dimethyltryptamine
(5-MeO) to wild-type (WT) and 5-HT2AR knock-
out (KO) mice (20) and assessed structural and
functional changes in layer 5 pyramidal neurons
of the prefrontal cortex (PFC) 24 hours later
(fig. S1, A to C). Golgi-Cox staining revealed
that 5-MeO increased spine density in both
male and female animals, and this effect was
absent in 5-HT2AR KO mice (fig. S1, A and B).
Furthermore, ex vivo electrophysiology con-
firmed that 5-HT2AR activation is necessary
for 5-MeO to produce sustained increases in
both the frequency and amplitude of sponta-
neous excitatory postsynaptic currents (sEPSCs)
(fig. S1C).
Next, we determined how the structures of
5-HT2AR ligands affect their abilities to pro-
mote neuronal growth by treating embryonic
rat cortical neurons with serotonin, tryptamine
(TRY), 5-methoxytryptamine (5-MeO–TRY), and
their corresponding N-methyl and N,N-dimethyl
congeners (Fig. 1A) before assessing neuronal
morphology by means of Sholl analysis (21).
Ketamine was used as a positive control be-
cause of its known ability to induce structural
plasticity in this assay (9). These structure-
activity relationship (SAR) studies revealed
that increasing N-methylation led to an en-
hanced ability to promote neuronal growth,
with the N,N-dimethyl compounds increasing
dendritic arbor complexity to the greatest ex-
tent (Fig. 1, B to D).
Increasing N-methylation is known to affect
the efficacy of 5-HT2AR signaling, so we at-
tempted to correlate psychoplastogenic effi-
cacy across a range of 5-HT2AR ligands with
efficacy in a traditional [3H]–inositol phosphates
(IP) accumulation assay (Fig. 1E) (22). Notably,
we did not observe a positive correlation be-
tween psychoplastogenic effects and ligand
efficacy. Indeed, there seemed to be a non-
significant inverse correlation between [3H]-IP
accumulation and dendritogenesis efficacy
(Fig. 1E). To avoid potential issues associated
with the amplification of secondary messengers,
we used psychLight2, a fluorescent biosensor
that is capable of directly detecting changes in
5-HT2AR conformation (14). PsychLight2 effi-
cacy closely mirrored that observed by using
[3H]-IP accumulation assays, although psy-
chLight2 efficacy exhibited an even stronger
anticorrelation with dendritogenesis efficacy
(P = 0.06) (Fig. 1F). Lastly, we used biolumines-
cence resonance energy transfer (BRET) assays
to directly measure Gq activation or b-arrestin-2
recruitment (fig. S2) (23). Both measures of
5-HT2AR efficacy exhibited a strong negative
correlation with psychoplastogenicity (Fig. 1,
G and H). Moreover, both Gq activation and
b-arrestin recruitment correlated well with
psychLight efficacy [coefficient of determina-
tion (R2) = 0.9; P < 0.0001 and P = 0.0006, re-
spectively] (Fig. 1I). Negative correlation between
5-HT2AR efficacy and psychoplastogenicity
should be interpreted with caution because
this relationship may only apply to tryptamine-
based ligands or compounds that exhibit a
threshold level of 5-HT2AR activation.
Vargas et al., Science 379, 700–706 (2023)
17 February 2023
1 of 7
RES EARCH | R E S E A R C H A R T I C L E
Nitrogen methylation of 5-HT2AR ligands
structurally related to serotonin is known to
result in partial agonism (22), but it also has
a profound effect on their physicochemical
properties. These compounds display a wide
range of lipophilicities that ranged from
highly polar molecules such as serotonin to
relatively nonpolar compounds such as N,N-
dimethyltryptamine (DMT). Using calculated
LogP (cLogP) values, we observed a significant
positive correlation with psychoplastogenic
effects—more lipophilic agonists exhibited
greater abilities to promote structural plas-
ticity than polar compounds (Fig. 1J). This
relationship was evident within the TRY,
serotonin, and 5-MeO–TRY scaffolds. The find-
ing that lipophilicity was a better predictor
of psychoplastogenicity than 5-HT2AR activa-
tion led us to hypothesize that an intracellular
pool of 5-HT2ARs in cortical neurons might
be responsible for psychedelic-induced neu-
ronal growth.
Primary localization of 5-HT2ARs in cortical
neurons is intracellular
Although most G protein–coupled receptors
(GPCRs) are believed to be localized primarily to
the plasma membrane, several exhibit substan-
tial intracellular localization (24–27). In vitro
and ex vivo experiments have also established
the existence of large intracellular pools of 5-
HT2ARs in various cell types in the absence of
a ligand (28–31). To compare 5-HT2AR local-
ization patterns between cell types, we expressed
a Myc–5-HT2AR–enhanced cyan fluorescent
protein (ECFP) construct in both human em-
bryonic kidney 293T (HEK293T) cells and cor-
tical neurons, performed live-cell imaging, and
assessed colocalization with a membrane dye
(Cellbrite Steady) that labels the extracellular
side of the plasma membrane. Because over-
expression of tagged receptor constructs might
alter trafficking and localization, we included
several controls. We used a b2 adrenergic re-
ceptor tagged with ECFP (b2AR-ECFP) as a
GPCR that is canonically described as being
plasma membrane bound, and we used an
ECFP construct to establish the localization of
a fluorescent protein that is not tagged to a
GPCR (32, 33).
When expressed in HEK293T cells that were
cultured in the absence of serum, 5-HT2ARs
and b2ARs exhibited similar cellular expres-
sion patterns (Fig. 2A) and possessed correla-
tion coefficients with the plasma membrane
marker that were not statistically different
(Fig. 2B). However, these two GPCRs displayed
distinct localization patterns in neurons. In
cortical neurons, b2AR expression was more
highly correlated with the plasma membrane
marker than was 5-HT2AR expression (Fig.
2B), which demonstrates that overexpression
of a GPCR in cortical neurons does not nec-
essarily lead to intracellular localization. The
extent of overexpression was similar for 5-
HT2ARs and b2ARs in both HEK293T cells
and cortical neurons (fig. S3A). We performed
these localization experiments without serum
in the culturing media because serum contains
serotonin, which can lead to agonist-induced
changes in trafficking and localization (fig. S3,
B and C). In both neurons and HEK293T cells,
C
Nmax
A
R
NH2
R
N H
Me
R
N
H
N
H
N Me
Me
N
H
R = H, TRY
R = OH, 5-HT
R = OMe, 5-MeO-TRY
R = H, NMT
R = OH, N-Me-5-HT
R = OMe, 5-MeO-NMT
R = H, DMT
R = OH, BUF
R = OMe, 5-MeO-DMT
B
i
s
g
n
s
s
o
r
C
f
O
r
e
b
m
u
N
8
6
4
2
0
Increasing N-methylation
DMT
VEH
TRY
VEH
i
s
g
n
s
s
o
r
C
f
O
r
e
b
m
u
N
8
6
4
2
0
NMT
VEH
i
s
g
n
s
s
o
r
C
f
O
r
e
b
m
u
N
8
6
4
2
0
VEH
KET
TRY
NMT
DMT
5-HT
N-Me-5-HT
BUF
25 50 75 100 125
Distance from Center (μM)
25 50 75 100 125
Distance from Center (μM)
25 50 75 100 125
Distance from Center (μM)
5-MeO-TRY
TRY
NMT
DMT
5-MeO-NMT
****
**
****
*
*
***
DMT
BUF
5-MeO-DMT
NMT
5-MeO-NMT
N-Me-5-HT
TRY
5-MeO-TRY
p = 0.2
R2 = 0.3
F
100
DMT
5-HT
y
c
a
c
i
f
f
E
%
80
60
40
20
0
BUF
5-MeO-DMT
NMT
5-MeO-NMT
5-HT
N-Me-5-HT
5-MeO-TRY
TRY
p = 0.06
R2 = 0.4
0
20
40
60
80
100
0
[3H]-IP accumulation Emax
5
10
psychLight 2 F/F
15
20
DMT
NMT
BUF
5-MeO-DMT
5-MeO-NMT
5-HT
5-MeO-TRY
TRY
N-Me-5-HT
p = 0.02
R2 = 0.6
H
100
y
c
a
c
i
f
f
E
%
80
60
40
20
0
95
90
85
Gq activation Emax
100 105
DMT
NMT
BUF
5-MeO-DMT
5-MeO-NMT
5-HT
N-Me-5-HT
5-MeO-TRY
TRY
p = 0.02
R2 = 0.6
60
80
40
120
-arrestin 2 recruitment Emax
100
5-MeO-DMT
****
5 6 7 8
Number of Crossings
5-HT
5-MeO-NMT
TRY
5-MeO-TRY
N-Me-5-HT
5-MeO-DMT
BUF
NMT
DMT
p = <0.001
R2 = 0.9
5-MeO-TRY
5-MeO-DMT
5-HT
TRY
5-MeO-NMT
N-Me-5-HT
DMT
NMT
BUF
p = 0.0006
R2 = 0.9
5
10
psychLight 2 F/F
15
20
DMT
I
x
a
m
E
q
G
100
95
90
85
80
x
a
m
E
2
n
i
t
s
e
r
r
a
-
120
100
80
60
40
0
J
100
y
c
a
c
i
f
f
E
%
80
60
40
20
0
BUF
5-MeO-DMT
NMT
5-HT
TRY
5-MeO-TRY 5-MeO-NMT
N-Me-5-HT
p = 0.04
R2 = 0.5
0.5
1.0
2.0
2.5
1.5
cLogP
Fig. 1. Compound-induced neuronal growth correlates with ligand lipophilicity. (A) Chemical structures of
serotonin, TRY, and 5-MeO–TRY as well as their corresponding N-methyl and N,N-dimethyl analogs. 5-HT,
serotonin; BUF, bufotenin; Me, methyl; N-Me-5-HT, N-methylserotonin; NMT, N-methyltryptamine. (B) Sholl
analysis demonstrates that increasing N-methylation leads to a concomitant increase in dendritic arbor complexity.
The shaded area represents 95% confidence intervals. Sholl plots were generated from rat embryonic cortical
neurons (DIV6) treated with compounds (10 mM). VEH, vehicle. (C) Maximum numbers of crossings (Nmax) of the
Sholl plots in (B) (N = 45 to 64 neurons per treatment). Error bars represent standard error of the mean. KET,
ketamine. (D) Widefield images of rat embryonic cortical neurons (DIV6) treated with compounds (10 mM). (E to
J) Correlation plots of Sholl analysis percent efficacy (Nmax values relative to 10 mM ketamine as the positive control)
versus [3H]-IP accumulation (E), activation of psychLight2 (F), Gq activation (G), b-arrestin recruitment (H), or
calculated LogP (J). PsychLight2 activation correlates well with both Gq activation and b-arrestin recruitment (I).
Compounds were treated at 10 mM. Data for [3H]-IP accumulation were obtained from literature values (22).
*p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001, as compared with VEH controls [one-way analysis of
variance (ANOVA) followed by Dunnett’s multiple comparisons test]. R2 values were calculated through simple liner
regression. DF/F, change in fluorescence intensity relative to the baseline fluorescence intensity; Emax, maximum effect.
D
VEH
20 μm
E
100
y
c
a
c
i
f
f
E
%
80
60
40
20
0
G
100
y
c
a
c
i
f
f
E
%
80
60
40
20
0
80
Vargas et al., Science 379, 700–706 (2023)
17 February 2023
2 of 7
RES EARCH | R E S E A R C H A R T I C L E
A
HEK293T
Neurons
Membrane
Membrane
Membrane
Membrane
Membrane
Membrane
10 μm
10 μm
5-HT2AR
β2AR
ECFP
5-HT2AR
β2AR
ECFP
Overlay
Overlay
Overlay
Overlay
Overlay
Overlay
HEK293T
****
ns
****
B
’
C
C
s
n
o
s
r
a
e
P
1.0
0.8
0.6
0.4
0.2
0.0
HEK293T
*
ns
**
C
C
’
s
r
e
d
n
a
M
1.0
0.8
0.6
0.4
0.2
0.0
Neurons
****
****
****
’
C
C
s
n
o
s
r
a
e
P
1.0
0.8
0.6
0.4
0.2
0.0
Neurons
****
***
****
C
C
’
s
r
e
d
n
a
M
1.0
0.8
0.6
0.4
0.2
0.0
5-HT2AR
2AR
CFP
5-HT2AR
2AR
CFP
2AR
CFP
5-HT2AR
2AR
CFP
5-HT2AR
Fig. 2. Cortical neurons express intracellular 5-HT2A receptors. (A) Live-cell images of HEK293T cells
and rat embryonic cortical neurons (DIV6) expressing Myc–5-HT2AR–CFP, b2AR-ECFP, or ECFP. Signals
from the fluorescent protein and fluorescent plasma membrane marker (Cellbrite Steady) are shown in cyan
and white, respectively. The expression patterns of 5-HT2ARs and b2ARs are comparable in HEK293T
cells. However, in neurons, b2ARs exhibit higher expression on the plasma membrane than 5-HT2ARs.
The expression of ECFP is diffuse in both HEK293T cells and cortical neurons. (B) Pearson’s correlation
coefficients (Pearson’s CCs) and Manders’ colocalization coefficients (Manders’ CCs) quantify the extent of
colocalization between MYC–5-HT2AR–CFP, b2AR-ECFP, or ECFP and the fluorescent plasma membrane
marker (N = 20 to 43 cells per group). Error bars represent standard error of the mean. ns is not significant,
*p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001; bars indicate comparisons between data (one-way
ANOVA followed by Tukey’s multiple comparisons test).
the expression patterns of the tagged GPCRs
were markedly distinct from ECFP, which con-
firms that the GPCR component of the con-
structs dictates cellular localization. Expression
of a FLAG–5-HT2AR construct in rat cortical
neurons produced a similar intracellular local-
ization pattern, which suggests that these tags
do not substantially alter the localization pat-
terns of the 5-HT2AR (34).
Bright punctate staining within neurons sug-
gested that 5-HT2ARs were localized within
intracellular compartments (Fig. 2A), so we
performed additional immunocytochemistry
experiments to assess the overlap with mark-
ers of various subcellular organelles (fig. S4).
We observed a high level of 5-HT2AR colocal-
ization with Rab5 and Rab7 in both HEK293T
cells and neurons. Ras-related guanosine tri-
phosphatases (GTPases) of the Rab family are
known to regulate intracellular transport of
GPCRs (35). We observed a very large differ-
ence in 5-HT2AR colocalization with the Golgi
apparatus between neurons and HEK293T
cells, with the former exhibiting substantially
higher correlation coefficients (fig. S4). The
Golgi apparatus is a key regulator of GPCR
signaling, and ligands can either passively
diffuse into Golgi-localized receptor pools or
gain access by facilitated transmembrane trans-
port (18, 36, 37). Signaling from within the
Golgi can be distinct, as is the case for opioid
receptors (38).
To ensure that high 5-HT2AR colocalization
with the Golgi was not an artifact of overex-
pression, we assessed native 5-HT2AR expres-
sion in neurons using one of the few validated
5-HT2AR antibodies (30). To confirm the anti-
body’s specificity, we performed in-house vali-
dation by overexpressing 5-HT2ARs in HEK293T
cells (fig. S5A) and we used mouse 5-HT2AR
KO cortical neurons (fig. S5B). Next, we imaged
rat embryonic cortical neurons that expressed
only native 5-HT2ARs or overexpressed a Myc–
5-HT2AR–ECFP construct. Longitudinal and
transverse line scans indicated that Myc–5-
HT2AR–ECFP expression closely mirrored the
native localization pattern of 5-HT2ARs (fig.
S5, C and D).
Membrane permeability is required
for psychedelic-induced neuroplasticity
Given that cortical neurons have a large pool
of intracellular 5-HT2ARs, we next used chem-
ical tools to determine whether activation of
this intracellular population was essential for
psychedelics to promote structural neuroplas-
ticity. Chemical modification of the membrane-
permeable ligands DMT, psilocin (PSI), and
ketanserin (KTSN) converted them into highly
charged membrane-impermeable congeners
N,N,N-trimethyltryptamine (TMT), psilocybin
(PSY), and methylated ketanserin (MKTSN),
respectively (Fig. 3A). All of the charged species
exhibited negative cLogP scores (fig. S6A) but
retained affinity for 5-HT2ARs, as determined by
radioligand competition binding experiments
(fig. S6B). In psychLight2 assays, the membrane-
impermeable analogs displayed comparable
efficacies to their uncharged parent molecules
with reduced potencies (fig. S6, C and D).
Next, we treated freshly dissected rat embry-
onic cortical neurons with DMT (1 mM) and PSI
(1 mM) as well as their membrane-impermeable
congeners in the presence and absence of elec-
troporation. Electroporation creates tempo-
rary openings in the plasma membrane, which
enables highly charged molecules to pass
through and access the intracellular space.
Although the membrane-permeable 5-HT2AR
agonists were able to promote dendritogenesis
regardless of whether electroporation was
applied, the membrane-impermeable com-
pounds could only promote neuronal growth
when applied with electroporation (Fig. 3, B
and C). Similarly, the membrane-permeable
5-HT2AR antagonist KTSN (treated in 10-fold
excess at 10 mM) blocked DMT-induced plas-
ticity both with and without electroporation;
however, the membrane-impermeable antag-
onist MKTSN (10 mM) could only block DMT-
induced neuronal growth when it was applied
with electroporation (Fig. 3, B and C). By using
more mature neurons (DIV15), we demonstrated
that the membrane-permeable 5-HT2AR ago-
nists increased dendritic spine density—another
measure of structural neural plasticity—whereas
the membrane-impermeable agonists did not
(Fig. 3, D and E). KTSN was able to inhibit the
effects of DMT and PSI on dendritic spine den-
sity, whereas MKTSN was not (fig. S7).
To further establish that membrane-permeable
and -impermeable ligands target different
populations of 5-HT2ARs, we performed an
experiment in HEK293T cells that express
psychLight1. The psychLight1 construct was
chosen over psychLight2 because the former
lacks an endoplasmic reticulum (ER) export
sequence, which results in greater intracel-
lular localization (14). PsychLight1-expressing
HEK293T cells were pretreated with either
Vargas et al., Science 379, 700–706 (2023)
17 February 2023
3 of 7
RES EARCH | R E S E A R C H A R T I C L E
A
N Me
Me
DMT
N Me
Me
Me
I
TMT
N
H
N
H
OH
N Me
Me
PSI
N
H
O
P
OH
O
O
N Me
H
Me
PSY
C
N
H
O
KTSN
O
MKTSN
O
N
H
O
N
H
N
N
O
I
O
N
N
Me
D
VEH
DMT
F
TMT
PSI
PSY
****
****
F
E
0 2 4 6 8 10
spines / 10 μm
VEH (+)
– Electroporation
+ Electroporation
B
VEH (–)
20 μm
DMT (–)
DMT (+)
TMT (–)
TMT (+)
PSI (–)
PSI (+)
PSY (–)
VEH
PSY (+)
5-HT
VEH
DMT
TMT
PSI
PSY
KTSN
MKTSN
KTSN
+ DMT
MKTSN
+ DMT
VEH
DMT
TMT
PSI
****
****
*
*
****
**
PSY
1 μm
F
12
5-HT
*
5-MeO-DMT
**
)
%
(
F
F
/
9
6
3
0
VEH
MKTSN
Pre-Treatment
****
7 8 9
Number of Crossings
7 8 9
Number of Crossings
Fig. 3. Intracellular 5-HT2ARs mediate structural plasticity induced by serotonergic psychoplastogens.
(A) Structures of membrane-permeable (blue) and impermeable (red) compounds. (B) Widefield images
of rat embryonic cortical neurons (DIV6) that were administered compounds with (+) and without (−)
electroporation. (C) Nmax values obtained from Sholl plots. Membrane-impermeable analogs of
psychedelics (1 mM) only promote growth when administered with electroporation (N = 35 to 110 neurons
per treatment). Unlike KTSN (10 mM), MKTSN (10 mM) only blocks psychedelic-induced plasticity when
administered with electroporation (N = 45 to 109 neurons per treatment). (D) Membrane-impermeable
agonists of 5-HT2ARs (1 mM) cannot promote spinogenesis in cultured embryonic rat cortical neurons
(DIV15) (N = 30 to 34 neurons per treatment). (E) Confocal images of treated cortical neurons (DIV15).
(F) HEK293T cells that expressed psychLight1 were pretreated with VEH or MKTSN (10 mM) before the
administration of serotonin (10 mM) or 5-MeO–DMT (10 mM) (N = 41 to 54 cells per treatment). Error bars
in (C) and (F) represent standard error of the mean. *p < 0.05, **p < 0.01, and ****p < 0.0001, as
compared with VEH controls in (D) and (C) (one-way ANOVA followed by Dunnett’s multiple comparisons test)
or between indicated pairs of data in (F) (two-way ANOVA followed by Šídák's multiple comparisons test).
For box-and-whisker plots in (D), the center line represents the median, box limits are upper and lower quartiles,
and whiskers are minimum and maximum values.
vehicle or MKTSN to selectively antagonize
the population of 5-HT2ARs that were ex-
pressed on the plasma membrane. Next, changes
in psychLight1 fluorescence were measured after
administration of the membrane-impermeable
ligand serotonin or its membrane-permeable
congener 5-MeO–DMT. Because serotonin is
a full agonist, it induced a stronger response
in the absence of antagonist compared with
the partial agonist 5-MeO–DMT. Pretreatment
with MKTSN resulted in a much larger re-
duction in the serotonin-induced psychLight1
signal compared with that induced by 5-MeO–
DMT (Fig. 3F). Pretreatment with MKTSN
could nearly fully antagonize the effect of
serotonin. In sharp contrast, MKTSN only par-
tially blocked the ability of 5-MeO–DMT to
turn on psychLight1 fluorescence.
Because serotonin and 5-MeO–DMT exhibit
different lipophilicities (cLogP values of 0.57
and 2.33, respectively), we reasoned that their
abilities to displace [3H]–D-lysergic acid diethyl-
amide (LSD) bound to the 5-HT2AR would de-
pend on whether those receptors were exposed
to the extracellular environment. Thus, we per-
formed radioligand competition binding ex-
periments using intact HEK293T cells that
expressed Myc–5-HT2AR or psychLight2 as
well as membrane preparations obtained from
these systems. The inhibition constant (Ki)
values for serotonin and 5-MeO–DMT were
nearly identical when using membrane prep-
arations or intact PSYLI2 cells, with HEK293T
cells that stably expressed a 5-HT2AR con-
struct with an ER export sequence resulting in
a large proportion of the 5-HT2ARs being ex-
posed to the extracellular environment (fig. S8).
However, 5-MeO–DMT was an order of magni-
tude more potent than serotonin when using
intact HEK293T cells that expressed large
populations of both plasma membrane–bound
and intracellular 5-HT2ARs (fig. S8).
To confirm that serotonin and 5-MeO–DMT
can target distinct populations of 5-HT2ARs,
we performed inositol monophosphate (IP1) as-
says in cortical neurons and HEK293T cells that
expressed the Myc–5-HT2AR–ECFP construct.
When the assay was performed in HEK293T
cells that expressed Myc–5-HT2AR–ECFP, which
display a large proportion of plasma membrane–
bound 5-HT2ARs, serotonin and 5-MeO–DMT
had comparable potencies and efficacies (fig.
S9A). When the same experiment was performed
in rat cortical neurons, serotonin failed to elicit
an agonist response, although 5-MeO–DMT re-
mained a potent agonist (fig. S9B).
Cellular import of serotonin leads to structural
plasticity and antidepressant-like effects
If activation of intracellular 5-HT2ARs in cor-
tical neurons is sufficient to promote structural
plasticity, we hypothesized that serotonin should
be able to promote cortical neuron growth if
given access to the intracellular space. To test
this hypothesis, we first treated cortical neu-
rons with serotonin (1 mM) in the presence and
absence of electroporation. Unlike ketamine
(1 mM), serotonin was only able to promote cor-
tical neuron growth when applied with electro-
poration (Fig. 4A). Next, we took advantage
of the serotonin transporter (SERT), which
can import serotonin from the extracellular
environment (39). Endogenous expression of
SERT is typically restricted to presynaptic ter-
minals of neurons that emanate from the raphe,
and thus, rat cortical neurons do not express
appreciable levels of the transporter (40, 41).
Embryonic rat cortical neurons were electro-
porated with an enhanced yellow fluorescent
protein (EYFP)–tagged SERT construct under
the control of the calcium/calmodulin-dependent
protein kinase II (CaMKII) promoter to re-
strict expression to excitatory pyramidal neu-
rons. Sparse transfection resulted in cultures
Vargas et al., Science 379, 700–706 (2023)
17 February 2023
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
A
VEH
KET
5-HT
2
B
MAP2
20 μm
SERT
– Electroporation
+ Electroporation
****
VEH
KET
5-HT
4
6
Number of Crossings
8
10
2
4
VEH
DMT
5-HT
MAP2
MAP2
6
Number of Crossings
C
SERT -
SERT
SERT
VEH
DMT
Overlay
Overlay
Overlay
SERT +
5-HT
SERT –
SERT –
SERT –
SERT +
5
6
7
8
Number of Crossings
Blocking 5-HT Effects
Blocking DMT Effects
SERT +
D
5-HT
–
KTSN
–
CIT
–
+
–
+
–
+
–
+
+
–
–
–
–
–
+
+
DMT
–
KTSN
–
CIT
–
ns
**** ****
ns
ns
ns
ns
+
–
+
–
+
–
+
+
–
–
–
–
–
+
+
****
****
8
10
SERT +
ns
***
ns
**
*
*
ns
ns
ns
ns
ns
ns
****
****
****
****
8
4
5
7
6
Number of Crossings
8
4
5
7
6
Number of Crossings
SERT -
SERT +
SERT -
SERT +
E
VEH
DMT
5-HT
F
VEH
DMT
5-HT
ns
****
****
ns
****
****
0 2 4 6 8 10
spines / 10 μm
SERT -
SERT +
2 μm
SERT –
SERT +
SERT –
SERT +
SERT –
SERT +
Fig. 4. Cellular uptake of serotonin induces structural plasticity in vitro. (A) Rat embryonic cortical neurons
were administered compounds (1 mM) with (+) and without (−) electroporation. Nmax values obtained from Sholl
plots (DIV6) demonstrates that unlike ketamine, serotonin only promotes growth when administered with
electroporation (N = 49 to 59 neurons per treatment). (B) Widefield images of embryonic rat cortical cultures (DIV6)
sparsely transfected with CaMKII-SERT-EYFP and treated with compounds. (C) Nmax values obtained from Sholl plots
demonstrate that serotonin (10 mM) can only increase the dendritic arbor complexity of SERT-positive neurons,
whereas DMT (10 mM) can promote the growth of both SERT-positive and SERT-negative neurons (N = 69 to
93 neurons per treatment). (D) KTSN pretreatment (10 mM) blocks the plasticity-promoting effects of serotonin
(1 mM) in SERT-positive neurons and DMT (1 mM) in both SERT-positive and SERT-negative neurons. CIT (10 mM)
only blocks the plasticity-promoting effects of serotonin (1 mM) in SERT-positive neurons (N = 59 to 100 neurons per
treatment). (E) Confocal images of CaMKII-SERT-EYFP positive and negative dendrites (DIV15) treated with
compounds (10 mM). (F) Serotonin only promotes spine growth in SERT-positive neurons (N = 11 to 35 neurons per
treatment). Error bars in (A), (C), and (D) represent standard error of the mean. ns is not significant, *p < 0.05,
**p < 0.01, ***p < 0.001, and ****p < 0.0001, as compared with VEH controls in (A) (one-way ANOVA followed by
Dunnett’s multiple comparisons test) or compared with VEH controls from the same genotype in (C), (D), and (F)
(two-way ANOVA followed by Tukey’s multiple comparisons test). For box-and-whisker plots in (F), the center line
represents the median, box limits are upper and lower quartiles, and whiskers are minimum and maximum values.
that expressed both SERT-positive and SERT-
negative neurons (Fig. 4B), which enabled us
to compare the effects of serotonin on neurons
capable of importing the monoamine with those
that could not within the same cultures.
Treatment with the membrane-permeable
psychedelic DMT (10 mM) resulted in greater
dendritic arbor complexity and increased spine
density in both SERT-positive and SERT-negative
neurons (Fig. 4, B to F). Only SERT-positive
neurons treated with serotonin (10 mM) displayed
increased dendritogenesis and spinogenesis
(Fig. 4, B to F). To ensure that serotonin-
induced changes in structural neural plasticity
were due to the engagement of intracellular
5-HT2Rs through serotonin importation from
SERT, we performed similar experiments in
the presence of the selective SERT inhibitor
citalopram (CIT) and KTSN. When these ex-
periments were performed in the presence of
CIT (10 mM), the plasticity-promoting effects of
serotonin (1 mM) on SERT-positive neurons
were blocked (Fig. 4D). By contrast, CIT had no
effect on the ability of DMT (1 mM) to promote
the growth of SERT-positive or SERT-negative
cortical neurons (Fig. 4D). In the presence of
KTSN (10 mM), neither serotonin nor DMT could
promote neuronal growth, which confirms that
DMT and intracellular serotonin promote plas-
ticity by means of 5-HT2AR receptors in vitro
(Fig. 4D).
To determine whether intracellular serotonin
could promote the growth of cortical neurons
in vivo, we injected the medial PFC (mPFC) of
Thy1–enhanced green fluorescent protein (EGFP)
mice with either CaMKII-SERT-mCherry or
CaMKII-mCherry. The mPFC was chosen as the
injection site because it exhibits high levels of
5-HT2AR expression (28) and has been impli-
cated in the antidepressant-like effects of psycho-
plastogens (42). After 3 weeks to enable construct
expression (Fig. 5, A and B), both groups were
administered (±)-para-chloroamphetamine
[PCA, 5 mg/kg intraperitoneally (ip)], a selec-
tive serotonin-releasing agent (43). Dendritic
spine density on mPFC pyramidal neurons was
assessed 24 hours later. Animals that expressed
SERT in the mPFC displayed significantly higher
densities of dendritic spines after PCA admin-
istration as compared with mCherry controls
(Fig. 5, C and D). Notably, PCA is not a 5-HT2AR
agonist (fig. S10A), does not directly promote
the growth of SERT-positive or SERT-negative
cortical neurons in culture (fig. S10B), and does
not induce a head-twitch response (HTR) in WT
mice (fig. S10C).
Evidence suggests that psychoplastogen-
induced structural plasticity in the mPFC might
be related to sustained antidepressant-like effects
in rodents (44). To probe for an antidepressant-
like response that might be linked to neuro-
plasticity, we injected an adeno-associated virus
(AAV) that contained a CaMKII-SERT-EYFP
or CaMKII–green fluorescent protein (GFP)
Vargas et al., Science 379, 700–706 (2023)
17 February 2023
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
construct into the mPFC of WT C57BL/6J mice
(Fig. 5E and fig. S10D). After 3 weeks to allow
for construct expression, both groups of mice
underwent a novelty induced locomotion
(NIL) test, and no differences in locomotion
were observed (Fig. 5F). After 24 hours, mice
were subjected to a forced swim test (FST)
(7, 45). Again, we observed no differences
between the mice that expressed SERT and
those that expressed the GFP control (Fig. 5G).
Two days later, we administered PCA (5 mg/
kg ip), waited 24 hours, and then performed
a second FST (Fig. 5G). The SERT-expressing
mice exhibited a statistically significant HTR
immediately after the administration of PCA
as compared with the GFP control mice (fig.
S10E), and they also displayed a significant
reduction in immobility in the FST 24 hours
after administration (Fig. 5G).
Discussion
Although GPCRs are traditionally viewed as
initiators of signal transduction that originates
at the plasma membrane, increasing evidence
suggests that GPCR signaling from intracellu-
lar compartments can play important roles in
cellular responses to drugs. Recently, location
bias has been proposed to explain signaling
differences between endogenous membrane-
impermeable peptide ligands and membrane-
permeable ligands of opioid receptors (38).
Moreover, distinct ligand-induced signaling
has been observed for plasma membrane–
localized and intracellular populations of
d-opioid receptors (46). Here, we extend the
concept of location bias to ligands of the
5-HT2AR.
A substantial proportion of 5-HT2ARs in
cortical neurons are localized to the Golgi, and
intracellular compartments such as the Golgi
are slightly acidic compared with the cytosol
and extracellular space. Thus, it is possible that
protonation of psychedelics within the Golgi
leads to retention and sustained signaling,
which results in neuronal growth, even after
transient stimulation (47). Persistent growth
after the drugs have been removed from the
extracellular space is a hallmark of serotoner-
gic psychoplastogens (47, 48). Although the
mechanistic details that link intracellular 5-
HT2AR activation to cortical neuron growth
have not been fully elucidated, they are likely
to involve AMPA receptor, TrkB, and mTOR
signaling, as previously established (9). Future
studies should examine the detailed signaling
interplay between these proteins.
In addition to promoting psychedelic-induced
structural neuroplasticity, the intracellular
population of 5-HT2ARs might also contribute
to the hallucinogenic effects of psychedelics.
When we administered a serotonin-releasing
agent to WT mice, we did not observe a HTR.
However, the same drug was able to induce a
HTR in mice that expressed SERT on cortical
Fig. 5. Cellular uptake of serotonin produces antidepressant-like effects in vivo. (A) Schematic that
displays experimental design for measuring spine density in Thy1-EGFP mice after administration of
a serotonin-releasing agent. (B) Histology images of the mPFC of Thy1-EGFP mice that express CaMKII-
mCherry or CaMKII-SERT-mCherry. (C) Confocal images of dendritic spines in the mPFC of mice treated with
PCA (5 mg/kg ip). (D) Mice that express CaMKII-SERT-mCherry display increased dendritic spine density
after PCA treatment. (E) Schematic that displays experimental design for determining the sustained
antidepressant-like effects of serotonin in mice that express SERT in the mPFC. (F) NIL demonstrates no
difference between CaMKII-SERT-EYFP– and CaMKII-GFP–expressing mice. (G) CaMKII-SERT-EYFP– and
CaMKII-GFP–expressing mice exhibit no differences in the FST. After PCA (5 mg/kg ip) administration,
CaMKII-SERT-EYFP–expressing mice display a sustained antidepressant-like effect. ns is not significant, *p <
0.05, and **p < 0.01, as compared between indicated pairs of data in (D) and (F) (two-tailed unpaired
Student’s t test) or (G) (two-way repeated measures ANOVA followed by Šídák’s multiple comparisons test).
Error bars in (D), (F), and (G) represent standard error of the mean.
neurons of the mPFC, a brain region that is
known to be essential for the HTR (49). Thus,
activation of intracellular cortical 5-HT2ARs
may play a role in the subjective effects of
psychedelics. This hypothesis is further sup-
ported by previous work that demonstrates
that a high dose of the serotonin precursor 5-
hydroxytryptophan (5-HTP) induces a HTR
in WT mice, which can be blocked by an N-
methyltransferase inhibitor that prevents the
metabolism of 5-HTP to N-methyltryptamines
(50). Inhibition of N-methyltransferase failed
to block the HTR induced by 5-MeO–DMT (50).
Taken together, this work emphasizes that ac-
cessing intracellular 5-HT2ARs is important for
5-HT2AR agonists to produce a HTR.
Our results demonstrate that membrane per-
meability is essential for a ligand to activate
5-HT2ARs in cortical neurons; however, our
experiments did not distinguish between in-
tracellular signaling or the possibility of psy-
chedelics acting as pharmacological chaperones.
Others have hypothesized that GPCR ligands
may act as pharmacological chaperones, which
facilitates their export to the plasma membrane
where they could presumably engage in canon-
ical signaling (51). Thus, future studies should
determine whether intracellular 5-HT2AR sig-
naling is distinct from 5-HT2AR signaling at
the plasma membrane.
Although intracellular expression of 5-HT2ARs
within cortical pyramidal neurons has been
known for some time, it is unclear at present
why the subcellular localization of these re-
ceptors differs greatly between neurons and
HEK293T cells (28, 29, 31). One possibility is
that 5-HT2ARs form heterodimeric complexes
that affect cellular trafficking (31). Thus, by
dictating 5-HT2AR localization, cellular con-
text could influence responses to endogenous
neuromodulators and/or exogenous drugs,
which potentially results in circuit-specific ef-
fects of 5-HT2AR ligands.
Intracellular signaling has been hypothesized
to contribute to the pharmacological proper-
ties of a diverse range of compounds that in-
cludes nicotine, ketamine, and SSRIs (51–53).
Like psychedelics, these compounds are weak
bases with pKa (where Ka is the acid dissoci-
ation constant) values ranging from 7 to 10.
Given that the antidepressant mechanisms of
ketamine and SSRIs have not been definitively
established, it is intriguing to speculate that
they also might promote cortical neuron growth
Vargas et al., Science 379, 700–706 (2023)
17 February 2023
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
by binding to intracellular targets. Perhaps other
antidepressants affect the function of scaffold-
ing proteins within the cell interior to modu-
late neuronal growth phenotypes.
Without facilitated transport across the
plasma membrane, serotonin cannot induce
psychedelic-like effects on neuronal morphol-
ogy. Although it is possible that serotonin could
alter cortical neuron physiology by activating
cell-surface 5-HT2ARs, this receptor pool does
not seem to be involved in 5-HT2AR–induced
structural plasticity. Our results raise the in-
triguing possibility that serotonin may not be
the endogenous ligand for the population of
5-HT2ARs expressed inside cortical neurons.
Alternative ligands could include methylated
congeners of serotonin or TRY because these
compounds have greater abilities to cross non-
polar membranes. Endogenous psychedelics
such as DMT, 5-MeO–DMT, and bufotenin
have been identified in a variety of species, in-
cluding humans, and have long been hypothesized
to play roles in diseases such as schizophrenia
(54). However, they are rapidly degraded in vivo,
which makes their detection by classic an-
alytical methods quite challenging. The use
of more modern analytical techniques has
improved detection of these analytes (55), with
a recent study demonstrating that the concen-
tration of DMT in the cortex was comparable
to that of serotonin (56). The possibility that
endogenous psychedelics play a role in health
or disease should therefore be thoroughly
investigated.
RE FE RENCES AND N OT ES
1. R. S. Duman, G. K. Aghajanian, G. Sanacora, J. H. Krystal,
Nat. Med. 22, 238–249 (2016).
2. H. Qiao et al., Neural Plast. 2016, 8056370 (2016).
3. R. M. Shansky, C. Hamo, P. R. Hof, B. S. McEwen,
J. H. Morrison, Cereb. Cortex 19, 2479–2484 (2009).
4. T. Rantamäki et al., Neuropsychopharmacology 32, 2152–2162
(2007).
5. P. C. Casarotto et al., Cell 184, 1299–1313.e19 (2021).
6. D. E. Olson, J. Exp. Neurosci. 12, 1179069518800508 (2018).
7. M. V. Vargas, R. Meyer, A. A. Avanes, M. Rus, D. E. Olson,
Front. Psychiatry 12, 727117 (2021).
8. A. C. Kwan, D. E. Olson, K. H. Preller, B. L. Roth, Nat. Neurosci.
25, 1407–1419 (2022).
9. C. Ly et al., Cell Rep. 23, 3170–3182 (2018).
10. M. de la Fuente Revenga et al., Cell Rep. 37, 109836 (2021).
11. D. E. Olson, Biochemistry 61, 127–136 (2021).
12. L. E. Dunlap et al., J. Med. Chem. 63, 1142–1155 (2020).
13. L. P. Cameron et al., Nature 589, 474–479 (2021).
14. C. Dong et al., Cell 184, 2779–2792.e18 (2021).
15. D. Cao et al., Science 375, 403–411 (2022).
16. L. M. Luttrell, S. Maudsley, L. M. Bohn, Mol. Pharmacol. 88,
579–588 (2015).
17. K. Kim et al., Cell 182, 1574–1588.e19 (2020).
18. R. Irannejad et al., Nat. Chem. Biol. 13, 799–806 (2017).
19. M. A. Mohammad Nexhady, J. C. Nezhady, S. Chemtob,
iScience 23, 101643 (2020).
20. N. V. Weisstaub et al., Science 313, 536–540 (2006).
21. D. Ristanović, N. T. Milosević, V. Stulić, J. Neurosci. Methods
158, 212–218 (2006).
22. B. J. Ebersole, I. Visiers, H. Weinstein, S. C. Sealfon, Mol. Pharmacol.
63, 36–43 (2003).
23. R. H. J. Olsen et al., Nat. Chem. Biol. 16, 841–849 (2020).
24. J. C. Bermak, M. Li, C. Bullock, Q. Y. Zhou, Nat. Cell Biol. 3,
492–498 (2001).
25. U. E. Petäjä-Repo, M. Hogue, A. Laperriere, P. Walker,
M. Bouvier, J. Biol. Chem. 275, 13727–13736 (2000).
26. S. E. Jacobsen et al., J. Biol. Chem. 292, 6910–6926
(2017).
27. C. A. Purgert et al., J. Neurosci. 34, 4589–4598 (2014).
28. V. Cornea-Hébert, M. Riad, C. Wu, S. K. Singh, L. Descarries,
J. Comp. Neurol. 409, 187–209 (1999).
29. V. Cornea-Hébert et al., Neuroscience 113, 23–35 (2002).
30. C. L. Schmid, K. M. Raehal, L. M. Bohn, Proc. Natl. Acad. Sci. U.S.A.
105, 1079–1084 (2008).
31. R. Toneatti et al., Sci. Signal. 13, eaaw3122 (2020).
32. J. E. Saffitz, S. B. Liggett, Circ. Res. 70, 1320–1325 (1992).
33. Q. Fu, Y. K. Xiang, Prog. Mol. Biol. Transl. Sci. 132, 151–188
(2015).
34. A. C. Magalhaes et al., Nat. Neurosci. 13, 622–629 (2010).
35. J. L. Seachrist, S. S. Ferguson, Life Sci. 74, 225–235
(2003).
36. C. A. Nash, W. Wei, R. Irannejad, A. V. Smrcka, eLife 8, e48167
(2019).
37. R. Irannejad et al., Nature 495, 534–538 (2013).
38. M. Stoeber et al., Neuron 98, 963–976.e5 (2018).
39. S. Aggarwal, O. V. Mortensen, Curr. Protocols Pharmacol. 79,
12.16.1–12.16.17 (2017).
40. F. C. Zhou, Y. Sari, J. K. Zhang, Brain Res. Dev. Brain Res. 119,
33–45 (2000).
41. C. Lebrand et al., J. Comp. Neurol. 401, 506–524 (1998).
42. M. Fuchikami et al., Proc. Natl. Acad. Sci. U.S.A. 112, 8106–8111
(2015).
43. T. Sharp, T. Zetterström, L. Christmanson, U. Ungerstedt,
Neurosci. Lett. 72, 320–324 (1986).
44. R. N. Moda-Sava et al., Science 364, eaat8078 (2019).
45. D. A. Slattery, J. F. Cryan, Nat. Protoc. 7, 1009–1014 (2012).
46. S. E. Crilly, W. Ko, Z. Y. Weinberg, M. A. Puthenveedu, eLife 10,
e67478 (2021).
47. C. Ly et al., ACS Pharmacol. Transl. Sci. 4, 452–460 (2020).
48. L. X. Shao et al., Neuron 109, 2535–2544.e4 (2021).
49. J. González-Maeso et al., Neuron 53, 439–452 (2007).
50. C. L. Schmid, L. M. Bohn, J. Neurosci. 30, 13513–13524
(2010).
51. H. A. Lester, J. M. Miwa, R. Srinivasan, Biol. Psychiatry 72,
907–915 (2012).
52. B. J. Henderson, H. A. Lester, Neuropharmacology 96, 178–193
(2015).
53. H. A. Lester, L. D. Lavis, D. A. Dougherty, Am. J. Psychiatry 172,
1064–1066 (2015).
54. S. A. Barker, E. H. McIlhenny, R. Strassman, Drug Test. Anal. 4,
617–635 (2012).
55. S. A. Barker, J. Borjigin, I. Lomnicka, R. Strassman,
Biomed. Chromatogr. 27, 1690–1700 (2013).
56. J. G. Dean et al., Sci. Rep. 9, 9333 (2019).
57. M. V. Vargas et al., Data set for Intracellular 5-HT2A Paper,
Figshare (2022); https://doi.org/10.6084/m9.figshare.21669635.
ACKN OWLED GMEN TS
We thank C. Ly, A. C. Greb, L. P. Cameron, and W. C. Duim for
performing early pilot studies; A. Avanes for assistance with
radioligand binding studies; K. Zito for providing Thy1-GFP
breeders; J. González-Maeso for providing the Myc–5-HT2AR–
ECFP plasmid; R. Iyer for performing high-resolution mass
spectrometry analysis; and P. Beal for use of his CLARIOStar plate
reader. We also thank C. Nichols for advice about the radioligand
binding experiments. Funding: This work was supported by
funds from the National Institutes of Health (NIH) (R01GM128997
to D.E.O., R35GM133421 to J.D.M., and U01NS120820,
U01NS115579, and 2R01MH101214-06 to L.T.), three NIH training
grants (T32GM099608 to M.V.V., T32GM113770 to R.J.T., and
T32MH112507 to H.N.S.), Human Frontier (L.T.), the Camille and
Henry Dreyfus Foundation (D.E.O.), a sponsored research
agreement with Delix Therapeutics (D.E.O.), and a University of
California (UC) Davis Provost’s Undergraduate Fellowship (S.J.C.).
This project used the Biological Analysis Core of the UC Davis
MIND Institute Intellectual and Development Disabilities Research
Center (U54 HD079125). The Nikon High Content Analysis
Spinning Disk Confocal microscope used in this study was
purchased using NIH Shared Instrumentation Grant 1S10OD019980-
01A1. We thank the MCB Light Microscopy Imaging Facility,
which is a UC Davis Campus Core Research Facility, for the use of
this microscope. Funding for the nuclear magnetic resonance
spectrometers was provided by the National Science Foundation
(NSF CHE-04-43516) and NIH (08P0ES 05707C). Analysis
for this project was performed in the UC Davis Campus Mass
Spectrometry Facilities, with instrument funding provided by the
NIH (1S10OD025271-01A1). Several of the drugs used in this study
were provided by the National Institute on Drug Abuse (NIDA)
Drug Supply Program. Author contributions: M.V.V. performed
most of the in vitro experiments, including the dendritogenesis,
spinogenesis, subcellular colocalization, IP1, neuromics antibody
validation, and psychLight assays. C.D. cloned and validated key
reagents for the in vivo experiments, performed the surgeries, and
the perfusions. M.V.V. performed the small-molecule electroporation
experiments, key pilot experiments imaging HEK293T cells and
neurons, and brain-slice imaging with assistance from C.D., and
M.V.V. performed the behavioral experiments. L.E.D. performed the
N-methylation SAR dendritogenesis experiments, calculated cLogP
values, and synthesized TMT and MKTSN. R.J.T. and S.J.C.
performed the radioligand binding studies. H.N.S. performed
culturing of 5-HT2AR KO cultures. L.P.C. and S.D.P. performed the
Golgi staining experiments. S.A.J. performed the electrophysiology
experiments. J.J.H. performed BRET assays of 5-HT2AR activation.
J.D.M., J.A.G., L.T., and D.E.O. supervised various aspects of this project
and assisted with data analysis. D.E.O. conceived the project and
wrote the manuscript with input from all authors. Competing
interests: D.E.O. is a co-founder of Delix Therapeutics, Inc., serves as
the chief innovation officer and head of the scientific advisory board,
and has sponsored research agreements with Delix Therapeutics.
Delix Therapeutics has licensed technology from UC Davis. The
sponsors of this research were not involved in the conceptualization,
design, decision to publish, or preparation of the manuscript.
Data and materials availability: Data are available at Figshare (57).
Custom written data analysis codes are available through GitHub
(see methods for details). All materials are available upon request or
are commercially available. TMT and MKTSN are available from D.E.O.
and psychLight is available from L.T. under material agreements
with UC Davis. License information: Copyright © 2023 the authors,
some rights reserved; exclusive licensee American Association for
the Advancement of Science. No claim to original US government works.
https://www.science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adf0435
Materials and Methods
Figs. S1 to S10
References (58, 59)
MDAR Reproducibility Checklist
View/request a protocol for this paper from Bio-protocol.
Submitted 27 September 2022; accepted 9 January 2023
10.1126/science.adf0435
Vargas et al., Science 379, 700–706 (2023)
17 February 2023
7 of 7
|
10.1126_science.ade9858
|
RES EARCH
QUANTUM SENSING
Nanoscale covariance magnetometry with diamond
quantum sensors
Jared Rovny1, Zhiyang Yuan1, Mattias Fitzpatrick1†, Ahmed I. Abdalla1‡, Laura Futamura1§,
Carter Fox2, Matthew Carl Cambria2, Shimon Kolkowitz2, Nathalie P. de Leon1*
Nitrogen vacancy (NV) centers in diamond are atom-scale defects that can be used to sense magnetic
fields with high sensitivity and spatial resolution. Typically, the magnetic field is measured by averaging
sequential measurements of single NV centers, or by spatial averaging over ensembles of many NV
centers, which provides mean values that contain no nonlocal information about the relationship
between two points separated in space or time. Here, we propose and implement a sensing modality
whereby two or more NV centers are measured simultaneously, and we extract temporal and spatial
correlations in their signals that would otherwise be inaccessible. We demonstrate measurements of
correlated applied noise using spin-to-charge readout of two NV centers and implement a spectral
reconstruction protocol for disentangling local and nonlocal noise sources.
C orrelated phenomena play a central role
in condensed matter physics and have
been studied in many contexts, includ-
ing phase transitions (1, 2), many-body
interactions and entanglement (3, 4),
and magnetic ordering (5, 6), as well as in the
context of fluctuating electromagnetic fields,
where two-point correlators are central to
characterizing field statistics (7, 8). Recent ef-
forts toward improving quantum devices have
also explored correlated noise in superconduct-
ing quantum interference devices (9) and
qubits (10–12). Nitrogen vacancy (NV) centers
in diamond are a promising sensing platform
for detecting correlations because they are
robust, noninvasive, and capable of measuring
weak signals with nanoscale resolution (13).
These advantages have made them a useful tool
for studying many condensed matter sys-
tems, including magnetic systems such as two-
dimensional (2D) van der Waals materials
(14, 15), magnons (16), and skyrmions (17, 18);
and transport phenomena such as Johnson
noise (19), hydrodynamic flow (20–22), and
electron-phonon interactions in graphene
(23). These applications are powerful but have
so far been limited to signals that are averaged
over space or time—more information is po-
tentially available by studying spatial and tem-
poral correlations in the system. Advances in
nanoscale spectroscopy have already been
made by studying correlations from a single
NV center at different points in time (24–26),
where temporal correlations are calculated be-
1Department of Electrical and Computer Engineering,
Princeton University, Princeton, NJ 08544, USA.
2Department of Physics, University of Wisconsin–Madison,
Madison, WI 53706, USA.
*Corresponding author. Email: npdeleon@princeton.edu
†Present address: Thayer School of Engineering, Dartmouth
College, Hanover, NH 03755, USA.
‡Present address: Department of Electrical Engineering, Stanford
University, Stanford, CA 94305, USA.
§Present address: Department of Physics, Stanford University,
Stanford, CA 94305, USA.
tween subsequent measurements before sig-
nal averaging. Here, we extend this technique
to measuring correlations in both space and
time between individual measurements with
pairs of NV centers before signal averaging;
measuring correlated dynamics between two
different NV centers provides simultaneous
information at length scales ranging from
the diffraction limit to the full field of view
(~0.1- to 100-mm length scales) and at two
different sensing times limited only by the
experimental clock cycle (~1-ns resolution).
Measurements of spatiotemporal correlations
at these length and time scales would provide
useful information about the dynamics of the
target system, including the electron mean free
path, signatures of hydrodynamic flow (27), or
the microscopic nature of local NV center noise
sources such as surface spins (28, 29).
Determining correlations between NV centers
We consider two NV centers that do not directly
interact with each other but experience a shared
classical magnetic field, whose amplitude is cor-
related at the locations of the two NV centers
(Fig. 1A). Each NV center also sees a distinct
local magnetic field that is uncorrelated between
the two locations. These fields are detected using
a Ramsey-type experiment that addresses the
ms = 0 and ms = +1 (or −1) spin sublevels of
the NV center (referred to as states 0 and 1,
respectively) (Fig. 1, B to D). After many
repeated measurements, we accumulated a
list of signals S1 = {s1,i} and S2 = {s2,i} from NV1
and NV2, respectively, where s is the number
of photons detected in a single experiment
and i = 1…N indexes the N total experiments.
Though similar to a typical Ramsey-type
variance detection sequence (30), we empha-
size two modifications for covariance detec-
tion. First, despite detecting zero-mean noise,
we chose a final pulse that is 90° out of phase with
the initial pulse, such that for high-frequency
noise detection, the final spin state is equally
likely to be 0 or 1 (Fig. 1, B and C), which maxi-
mizes our sensitivity to correlations. This is not
done in conventional noise detection using
variance magnetometry, because straightforward
signal averaging would then always produce
i ¼ 0:5. Second, we did not
the same result msi
compute the average value of this signal but rather
computed the shot-to-shot cross-correlation be-
tween the raw signals S1 and S2 (Fig. 1D).
h
Whereas conventional variance measure-
ments provide spectral densities with no spa-
tial information (top row of Fig. 1E), the addition
of covariance information allows us to identify
which spectral components are common be-
tween two NV centers and which are specific to
each (bottom row of Fig. 1E). Throughout this
work, we focus on the measured Pearson corre-
lation r = Cov(S1, S2)/(σ1σ2), where Cov is the
covariance and σ1 and σ2 are the standard
deviations of S1 and S2.
Experimental implementation of
covariance measurements
To demonstrate our protocol, we used an ex-
ternal radiofrequency coil or stripline to apply
a global, random phase ac signal to two shal-
low NV centers ~10 nm from the diamond sur-
face. Here, the two NV centers share the same
magnetic resonance frequency, so that all micro-
wave pulses address both. They are spatially
resolved, which allows for separate excitation
and readout using two independent optical paths
(31). To boost the sensitivity of our readout, we
used a simultaneous spin-to-charge conver-
sion (SCC) protocol (32, 33) on each NV center
separately. We used an XY8 sensing protocol
for each NV center to maximize sensitivity to
the applied ac signal (34) (Fig. 2A). We ob-
served correlations that are maximized when
the interpulse spacing matches the frequency
of the global signal (blue circles in Fig. 2B). The
correlations are apparent in the photon count
statistics [bottom panel (ii) of Fig. 2B]; when
one or more photons are detected from NV1, we
observed a higher likelihood of also detecting a
photon from NV2. To confirm that we were
indeed detecting correlations in the spin state
of the NV centers rather than spurious techni-
cal correlations (31), we could also initialize the
two NV centers on opposite sides of the Bloch
sphere before applying the XY8 sequence (Fig.
2A). The phase accumulation step then results
in a final state that is anticorrelated between
the two NV centers (red squares in Fig. 2B).
The sensitivity of a covariance measurement
differs from that of a traditional magnetometry
measurement because it requires simultaneous
signals from two NV centers. Assuming that the
detected phases are statistically even, as for a
noisy or random-phase signal, we find (31) the
Pearson correlation
r ¼
½
e(cid:2) ˜c1
t1ð Þ þ ˜c2
σR2
σR1
(cid:3)
t2ð Þ
(cid:4)
(cid:1)
sin fC1
t1ð Þ
(cid:3)
(cid:1)
sin fC2
t2ð Þ
(cid:3)
i
ð1Þ
Rovny et al., Science 378, 1301–1305 (2022)
23 December 2022
1 of 5
RES EARCH | R E S E A R C H A R T I C L E
¼
p
where the subscripts 1 and 2 denote NV1 and
NV2, respectively; the decoherence function
tð Þ describes the “typical” coherence decay
˜c1;2
of the NV centers due to the local fields (35);
are the phases accumulated by the NV cen-
fC1;2
ters due to the correlated field; and the read-
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Þ2
ð
1 þ 2 a0 þ a1
out noise σR1;2
characterizes the fidelity of a photon-counting
experiment with mean detected photon num-
bers a0 and a1 for spin states 0 and 1, respec-
tively (30). For thresholding, the readout noise
instead depends on the fidelity of the spin-
state assignment. Readout noise may be gen-
eralized to include non-Poisson statistics that
result from errors in charge-state initialization
or ionization (31).
Þ= a0 (cid:2) a1
ð
Note that the detectable correlation depends
quadratically on the readout noise, which makes
readout fidelity especially important for de-
tecting correlations; this key fact is implicit in
prior calculations of single-NV center two-point
correlators derived in the context of repeated
weak measurements (25). This may be intui-
tively understood from Fig. 2C, which shows the
raw photon counts for conventional versus
SCC readout methods. Using conventional read-
out, only ~0.01 photons are detected per mea-
surement, such that detecting simultaneous
counts from both NV centers is extremely un-
likely. Using SCC readout substantially increases
our ability to detect coincident events and has
a greater effect on covariance measurements
than on conventional single–NV center mea-
surements. From the independently measured
values for each term on the right-hand side of
Eq. 1 (31), we expect the detectable correlation
in our experiment to be approximately bounded
by r ≈ 0.01, in good agreement with the max-
imum correlation r ≈ 0.008 that we detect here
(Fig. 2B). The remaining discrepancy is likely
due to experimental imperfections such as sam-
ple drift or pulse miscalibration over time.
Because readout noise plays an amplified
role in covariance detection, covariance mea-
surements can become prohibitively long with-
out optimizing sensitivity, for which we require
a detailed understanding of the signal-to-noise
ratio (SNR). The sensitivity (minimum noise
amplitude sBmin with SNR = 1) of an experiment
that detects Gaussian noise is given by (31)
!
s2
B;min ¼
(cid:2)p⋅Hz
2t
4ge
ln 1 (cid:2)
p
2e2t=T2
2sR
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
T = t þ tR
Þ
ð
ð2Þ
where ge is the electron gyromagnetic ratio, t
is the phase integration time, T2 is the co-
herence time, tR is the readout time, and T ≈
(t + tR)N is the total experiment time ig-
noring initialization. This is shown in the
bottom graph of Fig. 2C for three different
readout methods: conventional (sR = 35),
SCC (sR = 4), and single-shot readout with
perfect fidelity (sR = 1), which is ultimately
Fig. 1. Covariance noise sensing. (A) Diagram of a diamond with two near-surface NV centers that are
experiencing uncorrelated local fields and a correlated common field. (B) Bloch sphere representations of
each qubit state during sensing, with the states prepared along x followed by a phase accumulation that will
be different in each experiment, resulting in a distribution of phases. At the end of each experiment, a final
p/2 pulse maps these phases to populations. (C and D) Pulse sequence diagrams showing the sensing (XY8)
and measurement (SCC) sequence for each NV center. The measurement is repeated many times, retaining
the photon counts from each measurement without signal averaging; we instead measured the correlation
between the resulting lists Si. (E) Using conventional detection of single NV centers (top row), the coherence
decay gives access to the noise spectral density S(f) but provides no spatial information. Covariance
magnetometry measuring two NV centers (bottom row) provides information about which spectral features
are correlated and which are uncorrelated.
limited by quantum projection noise. Achiev-
ing an SNR equal to 1 for these three scenarios
when σB = 1 nT requires total experiment
times on the order of 300 hours, 3 hours, and
10 s, respectively. Whereas detecting correla-
tions is extremely inefficient using conventional
readout, enhanced readout protocols like SCC
(32, 36) allow for substantially lower readout
noise, making covariance magnetometry pos-
sible to implement in practice.
Disentangling correlated and uncorrelated
noise sources
Detecting cross-correlations in pure noise re-
veals previously hidden information about
the spatial structure of the noise, which we
now demonstrate using two NV centers that
sense both local and nonlocal magnetic fields.
We first measured the spectral density S(f),
where f is frequency, using a conventional
variance magnetometry measurement of two
different NV centers (Fig. 3A, top). These indi-
vidual spectra reveal that there are two fre-
quencies where signals are seen by both NV
centers but cannot provide simultaneous non-
local spatial information about that signal. Using
covariance magnetometry over the same fre-
quency range (Fig. 3A, bottom) shows only the
higher-frequency feature, which clearly reveals
that the higher-frequency feature is caused by a
noise signal common to each NV center, whereas
the lower-frequency feature is instead caused by
local noise sources specific to each NV center.
This ability to distinguish correlated and un-
correlated features enables spatially resolved
spectral decomposition, which allows us to dis-
tinguish spectral components that are shared
from those that are local. For phases that are
Rovny et al., Science 378, 1301–1305 (2022)
23 December 2022
2 of 5
RES EARCH | R E S E A R C H A R T I C L E
A
NV1
NV2
NV1,2
NV1
NV2
B
)
3
0
1
(
n
o
i
t
a
e
r
r
o
C
l
8
4
0
4
8
(i)
(ii)
(iii)
Init.
Reset XY8
SCC
200
225
250
275
Interpulse spacing τ (ns)
300
(i)
(ii)
(iii)
s2
1
0
0
1
0
1
0
1
NV1 photon count s1
1.0
0.5
0.0
0.5
1.0
)
3
0
1
(
b
a
P
C
t
n
u
o
c
n
o
t
o
h
P
Conventional
SCC
NV1
NV2
8
4
0
4
0
0
100 200 0
Experiment number
100 200
SCC
optimal
conv.
)
T
n
(
.
i
n
m
,
B
10.0
1.0
0.1
10 2 100
102
106
Total experiment time (s)
104
ð
ð
ð
Þ (cid:2) P s1 ¼ a
points in the top panel show no correlation (i), positive correlation (ii), or
negative correlation (iii), where the color indicates the joint detection probability
~
Pab (cid:4) P s1 ¼ a; s2 ¼ b
Þ. (C) Comparison of shot-to-shot
ÞP s2 ¼ b
photon counts during averaging for conventional readout (top left) and SCC
readout (top right). Minimum magnetic field amplitude to detect correlations with
SNR = 1 for Gaussian noise is shown at the bottom. Here, we have assumed
T2 = 100 μs and the phase integration time t = T2/2 = 50 μs, as well as a
readout time of 300 ns for conventional readout and 1 ms for SCC and optimal
readout. Initialization time was ignored. The horizontal and vertical gray lines are
guides to the eye.
Fig. 2. Detecting correlations and anticorrelations. (A) Pulse sequence and
final Bloch sphere mapping for correlation (top left) or anticorrelation (top right)
measurements using global microwave control. For anticorrelations, an extra p
pulse and spatially selective NV polarization optical pulse (“reset”) are added
during initialization (bottom, gray box). (B) Correlation detected from a 2-MHz ac
signal whose phase is randomized with 1-MHz bandwidth Gaussian noise.
The measured correlations are positive when the NV centers are initialized
parallel to one another (blue circles) and negative when they are initialized
antiparallel to one another (red squares). Lines indicate the predicted correlation
shape (31). Raw photon count statistics (bottom) taken from the marked data
Fig. 3. Disentangling correlated and uncorrelated signals.
(A) Single-NV noise spectra derived from conventional
XY8 variance magnetometry (top) of two NV centers (orange
open markers and gray filled markers with Gaussian fit
lines, arbitrarily offset). Each NV center detects signals at two
common frequencies, but it is impossible to directly
determine whether the sources are local or nonlocal. Spectral
decomposition (bottom) using covariance magnetometry
(Eq. 3) reveals that the higher-frequency peak is caused by
a shared noise source. Here, the shared noise feature is
engineered using an applied global 1.75-MHz ac signal,
whereas the local feature is caused by the 15N nuclear spin
intrinsic to each NV center. The line indicates the predicted
correlation shape. The light red vertical lines indicate the
expected peak locations (31). (B) In a broadband correlated
noise environment, the two NV centers rapidly decohere (orange
open markers and gray filled markers). Lines are exponential
fits. (C) Covariance magnetometry for evolution times indicated
by the gray region in (B) reveals a dip in the Pearson correlation
around t = 1800 ns that arises from the uncorrelated 15N
nuclear spins intrinsic to each NV center. The line indicates the
predicted correlation shape. The broadband noise is correlated,
which allows for the observation of spectral features from local
signals even at evolution times beyond the coherence time of both NV centers.
Gaussian-distributed or small ( f ≪ p ), we
can find (31) the correlated noise spectrum
SC(f) if we have access to both the two-NV
correlation r as well as each NV center’s
coherence decay Ci(t) = e−c(t) [note that Ci(t)
includes both the correlated and uncorrelated
noise sources]:
SC fð Þ ¼ p
2t
sinh(cid:2)1
(cid:6)
(cid:7)
2r
sR
C1ðtÞC2ðtÞ
ð3Þ
where t = n/(2f ) and n is the total number of
applied XY8 pulses. This equation is used
to obtain the correlated spectrum from the
measured correlation and single–NV center
coherence decays, as shown in Fig. 3A. The
Rovny et al., Science 378, 1301–1305 (2022)
23 December 2022
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Temporal structure in correlations using independent control.
(A) Confocal image showing the two NV centers used for these experiments
(left). Optically detected magnetic resonance spectrum (middle) showing optical
contrast as a function of microwave drive frequency displays two distinct sets of
transitions corresponding to NV1 and NV2, with assignments (right). The NV
centers are driven independently on either the (0, −1) transitions for both NVs,
labeled {−, −}, or the (0, −1) and (0, +1) transitions for NV1 and NV2, respectively,
labeled {−, +}. (B) Diagram of the pulse sequence used to probe temporal
correlations. After initialization, the start of the XY8 pulse sequence applied to
NV2 is delayed by time tdelay from the start of the pulses on NV1. A f0 = 3.125 MHz
global ac signal is applied, making the resonant XY8 interpulse spacing t = 160 ns.
(C) Correlations for the case in which the NV centers are addressed on the same
transitions ({−, −}, blue circles) oscillate as a function of tdelay at the ac signal
frequency of 3.125 MHz. The correlations invert (red squares) when the two NV
centers are addressed on different transitions ({−, +}), because they now accumulate
opposite phases for the same signal. (D) With added phase noise, the time-domain
dephasing of the ac signal is resolvable, despite having a short coherence time
(less than 2 μs) compared with the XY8 sequence time.
local spectrum for each NV center SL1;2
ðf Þ may
also be found from each individual NV center’s
total spectrum SL1;2
ðf Þ ¼ S1;2ðf Þ (cid:2) SCðf Þ.
So far, we have analyzed the case where shared
and local features are spectrally resolved, but
an interesting scenario arises when a shared
signal decoheres each NV center at frequen-
cies coincident with local noise sources. To
probe this case, we applied a global broad-
band Gaussian noise signal, decohering both
NV centers while inducing broadband cor-
relations in their phases (Fig. 3, B and C).
Beyond the coherence time of each NV cen-
ter, conventional variance detection cannot
reveal any information (gray region in Fig. 3B).
However, covariance magnetometry (Fig. 3C)
measures the broadband correlation in the
random phases of the decohered NV centers;
this correlation will dip if either NV center
interacts with a local noise source in its
vicinity, because the local signal induces a
phase that is specific to that NV center. The
covariance magnetometry spectrum there-
fore reveals a feature that is hidden in the
single-NV spectra.
Temporal structure of correlations
Covariance magnetometry also enables mea-
surements of the temporal structure of the
two-point correlator B r1; t1
i sepa-
h
rated in time as well as space for short time
scales where t2 – t1 < t + tR, which is not possible
with single–NV center correlation measure-
ð
ÞB r2; t2
ð
Þ
ments (24–26). To perform this measurement,
independent control of each NV center is
required. We accomplished this by choosing
two NV centers with different orientations
at low magnetic fields (Fig. 4A), such that
the 0→−1 transition of the NV center that is
aligned with the magnetic field is detuned
by 70 MHz from that of the misaligned NV
center. We then offset the beginning of the
XY8 sequence applied to NV2 by time tdelay
(Fig. 4B) and measured an applied ac field at
frequency f0 = 3.125 MHz. As we swept tdelay,
the correlations oscillated at frequency f0
(Fig. 4C), as expected for a random-phase ac
signal (26, 37). Independent control also al-
lowed us to simultaneously address opposite
spin transitions for each NV center (Fig. 4A,
right). Because the two NV centers then accu-
mulate opposite phases from the ac field, we
observed anticorrelations with the same fre-
quency (red squares in Fig. 4C).
Because the two NV centers are manipu-
lated independently, there are no fundamental
constraints on the length of tdelay. This allowed
us to directly measure time-domain structure
on the nanosecond time scale at two points in
space, despite using p pulses with 60-ns du-
ration. When we measured the correlations
between two NV centers experiencing a shared
ac signal with added phase noise (Fig. 4D), we
could directly resolve the temporal structure
of the ac signal despite its short coherence time
of less than 2 ms, without making use of spectral
deconvolution. We emphasize that this techni-
que is very general and is thus applicable to any
time-varying signal with a nonzero correla-
tion time that can be detected with NV centers.
Correlations will remain detectable on the
time scale of the underlying signal correla-
tion time, even when the signal phase is
completely random from one experiment to
the next (as in Fig. 3C).
Concluding remarks
Our demonstration of simultaneous control
and readout of two spatially resolved NV cen-
ters shows that nanoscale magnetometry of
two-point spatiotemporal field correlators
that would normally be discarded using con-
ventional NV center magnetometry is possi-
ble. Spatiotemporal correlations of any signal
that can be imprinted as a phase on the NV
centers can be sensed with this technique, pro-
vided that the statistics of the signal remain
sufficiently stationary over the course of the
experiment. Our approach has many potential
applications; measurements of these two-point
correlators can reveal the underlying length
and time scales of fluctuating electromagnetic
fields near surfaces (7, 8), which provides in-
formation about nonequilibrium transport dy-
namics (38) and condensed matter phenomena
like magnetic ordering in low-dimensional
systems (5, 15). For example, there has been
considerable recent interest in studying hydro-
dynamic flow in 2D materials (20–22), but it
Rovny et al., Science 378, 1301–1305 (2022)
23 December 2022
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
is challenging to directly observe the hydro-
dynamic transition—covariance noise sensing
could provide new quantitative information
about these dynamics. Also, magnetic excita-
tions such as magnons can have micron-scale
dynamics, a natural length scale for covariance
magnetometry with pairs of NV centers. Fu-
ture extensions of the current demonstration
include using photonic structures to improve
photon collection efficiency (33), applying dif-
ferent pulse sequences to each NV center to
probe the correlations between signals at dif-
ferent frequencies or phases (39), measuring
more NV centers to measure higher-order
joint cumulants (31), and using detector arrays
to perform simultaneous readout of many
pairs of NV centers.
RE FE RENCES AND N OT ES
1. H. Bernien et al., Nature 551, 579–584 (2017).
J. Zhang et al., Nature 551, 601–604 (2017).
2.
3. X.-L. Deng, D. Porras, J. I. Cirac, Phys. Rev. A 72, 063407
(2005).
4. M. L. Baez et al., Proc. Natl. Acad. Sci. U.S.A. 117, 26123–26134
(2020).
5. A. Mazurenko et al., Nature 545, 462–466 (2017).
J. Simon et al., Nature 472, 307–312 (2011).
6.
7. V. N. Premakumar, M. G. Vavilov, R. Joynt, Quantum Sci. Technol. 3,
015001 (2017).
8. K. Agarwal et al., Phys. Rev. B 95, 155107 (2017).
9. S. Gustavsson et al., Phys. Rev. B 84, 014525 (2011).
10. C. D. Wilen et al., Nature 594, 369–373 (2021).
11. U. von Lüpke et al., PRX Quantum 1, 010305 (2020).
12. P. Szańkowski, M. Trippenbach, Ł. Cywiński, Phys. Rev. A 94,
012109 (2016).
13. F. Casola, T. van der Sar, A. Yacoby, Nat. Rev. Mater. 3, 17088
40. J. Rovny et al., Nanoscale covariance magnetometry with
diamond quantum sensors, Harvard Dataverse (2022);
https://doi.org/10.7910/DVN/XMEIHR.
(2018).
14. Q.-C. Sun et al., Nat. Commun. 12, 1989 (2021).
15. L. Thiel et al., Science 364, 973–976 (2019).
16. E. Lee-Wong et al., Nano Lett. 20, 3284–3290 (2020).
17. A. Jenkins et al., Phys. Rev. Mater. 3, 083801 (2019).
18. Y. Dovzhenko et al., Nat. Commun. 9, 2712 (2018).
19. S. Kolkowitz et al., Science 347, 1129–1132 (2015).
20. M. J. H. Ku et al., Nature 583, 537–541 (2020).
21. U. Vool et al., Nat. Phys. 17, 1216–1220 (2021).
22. A. Jenkins et al., Phys. Rev. Lett. 129, 087701 (2022).
23. T. I. Andersen et al., Science 364, 154–157 (2019).
24. J. M. Boss, K. S. Cujia, J. Zopes, C. L. Degen, Science 356,
837–840 (2017).
25. M. Pfender et al., Nat. Commun. 10, 594 (2019).
26. A. Laraoui et al., Nat. Commun. 4, 1651 (2013).
27. L. Levitov, G. Falkovich, Nat. Phys. 12, 672–676 (2016).
28. B. L. Dwyer et al., arXiv:2103.12757 [quant-ph] (2021).
29. S. Sangtawesin et al., Phys. Rev. X 9, 031052 (2019).
30. J. M. Taylor et al., Nat. Phys. 4, 810–816 (2008).
31. See supplementary materials.
32. B. J. Shields, Q. P. Unterreithmeier, N. P. de Leon, H. Park,
M. D. Lukin, Phys. Rev. Lett. 114, 136402 (2015).
33. J. F. Barry et al., Rev. Mod. Phys. 92, 015004 (2020).
34. T. Gullion, D. B. Baker, M. S. Conradi, J. Magn. Reson. 89,
479–484 (1990).
35. Ł. Cywiński, R. M. Lutchyn, C. P. Nave, S. Das Sarma,
Phys. Rev. B 77, 174509 (2008).
AC KNOWLED GME NTS
We acknowledge helpful conversations with A. Burgers, S. Gopalakrishnan,
and J. Thompson. Funding: This work was funded by NSF
CAREER grant DMR-1752047; the Princeton Catalysis Initiative;
the US Department of Energy, Office of Science, Office of Basic
Energy Sciences, under award no. DE-SC0018978; the US
Department of Energy Office of Science National Quantum
Information Science Research Centers; a Princeton Quantum
Initiative Postdoctoral Fellowship (J.R.); and the Intelligence
Community Postdoctoral Research Fellowship Program by the Oak
Ridge Institute for Science and Education (ORISE) through an
interagency agreement between the US Department of Energy and
the Office of the Director of National Intelligence (ODNI) (M.F.).
Author contributions: Theoretical framework: J.R., M.F., A.I.A.,
C.F., M.C.C., S.K., N.P.d.L.; Experiment: J.R., Z.Y., L.F.; Sensing
technique, experimental design, data analysis: J.R., C.F., M.C.C.,
S.K., N.P.d.L.; Writing: J.R., C.F., M.C.C., S.K., N.P.d.L. Competing
interests: The authors declare that they have no competing
interests. Data and materials availability: All data needed to
evaluate the conclusions in this paper are present in the paper or
the supplementary materials and are publicly available at Harvard
Dataverse (40). License information: Copyright © 2022 the
authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.science.org/about/science-
licenses-journal-article-reuse
36. D. A. Hopper, H. J. Shulevitz, L. C. Bassett, Micromachines 9,
SUPPLEMENTARY MATERIALS
437 (2018).
37. C. L. Degen, F. Reinhard, P. Cappellaro, Rev. Mod. Phys. 89,
035002 (2017).
38. P. E. Dolgirev et al., Phys. Rev. B 105, 024507 (2022).
39. P. Szańkowski, G. Ramon, J. Krzywda, D. Kwiatkowski,
Ł. Cywiński, J. Phys. Condens. Matter 29, 333001
(2017).
science.org/doi/10.1126/science.ade9858
Materials and Methods
Figs. S1 and S2
References (41–45)
Submitted 20 September 2022; accepted 25 November 2022
10.1126/science.ade9858
Rovny et al., Science 378, 1301–1305 (2022)
23 December 2022
5 of 5
|
10.1126_science.adf0460
|
RES EARCH
R E S E A R C H A R T I C L E
◥
NEUROSCIENCE
Cross-modal representation of identity in the
primate hippocampus
Timothy J. Tyree1,2, Michael Metke1,3, Cory T. Miller1,3*
Faces and voices are the dominant social signals used to recognize individuals among primates. Yet, it
is not known how these signals are integrated into a cross-modal representation of individual identity
in the primate brain. We discovered that, although single neurons in the marmoset hippocampus
exhibited selective responses when presented with the face or voice of a specific individual, a parallel
mechanism for representing the cross-modal identities for multiple individuals was evident within
single neurons and at the population level. Manifold projections likewise showed the separability of
individuals as well as clustering for others’ families, which suggests that multiple learned social
categories are encoded as related dimensions of identity in the hippocampus. Neural representations
of identity in the hippocampus are thus both modality independent and reflect the primate social network.
N avigating complex primate societies re-
lies on learning the identity of each indi-
vidual in the group and their respective
social relationships (1). Neurons in the
brains of primates and other mammals
selectively respond to the identity when view-
ing the face or hearing the voice of a specific
individual as unimodal signals (2–8). How-
ever, data showing that single neurons are
responsive to both the face and voice of an
individual—a cross-modal representation of
identity—are limited to “concept cells” in the
human hippocampus (9–11). These neurons
are notable for several reasons, including their
putative role in memory functions (12) and
potential uniqueness to humans (13). We in-
vestigated whether cross-modal representations
of identity are evident in the hippocampus of
marmoset monkeys by recording single hippo-
campal neurons (14) while presenting subjects
with multiple exemplars of individual marmo-
set faces (from different viewpoints) and voices
as unimodal stimuli (4), as well as concurrently
by presenting the faces and voices from the same
or different individuals—i.e., match versus mis-
match (MvMM). Visual stimuli were presented
from a monitor directly in front of the animal
while a speaker positioned directly below the
screen broadcast the acoustic stimuli. Subjects
were only presented with familiar conspecifics
housed in the same colony who differed in
their respective social relatedness (11).
Identity-selective neurons
To first test whether cross-modal representa-
tions of identity are evident in the hippocampus
1Cortical Systems and Behavior Laboratory, University of
California San Diego, La Jolla, CA 92039, USA. 2Department
of Physics, University of California San Diego, La Jolla, CA
92039, USA. 3Neurosciences Graduate Program, University
of California San Diego, La Jolla, CA 92039, USA.
*Corresponding author. Email: corymiller@ucsd.edu
of a nonhuman primate, we performed the same
receiver operator characteristic (ROC) selectivity
analysis described previously in humans (9–11)
and detected a population of cross-modal in-
variant neurons for individual identity when
observing marmoset faces or voices (Fig. 1A
and fig. S1A), as well as neurons selective for
individual identity when viewing only their
faces (Fig. 1B and fig. S1B) or hearing only
their voices (Fig. 1C and fig. S1C). These iden-
tity neurons were confirmed in all hippocam-
pal subfields (Fig. 1D). Notably, only neurons
that exhibited a mean peak firing rate 2 SDs
above baseline qualified for this analysis, which
supports P < 0.01, and differed from the 5 SDs
used previously in humans (9–11). Responses
were determined from the mean firing rate
during a 500-ms continuous sliding window
maximized over the duration of the 3500-ms
stimulus. Overall, we observed that N = 148
(9.2%) of N = 1602 qualifying neurons dem-
onstrated selectivity for a single preferred
individual (Fig. 1E), with different neurons
selective for faces (N = 52), voices (N = 39), or
both faces and voices (N = 57) (Fig. 1F). The
mean area under the ROC curve (AUC) of iden-
tity neurons (AUC = 0.902 ± 0.014) was sig-
nificantly above chance (P < 0.001) (Fig. 1G).
Although these neurons in marmosets were
overall less selective than in humans (9–11), this
disparity may reflect species differences in hip-
pocampus properties that affect neural coding
mechanisms for identity. Baseline hippocampal
activity, for example, was considerably higher
in the current study (mean 6.47 Hz; N = 2358
neurons) (fig. S2) than has been reported in
humans (15), although a more comprehensive
comparative analysis of physiological differ-
ences is needed to better understand how
such differences affect hippocampal functions.
Analysis of eye movements (Fig. 1H) revealed
that marmosets’ visual behavior and neural
activity were differentially affected by modal-
ity and identity. Marmosets exhibited signifi-
cantly shorter fixations (P < 0.001, Nfixations =
18,965) (Fig. 1I) and significantly more sac-
cades (P < 0.001, Nsaccades = 2203) during trials
with face-only relative to the voice-only trials
(Fig. 1J). These monkeys were also highly
focused on faces during stimulus presenta-
tions, with faces accounting for 77.9% of view-
ing time and eyes specifically accounting for
37.6% of viewing time. The firing rate of iden-
tity neurons was significantly greater than the
remaining neurons when subjects were look-
ing at the eyes or face (both P < 0.001) (Fig.
1K). This was not, however, a broad atten-
tional effect (16) because the firing rate of simul-
taneously recorded nonidentity neurons did
not show the same increased firing rate when
gazing at faces or eyes.
Multiple identities are represented in
single neurons
A potential parallel mechanism to highly se-
lective concept cells is for individual cells to
contribute to multiple functions (17, 18), such
as single neurons being sensitive to the cross-
modal identity of multiple conspecifics. Hip-
pocampal neurons are sensitive to mismatches
between the features of a particular stimulus
and a previously learned category (19, 20). To
test whether a similar mechanism is evident
for the learned social identities of conspecifics
in marmoset hippocampus, we tested whether
neurons would respond differently when simul-
taneously observing the face and voice from
the same (identity match) or different (iden-
tity mismatch) individuals. By presenting a
face and voice in all identity MvMM trials, we
controlled for the potential effects of multi-
modal integration (fig. S3A) and instead tested
whether a subordinate category, identity,
elicited changes in neural activity. Indeed, a
subpopulation of units, MvMM neurons, ex-
hibited a significant firing rate preference for
either match trials (Fig. 2A) or mismatch trials
(Fig. 2B), with some neurons modulated only
by this category distinction (Fig. 2A) and others
more generally stimulus driven (Fig. 2B). Over-
all, 21.7% of neurons (N = 511 of 2358) exhibited
a significant response during MvMM trials,
with significantly more units exhibiting a higher
firing rate during match (N = 401) than mis-
match (N = 110) trials (P < 0.001) (Fig. 2C and
fig. S3B). MvMM neurons were largely distinct
from the identity neurons described above
(Fig. 2D and fig. S3C). Notably, 56% of the
neurons observed in both populations whose
anatomical location could be confirmed were
recorded in CA1. In contrast to identity neu-
rons, MvMM neurons were biased to CA1 (Fig.
2E), with N = 155 (44.3%) out of 350 neurons
confirmed in the CA1 qualifying as MvMM
neurons. In CA1, significantly more MvMM
neurons (N = 129/155, 83.2%) preferred match
Tyree et al., Science 382, 417–423 (2023)
27 October 2023
1 of 7
RES EARCH | R E S E A R C H A R T I C L E
trials to mismatch trials (P < 0.001). MvMM
neurons exhibited significantly higher median
firing rate while the subject was looking at
the eyes or face (P < 0.001, N = 511) (Fig. 2F).
Marmosets exhibited significantly more sac-
cadic eye movements during mismatch trials
(Fig. 2G), and this difference in behavior was
most prominent 1 to 2 s after stimulus onset
(P < 0.05, Nsaccades = 4603) (Fig. 2H).
Encoding cross-modal identity in
neuron populations
These findings suggest that two seemingly dis-
tinct mechanisms for representing cross-modal
identity are evident in primate hippocampus.
We conjectured that more temporally selective
coding mechanisms in hippocampus may in-
form how these two processes for encoding
identity are integrated at a population level.
Therefore, we developed an algorithm to iden-
tify intervals of time during which individual
Fig. 1. Putative concept cells in marmoset
hippocampus. (A to C) Top row: Subset of
stimuli shown above raster and peristimulus
time histogram (PSTH). Bottom row: Spike
waveform density; normalized PSTH to all
stimuli (preferred: red, nonpreferred: black),
indicated are time points that show
significant difference (P < 0.05); median
number of spikes for unimodal stimuli
(gray/black indicate nonpreferred individuals);
ROC curve (shuffled controls shown in black).
Exemplar identity neurons responding
selectively to the face and voice of a
preferred conspecific (red) (A), the face
only (B), and the voice only (C) are shown.
(D) Anatomical distribution of identity
neurons (red) in hippocampal subfields
relative to neurons remaining that responded
to any stimulus (white). Black shadow indicates
the electrode array track with magnetic
resonance imaging distortion artifact.
DG, dentate gyrus; Sub, subiculum. (E) Pie
chart showing the abundance of identity
neurons in red with the number of remaining
neurons that qualified for the ROC
selectivity analysis in white. (F) Mode
distribution of identity neurons. Modes
included face (light blue), voice (dark blue),
and both (orange). (G) Histogram showing
the distribution of AUCs comparable
with red ROC curves in (A) to (C). Colors
are as in (F). Black dotted line is the mean,
and red dotted line is the mean of 10,000
random shuffles of the labels. (H) Exemplar eye
movements (yellow) with fixations indicated
(red). (I) Distribution of eye fixation
durations for unimodal trials. (J) Distribution
of apparent saccade number for unimodal
trials. (K) Distribution of median firing
rates while observer was looking at eyes
(left) and face (right) for identity neurons
(black) versus remaining neurons (white).
Significant median differences, ***P < 0.001.
Tyree et al., Science 382, 417–423 (2023)
27 October 2023
2 of 7
Fig. 2. Single neurons in hippocampus
represent multiple individuals. (A and B)
PSTH normalized by the prestimulus baseline
(top) and spike raster (bottom) for two
exemplar MvMM neurons. Black indicates
match trials, and red indicates mismatch
trials. Vertical line indicates stimulus
onset. Inset shows spike waveform density.
Significant time points, *P < 0.05.
Exemplar neuron with higher firing rate
for match (A) and mismatch (B) trials.
(C) Pie chart showing the number of
neurons that responded significantly
more for match (black) or mismatch
(red) trials. (D) Venn diagram showing
the number of MvMM neurons (black)
in common with identity neurons (red).
(E) Relative abundance of MvMM
neurons in each hippocampal subfield.
(F) Distribution of median firing rate
while looking at the eyes (left) and face
(right) for MvMM neurons (black) versus
remaining neurons (white). (G) Probability
density of saccadic eye movements
directed toward the eyes for match
(black) and mismatch (red) trials.
Indicated are the time points in (H).
(H) Distribution of apparent number
of saccades to eyes. Significant median
differences, *P < 0.05.
RES EARCH | R E S E A R C H A R T I C L E
A
B
C
E
G
D
F
H
Tyree et al., Science 382, 417–423 (2023)
27 October 2023
3 of 7
RES EARCH | R E S E A R C H A R T I C L E
neurons exhibited significant differences in
median firing rate for a specific category (P <
0.05), which we labeled as predictive time bins
(fig. S4). This algorithm was applied to all
neurons in the population, not only those
classified as identity-selective or MvMM neu-
rons. We first identified predictive time bins
selective for specific individuals when observing
their face or voice. A pair of exemplar neurons
that exhibited separate predictive time bins for
two different individuals is shown (Fig. 3A and
fig. S5). Out of 2358 hippocampal neurons,
1634 (69.3%) exhibited at least one identity-
specific predictive time bin, with most exhibiting
predictive time bins for two or more individ-
uals (Fig. 3B). Identity-specific predictive time
bins exhibited a mean AUC (0.802 ± 0.003)
that was significantly above chance (P < 0.001,
Nbins = 3958) (Fig. 3C). Neurons that had
identity-specific predictive time bins exhibited
a significantly greater median firing rate when
subjects were looking at the face of a preferred
individual (P < 0.001) (fig. S6A), although sig-
nificant suppression was observed relative to
F
E
A
B
C
D
Fig. 3. Cross-modal decoding of
identity. (A) PSTH of two exemplar
predictive neurons. Colored traces
average over trials involving preferred
individual, and the gray shaded regions
indicate 95% confidence intervals.
Colored regions indicate identity-
specific time bins. (B) Pie chart
showing number of identity-specific
predictive neurons that prefer one
(white), two (gray), and three or
more individuals (red). (C) Histogram
showing AUC distribution of
identity-specific time bins with
colors indicating preferred individuals
in legend. Dotted lines indicate
the mean (black) and the control
(red). (D) Distribution of median firing
rates while the observer was looking
at the face for the identity-specific
predictive neurons compared with
the remaining neurons. (E) Histogram
showing AUC distribution of MvMM
time bins. Dashed lines indicate
the mean (black) and the control
(red). (F) Distribution of median firing
rates while the observer was looking
at the face for the MvMM predictive
neurons compared with the remaining
neurons. (G) ROC curves for the
detection of face or voice of individu-
als. Firing rates were considered
from MvMM time bins (green, AUC =
0.536) and identity-specific time bins
(black, AUC = 0.779) similarly aver-
aged over individuals. Thinner colored
lines indicate individuals as in (C).
(H) ROC curves for the detection of
match trials. Firing rates from MvMM
time bins (green, AUC = 0.782)
and from identity-specific time bins
(black, AUC = 0.516). (I) ROC curves
for the detection of both face and
voice of individuals from same
19 recording sessions as in (G) and
(H). Firing rates from MvMM time bins
(green, AUC = 0.615), identity-specific
time bins (black, AUC = 0.622), and
the INM (gray, AUC = 0.818) are similarly averaged over individuals. Results of the INM for individuals are shown by thin lines colored as in the legend of (C).
Red dotted line indicates random as in (G) and (H). (J) Bar plot showing true positive rates predicted by a winner-take-all model that considered predictions from the
INM specific to 12 individuals. Indicated is the mean of the shuffled labels (red) and 5× that value (black). Bar plots summarize the trials from the testing sets of
33 recording sessions (Ntrials = 454). (K) Bar plot showing mean AUC with identity neurons removed (light gray) versus the control randomly removing an equal number of
bins from the remaining cells (dark gray). Uncertainty indicates 95% confidence of the mean. No significant difference was observed across recording sessions for any of the
three qualifying subjects (Archie, P = 0.81, Nidentities = 14; Baloo, P = 0.58, Nidentities = 9; Hades, P = 0.50, Nidentities = 12). ***P < 0.001; n.s., not significant.
H
G
K
J
I
Tyree et al., Science 382, 417–423 (2023)
27 October 2023
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
the other neurons when normalizing by the
background, which was averaged from t =
−0.8 s to t = −0.3 s (P < 0.001) (Fig. 3D). We
applied the same algorithm to test for pre-
dictive time bins that distinguished MvMM
trials and found a similar result (Fig. 3E), with
1455 neurons exhibiting MvMM predictive
time bins. Neurons with predictive time bins
for MvMM exhibited a significantly greater
median firing rate when subjects looked at the
face (P < 0.001) (fig. S6B), although signif-
icant suppression was observed relative to the
Fig. 4. Cross-modal representation of identity using rate and event codes.
(A) Two-dimensional manifold projection of our rate-coded representation
computed from firing rates of identity-specific time bins. One identity-match
trial was equivalent to one presentation of the stimulus as face and voice
matched. Each identity-match trial represented a different, randomly selected
face and/or voice stimulus. Firing rates were computed from each identity-
specific predictive time bins and then concatenated into a feature vector for
each identity-match trial. This feature vector was then projected onto the
manifold and plotted as one symbol per identity-match trial in one of the
scatter plots. Indicated is the mean (black). Colors in legend correspond to
individuals. UMAP, uniform manifold approximation and projection. (B) Schematic
illustrating the hindsight delay to a given neuron (left), used to generate histograms
of signed connection rates to three neurons (right). PDF, probability density function.
(C) Two-dimensional manifold projection of our event-coded representation of
identity computed as the manifold projection of signed connection rates of all
neurons in the same exemplar recording session. One symbol represents
one spike. Indicated is the mean (black). (D) Boxplots of MSR showing significantly
different values when subjects observed family of other subjects. Shown is Archie
observing family of Hades (top left, P < 0.001, Nidentities ≥ 23), Buck observing family
of Hades (top right, P = 0.003, Nidentities ≥ 26), Archie observing family of Baloo
(bottom left, P = 0.017, Nidentities ≥ 30), and Buck observing family of Baloo
(bottom right, P = 0.828, Nidentities ≥ 37). Significance was computed according to
Student’s t test. (E) Latent activity averaged over all recording sessions from
subjects Archie (left) and Buck (right). Colors indicate average over the family of
Baloo (blue) and Hades (orange) relative to all conspecifics (gray). Shaded regions
indicate 95% confidence of the mean estimated through bootstrap. (F) Graph of
connections bundled between individuals. Triangles in legend indicate family
members as in (A) and (C).
Tyree et al., Science 382, 417–423 (2023)
27 October 2023
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
other neurons when normalizing by the same
background (P < 0.001) (Fig. 3F). Instances
of face and eye viewing were highly variable
and not limited to the timing of predictive time
bins, which suggests that attentional effects
from visual behavior were not likely driving
neural activity during these periods (fig. S7).
We observed considerable overlap between
neurons with identity-specific and MvMM
predictive time bins because 82.2% (N = 1196;
fig. S8A) exhibited predictive time bins in both
analyses.
Identity network model
We developed a stable neural decoder by com-
bining the firing rates of predictive time bins
using an ensemble of gradient-boosted deci-
sions trees (21). When using identity-specific
time bins, we could reliably decode the iden-
tity of all marmosets when subjects observed
their face or voice (accuracy: 77.4%) (Fig. 3G).
The same approach could successfully decode
MvMM trials when using MvMM time bins
(accuracy: 75.7%) (Fig. 3H). The two kinds of
decoders used mostly different time points,
with only 24.6% ± 1.5% of identity-specific
time bins overlapping with MvMM time bins
within the same neurons (fig. S8B).
To test whether the same population could
represent multiple cross-modal identities, we
developed the identity network model (INM),
which integrates these two decoding approaches.
The first approach was identical to the identity-
specific decoder described above, which re-
sulted in accurate decoding for each individual’s
face or voice. The second approach classified
MvMM trials as either match or mismatch but
was anonymized to individual identity. Our INM
combined these two approaches to achieve cross-
modal decoding of individual identity (fig. S9).
This combination was critical because the
identity-specific predictive population was only
accurate for individual identity but performed
poorly for classifying MvMM (Fig. 3G), whereas
the MvMM predictive population was the op-
posite (Fig. 3H). When combined across in-
dividuals, the INM successfully decoded the
cross-modal identity of all 12 individuals (ac-
curacy: 84.5%) (Fig. 3I). Decoding perform-
ance tested at least 5× above chance when
distinguishing all individuals (Fig. 3J and
fig. S10).
Because identity neurons were included in
decoding, we investigated whether their ex-
planatory contribution was disproportionate
to their sparse distribution. We compared
INM performance when these neurons were
removed from the analysis and separately
used only in the analysis versus an equal num-
ber of other neurons. We observed no signif-
icant effect on decoding performance despite
the consideration of only individuals preferred
by identity neurons (Fig. 3K and figs. S11 and
S12), which suggests that these highly selective
neurons are no more notable for decoding the
cross-modal identity of familiar individuals
in the hippocampus than other neurons in the
same population. Furthermore, no significant
effect of identity neurons on decoding was dem-
onstrated at a larger 5-SD response threshold.
Social category representations in hippocampus
An individual’s identity is also coupled to their
social relationships, such as their family. To
test whether hippocampus encodes categor-
ical attributes of social identity, we applied
nonlinear dimensionality reduction techniques
(22). Using mean firing rates consistent with
studies of face and voice processing in the
primate brain (23), we first verified that these
reduction techniques were capable of separat-
ing the stimulus categories at multiple probe
locations along the anterior-posterior axis (fig.
S13). We next replicated the findings of the
INM using the same identity-specific predic-
tive time bins for marmoset faces and voices
drawn from the entire hippocampal popula-
tion and showed that manifold projections
similarly separated individuals (Fig. 4A and
fig. S14), including for different subpopula-
tions of neurons (fig. S14G). This suggests that
cross-modal identity representations are evi-
dent in the population activity of marmoset
hippocampus.
To investigate whether representation of
identity can be described by the relative timing
of spikes, we computed manifold projections
of spike times recorded during identity-match
trials (Fig. 4B, left) using parameterless signed
connection rate features. The signed connec-
tion rate from one neuron to another describes
how they interact, which reveals statistical dis-
tributions specific to any given pair of neurons
(Fig. 4B, right)—a facet of neural activity dis-
tinct from the firing rate of any single neuron.
Each spike was concatenated into a feature
vector, which was then projected onto the
manifold as is shown by one symbol (Fig. 4C).
The feature vector was computed as the signed
connection rate to each neuron at each ob-
servation time. Each observation time was a
spike time of the neuron with the greatest
number of spikes. Results using this event-
coded measure revealed excellent separability
for identity-match trials (Fig. 4C), which there-
by replicated the effect observed with the INM
using a distinct facet of neural activity.
We next investigated whether social cate-
gories other than identity may likewise be
represented in event-coded hippocampal ac-
tivity. We tested whether representations of
other marmosets’ family members were dis-
tinct from nonfamily members for the two
marmosets whose families were not included
in the stimulus sets using two distinct quan-
tifications of manifold projections, although
the pattern was consistent for all subjects.
First, results revealed a significant difference
in the mean square range (MSR) of the mani-
fold projections along this category boundary
(Fig. 4D and fig. S15A), which suggests that a
larger event-coded state space was occupied
while observing family members (fig. S15B).
Although these projections were supervised,
the clustering that emerged on the basis of
respective social relatedness was unsuper-
vised. Second, we computed the unsupervised
latent firing rate as the manifold projection
of the absolute value of signed connection
rate. Although individual identities did not
separate (fig. S16A), we found trajectories that
appeared stable in time and comparable across
trials (fig. S16B). The motion of mean latent
firing rate significantly separated social cate-
gories at multiple time points for all subjects
(Fig. 4E and fig. S16, C and D). Together, these
results demonstrate that neural representa-
tions of social identity in primate hippocampus
are not only invariant to the sensory modality
and comparable over time (fig. S17) but that
low-dimensional manifolds (Fig. 4F) can describe
relationships between different social catego-
ries (e.g., individual identity, family groups).
Conclusions
We showed that the cross-modal identity of
multiple conspecifics is represented in the pri-
mate hippocampus. Although we identified
putative concept cells similarly to human
studies (9, 12), we discovered that this popu-
lation of highly selective neurons is not the
only mechanism for representing concepts
of individuals. Rather, both single neurons
and the broader population in hippocampus
encode cross-modal identity of multiple con-
specifics, similar to what has been reported
for objects (24), which suggests that the sparse
representations of concept cells may not be
the only mechanism to represent semantic
memory in hippocampus. An important caveat
to these findings, however, is that our criteria
for determining the responsiveness of putative
concept cells differed somewhat from previous
studies in humans (9–11), as described above.
It is possible, therefore, that the concept cells
in marmoset and human hippocampus are not
strictly analogous. Ultimately, whether these
neurons are equivalent is not determined solely
by the physiological properties used to classify
them in analyses but their functional role in
memory. To directly address this issue, com-
parative experiments examining the com-
putational contributions of concept cells in
hippocampus across species during memory
are critical, as such data are currently lacking.
In addition to these findings at the single-
neuron level, a population-level code repre-
senting not only the cross-modal identity of
multiple familiar individuals but information
pertinent to social categories was likewise re-
ported in this study. Information from both
the putative concept cells and those neurons
Tyree et al., Science 382, 417–423 (2023)
27 October 2023
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
that encoded multiple identities were inte-
grated, which suggests that cross-modal iden-
tity in hippocampus is evident in the ensemble
activity of this keystone brain structure. Sim-
ilar to the role of hippocampus in other con-
texts (fig. S18) (25), the cross-modal identity
representations revealed in this experiment
may support a learned schema that here ap-
plies to social identity (26, 27). That these
experiments were performed in a highly con-
strained context limits our ability to deter-
mine whether such a schema would indeed be
leveraged in more naturalistic contexts during
which social decisions on the basis of con-
specifics’ identity are made continuously (6, 8).
The presence of unimodal representations of
identity in the primate frontal and temporal
cortex (2, 8, 28), amygdala (5, 29), and the
medial temporal lobe (30) and representations
of social dominance in the amygdala (31) may
reflect an integrative social recognition circuit
in which substrates in the broader network
play distinct but complementary roles that col-
lectively govern natural primate social brain
functions (32).
RE FE RENCES AND N OT ES
3. P. Belin, C. Bodin, V. Aglieri, Hear. Res. 366, 65–74 (2018).
4.
J. Sliwa, A. Planté, J.-R. Duhamel, S. Wirth, Cereb. Cortex 26,
950–966 (2016).
5. U. Rutishauser et al., Curr. Biol. 21, 1654–1660 (2011).
6. M. C. Rose, B. Styr, T. A. Schmid, J. E. Elie, M. M. Yartsev,
Science 374, eaba9584 (2021).
7. L. Chang, D. Y. Tsao, Cell 169, 1013–1028.e14 (2017).
8. R. Báez-Mendoza, E. P. Mastrobattista, A. J. Wang,
Z. M. Williams, Science 374, eabb4149 (2021).
9. R. Q. Quiroga, L. Reddy, G. Kreiman, C. Koch, I. Fried, Nature
435, 1102–1107 (2005).
10. R. Quian Quiroga, A. Kraskov, C. Koch, I. Fried, Curr. Biol. 19,
11.
1308–1313 (2009).
I. V. Viskontas, R. Q. Quiroga, I. Fried, Proc. Natl. Acad. Sci. U.S.A.
106, 21329–21334 (2009).
12. R. Q. Quiroga, Nat. Rev. Neurosci. 13, 587–597 (2012).
13. R. Quian Quiroga, Hippocampus 33, 616–634 (2023).
14. H. S. Courellis et al., PLOS Biol. 17, e3000546 (2019).
15. F. Mormann et al., J. Neurosci. 28, 8865–8872 (2008).
16. J. Minxha et al., Cell Rep. 18, 878–891 (2017).
17. M. Rigotti et al., Nature 497, 585–590 (2013).
18. C. T. Miller et al., Curr. Biol. 32, R482–R493 (2022).
19. D. Kumaran, E. A. Maguire, J. Neurosci. 27, 8517–8524
(2007).
20. M. Fyhn, S. Molden, S. Hollup, M.-B. Moser, E. Moser, Neuron
35, 555–566 (2002).
21. T. Chen, C. Guestrin, in KDD ’16: Proceedings of the 22nd
ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining (Association for Computing Machinery, 2016),
pp. 785–794.
22. E. H. Nieh et al., Nature 595, 80–84 (2021).
23. W. A. Freiwald, D. Y. Tsao, Science 330, 845–851 (2010).
24. T. P. Reber et al., PLOS Biol. 17, e3000290 (2019).
25. P. Baraduc, J.-R. Duhamel, S. Wirth, Science 363, 635–639
(2019).
1. R. M. Seyfarth, D. L. Cheney, in The Evolution of Primate
Societies, J. Mitani, J. Call, P. M. Kappeler, R. Palombit,
J. B. Silk, Eds. (Univ. Chicago Press, 2012), pp. 629–642.
2. C. Perrodin, C. Kayser, N. K. Logothetis, C. I. Petkov, Curr. Biol.
26. J. Sliwa, J.-R. Duhamel, O. Pascalis, S. Wirth, Proc. Natl. Acad.
Sci. U.S.A. 108, 1735–1740 (2011).
27. I. Adachi, R. R. Hampton, PLOS ONE 6, e23345 (2011).
28. D. Y. Tsao, M. S. Livingstone, Annu. Rev. Neurosci. 31, 411–437
21, 1408–1415 (2011).
(2008).
29. K. M. Gothard, F. P. Battaglia, C. A. Erickson, K. M. Spitler,
D. G. Amaral, J. Neurophysiol. 97, 1671–1683 (2007).
30. S. M. Landi, P. Viswanathan, S. Serene, W. A. Freiwald, Science
373, 581–585 (2021).
31. J. Munuera, M. Rigotti, C. D. Salzman, Nat. Neurosci. 21,
415–423 (2018).
32. W. A. Freiwald, Curr. Opin. Neurobiol. 65, 49–58 (2020).
33. T. Tyree, M. Metke, C. Miller, Cross-modal representation of
identity in primate hippocampus, Dataset, Dryad (2022).
AC KNOWLED GME NTS
We thank H. Courellis for assistance with data collection and
D. Leopold for comments on a previous version of this manuscript.
Funding: This study was supported by National Institutes of
Health grant R01 NS109294 (to C.T.M.) and National Institutes of
Health grant R01 DC012087 (to C.T.M.). Author contributions:
Conceptualization: C.T.M.; Methodology: T.J.T., M.M., and C.T.M.;
Investigation: M.M. and T.J.T.; Analysis: T.J.T.; Visualization:
T.J.T.; Funding acquisition: C.T.M.; Supervision: C.T.M.; Writing –
original draft: T.J.T. and C.T.M.; Writing – review & editing:
T.J.T. and C.T.M. Competing interests: The authors declare no
competing interests. Data and materials availability: All data
are available in the manuscript or the supplementary materials
or are deposited at Dryad (33). License information: Copyright ©
2023 the authors, some rights reserved; exclusive licensee
American Association for the Advancement of Science. No claim to
original US government works. https://www.science.org/about/
science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adf0460
Materials and Methods
Figs. S1 to S20
Table S1
References (34–38)
MDAR Reproducibility Checklist
Submitted 25 September 2022; resubmitted 30 May 2023
Accepted 1 September 2023
10.1126/science.adf0460
Tyree et al., Science 382, 417–423 (2023)
27 October 2023
7 of 7
|
10.1126_science.adf0752
|
RES EARCH
PLANT SCIENCE
Brassinosteroid coordinates cell layer interactions in
plants via cell wall and tissue mechanics
Robert Kelly-Bellow1*†, Karen Lee1†, Richard Kennaway1, J. Elaine Barclay1, Annabel Whibley1,
Claire Bushell1, Jamie Spooner1, Man Yu1, Paul Brett2, Baldeep Kular2, Shujing Cheng3,
Jinfang Chu3,4, Ting Xu5, Brendan Lane6, James Fitzsimons7, Yongbiao Xue5, Richard S. Smith6*,
Christopher D. Whitewoods1,7*, Enrico Coen1*
Growth coordination between cell layers is essential for development of most multicellular organisms.
Coordination may be mediated by molecular signaling and/or mechanical connectivity between cells,
but how genes modify mechanical interactions between layers is unknown. Here we show that genes
driving brassinosteroid synthesis promote growth of internal tissue, at least in part, by reducing
mechanical epidermal constraint. We identified a brassinosteroid-deficient dwarf mutant in the aquatic
plant Utricularia gibba with twisted internal tissue, likely caused by mechanical constraint from a slow-
growing epidermis. We tested this hypothesis by showing that a brassinosteroid mutant in Arabidopsis
enhances epidermal crack formation, indicative of increased tissue stress. We propose that by
remodeling cell walls, brassinosteroids reduce epidermal constraint, showing how genes can control
growth coordination between layers by means of mechanics.
produce axial growth. There are no growth
conflicts between cells. However, if the epi-
dermal walls (Fig. 1B, purple) yield less to
turgor (e.g., because they are thicker or less
extensible than inner walls), an epidermal
growth constraint is generated, and load is
transferred from inner to outer walls. Each
cell experiences mechanical stresses caused by
the cell’s own turgor (cell-autonomous stresses)
and by mechanical effects from surrounding
tissue (non–cell-autonomous stresses) (7) called
tissue stresses (5). Whereas the cell-autonomous
stress is always tensile, by our definition, tis-
sue stresses can be tensile or compressive; the
epidermis is under tissue tension (Fig. 1B, di-
vergent red arrows), whereas internal regions
are under tissue compression (convergent blue
arrows).
Tissue stresses can be revealed by immedi-
ate outward recurvature of median slices through
internodes or by the formation of epidermal
cracks when adhesion between cells is weak-
ened (6, 8, 9). They can be quantified by stretch-
ing detached epidermal tissue to the point that
its original length is restored (10, 11). However,
little is known about how tissue stresses are
controlled genetically and thus the role they
may play in non–cell-autonomous gene action.
In this study, we addressed this problem through
the analysis of dwarf mutants in the aquatic
plant Utricularia gibba and the terrestrial
plant Arabidopsis thaliana.
U. gibba dwarf has twisted internal tissue
U. gibba is a carnivorous plant with a spiral
vegetative growing tip, comprising an apex
that produces stolons bearing filiform leaves
and traps (12) (Fig. 2A). The stolons and leaves
have internal air spaces that allow the plant to
float just below the water surface. To obtain
developmental mutants in U. gibba, we carried
out ethyl methanesulfonate mutagenesis.
Obtaining large numbers of progeny proved
difficult because of poor seed set and germi-
nation rates. Rather than mutagenizing seed,
we therefore mutagenized small stolon explants
and grew each on to flowering (see methods in
the supplementary materials for details). M1
seed was collected from 441 explants and gave
M any multicellular organisms are formed
from multiple cell layers, raising the
question of how growth is coordinated
between layers to produce an integrated
final form. In plants, evidence from
genetic chimeras and from layer-specific modi-
fication of gene function shows that genes
active in one layer can act nonautonomously
to influence growth in other layers (1–4).
Nonautonomy could be explained through
chemical signaling between layers and/or
mechanical interactions.
Mechanics may act nonautonomously through
the generation of tissue stresses (5), as demon-
strated experimentally by Hofmeister more
than 150 years ago (6). To understand the
origin of tissue stresses, consider a cylindrical
tissue in which cells are tightly stuck together,
with all cells having the same size, turgor, wall
material properties, and wall thickness (Fig.
1A). If cell walls are anisotropic, such that they
yield more readily in the vertical orientation,
the vertical component of turgor forces in each
cell can cause stresses (highlighted for three
cells with black double-headed arrows) that
1Department of Cell and Developmental Biology, John Innes
Centre, Norwich NR4 7UH, UK. 2Department of Biochemistry
and Metabolism, John Innes Centre, Norwich NR4 7UH, UK.
3National Centre for Plant Gene Research (Beijing), Institute
of Genetics and Developmental Biology, Chinese Academy of
Sciences, Beijing 100101, China. 4College of Advanced
Agricultural Sciences, University of Chinese Academy of
Sciences, Beijing 100039, China. 5State Key Laboratory of
Plant Cell and Chromosome Engineering, Institute of
Genetics and Developmental Biology, Chinese Academy of
Sciences, Beijing 100101, China. 6Department of
Computational and Systems Biology, John Innes Centre,
Norwich NR4 7UH, UK. 7Sainsbury Laboratory, University of
Cambridge, Cambridge CB2 1LR, UK.
*Corresponding author. Email: robert.bellow@jic.ac.uk (R.K.-B.);
richard.smith@jic.ac.uk (R.S.S.); chris.whitewoods@slcu.cam.ac.uk
(C.D.W.); enrico.coen@jic.ac.uk (E.C.)
†These authors contributed equally to this work.
Fig. 1. Origin of tissue stresses. (A) Cross section of a stem, with epidermal cells in gray and cell walls
in black. All cell walls have the same material properties. Vertical component of turgor forces autonomous to
each cell causes stresses (highlighted for three cells with black double-headed arrows) that produce
axial growth. There are no growth conflicts between cells. (B) If the epidermal walls (purple) yield less to
turgor, an epidermal growth constraint is generated. Each cell now experiences two types of stress: cell-
autonomous stress caused by the cell’s own turgor (black double-headed arrows) and tissue stress caused
by mechanical effects from surrounding tissue. Tissue stresses can be tensile (divergent red arrows) or
compressive (convergent blue arrows).
Kelly-Bellow et al., Science 380, 1275–1281 (2023)
23 June 2023
1 of 7
RES EARCH | R E S E A R C H A R T I C L E
M2 phenotypes including altered traps, absent
traps, reduced leaf and stolon growth, long
flower spurs, spiky leaves, multiple traps on
leaves, and fasciation. One M2 family con-
tained two dwarf plants, and self-seed from a
wild-type sib gave 37 wild-type, 9 dwarf, and
3 extreme-dwarf plants (Fig. 2, A to C), con-
sistent with segregation of two recessive muta-
tions: dwarf and enhancer of dwarf.
Both the dwarf and extreme-dwarf plants
had short internodes, short leaves, and small
traps (Fig. 2, A to D). To follow their devel-
opment, we numbered internodes sequentially
relative to the spiral apex, with internode 1 cor-
responding to the first clearly visible in-
ternode to emerge from the apex (fig. S1).
Wild-type internode length increased until
about internode 4, after which it plateaued
to give a mature internode length of ~2 mm
(Fig. 2E). By contrast, dwarf and extreme-dwarf
plants exhibited very little growth after inter-
node 1, generating mature internode lengths
of ~0.7 and ~0.3 mm, respectively (Fig. 2E).
Epidermal cells of mutant stolons were shorter
and smaller than those of wild type (Fig. 2, F
to J). Measurements of cell lengths parallel
to the stolon axis indicated that 70% of the
reduction in dwarf internode length was
caused by reduced longitudinal growth after
cell division arrest (fig. S2A). Further reduc-
tion in internode length in extreme dwarf
was caused by reduced growth before di-
vision arrest. In addition to reduced internode
length, both dwarf and extreme dwarf ex-
hibited a significant increase in stolon cir-
cumference and number of epidermal cells
in transverse sections compared with wild
type, indicating increased radial and circum-
ferential growth before division arrest (fig.
S2, B and C).
We next determined the phenotype of inter-
nal tissues. Wild-type stolons had a cylindrical
epidermis (Fig. 3C, purple) connected by five
or six straight “blades” (cyan) to an axial cylinder
of large cells (yellow) surrounding a vascular
bundle (orange), with air spaces (magenta) be-
tween the blades (Fig. 3, A to C). Dwarf stolons
had smaller air spaces, twisted blades, and a
sinuous contorted vascular bundle (Fig. 3, D
to F). Extreme-dwarf stolons had smaller air
spaces and less-twisted vasculature than dwarf
plants (Fig. 3, G to I). Both dwarf and extreme-
dwarf plants sank in water, presumably be-
cause of their reduced air spaces.
The twisted internal tissue of the dwarf
plants might be caused by a contorted pattern
of early vascular and blade cell-type specifica-
tion or by altered tissue growth after specifi-
cation has occurred. To distinguish between
these possibilities, we determined the devel-
opmental timing of the twisted phenotype in
dwarf plants. Straight vasculature cell types
surrounded by blade and air spaces were
evident at internodes 0 in dwarf (Fig. 3J),
Fig. 2. External phenotype of U. gibba wild type and dwarf mutants. (A to C) U. gibba vegetative
plants comprise a spiral apex (ap), filiform leaves (l), stolons (st), and traps (t). (A) Wild type. (B) Dwarf.
(C) Extreme dwarf. Scale bar, 1 mm. (D) Violin plots of wild type [(cid:1)x = 3.07 mm ± 0.11 (SEM), n = 10], dwarf
[(cid:1)x = 0.72 mm ± 0.02 (SEM), n = 10], and extreme dwarf [(cid:1)x = 0.29 mm ± 0.01 (SEM), n = 13] mature internode
lengths from plants grown in continuous culture. Block indicates interquartile range, and horizontal line
the mean. Both mutants have reduced lengths compared with wild type (***P < 0.001). (E) Internode
lengths from growing explants of wild type (red), dwarf (orange) and extreme dwarf (blue) plotted
against internode number. Dashed line shows mean from internode 10 onwards (n > 4 plants). (F to H) Cell
areas in mature stolons of wild type (F), dwarf (G), and extreme dwarf (H), color coded according to the
color scale shown at the left of (F). Scale bar, 100 mm. (I and J) Violin plots of cell area (I) and cell anisotropy
[cell maximum length/(cell maximum length + cell minimum length)] (J) of mature stolons of wild type
(n = 1817 cells from eight explants), dwarf (n = 2289 cells from five explants), and extreme dwarf (n = 1494
cells from four explants). Both mutants have significantly lower values than wild type.
as in wild type (fig. S3A). Twisted vascular
tissue in dwarf plants was only observed from
internode 4 onward (Fig. 3K, orange line). Con-
tortions of the blade were evident in dwarf
mutants earlier, at internode 1, as tissue strips
running perpendicular to the vascular axis in
longitudinal sections (Fig. 3J, blue arrow).
These blade contortions were only seen after
air spaces had formed. Thus, contortion and
twisting of internal tissue in dwarf plants
arose through altered growth after cell type
specification and air space formation, lead-
ing to excess vascular length compared with
epidermal length (fig. S2D). In extreme-dwarf
plants, which showed little contortion, orga-
nized vasculature and surrounding tissue were
evident in early internodes, but air spaces
were not (fig. S3B).
Twisted dwarf phenotype explained by
epidermal constraint
To evaluate hypotheses that might account
for both the internal twisting and shortened
internode length of dwarf mutants, we modeled
tissue growth using continuum mechanics.
For these purposes, we distinguished between
two types of regional growth: specified and
resultant (13). Specified growth corresponds
to the growth driven by a cell’s own turgor,
in mechanical isolation from other cells. Re-
sultant growth corresponds to the growth
generated when tissue stresses, which act
non–cell-autonomously, are also factored in.
Computational models allow tissue stresses
and resultant growth to be calculated from
an input pattern of specified growth rates and
orientations.
We modeled a small length of U. gibba
stolon as a stiff cylindrical epidermal sheet
connected by blades to an axial core (Fig. 4,
A to C, and fig. S9, A and B). Specified growth
was oriented parallel to an axial (initially vertical)
polarity field (Fig. 4A, arrows). To reduce
boundary effects, each stolon end was con-
strained to remain flat and horizontal.
If all regions had the same specified growth
rate, the cylinder elongated without genera-
tion of tissue stresses or twisting of internal
Kelly-Bellow et al., Science 380, 1275–1281 (2023)
23 June 2023
2 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Internal phenotype of U. gibba wild type and dwarf mutants. Wild type (A to C), dwarf (D to F),
and extreme dwarf (G to I) longitudinal confocal sections [(A), (D), and (G)], freeze-fracture scanning
electron microscopy [(B), (E), and (H)], and toluidine blue–stained transverse sections [(C), (F), and
(I)]. Arrows and cells are color coded purple for epidermis (e), cyan for blades (b), magenta for air
spaces (s), yellow for axial core (a), and orange for vasculature (v). Scale bars, 50 mm. (J and K) Confocal
Z-slice of early dwarf internodes 0 to 4. Scale bars, 100 mm.
tissue (Fig. 4, D to F). If specified growth rate
was set to zero in the epidermis, the epidermal
constraint caused a dwarf phenotype (Fig. 4, G
and H). Tissue tension was generated in the
epidermis (Fig. 4I, red), and tissue compres-
sion was generated in the blades and core (Fig.
4, I and J, blue). The tissue tension caused the
epidermis to grow to some extent, despite
its specified growth rate being zero [com-
pare epidermal resultant growth rate (Fig.
4K) with specified growth rate (Fig. 4G)].
Conversely, tissue compression in blades and
core caused a lower resultant growth rate than
specified (compare Fig. 4L with Fig. 4H), but
higher than in the epidermis (zero). The tissue
stresses also caused twisting of blades and
core (Fig. 4, M to P, and movie S1). Thus, re-
duced specified growth rate in the epidermis
captured both the dwarf phenotype and inter-
nal contortion.
Twisting of the axial core still occurred
when blades were removed from a middle
segment of the cylinder (fig. S4, A to F),
showing that tissue compression could be
transmitted to the core from above and below.
No twisting occurred if the cylinder was solid
(fig. S4, G to K), showing that air spaces were
needed to accommodate buckling, and ac-
counting for the reduced twisting observed
in extreme dwarf plants. Restricting specified
growth to the axial core led to a dwarf pheno-
type and a sinuous core but little twisting of
the blades (fig. S4, L and M). Restricting spec-
ified growth to the blades gave a dwarf pheno-
type with twisted blades but little twisting
of the core (fig. S4, N to P). Radial specified
growth of the blades led to blade twisting,
but cylinder elongation and axial core straight-
ness were not affected (fig. S4, Q to V). Thus,
both the dwarfism and internal axial and blade
twisting could be most readily accounted for
by reduced specified growth rate of the epi-
dermis alone.
DWARF encodes a brassinosteroid
biosynthetic enzyme
To understand the molecular basis of the
dwarf mutant, we sequenced the wild-type
progenitor and 33 wild-type, 10 dwarf, and 3
extreme-dwarf segregants. Only one single-
nucleotide polymorphism (SNP) was absent
from the progenitor, heterozygous or absent
in wild-type segregants, and homozygous in
all mutants, indicating that it was located in
the DWARF gene. Extreme-dwarf plants carried
four additional SNPs absent from the progeni-
tor (table S1) that were candidate mutations in
ENHANCER OF DWARF. Plants homozygous
for enhancer of dwarf and heterozygous or
homozygous for DWARF were scored as wild
type, suggesting that the enhancer of dwarf
mutation alone did not have a strong pheno-
typic effect. However, the mutation may have
caused a subtle phenotype that we missed
when initially scoring the families.
The DWARF SNP introduced an early stop
codon in a gene encoding a cytochrome P450
90B1 enzyme, which catalyzes the C22-alpha-
hydroxylation step in the brassinosteroid bio-
synthesis pathway (14). This gene is homologous
to DWARF4 (DWF4) in Arabidopsis, which
affects cell area and cell anisotropy in a similar
way to U. gibba DWARF (15, 16). Brassinosteroid
precursors after the C22-alpha-hydroxylation
step were undetectable or at a low level in
dwarf mutants, whereas a precursor before the
step was present (fig. S5). Inhibiting brassi-
nosteroid biosynthesis in wild type using brassi-
nazole led to short stolons, smaller cells, and
contorted vasculature, similar to dwarf mutants
(fig. S6). Adding brassinosteroid, by growing
mutants in epibrassinolide, rescued dwarf
plants and partially rescued extreme dwarf
plants (fig. S6). Thus, DWARF likely encodes a
brassinosteroid biosynthesis gene.
To determine the timing of brassinosteroid
action, we tracked dwarf stolons after treat-
ment with epibrassinolide (Fig. 4S). Inter-
nodes that were not readily visible when the
treatment began, because they were concealed
Kelly-Bellow et al., Science 380, 1275–1281 (2023)
23 June 2023
3 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Simulations of U. gibba wild type and dwarf mutant and timing of brassinosteroid action.
(A) Initial state for wild-type and dwarf mutant models [epidermis (purple), blades (cyan), and axial core (yellow)].
Arrows indicate polarity. (B) Initial state without epidermis. (C) Transverse slice of (A). (D) Final state of wild-
type simulation, color coded for specified growth rate, which is uniformly high. (E) Same as (D), but color coded for
tissue type. (F) Same as (E), with epidermis removed. (G) Final state for simulation of dwarf mutant, color coded
for specified growth rate, which is excluded from the epidermis and gives reduced elongation. (I) Same as (G), but
color coded for tissue stresses. (K) Same as (G), but color coded for resultant growth rate. (M) Same as (G),
but color coded for tissue type. (H, J, L, and N) Same as (G), (I), (K), and (M), respectively, but with epidermis
clipped away. (O) Transverse slice of (M). (P) Same as (M), but showing the axial core only. (Q) Color scale for
specified and resultant growth rates, in strain per time step of simulation. (R) Color scale for tissue stresses, with red
indicating tension (t) and blue indicating compression (c). (S) A dwarf explant imaged on days 0 and 14 after
treatment with 0.01 mM epibrassinolide. Internode numbers labeled on day 14. Scale bar, 5 mm. (T) Average
internode lengths of dwarf explants before treatment (day 0, solid orange line) and at 14 days (day 14, solid brown
line) (n ≥ 10 explants). Day 14 internode −10 to 0 lengths were not significantly different from the mean of Fig. 2E
(red dashed line). Error bars show SEM.
within the spiral vegetative shoot tip or had
not yet initiated, were assigned consecutive
negative numbers, starting from 0. These in-
ternodes grew to a length similar to those of
mature wild type (Fig. 4T). Internodes 1 to 5
also showed a significant length increase in
response to treatment (P < 0.05), with the
magnitude of the increase declining with in-
ternode number. Thus, brassinosteroid likely
acts from around internode 0, when cell divi-
sion is nearing arrest, until around internode 5,
by which stage cell elongation has arrested
in wild type. However, we cannot rule out the
possibility that internodes above 5 are imper-
meable to exogenous brassinosteroid.
Arabidopsis brassinosteroid mutant has
elevated tissue stresses
Our experimental and modeling results indi-
cate that brassinosteroid promotes U. gibba
stolon growth starting just before cell division
arrest by counteracting an epidermal con-
straint, thus reducing tissue stresses. If gener-
ally applicable, this hypothesis predicts that
Arabidopsis dwf4 mutants should also exhibit
elevated tissue stresses. However, the effect
of these stresses might be masked because
Arabidopsis stems are solid and therefore lack
air spaces to accommodate buckling (fig. S4, G
to K). To determine whether tissue stresses are
enhanced in dwf4 mutants, we therefore ex-
ploited the quasimodo2-1 (qua2-1) mutation,
which weakens cell-cell adhesion (17). As illus-
trated in Fig. 1B (red arrows), tissue stresses
generated a force that acts to pull epidermal
cells apart. In wild-type Arabidopsis, cell-cell
adhesion is strong enough to resist this force,
but in qua2-1 mutants, epidermal cracks are
observed between cells in dark-grown hypo-
cotyls, confirming that epidermal tissue tension
is present (9, 18). If brassinosteroid normally
acts to reduce tissue tension, cracks are pre-
dicted to be exacerbated in qua2-1 dwf4 double
mutants, or in qua2-1 mutants treated with a
brassinosteroid inhibitor.
To test these predictions, we intercrossed
dwf4 and qua2-1 mutant lines. About 1/16
(58/885) of the F2 dark-grown seedlings ex-
hibited a distinctive novel phenotype: Hypocotyls
were dwarf and seemed devoid of epidermis
(Fig. 5, A and B), unlike qua2-1 single mutants,
which showed small epidermal cracks at a
similar stage (Fig. 5C). To clarify the devel-
opmental origin of the double-mutant pheno-
type, we imaged seedlings at different days
after germination. Seedlings of dwf4 qua2-1 were
indistinguishable from dwf4 seedlings until
~3 days after stratification, when wide cracks
appeared in the double mutant (Fig. 5D). These
cracks were much larger than those observed in
qua2-1 single mutants at the same stage (Fig.
5E). By 5 days after stratification, the cracks in
dwf4 qua2-1 had enlarged to the extent that
much of the epidermis was no longer evident
(Fig. 5B). Crack formation was also enhanced
when qua2-1 single mutants were grown in
the presence of a brassinosteroid inhibitor,
brassinazole (Fig. 5, F to H). These results thus
support the hypothesis that brassinosteroid
promotes stem growth by counteracting an
epidermal constraint.
To further validate this interpretation, we
modeled the growth of a solid cylinder with a
stiff epidermis in which cracks can form when
tension exceeds a threshold value. Uniform
specified growth rate gave elongation without
Kelly-Bellow et al., Science 380, 1275–1281 (2023)
23 June 2023
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
Kelly-Bellow et al., Science 380, 1275–1281 (2023)
23 June 2023
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. Effect of reduced brassinosteroid on epidermal cracks in
Arabidopsis qua2-1 and explanatory computational models. (A to C)
Confocal images 5 days after stratification. (A) qua2-1 dwf4 double mutant.
(B) qua2-1 dwf4 shown in (A), with cells artificially colored for clarity (purple
indicates epidermal cells, and cyan indicates interior cells). It is possible that the
cyan-colored cells include some disorganized epidermal cells. Close-up shown on
the right. (C) qua2-1 single mutant with a close up of a region with cracks.
(D and E) Confocal images 3 days after stratification. (D) qua2-1 dwf4 double
mutant. (E) qua2-1 single mutant. Scale bars, 100 mm. Purple arrows highlight
curled free ends of epidermal cells. (F to H) Phenotypes of qua2-1 Arabidopsis
hypocotyls without or with treatment with brassinosteroid inhibitor [1 mM
brassinazole (BRZ)]. [(F) and (G)] Confocal images of seedlings after 9 days
growth in the dark. (F) qua2-1 with close up of a region with cracks selected for
magnification. (G) qua2-1 grown on 1 mM BRZ with close up. Curved cells at
crack boundaries are indicated with an arrow. Epidermal cells are in purple,
internal cells are in cyan. Scale bars, 100 mm except in (F) (gray scale bar),
which is 1000 mm. (H) Violin plots of crack widths measured by number of cells.
Mean crack width covers more cell files in qua2-1 + BRZ [(cid:1)x = 4.089 ± 0.38
(SEM)] than qua2-1 untreated [(cid:1)x = 1.611 ± 0.14 (SEM)], but not significantly
more (P = 0.0531, untreated hypocotyls n = 74 cracks from two plants,
BRZ-treated hypocotyls n = 83 cracks from six plants). (I to L) Tissue-level
computer simulations. (I) Initial state, with epidermis in purple and inner regions
in cyan. Arrows indicate polarity. (J) Final state with all regions having the same
specified growth parallel to polarity, leading to elongation without epidermal
cracking. (K) Final state with reduced specified growth in the epidermis leads to
a shorter cylinder and epidermal cracks. (L) Longitudinal section through (K),
showing longitudinal tissue tension (t) in red and compression (c) in blue.
(M to Z) Cellular-level computer simulations. (M) Initial state, with outer epidermal
wall in dark purple, epidermis in purple, and inner tissue in cyan. [(N) to (Y)] Final
state. [(N) to (Q)] All cell walls have the same thickness and material properties. [(R)
to (U)] Outer epidermal wall is 10 times thicker. [(V) to (Y)] Outer wall is 10 times
thicker, with higher extensibility and reduced yield threshold. [(Q), (U), and (Y)]
Epidermal fracture introduced at an early stage. White arrow in (Q) indicates position
of fracture. Fracture is two cells wide for (Q) and (U) and eight cells wide for (Y).
[(N), (R), and (V)] Specified growth rate. [(O), (S), (W), (Q), (U), and (Y)] Tissue
stresses. [(P), (T), and (X)] Resultant longitudinal wall stresses. (Z) Color scales. (Top)
Specified growth rate (7% per time step). (Middle) Tissue stresses (23). (Bottom)
Longitudinal wall stresses (23). Scale bar, 50 mm.
tissue stresses or cracks (Fig. 5, I and J), whereas
low epidermal specified growth led to reduced
elongation, elevated tissue stresses, and crack
formation (Fig. 5, K and L).
Release from epidermal constraint by
wall remodeling
The above results raise the question of how
brassinosteroid reduces epidermal constraint.
The most obvious source of epidermal con-
straint is the thick outer wall of the epidermal
cells (6). A constraining outer wall is also con-
sistent with the concave shape of the outer
wall in epidermal cells released by crack for-
mation (Fig. 5, C to E, purple arrows). Outer
epidermal walls of U. gibba dwarf stolons
were about two to three times thicker than
inner walls at internode 1 (fig. S7, D and F),
by which time growth had ceased (Fig. 2E).
Outer epidermal walls of A. thaliana dwf4
dark-grown hypocotyls were ~20 times thicker
than inner walls at 4 days after stratification
(fig. S7, C and E), by which time growth had
largely ceased (fig. S8). Similar wall thicknesses
were observed for wild types at comparable
stages (fig. S7, A, B, E, and F), even though
growth continued afterwards, suggesting that
brassinosteroid does not reduce epidermal
constraint primarily by altering wall thickness.
A possible mechanism for reduction of
epidermal constraint is wall loosening. Brassi-
nosteroids promote hypocotyl elongation with-
in 6 hours of application through increased
wall relaxation properties (i.e., wall loosening)
(19, 20), possibly through phosphorylation of
plasma membrane proton adenosine triphos-
phatase (21). To explore the possible contribu-
tion of wall loosening, we modeled hypocotyl
tissue growth at the cellular level. A segment
of hypocotyl was modeled as a vertical cylinder
of tightly attached cells of similar size and
under the same turgor (Fig. 5M). Wall growth
by means of creep (22) was simulated by con-
verting a proportion of reversible elastic wall
strain, above a yield threshold, into irreversible
strain at each time step. The proportion corre-
sponded to the extensibility of the wall. Walls
were seven times stiffer (a sevenfold larger
Young’s modulus) in the transverse compared
with longitudinal orientation, leading to ver-
tical specified growth.
If all cell walls had the same material prop-
erties, the cylinder elongated with uniform
specified growth rates (Fig. 5N), low tissue
stresses (Fig. 5O), and uniform longitudinal
wall stresses (Fig. 5P). Introducing an epi-
dermal fracture caused rounding of the cell
ends but did not cause further cell separa-
tion (Fig. 5Q, arrow indicates the position of
fracture). Setting the outer epidermal wall to be
10 times thicker than the inner walls lowered
the epidermal specified growth rate (Fig. 5R).
The growth constraint generated longitudinal
epidermal tissue tension and internal tissue
compression (Fig. 5S). Resultant longitudinal
wall stresses were uniform but lower than if
all cell walls had the same material properties
(compare Fig. 5T with Fig. 5P). The cylinder
therefore grew less, capturing the dwf4 pheno-
type. Introducing a wide epidermal fracture
(eight cells wide) released epidermal cell ends
to peel back (Fig. 5U and movie S2), capturing
the dwf4 qua2-1 phenotype (Fig. 5D).
To simulate wild type, the thick outer wall
was loosened by increasing its extensibility
and reducing its yield threshold. This modi-
fication increased the epidermal specified
growth rate (compare Fig. 5V with Fig. 5R),
lowered tissue stresses (compare Fig. 5W with
Fig. 5S), and raised longitudinal wall stresses
of internal tissue (compare Fig. 5X with Fig. 5T).
The cylinder therefore elongated more than in
the dwf4 simulation, capturing the wild-type
phenotype. Introducing a narrow epidermal
fracture (two cells wide) led to released epi-
dermal cell ends peeling back (they curve
because the thick outer wall grows more slowly)
(Fig. 5Y), capturing the qua2-1 phenotype (Fig.
5C). Thus, brassinosteroid likely acts, at least
in part, by loosening of the thick outer wall,
counteracting the epidermal constraint.
In addition to wall thickness, epidermal
constraint may be further enhanced by the
orientation of microfibrils, which are less
transverse in outer than in inner walls for
wild-type Arabidopsis hypocotyls (23). Brassi-
nosteroid treatment can cause microtubules
of the outer epidermal plasma membrane to
orient more transversely (24, 25). Thus, brassi-
nosteroid may reduce epidermal constraint by
remodeling the thick outer wall in two ways:
wall loosening and reducing the proportion of
longitudinally oriented microfibrils. Such an
effect on microfibril orientation might explain
why Utricularia dwarf mutants have wider
stolons (fig. S2, B and C). The differential
properties of the outer epidermal wall (e.g.,
thickness, extensibility, microfibril orientation)
may depend on cell polarity factors that confer
differences between outer and inner cell faces
(26–28). Brassinosteroids may also reduce
epidermal constraint by increasing turgor in
epidermal cells, although there is currently no
experimental evidence to support this possibility.
Conclusions
When brassinosteroid synthesis or perception
genes are expressed only in the epidermal cell
layer of Arabidopsis brassinosteroid mutants,
a near–wild-type phenotype is generated, even
though these genes are normally expressed in
both epidermis and ground tissue (4, 29). Our
results indicate that this nonautonomous effect
of epidermal brassinosteroid gene expression
on resultant growth of internal tissue involves
release of internal tissue from epidermal me-
chanical constraint, although this does not
preclude additional contributions from mo-
lecular signaling. Mechanical interactions
Kelly-Bellow et al., Science 380, 1275–1281 (2023)
23 June 2023
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
between cell layers also play a role in animal
development, such as formation of crocodile
skin cracks (30) and intestinal villi (31). Here
we show how genes may modify tissue layer
interactions by changing cellular growth prop-
erties and thus tissue stresses. Gene activity
may therefore have coordinated effects on
tissue development not only via molecular sig-
naling but also via mechanics.
RE FE RENCES AND N OT ES
1. S. Hake, B. R. Char, Genes Dev. 11, 1087–1097 (1997).
2. M. H. Frank, D. H. Chitwood, Dev. Biol. 419, 41–53 (2016).
3. R. A. Tilney-Bassett, Plant Chimeras (Edward Arnold Publishers
Ltd., 1986).
4. S. Savaldi-Goldstein, C. Peto, J. Chory, Nature 446, 199–202
(2007).
5. Z. Hejnowicz, A. Sievers, J. Exp. Bot. 46, 1035–1043
(1995).
6. U. Kutschera, K. J. Niklas, J. Plant Physiol. 164, 1395–1409
(2007).
7. F. Boudon et al., PLOS Comput. Biol. 11, e1003950 (2015).
8. W. S. Peters, A. D. Tomos, Ann. Bot. 77, 657–665 (1996).
9. S. Verger, Y. Long, A. Boudaoud, O. Hamant, eLife 7, e34460
(2018).
10. Z. Hejnowicz, A. Sievers, J. Exp. Bot. 46, 1045–1053 (1995).
11. U. Kutschera, J. Plant Physiol. 138, 460–465 (1991).
12. P. Taylor, The Genus Utricularia: A Taxonomic Monograph
(Royal Botanic Gardens, 1989).
13. E. Coen, R. Kennaway, C. Whitewoods, Development 144,
4203–4213 (2017).
14. S. Fujioka et al., Biosci. Biotechnol. Biochem. 59, 1543–1547 (1995).
15. R. Azpiroz, Y. Wu, J. C. LoCascio, K. A. Feldmann, Plant Cell 10,
219–230 (1998).
16. S. Choe et al., Plant Cell 10, 231–243 (1998).
17. S. Verger, S. Chabout, E. Gineau, G. Mouille, Development 143,
2536–2540 (2016).
18. G. Mouille et al., Plant J. 50, 605–614 (2007).
19. D. M. Zurek, D. L. Rayle, T. C. McMorris, S. D. Clouse,
Plant Physiol. 104, 505–513 (1994).
20. T. W. Wang, D. J. Cosgrove, R. N. Arteca, Plant Physiol. 101,
965–968 (1993).
21. A. Minami, K. Takahashi, S. I. Inoue, Y. Tada, T. Kinoshita,
Plant Cell Physiol. 60, 935–944 (2019).
22. E. Coen, D. J. Cosgrove, Science 379, eade8055 (2023).
23. E. F. Crowell et al., Plant Cell 23, 2592–2605 (2011).
24. X. Wang et al., Plant Cell 24, 4012–4025 (2012).
25. M. Catterou et al., Planta 212, 673–683 (2001).
26. J. Takano et al., Proc. Natl. Acad. Sci. U.S.A. 107, 5220–5225
(2010).
27. L. Łangowski, K. Růzicka, S. Naramoto, J. Kleine-Vehn, J. Friml,
Curr. Biol. 20, 904–908 (2010).
28. V. Gorelova, J. Sprakel, D. Weijers, Nat. Plants 7, 1548–1559
(2021).
29. D. M. Friedrichsen, C. A. Joazeiro, J. Li, T. Hunter, J. Chory,
Plant Physiol. 123, 1247–1256 (2000).
30. M. C. Milinkovitch et al., Science 339, 78–81 (2013).
31. A. E. Shyer et al., Science 342, 212–218 (2013).
32. R. Kennaway, JIC-Enrico-Coen/Cell_Layer_Interactions_2023:
GFtbox_Cell_Layer_Interactions_2023 (v1.0.1), Zenodo (2023);
https://doi.org/10.5281/zenodo.7913146.
ACKN OWLED GMEN TS
We thank B. and P. Steward at The Fly Trap Plants and T. Bailey
from the Carnivorous Plant Society for plants, seeds, and advice.
We thank M. Majda for qua2-1 seeds, E. Wegel and S. Lopez of
John Innes Centre (JIC) Bioimaging for help with light microscopy,
R. Wightman for help with freeze-fracture scanning electron
microscopy, L. Perkins and the JIC horticulture team for large-
scale U. gibba cultivation, G. Mosca for MorphoMechanX, and
D. Bradley for critical reading of the manuscript. Funding: This
work was supported by a European Research Council grant
(323028-CarnoMorph) and Biotechnology and Biological Sciences
Research Council grants (BBS/E/J/000PR9787, BB/M023117/1,
BB/L008920/1, and BB/X01102X/1) awarded to E.C. Author
contributions: Biological experiments, data analysis, and
conceptualization: R.K.-B., K.L., J.E.B., M.Y., P.B., B.K., S.C., J.C.,
T.X., B.L., J.F., Y.X., R.S., and C.D.W. Computational modeling: R.K.,
R.S.S., and E.C. Software development: R.K. and R.S.S., Development
of U. gibba resources: K.L., C.B., J.S., M.Y., and C.D.W. Bioinformatic
analysis: R.K.-B. and A.W. Supervision, funding acquisition, and
conceptualization: E.C. Competing interests: The authors declare
that they have no competing interests. Data and materials
availability: All data are available in the manuscript or the
supplementary materials or are deposited at Zenodo (32). License
information: Copyright © 2023 the authors, some rights reserved;
exclusive licensee American Association for the Advancement of
Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adf0752
Materials and Methods
Supplementary Text
Figs. S1 to S9
Tables S1 and S2
References (33–53)
MDAR Reproducibility Checklist
Movies S1 and S2
View/request a protocol for this paper from Bio-protocol.
Submitted 5 October 2022; accepted 18 May 2023
10.1126/science.adf0752
Kelly-Bellow et al., Science 380, 1275–1281 (2023)
23 June 2023
7 of 7
|
10.1126_science.adf1506
|
Corrected 21 August 2023. See full text.
RES EARCH
R E S E A R C H A R T I C L E
◥
TOPOLOGICAL MATTER
Quantum metric nonlinear Hall effect in a topological
antiferromagnetic heterostructure
Anyuan Gao1, Yu-Fei Liu1,2, Jian-Xiang Qiu1, Barun Ghosh3, Thaís V. Trevisan4,5, Yugo Onishi6, Chaowei Hu7,
Tiema Qian7, Hung-Ju Tien8, Shao-Wen Chen2, Mengqi Huang9, Damien Bérubé1, Houchen Li1,
Christian Tzschaschel1, Thao Dinh1,2, Zhe Sun1,10, Sheng-Chin Ho1, Shang-Wei Lien8, Bahadur Singh11,
Kenji Watanabe12, Takashi Taniguchi12, David C. Bell13,14, Hsin Lin15, Tay-Rong Chang8,16,17, Chunhui Rita Du9,
Arun Bansil3, Liang Fu6, Ni Ni7, Peter P. Orth4,5, Qiong Ma10,18, Su-Yang Xu1*
Quantum geometry in condensed-matter physics has two components: the real part quantum metric
and the imaginary part Berry curvature. Whereas the effects of Berry curvature have been observed
through phenomena such as the quantum Hall effect in two-dimensional electron gases and the
anomalous Hall effect (AHE) in ferromagnets, the quantum metric has rarely been explored. Here, we
report a nonlinear Hall effect induced by the quantum metric dipole by interfacing even-layered
MnBi2Te4 with black phosphorus. The quantum metric nonlinear Hall effect switches direction upon
reversing the antiferromagnetic (AFM) spins and exhibits distinct scaling that is independent of the
scattering time. Our results open the door to discovering quantum metric responses predicted
theoretically and pave the way for applications that bridge nonlinear electronics with AFM spintronics.
N onlinearities are crucial in many branches
of physics, ranging from atomic physics
to condensed-matter and complex dy-
namical systems. Nonlinear electrical
transport is the foundation of applica-
tions such as rectification and wave mixing.
Classically, the most well-known nonlinear
device is a PN diode (Fig. 1A). Noncentrosym-
metric polar materials (Fig. 1B) are similar to
PN diodes as they both possess an electric
dipole. They have recently been discovered to
show intrinsic nonlinear electrical transport,
which may not only lead to new nonlinear
applications but also provide a powerful probe
of the quantum geometry of the conduction
1Department of Chemistry and Chemical Biology, Harvard
University, Cambridge, MA 02138, USA. 2Department of
Physics, Harvard University, Cambridge, MA 02138, USA.
3Department of Physics, Northeastern University, Boston, MA
02115, USA. 4Department of Physics and Astronomy, Iowa
State University, Ames, IA 50011, USA. 5Ames National
Laboratory, Ames, IA 50011, USA. 6Department of Physics,
Massachusetts Institute of Technology, Cambridge, MA
02139, USA. 7Department of Physics and Astronomy and
California NanoSystems Institute, University of California, Los
Angeles, Los Angeles, CA 90095, USA. 8Department of
Physics, National Cheng Kung University, Tainan 701, Taiwan.
9Department of Physics, University of California San Diego,
La Jolla, CA, 92093, USA. 10Department of Physics,
Boston College, Chestnut Hill, MA, USA. 11Department of
Condensed Matter Physics and Materials Science, Tata
Institute of Fundamental Research, Colaba, Mumbai, India.
12International Center for Materials Nanoarchitectonics,
National Institute for Materials Science, 1-1 Namiki, Tsukuba
305-0044, Japan. 13Harvard John A. Paulson School of
Engineering and Applied Sciences, Harvard University,
Cambridge, MA 02138, USA. 14Center for Nanoscale
Systems, Harvard University, Cambridge, MA 02138, USA.
15Institute of Physics, Academia Sinica, Taipei 11529, Taiwan.
16Center for Quantum Frontiers of Research and Technology
(QFort), Tainan 701, Taiwan. 17Physics Division National
Center for Theoretical Sciences, Taipei 10617, Taiwan.
18Canadian Institute for Advanced Research, Toronto, Canada.
*Corresponding author. Email: suyangxu@fas.harvard.edu
electrons (1–16). Broadly, the nonlinear trans-
port in both diodes (Fig. 1A) and noncentro-
symmetric conductors (Fig. 1B) arises from an
inversion asymmetric charge distributions
(e.g., an electric dipole). Because the electron
has another fundamental degree of freedom,
spin, an interesting question is whether spin
can also lead to an electrical nonlinearity even
in a centrosymmetric lattice. One ideal platform
is the class of parity-inversion time-reversal
(PT)–symmetric antiferromagnets (AFMs) (17),
where only the spins feature a noncentrosym-
metric distribution (Fig. 1C).
Important clues can be drawn from previous
optical experiments, in which optical second-
harmonic generation (SHG) has been observed
in the PT-symmetric AFMs, including Cr2O3
and CrI3 (18). Nevertheless, nonlinear transport
is distinct because it directly probes the Fermi
surface electrons and in many cases their geo-
metrical properties (1, 2). As such, it enables a
probe of the quantum geometry (1, 2) of the
topological bands at the Fermi level.
The quantum geometry has two parts, T ¼
g-i=2W (1) (T is the quantum metric tensor). The
imaginary part is the well-known Berry curva-
(cid:3)
(cid:2)
hunji@ka umihumji@kb uni
ture Wab ¼ (cid:2)2Im
,
X
m≠n
which describes the curvature of wave func-
tion in Hilbert space (n, m are band indices
and a, b are spatial directions). Berry curva-
ture has been identified as the source of many
unusual electronic and optical responses. By
contrast, the real part is the quantum metric,
(cid:3)
(cid:2)
hunji@ka umihumji@kb uni
gab ¼ Re
, which
X
m≠n
measures the distance between neighboring
Bloch wave functions in Hilbert space [i.e.,
see section IV.1 of (19)]. Although equally impor-
tant, the quantum metric is much less explored.
There have been a few examples, including the
orbital magnetic susceptibilities (20), a third-
order Hall effect (13), and the quantum metric
in atomic physics (21). However, the way quan-
tum metric regulates electronic motion remains
largely unknown. Recently, theories have pre-
dicted a wide range of exotic quantum metric
responses (20, 22–34).
Here, we report the observation of the quan-
tum metric dipole–induced second-order anom-
alous Hall effect (AHE) (20, 22–25). In past
decades, there have been numerous studies of
the AHE (both linear and nonlinear) induced
by Berry curvature. Recent theoretical studies,
however, predict that the quantum metric can
also lead to AHE, thereby advancing our under-
standing of the fundamental origin of the AHE.
Distinct from the Berry curvature–induced
AHEs, this effect is predicted to exist in the
PT-symmetric AFMs (20, 22–25), where PT
forces the Berry curvature to vanish identically
but the effects of quantum metric can manifest
strongly. We design and fabricate a feasible
material platform and demonstrate the ap-
pearance of the quantum metric nonlinear
Hall effect (NHE). To conceptualize this type
of NHE, we draw comparison with the well-
known AHE in ferromagnetic metals (35), where
Berry curvature leads to the anomalous veloc-
ity and therefore the AHE, vanomalousº∫
Ejj(cid:3)
W (Ejj is the in-plane source-drain electric field).
By contrast, in a PT-symmetric AFM, Berry
curvature is zero as dictated by PT. However,
a nonzero quantum metric g can induce an
anomalous velocity to second order in Ejj,
vanomalousº∫
Ejj (cid:3) ∇k (cid:3) gEjj
, as proposed
(cid:3)
(cid:5)
(cid:4)
(cid:2)
k
k
in (20). This leads to an intrinsic second-order
Hall effect. From the expression above, one
can show that this effect is nonzero only when
the system breaks both P and T. Therefore, we
need PT-symmetric AFM conductors with a
large quantum metric on the Fermi surface. We
have carefully considered possible materials,
and identified two-dimensional (2D) even-
layered MnBi2Te4 (15, 36–46) as an ideal plat-
form. Even-layered MnBi2Te4 is a PT-symmetric
AFM. Moreover, its topological bands support
gate-tunable transport and a giant quantum
metric (19). However, its lattice has C3z rotational
symmetry (Fig. 1, D and E), which forces the
effect to vanish (22). To break C3z, we interface
MnBi2Te4 with black phosphorus (BP) (47).
Demonstration of rotational
symmetry breaking
We start by showing that interfacing MnBi2Te4
with BP indeed breaks its C3z rotational symme-
try. To this end, we study the directional depen-
dence of the resistance (6) of MnBi2Te4 without
and with BP. We fabricated a six-septuple-layer
Gao et al., Science 381, 181–186 (2023)
14 July 2023
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
(6SL) MnBi2Te4 device with radially distributed
electrical contacts (Device-BM1). The four-probe
resistance (T = 1.8 K) is found to be isotropic
(Fig. 1G, blue curve), consistent with the C3z
symmetry. We then stacked a BP layer (~10 nm)
onto this MnBi2Te4 sample and performed the
measurements again. The resistance develops
a clear anisotropy with a 180° periodicity (Fig.
1G, red curve), providing a clear signature of the
breaking of C3z [section I.4 of (19) shows that
the transport is dominated by the MnBi2Te4
layer of the heterostructure]. The transverse
resistance and two-probe resistance also show
the breaking of C3z [fig. S12 of (19)]. We further
substantiate the breaking of C3z symmetry by
an independent method, the optical SHG, at
room temperature. As shown in Fig. 1H, our
SHG data also show the clear breaking of C3z
symmetry [see detailed discussions in section
I.5 of (19)]. Our demonstration of C3z breaking
establishes the BP/MnBi2Te4 heterostructure
as an ideal platform to search for this effect.
Observation of the nonlinear Hall effect
To measure the linear and nonlinear electrical
transport, we pass a current at frequency w (I w)
and use the lock-in technique to detect linear
voltage V w and nonlinear voltage V 2w . We
describe the nonlinear voltage as V 2w
ijk , where i is
the direction of the nonlinear voltage V 2w and
j; k are the directions of the injected current I w.
All measurements are performed at B = 0.
Figure 1I shows the nonlinear Hall voltage
V 2w
yxx of the Device-BM1 before and after being
interfaced with BP. A prominent nonlinear
Hall signal emerges only after BP is intro-
duced. This is in sharp contrast to the linear
voltage (Fig. 1I, inset), which becomes even
slightly smaller upon the introduction of BP.
Such observation agrees well with the the-
oretical expectation of the intrinsic NHE in-
duced by a quantum metric dipole. To exclude
that the effect is caused by a Berry curvature
dipole (4, 6, 7, 9), we study the relationship
between the second-order NHE and the AFM
order in MnBi2Te4.
The AFM spin-induced nonlinearity
Overall, we have fabricated 30 BP/even-layered
MnBi2Te4 heterostructure devices [see section
I.0 of (19) for our systematic data that con-
firm the MnBi2Te4 thickness in our devices].
In all of the 30 devices, we have observed the
NHE with consistent behaviors as a function
of AFM order, spatial direction, scattering time,
vertical electric field, and doping [see fig. S22
and table S1 for a summary of all 30 devices
(19)]. Here, we focus on the Device-BMB1 (Fig.
2A), which has two-layer BP on both sides of
six-septuple-layer MnBi2Te4. Moreover, we have
ensured that the crystalline a axes of the BPs
and the MnBi2Te4 are aligned (Fig. 2A). Such a
carefully controlled configuration is important to
preserve MnBi2Te4’s PT symmetry, which en-
Corrected 21 August 2023. See full text.
Fig. 1. Spin-induced electrical nonlinearity in PT-symmetric antiferromagnets and introduction to
our sample. (A and B) Nonlinear electrical transport in PN junctions and noncentrosymmetric conductors
(charge-induced electrical nonlinearity). (C) Nonlinear electrical transport in PT-symmetric AFMs (spin-
induced electrical nonlinearity). (D to F) Lattice structures of the MnBi2Te4 and BP. (G and H) Angle-resolved
resistance and optical SHG measurements of a 6SL MnBi2Te4 before and after being interfaced with BP.
(I) The nonlinear Hall signal V2w
yxx before and after being interfaced with BP at B = 0 T. Inset: The linear
longitudinal voltage Vw
at temperature T ¼ 1:8 K unless otherwise noted. Data in Fig. 1 and Fig. 2B are from Device-BM1. Data in
Fig. 2, E to J, Fig. 3, and Fig. 4A are from Device-BMB1. Data in Fig. 4F are from Device-BM21.
xx before and after being interfaced with BP. All data in Figs. 1 to 4 were taken
sures that the Berry curvature and Berry cur-
vature dipole vanish. Figure 2B shows a large
transverse nonlinear response V 2w
yxx . We have
also measured the longitudinal nonlinear re-
sponse V 2w
xxx, which shows no observable signal.
Therefore, our data reveal an interesting “Hall
dominance” in the nonlinear transport.
We now focus on exploring how the non-
linear Hall signal depends on opposite AFM
states. In ferromagnets, the opposite ferromag-
netic states can be controlled by sweeping the B
field. In PT-symmetric AFMs including Cr2O3,
even-layered CrI3, and even-layered MnBi2Te4
(44, 48, 49), previous works have shown that
the opposite AFM states can be controlled by
sweeping the vertical Bz field under a fixed
vertical Ez field. Hence, we follow the pre-
viously established procedures (44): Under a
fixed Ez (Ez = −0.17 V/nm), we sweep Bz from
−8 T to 0 T or from +8 T to 0 T to prepare the
two AFM states (Fig. 2, C and D). We first
study the AFM I. The linear voltage V w
xx (Fig.
2E) exhibits a typical ohmic behavior. The
nonlinear voltage V 2w
yxx (Fig. 2G) is prominent
and its sign is positive. We then prepare AFM
II. The linear voltage V w
xx (Fig. 2F) remains
unchanged. In sharp contrast, the nonlinear
voltage V 2w
yxx (Fig. 2H) flips sign. For both
AFM I and II, the NHE is only present in the
AFM phase (Fig. 2, I and J). Our observation
that the nonlinear Hall signal flips sign upon
reversing the AFM order further demonstrates
its quantum metric dipole origin, because
the quantum metric dipole is theoretically ex-
pected to be opposite for the opposite AFM
domains [see section III of (19)]. Our non-
linear Hall signal measures an average over
all AFM domains. However, our experiments
suggest that our sample is prepared into pre-
dominantly one domain. If our sample con-
sisted of opposite domains with a 50%-50%
composition, then the measured nonlinear
Hall signal would average to zero. By contrast,
our data show a large nonzero nonlinear Hall
signal. Moreover, the sign of the observed signal
flips as we prepare the opposite AFM domain.
Gao et al., Science 381, 181–186 (2023)
14 July 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
Corrected 21 August 2023. See full text.
Fig. 2. Observation of the antiferromagnetic NHE. (A) Schematic illustration of 2L BP/6SL MnBi2Te4/2L
BP device (Device-BMB1). The crystalline a axes of the BPs and the MnBi2Te4 were all aligned [fig. S16
of (19)]. (B) The longitudinal (V2w
the procedures established by previous works (44): Under a fixed Ez (−0.17 V/nm), we sweep Bz from −8 T
to 0 T or from +8 T to 0 T to prepare the two AFM states. (E and F) Linear longitudinal voltage as a function of
incidence current for AFM I and AFM II. (G and I) Nonlinear Hall voltage as a function of incident current
and temperature of AFM I, respectively. (H and J) The same as panels (G) and (I) but for AFM II.
yxx) components of the nonlinear voltage. (C and D) We follow
xxx) and Hall (V2w
Further, the magnitude of the measured signal
is consistent with the theoretically calculated
value, which assumes a single domain. Spatially
resolved magnetic measurements will be needed
to determine the exact domain composition.
We now perform further systematic studies.
Because the nonlinear Hall current flips sign
upon reversing the AFM order, all the non-
linear Hall data (apart from Fig. 2) are ob-
tained by taking the difference between the
two AFM domains. First, the intrinsic NHE is
expected to be independent of the scattering
time, unlike other related responses. Similar
to the intrinsic AHE in ferromagnetic metals
Gao et al., Science 381, 181–186 (2023)
14 July 2023
(cid:5)
I w
x
2 ¼
2R3
(cid:4)
=
=Ew
x
xxw2d
¼ J 2w
yxx
(35), and in contrast to the quantum AHE, there
is still dissipation through the linear Drude
conductivity sxx , The nonlinear Hall conduc-
tivity can be directly extracted from our data
(cid:5)
(cid:4)
by s2w
yxxl3
V 2w
,
yxx
where l; w; d are the length, width, and thick-
ness of the sample. Previous experiments have
studied the scattering time t dependence of
various Hall effects (6, 9, 14, 35) by investigat-
ing the scaling between the corresponding
Hall conductivity and the Drude conductivity.
Following this established method, we study
the scaling betweens2w
yxx andsxx. Our data (Fig.
3A) show that s2w
yxx is independent of sxx below
~15 K. Moreover, consistent results have been
observed at multiple charge densities in the
same sample and from different samples [sec-
tion III.9 of (19)]. These systematic data point
to the conclusion that the s2w
yxx is independent
of scattering time t below ~15 K. Above ~15 K,
s2w
yxx vanishes quickly across Néel temperature
TN because the AFM order vanishes and our
NHE only exists in the AFM phase. Hence,
studying the t dependence at temperatures
near TN would require one to take the strong
influence of the AFM order near TN into ac-
count [see section III.8 of (19) for additional
measurements and analysis]. Second, the in-
trinsic NHE does not require a noncentrosym-
metric lattice or any explicit breaking of PT
symmetry. To test this, we explicitly break PT
by applying a vertical Ez field via dual gating.
The nonlinear Hall signal is already promi-
nent even at Ez = 0 (Fig. 3D), confirming that it
does not require any PT breaking. Moreover,
the nonlinear Hall signal is symmetric for ±Ez,
also consistent with the expectation [see
section IV.2 of (19)]. Third, the NHE is ex-
pected to be sensitive to the direction of the
incident current I w. In Fig. 3B, we measure the
nonlinear Hall conductivity as a function of
the direction of I w. Indeed, the signal is most
prominent when I w is along a particular in-
plane direction. In this way, we experimentally
mapped out the direction of the relevant geo-
metrical dipole (the quantum metric dipole in
our case, as we demonstrate next). Moreover,
the intrinsic NHE is found to be independent
of frequency [fig. S23 of (19)]. In principle, the
frequency independence is expected to per-
sist until w is large enough to induce an
interband transition (roughly terahertz or far-
infrared). Future experiments are needed to
test the NHE in that regime.
Excluding competing mechanisms
Although we tried to eliminate Berry curva-
ture dipole by aligning the crystalline a axes
between BPs and MnBi2Te4 to preserve PT
symmetry (Fig. 2A), let us assume that the
alignment is imperfect, so Berry curvature di-
pole is allowed. We now show that the observed
relationship between the nonlinear Hall sig-
nal and AFM order can distinguish between
Berry curvature dipole DBerry and quantum
metric dipole DMetric. DBerry can be understood
as a distribution of the Berry curvature around
the Fermi surface such that it is larger on one
side of the Fermi surface than on the opposite
side. A similar picture holds for DMetric (Fig. 3).
As we observe that the nonlinear Hall signal
changes sign upon the reversal of AFM order,
the dipole that causes our observed nonlinear
Hall signal must also flip. Let us assume that
the AFM I has DBerry > 0 and DMetric > 0 ,
which is visualized in a tilted gapped Dirac
band structure in Fig. 3, E and G. We now flip
the AFM order to the AFM II by performing
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
Corrected 21 August 2023. See full text.
ð
ð
ð
Þ ¼ DBerry AFMI
Þ ¼ (cid:2)DMetric AFMI
time reversal T. Under T, the bands are flipped
between Tk (Fig. 3, F and H), the Berry cur-
vature flips sign, but the quantum metric keeps
the same sign. Hence, from Fig. 3, F and H, one
Þ ,
ð
can see that, DBerry AFMII
Þ. There-
but DMetric AFMII
fore, our observation that the nonlinear Hall
signal flips sign upon reversing the AFM order
excludes the Berry curvature dipole mecha-
nism. In section II.1 of (19), we enumerate ex-
perimental results, including the relation with
AFM order, scaling, vertical electric field de-
pendence, and relation with mirror symmetry,
which corroborate that the Berry curvature di-
pole mechanism cannot account for our data.
Within the nonlinear effects that flip sign
upon reversing the AFM order, there is an-
other possibility, the second-order Drude ef-
fect (5, 12, 17, 22). This effect is expected to be
proportional to t2 (22) and therefore can be
ruled out on the basis of our scaling data in
Fig. 3A. Moreover, the NHE is antisymmetric
¼
(upon exchanging the first two indices), sNHE
abg
(cid:2)sNHE
bag , but the second-order Drude effect
(SODE) is symmetric, sSODE
(22). Using
abg
an electrical sum-frequency generation meth-
od [section II.2 of (19)], we showed that our
¼ (cid:2)s2w
signal is indeed antisymmetric, s2w
xyx,
yxx
which demonstrates that the SODE is insig-
nificant in our signal. In section II.2.3 of (19),
we present additional data that suggest that
the NHE is dominant over the SODE at dif-
ferent temperatures and charge densities. In
section III.5 of (19), we show that the nonlin-
ear Hall signal is negligibly small at T8 T be-
cause the forced ferromagnetic state recovers
inversion symmetry. We also carefully addressed
other competing origins such as thermal and
accidental diode junctions [section II.3 of (19)].
¼ sSODE
bag
Energy-resolved probe of quantum metric
in PT-symmetric antiferromagnets
We also study the evolution of the nonlinear
conductivity s2w
yxx with the charge density. The
nonlinear Hall signal is zero inside the charge
neutrality gap (Fig. 4A). This is consistent with
the expectation that the NHE is a Fermi sur-
face property. As we tune the Fermi energy
away from the charge neutrality, the nonlinear
Hall signal emerges. Notably, the conductiv-
ities in electron and hole regimes have the same
sign. As we go deeper into the electron-doped
regime, the signal reverses sign.
We now provide an intuitive physical picture
to understand the large quantum metric dipole
and its Fermi-level dependence. MnBi2Te4 fea-
tures Dirac surface states, which are gapped
owing to the AFM, leading to a large quantum
metric near the gap edge. Moreover, because
the AFM order breaks both T and P, the Dirac
bands are asymmetric about k ¼ 0 (Fig. 3G).
Hence, at a fixed energy, positive and negative
momenta have a different quantum metric,
leading to a nonzero quantum metric dipole.
Gao et al., Science 381, 181–186 (2023)
14 July 2023
(cid:5)
(cid:4)
Iw
x
2 ¼
2R3
(cid:4)
=
=Ew
x
yxxl3
V2w
¼ J2w
yxx
Fig. 3. Systematic investigation of the NHE. (A) The scaling between the nonlinear Hall conductivity
and the Drude conductivity sxx ¼ 1=Rxx. The nonlinear Hall conductivity can be directly extracted from the
(cid:5)
data as s2w
xxw2d
. (B) Angular dependence of the nonlinear Hall conductivity in
yxx
Device-BM1. (C) Dual-gated resistance map of the 2L BP/6SL MnBi2Te4/2L BP heterostructure (Device-
BMB1). The vertical electric field Ez and charge density dependence can be independently tuned by combining the
top and bottom gate voltages. (D) Ez dependence of the nonlinear Hall conductivity and linear longitudinal
resistance. Ez follows the dashed line in (C). (E to H) Schematic illustration of the Berry curvature dipole
(DBerry) and the quantum metric dipole (DMetric) for the AFM I and AFM II of the BP/6SL MnBi2Te4/BP
heterostructure (see text). Although we aligned the crystalline axes of BP and MnBi2Te4 in our Device-BMB1
(Fig. 2A), realistically it is difficult to make the alignment perfect. If the alignment is imperfect and PT
symmetry is broken, a Berry curvature dipole is allowed.
Intuitively, we can understand the sign of the
nonlinear Hall signal by considering whether
positive or negative momenta have a larger
quantum metric. We see from Fig. 3G that both
upper and lower parts of the Dirac cone have
Þ , suggesting that the non-
ð
g þkF
linear Hall signals should show the same sign in
electron and hole regimes, consistent with our
data (Fig. 4A). The sign change in the electron-
doped regime is beyond this simple picture.
ð
Þ > g (cid:2)kF
To achieve a more comprehensive under-
standing, we built an effective model of the
BP/6SL MnBi2Te4/BP heterostructure [sec-
tion IV.4-9 of (19)]. Owing to the incommen-
surability of the BP and MnBi2Te4 lattices,
we need to derive the coupling between the
Bloch states of the two materials in the real-
space continuum [i.e., within the extended
Brillouin zone (BZ)]. The low-energy bands
are located in the BZ center, so only Bloch
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
Corrected 21 August 2023. See full text.
by fitting the first-principle band structures.
The MnBi2Te4 and BP coupling parameters
were partly constrained by considering the
independent data of Rxx=Ryy so that an agree-
ment in the overall magnitude was achieved
independently. We adjusted the remaining free
coupling model parameters to match detailed
features in the charge density dependence [sec-
tion IV.9 of (19)].
ð
We first turn off the coupling between the
MnBi2Te4 and BP. The Fermi surface shown in
Fig. 4C (−50 meV) is C3z symmetric, and there is
already a large quantum metric (gxx and gyx )
around it. According to (22), the DMetric respon-
sible for the nonlinear Hall is given by DMetric ¼
∫
Þ (v is the Fermi ve-
ð
vygxx (cid:2) vxgyx
Þd e (cid:2) eF
k
locity). We plot the integral kernel in color in
Fig. 4D. Positive and negative contributions
around the contour exactly cancel each other
out because of C3z symmetry. So, the integral
goes to zero (Fig. 4D, left panel). We then turn on
the MnBi2Te4-BP couplings, which breaks C3z.
For the C3z-breaking contour, we observe un-
equal contributions from the two colors, lead-
ing to a nonzero DMetric [Fig. 4D, right panel;
see details in the caption and in section III.16
of (19)]. Figure 4E shows the band structure
of the BP/6SL MnBi2Te4/BP heterostructure,
from which we can compute the intrinsic non-
linear Hall conductivity s2w
yxx as a function of
chemical potential. In particular, near the
charge neutrality gap, we found that s2w
yxx in-
deed mainly comes from the quantum me-
tric of the Dirac surface states, consistent
with the intuitive picture above. The sign in-
version in the electron-doped regime mainly
comes from the quantum metric of the avoided
crossing inside conduction bands according
to our calculation [section IV.9 of (19)]. Ow-
ing to the multiband nature of our model, the
s2w
yxx was calculated by the general expression
(cid:6)
Re∫
vn
(cid:2)
y
(cid:7)
Þ (22). This gen-
ð
d en (cid:2) eF
eral expression can be decomposed into the
quantum metric dipole contribution plus
additional interband contributions (AIC)
hunji@kx umihumji@kx uni
en(cid:2)em
hunji@ky umihumji@kx uni
en(cid:2)em
¼ (cid:2) 2e3
s2w
yxx
em≠en
X
n;m
vn
x
k
s2w
yxx
¼ (cid:2)2e3
X
n
(cid:2) vn
ygn
vn
xx
en (cid:2) e(cid:2)n
xgn
yx
∫
k
ð
d en (cid:2) eF
Þ þ AIC
ð1Þ
where the first term is the quantum metric
dipole contribution, and the second term is
(cid:6)
em≠en;en
X
vn
y
hunji@kx umihumji@kx uni
en(cid:2)em
(cid:2)
AIC¼ (cid:2)2e3
Re ∫
k
(cid:7)
n;m
vn
x
ð
d en (cid:2) eF
hunji@ky umihumji@kx uni
en(cid:2)em
em(cid:2)e(cid:2)n
Þ ((cid:2)n is the
en(cid:2)e(cid:2)n
band whose energy is closest to n). In our BP/
6SL MnBi2Te4/BP system, we found that the
quantum metric dipole contribution strongly
dominates, whereas the AIC is small [details
Fig. 4. The quantum metric dipole as the microscopic geometrical origin. (A) Experimentally measured
nonlinear Hall conductivity as a function of charge density. (B) Theoretically calculated s2w
of ne based on the BP/6SL MnBi2Te4/BP band structure (see text). (C to E) The electronic structure of the
BP/6SL MnBi2Te4/BP heterostructure calculated with an effective model (see text). (C) Fermi surface at
−50 meV (the lower part of the surface Dirac cone). The coupling between MnBi2Te4 and BP is turned off, so
that contour respects C3z symmetry. The quantum metric gxx and gyx plotted around the Fermi surface.
(cid:4)
(D) The nonlinear Hall conductivity is given by the integral DMetric ¼ ∫
Þ around the
vygxx (cid:2) vxgyx
yxx as a function
ð
d e (cid:2) eF
(cid:5)
k
Fermi surface. With C3z symmetry (left panel), the integral goes to zero. After turning on the coupling
between MnBi2Te4 and BP (right panel), C3z is broken, making the integral around the Fermi contour nonzero.
To more clearly see how the integral changes to nonzero when C3z is broken, we rewrite DMetric as an integral
of the polar angle q, DMetric ¼ ∫
dq (dl is an infinitesimal length along the Fermi surface).
vygxx (cid:2) vxgyx
(cid:4)
(cid:5) dl
dq
FS
The inset presents the change of the above kernel [see section III.16 of (19) for details]. (E) Band structure of
BP/6SL MnBi2Te4/BP heterostructure as a function of ky for kx = 0. Color represents the quantum metric
of the bands. (F) Measured microwave rectification based on the intrinsic NHE. Inset: DC signal as a function
of microwave frequency. (G) Schematic illustration of microwave rectification. (H) Schematic illustration of
quantum metric–induced nonlinear responses. The horizontal axes are kx and ky. The two black arrows
represent the Bloch wave functions at two nearby k points. The two arrows point to different directions, illustrating
a finite distance between two wave functions (i.e., a finite quantum metric). This quantum metric leads to a NHE,
which can turn an external AC electric field (e.g., the microwave in the figure) into a DC signal.
bands with the same momentum hybridize.
The coupling amplitude depends only on
the characteristic decay length of the atomic
orbitals as any discrete lattice structure is
averaged out (47). The Hamiltonian reads
^U t
^hBP; t
0
^hBP; t bð Þ are Hamiltonians for 6SL MnBi2Te4
^U b
0
^hBP; b
C
A. ^hMBT and
^hMBT
^U
^U
^h kx; ky
ð
B
@
Þ ¼
†
t
†
b
0
1
and top (bottom) BP. The spin-orbit coupling
(SOC) in MnBi2Te4 is crucial for a nonzero NHE
because it allows the low-energy orbitals to
feel the symmetry breaking by the AFM order
[section III.15 of (19)]. In particular, the SOC
was included in the model following the orig-
inal work by (50). ^U t and ^U b denote the band
hybridization caused by nearest-neighbor cou-
pling between MnBi2Te4 and BP. The bare
MnBi2Te4 and BP parameters were obtained
Gao et al., Science 381, 181–186 (2023)
14 July 2023
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
in section IV.3 of (19)]. Therefore, our non-
linear Hall measurement is a powerful, energy-
resolved probe of the quantum metric.
Possible AFM spin-based wireless rectification
The second-order nonlinear effect enables not
only frequency doubling but also rectification.
We use the intrinsic AFM nonlinear Hall effect
to demonstrate wireless rectification with zero
external bias (battery-free) and without mag-
netic field. We inject microwave radiation and
measure the DC signal. We observe clear rec-
tification DC voltage in response to the micro-
wave radiation (Fig. 4F), which shows a broad
band response, including the Wi-Fi frequencies
(2.4 and 5 GHz) and even higher frequencies
[see fig. S35 and section V.2 of (19)]. In section
III.12 of (19), we show that the rectification
signal flips sign as we reverse the AFM state,
which suggests the intrinsic quantum metric
dipole origin. Apart from the intrinsic quan-
tum metric dipole, extrinsic sources such as the
Schottky diodes at the metal-MnBi2Te4 junction,
unintentional diodes inside the MnBi2Te4, and
the gap between the two gates can also lead to
microwave rectification. To unambiguously rule
out these extrinsic sources, future systematic
experiments will be needed [discussion in
section V.1 of (19)].
Discussion and outlook
The intrinsic second-order Hall effect observed
here realizes an electrical nonlinearity in-
duced by the AFM spins and provides a rare
example of a quantum metric response. As
highlighted by recent theoretical studies, the
influence of the quantum metric is expected
to span many different areas, ranging from
nonlinear responses in PT-symmetric AFMs
to flat-band conductivity, superconductivity
and charge orders in moiré systems, the frac-
tional Chern insulator, and k-space dual of
gravity (20, 22–34). Another interesting fu-
ture direction is to explore the nonlinear re-
sponses in canted AFM materials [section
V.3 of (19)], where nonzero Berry curvature
of higher order in magnetization has recent-
ly been observed (38, 46, 51). In terms of ma-
terials, we show that, beyond “band structure
engineering,” the van der Waals interfaces
can engineer the properties of the wave func-
tion i.e., “quantum geometry engineering” (47).
Our observations may enable the use of AFM
spins to harvest electromagnetic energy and
to realize self-powered AFM spintronic devices.
An exciting future breakthrough would be to
demonstrate room-temperature wireless rectifi-
cation based on the quantum metric NHE in a
PT-symmetric AFM material.
RE FE RENCES AND N OT ES
1.
J. Provost, G. Vallee, Commun. Math. Phys. 76, 289–301
(1980).
2. Y. Tokura, N. Nagaosa, Nat. Commun. 9, 3740 (2018).
Corrected 21 August 2023. See full text.
3.
J. E. Moore, J. Orenstein, Phys. Rev. Lett. 105, 026805
(2010).
I. Sodemann, L. Fu, Phys. Rev. Lett. 115, 216806 (2015).
4.
5. T. Ideue et al., Nat. Phys. 13, 578–583 (2017).
6. K. Kang, T. Li, E. Sohn, J. Shan, K. F. Mak, Nat. Mater. 18,
324–328 (2019).
7. Q. Ma et al., Nature 565, 337–342 (2019).
8. S. Dzsaber et al., Proc. Natl. Acad. Sci. U.S.A. 118, e2013386118
(2021).
9. D. Kumar et al., Nat. Nanotechnol. 16, 421–425 (2021).
10. K. Yasuda et al., Nat. Nanotechnol. 15, 831–835 (2020).
11. W. Zhao et al., Nat. Mater. 19, 503–507 (2020).
12. H. Isobe, S.-Y. Xu, L. Fu, Sci. Adv. 6, eaay2497 (2020).
13. S. Lai et al., Nat. Nanotechnol. 16, 869–873 (2021).
14. P. He et al., Nat. Nanotechnol. 17, 378–383 (2022).
15. Z. Zhang et al., Nat. Commun. 13, 6191 (2022).
16. S. Sinha et al., Nat. Phys. 18, 765–770 (2022).
17. N. J. Zhang et al., Momentum polarized superconductivity in
twisted trilayer graphene. Preprint at https://arxiv.org/abs/
2209.12964 (2022).
18. Z. Sun et al., Nature 572, 497–501 (2019).
19. See additional text and data in the supplementary materials.
20. Y. Gao, S. A. Yang, Q. Niu, Phys. Rev. Lett. 112, 166601
(2014).
21. A. Gianfrate et al., Nature 578, 381–385 (2020).
22. C. Wang, Y. Gao, D. Xiao, Phys. Rev. Lett. 127, 277201
(2021).
23. H. Liu et al., Phys. Rev. Lett. 127, 277202 (2021).
24. S. Lahiri, K. Das, D. Culcer, A. Agarwal, Intrinsic nonlinear
conductivity induced by the quantum metric dipole. Preprint at
https://arxiv.org/abs/2207.02178 (2022).
25. T. B. Smith, L. Pullasseri, A. Srivastava, Phys. Rev. Res. 4,
013217 (2022).
26. A. Arora, M. S. Rudner, J. C. W. Song, Nano Lett. 22,
9351–9357 (2022).
27. J. Mitscherling, T. Holder, Phys. Rev. B 105, 085154
(2022).
28. J.-W. Rhim, K. Kim, B.-J. Yang, Nature 584, 59–63
(2020).
29. P. J. Ledwith, G. Tarnopolsky, E. Khalaf, A. Vishwanath,
Phys. Rev. Res. 2, 023237 (2020).
30. T. Holder, D. Kaplan, B. Yan, Phys. Rev. Res. 2, 033100
(2020).
31. H. Watanabe, Y. Yanase, Phys. Rev. X 11, 011001 (2021).
32. K.-E. Huhtinen, J. Herzog-Arbeitman, A. Chew, B. A. Bernevig,
P. Törmä, Phys. Rev. B 106, 014518 (2022).
33. J. S. Hofmann, E. Berg, D. Chowdhury, Phys. Rev. Lett. 130,
226001 (2023).
34. X. Hu, T. Hyart, D. I. Pikulin, E. Rossi, Phys. Rev. B 105,
L140506 (2022).
35. N. Nagaosa, J. Sinova, S. Onoda, A. MacDonald, N. P. Ong, Rev.
Mod. Phys. 82, 1539–1592 (2010).
36. M. M. Otrokov et al., Nature 576, 416–422 (2019).
37. D. Zhang et al., Phys. Rev. Lett. 122, 206401 (2019).
38. S. H. Lee et al., Phys. Rev. Res. 1, 012011 (2019).
39. Y. Deng et al., Science 367, 895–900 (2020).
40. C. Liu et al., Nat. Mater. 19, 522–527 (2020).
41. H. Deng et al., Nat. Phys. 17, 36–42 (2021).
42. D. Ovchinnikov et al., Nano Lett. 21, 2544–2550 (2021).
43. S. Yang et al., Phys. Rev. X 11, 011003 (2021).
44. A. Gao et al., Nature 595, 521–525 (2021).
45. L. Tai et al., ACS Nano 16, 17336–17346 (2022).
46. S.-K. Bac et al., npj Quantum Mater. 7, 46 (2022).
47. T. Akamatsu et al., Science 372, 68–72 (2021).
48. A. Iyama, T. Kimura, Phys. Rev. B 87, 180408 (2013).
49. S. Jiang, J. Shan, K. F. Mak, Nat. Mater. 17, 406–410
(2018).
50. H. Zhang et al., Nat. Phys. 5, 438–442 (2009).
51. J. Kipp et al., Commun. Phys. 4, 99 (2021).
52. A. Gao, S. Xu, Quantum metric nonlinear Hall effect in a
topological antiferromagnetic heterostructure, Zenodo (2023);
https://doi.org/10.5281/zenodo.8002209.
ACKN OWLED GMEN TS
We are grateful to A. Yacoby for help with the microwave
measurements. We thank M. Wesson and N. Poniatowski for
technical support during the microwave measurements. We also
thank Y. Gao, J. Ahn, and P. Kim for helpful discussions. We are
grateful to L. Ye, M. Mogi, Y. Fujishiro, and T. Kurumaji for
extensive discussions on the scaling of AHE. Funding: Work in the
S.-Y.X. group was partly supported through the Center for the
Advancement of Topological Semimetals (CATS), an Energy
Frontier Research Center (EFRC) funded by the US Department of
Energy (DOE) Office of Science (fabrication and measurements),
through the Ames National Laboratory under contract DE-
AC0207CH11358, and partly through Air Force Office of Scientific
Research (AFOSR) grant FA9550-23-1-0040 (data analysis and
manuscript writing). S.-Y.X. acknowledges the Corning Fund for
Faculty Development. Q.M. and L.F. acknowledge support from the
NSF Convergence program (NSF ITE-2235945) and the CIFAR
program. S.-Y.X. and D.B. were supported by the NSF Career DMR-
2143177. C.T. and Z.S. acknowledge support from the Swiss
National Science Foundation under project P2EZP2 191801 and
P500PT 206914, respectively. Y.F.L., S.-Y.X., D.C.B., Y.O., and L.F.
were supported by the STC Center for Integrated Quantum
Materials (CIQM), NSF grant no. DMR-1231319. This work was
performed in part at the Center for Nanoscale Systems (CNS)
Harvard University, a member of the National Nanotechnology
Coordinated Infrastructure Network (NNCI), which is supported by
the National Science Foundation under NSF award no. 1541959.
Bulk single-crystal growth and characterization of MnBi2Te4 were
performed at UCLA and were supported by the DOE, Office of
Science, under award no. DE-SC0021117. The work at Northeastern
University was supported by the Air Force Office of Scientific
Research under award no. FA9550-20-1-0322, and it benefited
from the computational resources of Northeastern University’s
Advanced Scientific Computation Center (ASCC) and the Discovery
Cluster. The work in the QM group was partly supported through
the CATS, an EFRC funded by the DOE Office of Science, through
the Ames National Laboratory under contract DE-AC0207CH11358
(fabrication and measurements) and partly through NSF DMR-
2143426 (data analysis and manuscript writing). T.V.T. and P.P.O.
were supported by the CATS, an EFRC funded by the DOE Office of
Science, through the Ames National Laboratory under contract DE-
AC0207CH11358. T.R.C. was supported by the 2030 Cross-
Generation Young Scholars Program of the National Science and
Technology Council (NSTC) in Taiwan (program no. MOST111-
2628-M-006-003-MY3); National Cheng Kung University (NCKU),
Taiwan; and the National Center for Theoretical Sciences
(NCTS), Taiwan. This research was supported, in part, by the
Higher Education Sprout Project, Ministry of Education to the
Headquarters of University Advancement at NCKU. H.L.
acknowledges support from the National Science and Technology
Council (NSTC) in Taiwan under grant no. MOST 111-2112-M-001-
057-MY3. The work at TIFR Mumbai was supported by the
Department of Atomic Energy of the Government of India under
project no. 12-R&D-TFR-5.10-0100 and benefited from the
computational resources of TIFR Mumbai. K.W. and T.T.
acknowledge support from the JSPS KAKENHI (grant nos.
20H00354, 21H05233, and 23H02052) and World Premier
International Research Center Initiative (WPI), MEXT, Japan.
M.H. and C.R.D. were supported by the AFOSR under award
no. FA9550-20-1-0319. S.W.C. acknowledges partial support from
the Harvard Quantum Initiative in Science and Engineering.
Author contributions: S.-Y.X. conceived the experiments and
supervised the project. A.G. fabricated the devices, performed the
measurements, and analyzed data with help from Y.F.L., D.B.,
J.X.Q., H.C.L., C.T., T.D., Z.S., S.C.H., D.C.B., and Q.M. A.G. and
S.W.C. performed the microwave rectification experiments. C.H.,
T.Q., and N.N. grew the MnBi2Te4 single crystals. M.H. and C.R.D.
performed NV center magnetometry. B.G. made the theoretical
studies, including first-principles calculations and effective
modeling, with the help from T.V.T., Y.O., H.J.T., S.W.L., B.S., H.L.,
A.B., T.R.C., L.F., and P.P.O. T.V.T. developed the effective model
with help from B.G. under the guidance of P.P.O. K.W. and T.T.
grew the hBN single crystals. S.-Y.X., A.G., and Q.M. wrote the
manuscript with input from all authors. Competing interests: The
authors declare no competing financial interests. Data and
materials availability: Data in the main text and supplementary
materials, as well as the codes for theoretical calculations of the
quantum metric, are available from Zenodo (52). License
information: Copyright © 2023 the authors, some rights reserved;
exclusive licensee American Association for the Advancement of
Science. No claim to original US government works. https://www.
sciencemag.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adf1506
Materials and Methods
Supplementary Text
Figs. S1 to S50
Tables S1 to S3
References (53–71)
Submitted 5 October 2022; accepted 6 June 2023
Published online 15 June 2023
10.1126/science.adf1506
Gao et al., Science 381, 181–186 (2023)
14 July 2023
6 of 6
|
10.1126_science.adf0895
|
RES EARCH
CONSERVATION ECOLOGY
Ecosystem-based management outperforms
species-focused stocking for enhancing fish populations
Johannes Radinger1*†, Sven Matern1,2†, Thomas Klefoth3, Christian Wolter1, Fritz Feldhege1,2,
Christopher T. Monk1,4, Robert Arlinghaus1,2,5
Ecosystem-based management is costly. Therefore, without rigorously showing that it can outperform
traditional species-focused alternatives, its broad-scale adoption in conservation is unlikely. We present
a large-scale replicated and controlled set of whole-lake experiments in fish conservation (20 lakes
monitored over 6 years with more than 150,000 fish sampled) to examine the outcomes of ecosystem-
based habitat enhancement (coarse woody habitat addition and shallow littoral zone creation) versus
a widespread, species-focused alternative that has long dominated fisheries management practice
(i.e., fish stocking). Adding coarse woody habitats alone did not, on average, enhance fish abundance,
but creating shallow water habitat consistently did, especially for juvenile fish. Species-focused fish
stocking completely failed. We provide strong evidence questioning the performance of species-focused
conservation actions in aquatic ecosystems and instead recommend ecosystem-based management
of key habitats.
T here is a long-standing debate on how
effectively ecosystem-based management
can counter biodiversity loss, aid in the
conservation of imperiled species (1),
or sustain and rebuild fisheries (2, 3).
Ecosystem-based management targets im-
proving or reinstalling key ecological processes,
habitats, and species interactions rather than
focusing on removing single stressors or sup-
porting individual species (1). Globally, the ap-
plication of ecosystem-based management is
still in its infancy (2, 3), and constraints in-
clude the strong political and financial sup-
port needed (2, 4). Garnering such support is
challenging when there are still many un-
knowns about its effectiveness. Habitat man-
agement actions can fail, particularly if they
are not wide-ranging enough or do not ad-
dress key bottlenecks critical in the life cycle of
an organism (5). Further, experimenting at the
scale of natural ecosystems in a replicated
fashion is rarely done because it is often prac-
tically infeasible or too costly (6). Ecosystem-based
habitat management may be systematically
more effective and sustainable than traditional
single-species–oriented measures for achieving
conservation objectives because of its compre-
hensive consideration of the interconnections
among species, their environment, and hu-
mans (1). However, policy-makers are unlikely
1Department of Fish Biology, Fisheries and Aquaculture,
Leibniz Institute of Freshwater Ecology and Inland Fisheries,
Berlin, Germany. 2Division of Integrative Fisheries
Management, Faculty of Life Sciences, Humboldt-Universität
zu Berlin, Berlin, Germany. 3Ecology and Conservation,
Faculty of Nature and Engineering, Hochschule Bremen,
Bremen, Germany. 4GEOMAR Helmholtz Centre for Ocean
Research Kiel, Marine Evolutionary Ecology, Kiel, Germany.
5Integrative Research Institute on Transformations of
Human-Environmental Systems (IRI THESys), Humboldt-
Universität zu Berlin, Berlin, Germany.
*Corresponding author. Email: johannes.radinger@igb-berlin.de
†These authors contributed equally to this work.
to fully support implementing such practices
on a large scale until robust supporting evi-
dence accumulates.
When species decline, a common species-
focused mitigation measure receiving substan-
tial stakeholder and political support is releasing
wild-captured or hatchery-bred animals (7–9);
in fisheries, this practice is known as stocking
(8, 9). However, stocking can have lasting
negative ecological and evolutionary effects on
populations, food webs, and ecosystems, e.g.,
due to the spread of non-native genotypes or
species (8–11). Further, models and empirical
studies have demonstrated that releasing fish
often fails to increase populations (8, 12–14).
Nevertheless, species-focused management
through stocking in inland fisheries and fish
conservation continues to be a standard prac-
tice because of a range of psychological (e.g.,
norms and habits) and institutional (e.g., lack
of monitoring) factors (9, 15).
Ecosystem-based approaches to habitat man-
agement are promising alternatives to stock-
ing (9, 16). To be successful and to support
wild-living animals, management interven-
tions must effectively remediate current pop-
ulation constraints. For fish, key bottlenecks
in population dynamics relate to density-
dependent mortality in early juveniles, a critical
life stage that determines year-class strength
and adult abundance (Fig. 1) (13, 14). In par-
ticular, the smallest length classes of fish face
an important trade-off between securing suffi-
cient food resources to support growth beyond
their predators’ gape width and minimizing
exposure to predation (17). This trade-off is
shaped by intra- and interspecific competition
and the arrangement of profitable foraging
areas with higher predation risk versus refuges
(e.g., vegetation, structural habitats, or shallow
water) that offer protection at the expense of
food intake (18, 19) (Fig. 1). Thus, habitat en-
hancement can improve the growth-mortality
trade-off to allow a greater number of juvenile
fish to grow into the population, whereas stock-
ing fish into naturally reproducing populations
can increase competition or predation without
providing refuges from mortality (14).
Shallow littoral zone creation (e.g., by exca-
vating new shallow areas) is an ecosystem-
based management action that holds promise
for remediating habitat constraints (20) and
effectively addressing the growth-mortality
trade-off in juvenile fish. For many fish spe-
cies, shallow lake zones provide valuable spawn-
ing and nursery habitats (21) and constitute
foraging areas that spatially overlap with safe
refuge areas within submerged macrophytes,
thereby contributing to fish recruitment and
productivity (22). An alternative strategy is
directly managing habitat structure by intro-
ducing coarse woody habitats (23, 24), an im-
portant functional habitat for different life
stages in many fish species (25). However,
in lakes, it has remained unclear whether
adding coarse woody habitats can effectively
increase fish abundance through either im-
proved reproduction (26) or provision of refuge
benefits (23, 27) that reduce juvenile mortality
(28), or if the practice simply alters fish dis-
tributions by attraction effects (29) and habi-
tat partitioning (23) without increasing overall
abundance (23, 30).
Previous studies have addressed selected as-
pects of the ecology and conservation value of
fish stocking and habitat enhancement in lakes
[e.g., (12, 23)]. However, lack of controls and
insufficient replication (6) [but see (31)] have
limited inference regarding the success of these
management measures on broader scales (32).
Whole-lake experiments (6, 23) have a large
potential to systematically evaluate ecosystem-
based habitat enhancements versus species-
focused stocking, particularly when conducted
in a before-after-control-impact (BACI) design
(33). Small freshwater ecosystems offer ex-
cellent opportunities for experimentation and
replication (32).
We present a large-scale replicated and con-
trolled set of whole-lake experiments in a trans-
disciplinary setting with strong participatory
involvement of local angling communities (34).
Using 20 mesotrophic gravel pit lakes (average
size 7 ha; table S1), we tested for the potential
for fish abundance–enhancing effects of three
types of management measures: fish stocking
with five species in four lakes, habitat enhance-
ment through additions of coarse wood bun-
dles in eight lakes, and shallow littoral zone
creation by excavation of riparian banks in a
subset of four wood-supplemented lakes (34)
(figs. S1 to S3). We used a BACI experimen-
tal design (including eight control lakes) and
monitored the fish community over 6 years to
test the following hypotheses: (i) that creating
shallow littoral habitats would most effectively
Radinger et al., Science 379, 946–951 (2023)
3 March 2023
1 of 6
RES EARCH | R E S E A R C H A R T I C L E
Shallow littoral
zones
Coarse woody
habitats
Shallow littoral
zones
Resources to
growth
Predation
mortality
Reproduction
Reproduction
Stocking
Predation risk-
sensitive
foraging
Stocking
Open habitat
Refuge habitat
Vulnerable
population component
Invulnerable
population component
Arena dynamics of foraging & predation risk taking
Survival
Quantity and
quality of spatial
refuge
Shallow littoral
zones
Coarse woody
habitats
Shallow littoral zone creation
Coarse woody habitat addition
Stocking
Ecosystem-based
habitat management
Single structure-focused
management
Single species-focused
management
Fig. 1. Population dynamic mechanisms emerging from ecosystem-based
habitat management through creating shallow littoral zones and coarse woody
habitat addition and from species-focused stocking management. In fish,
population constraints related to the density-dependent mortality bottleneck in
early juveniles are especially important (13, 14). The yellow box indicates the
foraging arena involved in the trade-off between fish growth and mortality in
littoral zones of lakes. The smallest length classes of fish face an important
trade-off between securing sufficient food resources to support growth to lengths
beyond their predators’ gape width while minimizing exposure to predation (17).
Therefore, they adjust their foraging behavior based on the perceived risk of
predation (i.e., predation risk–sensitive foraging). This trade-off between predation
risk taking and foraging is shaped by intra- and interspecific competition and the
spatial distribution of food resources. Juvenile fish may forage in profitable but
risky areas outside the refuge (i.e., the vulnerable population component) or
move into refuges (e.g., vegetation, coarse woody structures, or shallow water)
to limit mortality (i.e., the invulnerable population component) but at the expense
of food intake (18, 19). Although habitat enhancement can improve this growth-
mortality trade-off to allow a greater number of juvenile fish growing into the
population, stocking of otherwise naturally reproducing species would not modify
the spatial configuration of foraging arenas; it would only elevate competition or
predation without an opportunity to find refuge from mortality. Left and center
photos courtesy of Florian Möllers/AVN.
increase fish abundance by providing addi-
tional spawning and nursery grounds while
simultaneously reducing predation risk be-
cause of the beneficial spatial interspersion of
foraging with refuge habitats; (ii) that adding
coarse woody habitats would create fish ag-
gregations, and a simultaneous attraction of
predators and prey to the new structures might
manifest as neutral effects in total fish abun-
dance; and (iii) that fish stocking into popu-
lations that naturally reproduce in the lakes
would not lead to additive effects on fish
Radinger et al., Science 379, 946–951 (2023)
3 March 2023
2 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Changes in total fish abundance in response to the three management measures and across three sampling methods. Colored circles indicate model-
predicted mean CPUE before and after the management intervention. Dark circles indicate model-predicted CPUE for the control lakes. Error bars refer to the corresponding
95% bootstrapped CIs of the mean. Effect size estimates refer to the rate ratio of a given BACI interaction term (with 95% CI in parentheses).
abundance because of density-dependent mor-
tality regulation.
Creating shallow littoral zones effectively
enhances fish populations
Our study was based on a replicated and con-
trolled set of experiments involving 20 gravel
pit lakes (table S1 and fig. S1) and a sample of
159,943 fish captured 2 years before and 4 years
after implementing three management inter-
ventions (34). Contrasting relative abundance
(hereafter, abundance) changes between treated
and unmanipulated control lakes (BACI design)
(34) revealed that the creation of shallow littoral
zones (12.5% increased littoral area on average;
table S2 and figs. S2 and S3) was the most
effective method to enhance fish populations
(Fig. 2). Standardized total fish abundance
(catch per unit effort, CPUE) as assessed by
electrofishing (CPUEE) increased significantly
after shallow littoral zone creation by a factor
of 2.71 compared with control lakes [general-
ized linear mixed-effects model (GLMM); con-
fidence interval (CI) = 1.01 to 6.76, P = 0.045;
effect controlled for coarse wood additions
(34); table S3]. A similarly positive but non-
significant trend was observed from gillnet-
based abundance data (CPUEN) in response
to littoral zone creation and relative to con-
trols (BACI effect = 1.68, CI = 0.84 to 3.44, P =
0.14; table S4). Juvenile fish (<100 mm) par-
ticularly benefited from the ecosystem-based
management intervention of shallow littoral
zone creation, showing a significant >5-fold
abundance (CPUEjuv) increase compared with
control lakes (BACI effect = 5.25, CI = 1.29 to
37.84, P = 0.041; table S5).
Our findings suggest that creating shallow
littoral zones bolstered recruitment. The value
of shallow littoral zones has long been recog-
nized, particularly their importance during
the life cycle of almost all temperate fishes
(21, 35, 36). In view of the small spatial extent
of littoral area enhancement (table S2), the
clearly positive outcome of this measure is
noteworthy. In particular, roach (Rutilus rutilus),
one of the most frequent and abundant fish
species of temperate European lakes, consist-
ently increased in abundance in response to
littoral zone creation, with significant effects
detected for CPUEE and CPUEN (GLMM P <
0.05; Fig. 3 and tables S3 and S4), and a pro-
nounced trend in CPUEjuv (GLMM BACI effect =
5.00, P = 0.095; table S5). These positive effects
associated with littoral zone creation were
likely the result of enhanced reproduction and
improved nursery function (37, 38). How-
ever, ecosystem-wide benefits of littoral areas
extend beyond providing suitable spawning
grounds. In the presence of piscivores, littoral
areas can provide beneficial foraging areas
supporting juvenile growth through enhanced
benthic production and warm water, and their
vegetation cover and shallowness can effectively
reduce vulnerability to predation (18, 21, 39).
The profitable spatial overlap of forage and
relatively predation-safe refuge areas likely
contributed to enhanced juvenile fish devel-
opment and the observed increase in total
Radinger et al., Science 379, 946–951 (2023)
3 March 2023
3 of 6
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Changes in abundance of perch and roach in response to the three management measures and across three sampling methods. Colored circles indicate
model-predicted mean CPUE before and after the management intervention. Dark circles indicate model-predicted CPUE for the control lakes. Error bars refer to the
corresponding 95% bootstrapped CIs of the mean. Effect size estimates refer to the rate ratio of a given BACI interaction term (with 95% CI in parentheses).
fish abundance after shallow littoral zone
creation.
Convoluted effects of woody habitat additions
Although shallow zone creation was highly ef-
fective, habitat enhancement through coarse
woody habitat additions to 21% of the shore-
line (table S2 and fig. S3) alone did not sig-
nificantly enhance total fish abundance or
that of juveniles on average across all lakes
(GLMM all P > 0.05; Fig. 2 and tables S3 to
S5). Previous studies have shown that adding
coarse wood to lakes is not necessarily associated
with a short-term enhancement of fish popula-
tions (23, 24). Responses to structural habi-
tat enhancements in our study differed across
sampling methods and lakes and between fish
species, particularly between the two domi-
nant fish species in gravel pit lakes, roach and
European perch (Perca fluviatilis; hereafter,
perch). In temperate lakes, both species form
an important predator-prey relationship, with
roach acting as an abundant zooplanktivo-
rous fish and (large) perch as a key predator. In
perch, coarse wood additions and the creation
of littoral zones did not lead to significant
abundance changes compared with the con-
trol lakes (GLMM all P > 0.05; Fig. 3 and tables
S3 to S5). However, there was a strong but
nonsignificant trend toward increasing perch
CPUEN in response to coarse woody habitat
additions (GLMM BACI effect = 2.40, CI = 0.90
to 6.56, P = 0.113; Fig. 3 and table S4). This
trend suggests that wood additions could have
increased perch mobility, as has been previ-
ously found in other piscivorous fish (40) [but
see (41)], and thereby increased perch vulner-
ability to be caught by the passive sampling
gear. Alternatively, coarse wood additions might
have induced a spatial shift in perch from the
littoral to the more open sublittoral caused by
enhanced cover and increased foraging oppor-
tunities on prey at the edge of these structures
(30, 40). Spatial aggregation of fish near coarse
woody habitats has repeatedly been observed
(23, 42) and was also suggested by our results
showing a trend of relatively larger increases
in perch abundance at sites that were closer to
supplemented wood structures (GLMM BACI
effect = 0.998, P = 0.173; table S6 and fig. S4).
In roach, coarse woody habitat additions
resulted in a significant decrease of abundance
in the sublittoral, as indicated by CPUEN com-
pared with control lakes (BACI effect = 0.19,
CI = 0.06 to 0.59, P = 0.004; Fig. 3 and table
S4). However, there was no such evidence in
roach CPUEE and CPUEjuv (GLMM both P >
0.05; Fig. 3 and tables S3 and S5). The ob-
served decline may have been caused by lower
roach activity or the reduced use of sublittoral
habitats after wood additions and in response
to the elevated numbers in predators (43). Ad-
ditionally, the refuge function of coarse woody
habitats (27) might be less strong or universal
and depends on its structural complexity (28, 43).
Supplemented woody habitats might have
even facilitated predation, because predator-
prey interactions and predation rates change
with habitat structure and are often concen-
trated at the edge of complex habitats (30, 44).
In fact, structural habitats that simultaneously
attract predator and prey might become eco-
logical traps for the latter (45). Although fish
might mistakenly perceive coarse wood as ben-
eficial protective habitat, net-positive effects on
Radinger et al., Science 379, 946–951 (2023)
3 March 2023
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
fish abundance may vanish because of increased
predation rates. The attractiveness of struc-
tural habitats for fish that form ecological
traps has previously been demonstrated, e.g.,
associated with artificial reefs (45).
Our study outcome might have been in-
fluenced by the uniformity of the supple-
mented wood bundles, which did not fully
resemble the complexity of woody material
originating from riparian trees (25, 42), and by
our rather short-term observation time frame.
The first 4 years after manipulation might
have been an insufficient amount of time to
develop the full effects emerging from the long-
lasting transformative processes of coarse
wood, its colonization by invertebrates and
periphyton, and its effects on stabilizing lit-
toral habitats and macrophyte growth (25).
Long-term processes might also explain why
experimental coarse woody habitat removals
often resulted in declines of single (prey) fish
species (37, 46) [but see (47)], whereas recipro-
cal whole-lake additions were frequently not
associated with the short-term enhancement
of fish populations [(23) and observed in this
study: see Fig. 2].
Furthermore, the responses reported refer to
average effects over multiple lakes that under-
went a certain management measure compared
with several control lakes, all characterized by
inherent natural variability between ecosys-
tems and in time. For example, the abundance
of roach and perch increased in certain lakes
after coarse wood additions, whereas it de-
creased in others (fig. S5), which resulted in a
neutral mean effect over the sum of all studied
lakes. Some have argued that whole-lake ex-
periments using a low number of replicates of
treatments with modest manipulative inter-
ventions but strong environmental noise might
bear the risk of erroneously accepting the
hypothesis of no treatment effect (48). This
limitation cannot be fully excluded here de-
spite the grand scale of our experiments with
four (or more) replicate lakes manipulated at
reasonable real-world levels. With our repli-
cated study design, we aimed to account for
lake- and year-specific effects of confounding
factors (34) (e.g., differences in trophic state or
specific weather events) that may have other-
wise affected the results when solely relying
on a single lake experiment. Nevertheless,
or maybe for this very reason, we consider
the positive effects emerging especially from
ecosystem-based habitat management of lakes
by shallow littoral zone creation as particularly
robust and convincing.
No abundance-enhancing effects
of fish stocking
Of the three management measures investi-
gated, fish stocking achieved the poorest re-
sults. Stocking with five fish species (including
prey and predator species) at 97 kg/ha did not
produce any enhancing effects on fish abun-
dance (Fig. 2), with total CPUEE, CPUEN, and
CPUEjuv remaining at similar levels after stock-
ing (GLMM all P > 0.05; Fig. 2 and tables S3
to S5). When compared with controls, total
CPUEN tended to even decrease after stock-
ing (GLMM BACI effect = 0.57, CI = 0.29 to
1.10, P = 0.073; Fig. 2). At the species level,
neither stocking of species reproducing [com-
mon bream (Abramis brama), tench (Tinca
tinca), roach, and northern pike (Esox lucius)]
nor of species not reproducing in the study
lakes [pikeperch (Sander lucioperca)] resulted
in increased fish abundance (GLMM all P >
0.05; table S7). Roach abundance CPUEN de-
clined after stocking and when compared with
unstocked control lakes (BACI effect = 0.25,
CI = 0.08 to 0.81, P = 0.024; Fig. 3). Perch (not
stocked) abundance did not change after stock-
ing of other species and when compared with
control lakes (P > 0.05; Fig. 3 and tables S3 to
S5). These outcomes strongly agree with pre-
vious studies showing the very low additive ef-
fects of stocking (12–14). More specifically, for
lakes that are at or bouncing around carrying
capacity for juveniles, stocking additional fish
simply increases competition for food and
habitat without enhancing refuge opportuni-
ties from predation. Stocking may further lead
to poorly conditioned fish or rapid recapture
of stocked fish (13, 14); replacement of wild
with less fit, hatchery-produced fish (suffering
from domestication selection) (8, 12); and com-
petitive disadvantages of stocked fishes, which
might reduce the productivity of stocking-
enhanced fish populations in the long term
(49). We acknowledge that, as for other man-
agement measures, responses to fish stocking
can be strongly context and species dependent
and conditional on the specific lake character-
istics and its species composition.
Conclusions
Our large-scale replicated study design that
created 120 lake-years of observations revealed
that ecosystem-based habitat management
strongly outperforms the traditional, single-
species–focused practice of fish stocking and
single-structure–oriented management to sup-
port fisheries. Ecosystem-based habitat man-
agement, e.g., through the creation of shallow
littoral zones (20, 34), is the most promising
management tool to enhance fish abundance
in small lakes, especially when it effectively
improves the growth-predation risk trade-off
in early juvenile life stages. Holistic ecosystem-
based management emphasizes the impor-
tance of improving both the essential ecosystem
components and the societal dimension and
decision-making processes (1). Our experiment
was transdisciplinary and involved manipula-
tions under community governance by local
recreational fishing clubs and anglers, which
likely contributed to a rethinking of stocking
and fostered acceptance of more sustainable,
ecosystem-based alternatives (50). In the UN
Decade on Ecosystem Restoration, the impli-
cations of our work extend beyond fisheries
toward conservation more generally. A focus
on reestablishing central ecological processes
and habitats is likely to have stronger long-
term effects for the rebuilding of imperiled
species than narrow, species-focused conser-
vation actions.
REFERENCES AND NOTES
1. T. G. O’Higgins, M. Lago, T. H. DeWitt, Ecosystem-Based
Management, Ecosystem Services and Aquatic Biodiversity:
Theory, Tools and Applications (Springer, 2020).
J. Link, Ecosystem-Based Fisheries Management: Confronting
Tradeoffs (Cambridge Univ. Press, 2010).
2.
3. M. Ruckelshaus, T. Klinger, N. Knowlton, D. P. DeMaster,
Bioscience 58, 53–63 (2008).
4. C. J. Walters, in Successes, Limitations, and Frontiers in Ecosystem
Science, M. L. Pace, P. M. Groffman, Eds. (Springer, 1998),
pp. 272–286.
5. M. Palmer, A. Ruhi, Science 365, eaaw2087 (2019).
6. S. R. Carpenter, in Successes, Limitations, and Frontiers in
Ecosystem Science, M. L. Pace, P. M. Groffman, Eds. (Springer,
1998), pp. 287–312.
7. D. P. Armstrong, P. J. Seddon, Trends Ecol. Evol. 23, 20–25
(2008).
8. K. Lorenzen, M. C. M. Beveridge, M. Mangel, Biol. Rev. Camb.
Philos. Soc. 87, 639–660 (2012).
9. G. G. Sass, A. L. Rypel, J. D. Stafford, Fisheries 42, 197–209
(2017).
10. L. A. Eby, W. J. Roach, L. B. Crowder, J. A. Stanford,
Trends Ecol. Evol. 21, 576–584 (2006).
11. L. Laikre, M. K. Schwartz, R. S. Waples,
N. Ryman; GeM Working Group, Trends Ecol. Evol. 25, 520–529
(2010).
12. D. Hühn, K. Lübke, C. Skov, R. Arlinghaus, Can. J. Fish.
Aquat. Sci. 71, 1508–1519 (2014).
13. F. D. Johnston et al., Ecol. Appl. 28, 2033–2054 (2018).
14. K. Lorenzen, Philos. Trans. R. Soc. London B Biol. Sci. 360,
171–189 (2005).
15. R. Arlinghaus, C. Riepe, S. Theis, T. Pagel, M. Fujitani, J. Outdoor
Recreat. Tour. 38, 100475 (2022).
16. R. Arlinghaus, K. Lorenzen, B. M. Johnson, S. J. Cooke,
I. G. Cowx, in Freshwater Fisheries Ecology, J. F. Craig, Ed.
(Wiley, 2016), pp. 557–579.
17. E. E. Werner, D. J. Hall, Ecology 69, 1352–1366 (1988).
18. R. N. M. Ahrens, C. J. Walters, V. Christensen, Fish Fish. 13,
41–59 (2012).
19. C. J. Walters, F. Juanes, Can. J. Fish. Aquat. Sci. 50,
2058–2070 (1993).
20. J. L. Fischer, E. F. Roseman, C. M. Mayer, S. Qian, Ecol. Eng.
123, 54–64 (2018).
I. J. Winfield, Limnologica 34, 124–131 (2004).
21.
22. R. G. Randall, C. K. Minns, V. W. Cairns, J. E. Moore, Can. J.
Fish. Aquat. Sci. 53 (S1), 35–44 (1996).
23. G. G. Sass, S. R. Carpenter, J. W. Gaeta, J. F. Kitchell,
T. D. Ahrenstorff, Aquat. Sci. 74, 255–266 (2012).
24. K. E. Smokorowski, T. C. Pratt, Environ. Rev. 15, 15–41
(2007).
25. M. Czarnecka, Hydrobiologia 767, 13–25 (2016).
26. J. J. Taylor et al., Environ. Evid. 8, 19 (2019).
27. R. A. Everett, G. M. Ruiz, Oecologia 93, 475–486 (1993).
28. J. P. Ziegler, C. J. Dassow, S. E. Jones, A. J. Ross,
C. T. Solomon, Can. J. Fish. Aquat. Sci. 76, 998–1005 (2019).
29. C. A. Layman, J. E. Allgeier, J. Appl. Ecol. 57, 2139–2148
(2020).
30. G. G. Sass, C. M. Gille, J. T. Hinke, J. F. Kitchell, Ecol. Freshwat.
Fish 15, 301–308 (2006).
31. M. H. Olson et al., Fisheries 23, 6–12 (1998).
32. G. J. A. Hansen, J. W. Gaeta, J. F. Hansen, S. R. Carpenter,
Fisheries 40, 56–64 (2015).
33. A. P. Christie et al., J. Appl. Ecol. 56, 2742–2754 (2019).
34. Materials and methods are available as supplementary
materials.
35. W.-C. Lewin, T. Mehner, D. Ritterbusch, U. Brämick,
Hydrobiologia 724, 293–306 (2014).
36. S. Matern, T. Klefoth, C. Wolter, R. Arlinghaus, Hydrobiologia
848, 2449–2471 (2021).
Radinger et al., Science 379, 946–951 (2023)
3 March 2023
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
37. J. W. Gaeta, G. G. Sass, S. R. Carpenter, . Can. J. Fish.
50. M. Fujitani, A. McFall, C. Randler, R. Arlinghaus, Sci. Adv. 3,
Aquat. Sci. 71, 315–325 (2014).
e1602516 (2017).
38. U. Kahl, S. Hülsmann, R. J. Radke, J. Benndorf, Limnologica 38,
258–268 (2008).
39. V. Gotceitas, P. Colgan, Oecologia 80, 158–166 (1989).
40. Q. C. Smith, G. G. Sass, T. R. Hrabik, S. L. Shaw, J. K. Raabe,
Ecol. Freshwat. Fish 31, 454–468 (2021).
41. T. D. Ahrenstorff, G. G. Sass, M. R. Helmus, Hydrobiologia 623,
223–233 (2009).
42. M. G. Newbrey, M. A. Bozek, M. J. Jennings, J. E. Cook, Can. J.
Fish. Aquat. Sci. 62, 2110–2123 (2005).
43. C. S. DeBoom, D. H. Wahl, Trans. Am. Fish. Soc. 142,
1602–1614 (2013).
44. J. F. Savino, R. A. Stein, Environ. Biol. Fishes 24, 287–293 (1989).
45. S. E. Swearer et al., Front. Ecol. Environ. 19, 234–242 (2021).
46. G. G. Sass et al., Fisheries 31, 321–330 (2006).
47. K. E. Smokorowski, J. L. Pearce, W. D. Geiling, T. C. Pratt,
N. Am. J. Fish. Manage. 41, 142–157 (2021).
48. S. R. Carpenter, Ecology 70, 453–463 (1989).
49. M. W. Chilcote, K. W. Goodson, M. R. Falcy, Can. J. Fish.
Aquat. Sci. 68, 511–522 (2011).
51. J. Radinger et al., Data for: Ecosystem-based management
outperforms species-focused stocking for enhancing fish
populations, Figshare (2023); https://doi.org/10.6084/
m9.figshare.20528295.
ACKN OWLED GMEN TS
We thank the Angler Association of Lower Saxony and all
participating angling clubs for their participation; the chemical
laboratory at IGB for analyses of water samples; the many
helpful students, interns, and colleagues, especially A. Türck
and R. Nikolaus, for their valuable assistance during field
work; and S. R. Carpenter and two anonymous reviewers for
their insightful feedback on the first draft of the manuscript.
Funding: This study was jointly financed by the German Federal
Ministry of Education and Research (BMBF) and the German
Federal Agency for Nature Conservation (BfN) with funds granted
by the German Federal Ministry for the Environment, Nature
Conservation and Nuclear Safety (BMU; grant no. 01LC1320A).
Author contributions: R.A., T.K., and C.W. conceived and designed
the study. S.M., F.F., and T.K. performed the research. J.R. and
C.T.M. analyzed and visualized the data. J.R. and R.A. wrote the
paper with substantial input from all coauthors. Competing interests:
The authors declare no competing interests. Data and materials
availability: All data and computer code used in the analysis
are publicly available at the Figshare repository (51). License
information: Copyright © 2023 the authors, some rights
reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government works.
https://www.science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adf0895
Materials and Methods
Figs. S1 to S5
Tables S1 to S7
References (52–70)
MDAR Reproducibility Checklist
Submitted 13 October 2022; accepted 6 February 2023
10.1126/science.adf0895
Radinger et al., Science 379, 946–951 (2023)
3 March 2023
6 of 6
|
10.1126_science.ade7759
|
RES EARCH
ACTIVE MATTER
Ultrafast reversible self-assembly of living
tangled matter
Vishal P. Patil1†, Harry Tuazon2†, Emily Kaufman2, Tuhin Chakrabortty2, David Qin3,
Jörn Dunkel4*, M. Saad Bhamla2*
Tangled active filaments are ubiquitous in nature, from chromosomal DNA and cilia carpets to root
networks and worm collectives. How activity and elasticity facilitate collective topological
transformations in living tangled matter is not well understood. We studied California blackworms
(Lumbriculus variegatus), which slowly form tangles in minutes but can untangle in milliseconds.
Combining ultrasound imaging, theoretical analysis, and simulations, we developed and validated a
mechanistic model that explains how the kinematics of individual active filaments determines
their emergent collective topological dynamics. The model reveals that resonantly alternating helical
waves enable both tangle formation and ultrafast untangling. By identifying generic dynamical
principles of topological self-transformations, our results can provide guidance for designing classes
of topologically tunable active materials.
K nots determine the robustness and func-
tion of filamentous matter across a wide
range of scales, from the intertwined
yarns in ropes and fabrics (1) to the tan-
gled polymers in rubbers (2, 3) and gels
(4). The extraordinary stability of knotted ma-
terials arises from the intricate interplay of
mutual mechanical obstruction (5) and con-
tact friction (6) between adjacent filaments
(7, 8). As any fisherman or long-haired crea-
ture can confirm, creating knotty structures
(9) is not difficult: When soft elastic fibers
are randomly mixed together (10), they nat-
urally tend to form a highly disordered tan-
gled state (11, 12). By contrast, untangling a
complex knot presents a daunting and his-
torically infamous (13) task. Certain biolog-
ical species such as the California blackworm
(Lumbriculus variegatus) (14) have evolved
to solve both the tangling and the untangling
problem with great efficiency by using only a
relatively basic set of neurons and muscles.
Exactly how they are able to do this remains
poorly understood.
When considered from an active matter per-
spective, worm tangles constitute an archetypal
example of an autonomous filamentous mate-
rial that can self-assemble, shape-shift, and ex-
hibit emergent collective functions (15, 16). In
minutes, a group of initially dispersed California
blackworms (14) can self-organize into a per-
sistent three-dimensional (3D) tangled structure,
but they require only a few tens of milliseconds
1Department of Bioengineering, Stanford University, 475 Via
Ortega, Stanford, CA 94305, USA. 2School of Chemical and
Biomolecular Engineering, Georgia Institute of Technology,
Atlanta, GA 30318, USA. 3Wallace H. Coulter Department of
Biomedical Engineering, Georgia Institute of Technology,
Atlanta, GA 30332, USA. 4Department of Mathematics,
Massachusetts Institute of Technology, 77 Massachusetts
Avenue, Cambridge, MA 02139, USA.
*Corresponding author. Email: dunkel@mit.edu (J.D.);
saadb@chbe.gatech.edu (M.S.B.)
†These authors contributed equally to this work.
to disentangle upon sensing danger (movie
S1). Blackworms, as well as some of their rela-
tives (17), use the tangled state to efficiently
execute a range of essential biological functions,
such as temperature maintenance, moisture
retention, and collective locomotion (18, 19).
Perhaps more importantly, the ability to es-
cape rapidly (20) from the tangle can often be
a lifesaving escape response from predators
(14) and environmental threats (16). Motivated
by an interest to understand the biophysical
mechanisms by which filamentous organisms
can achieve both robust tangling and ultrafast
untangling, we combined ultrasound imaging
experiments and elasticity theory to explain
how individual worm gaits give rise to col-
lective topological dynamics and transitions
between tangled and untangled states. By
mapping worm tangling to percolation (21)
and picture-hanging puzzles (22), we show
how resonantly tuned helical waves can en-
able self-assembly and rapid unknotting of
filamentous matter, thus revealing a generic
dynamical principle that can guide the de-
sign of new active materials.
Ultrasound experiments
Blackworms can assemble into topologically
intricate tangles consisting of anywhere from
5 to 50,000 worms (Fig. 1A) (16). Our ultrasound
experiments, conducted on worm tangles im-
mobilized in gelatin (movie S2), allowed for
the reconstruction of the 3D structure of a
living tangle (Fig. 1, B and C, and supplemen-
tary materials, materials and methods). This
revealed a picture of the tangle as a strongly
interacting system, in which the worms are
tightly packed (Fig. 1D) and most worms are in
contact with most other worms (Fig. 1E). In
addition to describing the arrangement of
contact, the nontopological structure of the
worm tangle can also be described on the basis
of the variation of geometric quantities both
within and between different worms. To ana-
lyze the tangle geometry, we approximated
each worm as a curve, x(s), parameterized by
arc length, s, which can be characterized by
local in-plane curvature, k(s), and an out-of-
plane 3D torsion, t(s). These geometric quanti-
ties give rise to bending strain, D ¼ kh (Fig. 1F),
and chirality, c = k2t (Fig. 1G), where h is the
worm radius (23). The 3D distributions of
both strain and chirality are primarily het-
erogeneous (Fig. 1, F and G) and decay rap-
idly as functions of the spatial separation,
x (cid:2) y
j (Fig. 1, H and I). For small values of
j
x (cid:2) y
j, the correlation functions are domi-
j
nated by intraworm interactions, but decorrela-
tion occurs once rC begins to include interworm
effects. In particular, rC ≈ 0 for both strain and
chirality once x (cid:2) y
j > 2:5h, which indicates
the existence of an effective radius, heff = 1.25h.
This effective radius is a signature of the ul-
trasound protocol (23), which requires the
tangles to undergo a small dilation. The rapid
decorrelation demonstrates that strain and
chirality are not described by 3D continuum
fields, illustrating the difficulty of constructing
a continuum theory for the living tangle. Un-
derstanding the mesoscale structure of the
tangle requires moving beyond purely geo-
metrical properties.
j
Topological analysis of the tangle geometry
allows us to distinguish between different forms
of contact. The intuitive notion that worms
that intertwine should interact more strongly
than worms that simply touch can be captured
by considering the linking number (24), Lk, of
the ith worm and the jth worm
½
Þ
ð1Þ
Lkij ¼
j½
(cid:5)= xi sð Þ(cid:2)
Þ ¼ xi sð Þ (cid:2) xj sð Þ
∫dsds Gij (cid:3) @sGij (cid:4) @sGij
ð
1
4p
where Gij s; sð
xj sð Þj(cid:5),
and xi and xj are the curves representing the
ith and jth worms. Although traditionally de-
fined only for closed curves, the linking num-
ber of open curves quantifies entanglement by
taking an average of the amount of intertwin-
ing in every 2D projection (23, 25). Visually,
j > 1=2 appear to wind
pairs of worms with Lkj
around each other (Fig. 2, A and B). However,
Lk is not sensitive to contact, which must ul-
timately mediate every worm–worm interac-
tion. Accordingly, we defined a more sensitive
measure called “contact link,” or cLk, by set-
ting cLk ¼ Lkj
j for worms in contact and
cLk = 0 otherwise. In contrast to the contact
matrix (Fig. 1D), the contact link matrix (Fig.
2C) identifies a far smaller number of key in-
teractions, thus providing a sparser represen-
tation of tangle state. This is evident from the
tangle graph (Fig. 2D), which shows worm–
worm interactions with cLk > 1=2. Despite
being a function of pairwise tangling as opposed
to a function of total entanglement, the ro-
bustness of contact link as a tangling measure
is evident through its behavior across different
Patil et al., Science 380, 392–398 (2023)
28 April 2023
1 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Three-dimensional ultrasound data reveal the mechanical structure
of active, biological worm tangles. (A) Topologically complex tangle formed
by Lumbriculus variegatus consisting of approximately 200 worms. Scale bar, 3 mm.
(B and C) Ultrasound imaging reveals the interior structure of a 12-worm tangle.
Scale bar, 5 mm. (D and E) The contact matrix and contact graph confirm that
the worm tangle is a strongly interacting system. (F and G) Three-dimensional
½
D xð Þ; D yð Þ
experimental data enable the visualization of strain D, and chirality c, fields within
the tangle, revealing that the worms form achiral tangles. (H and I) Decorrelation of
j ≈ 2:5h
strain, rC
(dotted lines) demonstrates the limits of a continuum elastic theory for worm
tangles. The decorrelation length scale indicates the existence of an effective radius,
heff ~ 1.25h, arising from the preparation of tangles for ultrasound (23).
(cid:5), and chirality, rC c xð Þ; c yð Þ
(cid:5), over distances of x (cid:2) y
j
½
ultrasound datasets. For example, the proba-
bility distribution of the contact link between
two worms, a measure of topological inter-
action strength, retains a characteristic shape
across worm tangles (Fig. 2E). Additionally,
the total contact link (23), obtained by sum-
ming all the pair contact links from Fig. 2C,
is sensitive to the contact structure of the tan-
gle. When treated as a collection of tubes, the
contact structure of a tangle can be altered by
modifying the tube radius. The total contact
link as a function of tube radius behaves sim-
ilarly across datasets as the tubes are thick-
ened from zero radius to larger radii (Fig. 2F).
Thus, by incorporating topological information
(25, 26) as well as geometric information, cLk
captures core structural motifs that are repro-
ducible across different experiments, enabling
us to compare experimentally observed worm
tangles with tangled structures generated from
dynamical simulations.
Worm dynamics
The ability of the blackworm to form tangles
in minutes (Fig. 3A) but rapidly unravel in
milliseconds (Fig. 3B) is a key biological and
topological puzzle (27, 28). To understand the
dynamical process that gives rise to tangle
formation, we experimentally studied the head
trajectories of single worms (Fig. 3, A to D,
and supplementary materials, materials and
methods). Because these experiments were per-
formed in a shallow fluid well (height ~2 mm),
the projection of the trajectories into 2D (Fig. 3,
j
x(cid:6) tð Þ
A to D) did not cause substantial information
loss. To capture the winding motions associ-
ated with tangling and untangling, we assumed
the worm head has a preferred speed, v ¼
j
i, and focused on the worm turning di-
h
rection, q tð Þ ¼ arg x(cid:6) tð Þ. The q trajectories can
be described approximately in terms of two pa-
(cid:6) ji
rameters, the average angular speed, a ¼ hjq
(Fig. 3, A and B), and the rate, l, at which q
changes sign. These quantities can be esti-
mated from the noisy trajectory data (23).
Although the characteristic timescales for
slow tangling and ultrafast untangling, a−1,
differ by two orders of magnitude, rescaling
the q trajectories for each gait by a−1 revealed
similar underlying dynamics (Fig. 3, A and B).
This similarity reflects the biological constraints
(cid:6)
Patil et al., Science 380, 392–398 (2023)
28 April 2023
2 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Topological structure of worm tangles. (A) Individual topological
interactions between chosen worms (solid color) mapped in detail by 3D ultrasound
reconstructions (as in Fig. 1, B and C). Scale bar, 5 mm. (B) Topological analysis
enables the classification of tangle structure by distinguishing between (left column)
contact and (right column) linking interactions, which are defined by having
j > 1=2. (C) Contact link, cLk, defined as the absolute value of the
linking number Lkj
link between worms separated by at most 2heff, identifies the strongest topological
interactions within the tangle. The contact link between nontouching worms is 0.
Pairs of worms with cLk > 1=2 are highlighted in red. (D) The tangle graph provides
a sparser representation of tangle state than does the contact graph. Edges are
j > 1=2 [red bordered squares in (C)]. (E) The probability
present between pairs of worms with cLk > 1=2, that is, worms that both touch
and have Lkj
distribution of the contact link between two worms is stable across ultrasound
datasets. Pairs of worms with contact link greater than 1=2 (dotted line) lead
to edges in the corresponding tangle graphs (inset), with edge thickness given
by the value of the contact link. (F) Increasing the tube radius of the worm
curves modifies the contact structure of the tangle and thus increases the total
contact link (23). The radius dependence of total contact link is similar across
different tangles and indicates the presence of an effective radius, as in Fig. 1,
H and I, that is distinct from the true radius, h.
on locomotion machinery (29) and indicates
that tangling and untangling can be captured
by the same mathematical model. To confirm
this, we first formulated a minimal 2D model
of worm-head dynamics, which we then gen-
eralized to a full 3D dynamical picture.
A minimal 2D model can be constructed by
focusing on the helical worm-head dynamics
that we identified experimentally (Fig. 3). The
quantities a, l, and v motivate the following
stochastic differential equation (SDE) model
for a worm-head trajectory (23)
x(cid:6) ¼ vnq þ xT ; q
(cid:6) ¼ s t; lð
Þa þ xR
ð2Þ
where xT and xR are noise terms, nq is a unit
vector in the q direction, and s(t; l) switches
between +1 and −1 at rate l. These trajecto-
ries can be further classified by dimensionless
parameters. The chirality number, g ¼ a=2pl,
distinguishes between the tangling and untan-
gling gaits (Fig. 3, A and B). This nondimen-
sional parameter corresponds to the average
number of right- or left-handed loops traced
out by the worm before changing direction
and provides an intuitive way of understand-
ing the topological properties of each gait.
When g is large, worms wind around each
other before switching direction, producing
a coherent tangle. By contrast, for small g,
the worms change direction before they are
able to wind around one another and so re-
main untangled. This relationship between
tangle state and chirality can be thought of as
a form of resonance. Our trajectory model
thus explains how the characteristic helical
waves produced by untangling worms medi-
ate topology (movie S3).
We next showed that these conclusions gen-
eralize to a full 3D mechanical model of worm
gaits. To model the worms, we performed
elastic-fiber simulations in which the worms
were treated as Kirchhoff filaments (5, 30–34)
with active head dynamics. The head motions
were prescribed by the SDE model (2) together
with additional 3D drift (23); the body re-
sponded elastically. The resulting worm col-
lectives could form 3D tangled structures
(Fig. 3E) consistent with those seen in our
experiments, as quantified by contact link
(Fig. 3F). The tangling and untangling be-
havior in these simulations appears to be a
function of the chirality number, g, further
confirming its importance (Fig. 3, E and F,
Patil et al., Science 380, 392–398 (2023)
28 April 2023
3 of 7
RES EARCH | R E S E A R C H A R T I C L E
1.6s
8.2s
14.7s
21.3s
6
10/
50/
90/
130/
0
-6
10
Time (s)
20
44ms
221ms
398ms
575ms
3
10/
50/
90/
130/
0
-3
300
Time (ms)
600
A
g
n
i
l
g
n
a
T
l
e
g
n
a
i
g
n
n
r
u
T
B
g
n
i
l
g
n
a
t
n
U
l
e
g
n
a
i
g
n
n
r
u
T
E
s
n
o
i
t
a
u
m
S
l
i
0
50
Time (1/ )
100
150
C
D
F
4
2
m
r
o
w
r
e
p
k
n
i
l
t
c
a
t
n
o
c
l
t
a
o
T
0
0
Start
End
0
0
Time (s)
Time (1/ )
24.5
150
0
0
Time (ms)
Time (1/ )
664
150
Time (1/ )
150
Fig. 3. Resonant helical worm-head dynamics give rise to numerically
reproducible weaving and unweaving gaits. (A and B) Experimentally
observed worm-head trajectories projected into 2D can be approximated by their
angular direction, q tð Þ ¼ arg x˙(t), in both the (A) tangling and (B) untangling
cases (movie S3). q is characterized by an average turning rate, a ¼ hjq
rate of switching from left turning (red points, q
points, q
weaving (g ¼ 0:68) and unweaving (g ¼ 0:36) gaits. a−1 defines an intrinsic
timescale for tangle assembly and disassembly. Scale bars, 3 mm. (C and D)
Experimentally measured head trajectories of three worms (different colors)
executing the (C) tangling and (D) untangling gaits demonstrate the (C)
(cid:6) < 0). The chirality number, g ¼ a=2pl, captures the difference between
(cid:6) > 0) to right turning (blue
(cid:6) ji, and a
formation or (D) removal of topological obstructions within a similar time in units
of a−1. Scale bars, 5 mm. (E) Simulations of active Kirchhoff filaments
demonstrate that the gaits described in (A) and (B) are sufficient for reversible
tangle self-assembly (movie S3). The topological state is quantified with tangle
graphs (inset). Tangling filaments have large g [(E), top row, and (A)], and
untangling filaments have small g [(E), bottom row, and (B)]. The initial tangled
state [(E), bottom row] is obtained from 3D ultrasound reconstruction. Average
worm lengths range from 40 mm (top row) to 28 mm (bottom row), with a radius of
0.5 mm throughout. Displayed worms are thickened to aid visualization. (F) The total
contact link per worm (Fig. 2) obtained from simulations reveals the rate at which
tangles form [(E), top row, purple dots] and unravel [(E), bottom row, green dots].
Patil et al., Science 380, 392–398 (2023)
28 April 2023
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Bioinspired tangling model reveals phase diagram underlying
topological assembly and manipulation of generic tangles. (A) Two-
dimensional cross sections of 3D ultrasound reconstructions indicate the
obstacle landscape faced by a worm exhibiting quasi-2D motion. (B) A 2D mean-
field tangling model measures the winding of a worm-head trajectory (purple and
green curves) around fixed obstacles in the plane (solid circles). Contact
winding, cWp, around obstacles that are far from the trajectory (23) is 0. Points
with cWp > 1 contribute to the tangling index, T , of a trajectory (Eq. 3).
Trajectories with small chirality number, g, have smaller overall contact winding.
(C) Measured values of g and R for blackworms undergoing tangling (purple disks)
or untangling (green disks) dynamics lie in regions of the tangle phase space
corresponding to tangling (red, T > 2) and untangling (blue, T < 2), where the
critical value T (cid:7) ¼ 2 corresponds to a connected tangle graph, and hence a
minimally tangled state. The untangling data consists of n ¼ 25 worms (small green
disks) from n ¼ 5 separate 12-worm untangling experiments, and the tangling
data consists of n ¼ 18 worms (small purple disks) from n ¼ 4 separate 5-worm
tangling experiments. The large disks show mean values of g and R obtained
by averaging over all worms in a given experiment (23). Error bars show
standard deviation. (D) Worm gaits predicted by the tangling phase diagram
enable robust control of topological transitions (movie S4). Tangle formation
and avoidance can be controlled at fixed R by varying g, both for low worm
speeds v, (middle, R ¼ 3:4) and high worm speeds (right, R ¼ 1:0). Worms have
a length of 40 mm and a radius of 0.5 mm. Displayed worms are thickened to aid
visualization. (E) Timescales of tangling and untangling from simulations in
(D) are set by a−1, which varies from the low v simulations (t < 200=a; a(cid:2)1 ≈ 0:1 s)
to the high v simulations (t < 200=a; a(cid:2)1 ≈ 4 ms). The largest cluster of touching
worms produced by the low v, large g simulation is used as the initial condition
for the high v simulations (23), causing an apparent jump in total contact
link per worm at t ¼ 200=a. Tangle graphs (insets) illustrate the topological
structure of the simulated tangles.
and movie S3). This formulation of a 3D dy-
namical model allows us to understand how
the dynamics of single worms produces worm
collectives with distinct topologies.
Mean-field theory
On the basis of our analysis of the worm tra-
jectories, we built a mean-field tangling model,
which establishes a mapping between tan-
gling and percolation (Fig. 4). To formulate
an analytically tractable model, we treat
the worm motion as essentially 2D, so each
worm effectively moves in a 2D slice of the
3D tangle (Fig. 4, A and B). As a given worm
moves in a plane, its head traces out a curve,
x(t) (Fig. 4B, purple and green curves), de-
scribed by Eq. 2. The worm can encounter a
set of obstacles, L, that indicate intersections
of the other worms with the given plane (Fig.
4B, colored circles). The 3D notion of contact
link between worms can be mapped to this 2D
picture (22) by considering the winding of the
trajectory, x(t), around the obstacles, p ∈ L. We
can assign a value to each obstacle, p, that
measures how much x(t) winds around p and
how close the trajectory gets to p (Fig. 4B).
We call this value the “contact winding” of x(t)
Patil et al., Science 380, 392–398 (2023)
28 April 2023
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
about p and denote it cWp (23). Thresholding
and averaging all the contact winding num-
bers yields a tangling index
T ¼
*
X
p∈L
+
ð
Q cWp (cid:2) 1
Þ
ð3Þ
where the step function Q returns 1 if cWp >
1 and 0 otherwise. The tangling index there-
fore counts the number of obstacles that a
worm winds around and illustrates that worm-
head trajectories with different chirality num-
ber, g, are topologically distinct (Fig. 4B). For
example, by changing direction frequently,
trajectories with small g have smaller overall
contact winding (Fig. 4B, bottom row). Because
the tangling index counts entanglements, it can
also be interpreted as a measure of the mean
degree of a tangle graph. Because connected
graphs asymptotically have a mean degree of
at least 2, we identify T (cid:7)≈ 2 as the critical
tangling index separating tangled states, with
T > 2, from loose states, with T < 2. Near-
critical trajectories (23) bear a notable resem-
blance to curves that solve the famous picture-
hanging puzzle (22), which asks how to hang a
picture on two pegs so that it falls if either peg
is removed. Critical worm gaits could there-
fore be associated with such topological quick-
release mechanisms; our tomographic recon-
structions do indicate that worms form near-
critical tangles (Fig. 2F), thus balancing tangle
stability with ability to disentangle rapidly.
The tangling index enables the topological
state to be predicted from worm motion and
spacing (Fig. 4C). Assuming small noise terms
(23), the worm-head trajectories are charac-
terized by speed v, turning rate a, and angular
switching rate l; ‘ captures the worm spacing.
This leads to two dimensionless quantities,
the chirality number, g ¼ a=2pl, and the loop
number, R ¼ v=a‘, which measures the size of
the loops produced by the worm trajectory in
units of ‘. The resulting phase diagram, T g; Rð
Þ,
explains the observed values of g and R for
worms executing tangling and untangling gaits
(Fig. 4C). The timescale of these topological
transformations depends on a−1, which can
take any value for fixed g and R. Because a(cid:2)1 ¼
R‘v(cid:2)1 , the associated topological transfor-
mation timescale is small for fast worms and
large for slow worms, which is in agreement
with observed worm behavior (Fig. 3, A and
B), provided that R and ‘ stay approximately
constant. The tangling phase diagram further
demonstrates that the loop number, R, can
also be used to control topological state. For
example, larger values of R allow a worm to
wind around more obstacles, increasing topol-
ogical complexity. However, for R > 0.5, the
chirality number, g, is the key determinant of
tangle state (Fig. 4C), indicating that tangle
topology can be controlled purely by chang-
ing the rate, l, at which the turning direction
switches. The validity of this intuitive picture
was confirmed with 3D simulations, demon-
strating that by tuning g, active filaments can
be programmed to reversibly tangle and un-
tangle at any head speed v (Fig. 4D and movie
S4). The phase diagram therefore reveals how
tangle topology can be robustly controlled by
manipulating only the chiral dynamics of the
constituent filaments (Fig. 4, D and E, and
movie S4).
Discussion
Blackworm locomotion lies close to the critical
tangling threshold (Fig. 4C), indicating that
blackworm gaits are mechanically optimized
for crossing the tangling–untangling barrier.
However, our mean-field tangling model pre-
dicts a large space of tangling and untangling
strategies, within which blackworms occupy
a relatively small region. In addition, at fixed
g and R, the tangling and untangling time-
scale, a−1, can take any value, underscoring the
size of the locomotion space. Accounting for
energetics helps identify the topological strat-
egies that are inefficient for blackworms. For
example, untangling with small R requires
forming small, energetically costly loops. Sim-
ilarly, untangling by means of the linear tra-
jectories corresponding to large R gaits requires
braids to be unraveled by pulling rather than
unweaving, a motion associated with a higher
friction penalty (7, 23). Furthermore, blackworm
dynamics are necessarily multifunctional, and
topological requirements must be balanced
with the need to support efficient, biologically
feasible locomotion (14, 32, 35). For example,
the helical waves of alternating chirality that
promote untangling have also been identified
in the context of worm swimming (14). How-
ever, the highly entangled region of phase
space with g > 1, R > 0.5 suggests that there are
stable tangle topologies not accessed by the
worm collectives. Such a tangle could contain
chiral filaments, in contrast to our observed
living worm tangles (Fig. 1G). The chirality
number and loop number thus demonstrate
how complex topologies may be created and
tested beyond the biologically feasible regime.
Active helical waves produced by the mo-
tion of individual worms facilitate collective
tangling and ultrafast untangling. Because the
underlying mechanisms are generic, and be-
cause the predictions of elasticity theory are
known to generalize across a wide range of
scales (31), it is relevant to ask whether the
results of our mean-field tangling model could
apply to other systems of packed and tangled
fibers. Our model additionally demonstrates
methods for fine control of tangle topology,
opening up the possibility of programming a
wide range of behaviors into a single topolog-
ically adaptive material by harnessing the large
internal state space of tangles. The framework
developed here could help in better understand-
ing the mechanical advantages of specific
classes of tangles and aid in the development
of multifunctional materials based on topol-
ogical properties.
REFERENCES AND NOTES
1. P. B. Warren, R. C. Ball, R. E. Goldstein, Phys. Rev. Lett. 120,
158001 (2018).
2. P.-G. de Gennes, J. Chem. Phys. 55, 572–579 (1971).
3. S. Edwards, T. A. Vilgis, Rep. Prog. Phys. 51, 243–297 (1988).
4. M. L. Gardel et al., Science 304, 1301–1305 (2004).
5. V. P. Patil, J. D. Sandt, M. Kolle, J. Dunkel, Science 367, 71–75
(2020).
6. C. A. Daily-Diamond, C. E. Gregg, O. M. O’Reilly, Proc. R. Soc.
London Ser. A473, 20160770 (2017).
7. T. G. Sano, P. Johanns, P. Grandgeorge, C. Baek, P. M. Reis,
Extreme Mech. Lett. 55, 101788 (2022).
8. P. Johanns et al., Extreme Mech. Lett. 43, 101172 (2021).
9. Z. Chen, U. Pace, J. Heldman, A. Shapira, D. Lancet, J. Neurosci.
6, 2146–2154 (1986).
10. D. M. Raymer, D. E. Smith, Proc. Natl. Acad. Sci. U.S.A. 104,
16432–16437 (2007).
11. A. Belmonte, M. J. Shelley, S. T. Eldakar, C. H. Wiggins,
Phys. Rev. Lett. 87, 114301 (2001).
12. B. W. Soh, I. R. Gengaro, A. R. Klotz, P. S. Doyle, Phys. Rev.
Res. 1, 033194 (2019).
13. L. H. Kauffman, S. Lambropoulou, in Introductory Lectures on
Knot Theory, vol. 46 of Series on Knots and Everything,
L. H. Kauffman, S. Lambropoulou, S. Jablan, J. H. Przytycki,
Eds. (World Scientific, 2012), pp. 187–247.
14. C. D. Drewes, Aquatic Oligochaetes (Springer, 1999), pp. 263–269.
15. T. Heeremans, A. Deblais, D. Bonn, S. Woutersen, Sci. Adv. 8,
eabj7918 (2022).
16. Y. Ozkan-Aydin, D. I. Goldman, M. S. Bhamla, Proc. Natl. Acad.
Sci. U.S.A. 118, e2010542118 (2021).
17. J. L. Kaster, Hydrobiologia 180, 191–193 (1989).
18. C. Nguyen et al., Front. Phys. (Lausanne) 9, 734499 (2021).
19. N. R. Franks et al., Proc. Biol. Sci. 283, 20152946 (2016).
20. A. Worley, A. B. Sendova-Franks, N. R. Franks, R. Soc. Open
Sci. 6, 181626 (2019).
21. A. J. T. M. Mathijssen, J. Culver, M. S. Bhamla, M. Prakash,
Nature 571, 560–564 (2019).
22. E. D. Demaine et al., Theory Comput. Syst. 54, 531–550
(2014).
23. Materials and methods are available as supplementary
materials.
24. A. Qu, D. L. James, ACM Trans. Graph. 40, 106 (2021).
25. E. Panagiotou, K. C. Millett, S. Lambropoulou, J. Phys. A Math.
Theor. 43, 045208 (2010).
26. E. Panagiotou, L. H. Kauffman, Proc. R. Soc. London Ser. A 477,
20210440 (2021).
27. L. Zirbes et al., PLOS ONE 7, e32564 (2012).
28. S. Kuei, A. M. Słowicka, M. L. Ekiel-Jeżewska, E. Wajnryb,
H. A. Stone, New J. Phys. 17, 053009 (2015).
29. D. L. Hu, J. Nirody, T. Scott, M. J. Shelley, Proc. Natl. Acad. Sci.
U.S.A. 106, 10081–10085 (2009).
30. M. Bergou, M. Wardetzky, S. Robinson, B. Audoly, E. Grinspun,
ACM Trans. Graph. 27, 1–12 (2008).
31. B. Audoly, Y. Pomeau, Elasticity and Geometry: From Hair Curls
to the Non-linear Response of Shells (Oxford University Press,
2010).
32. C. W. Wolgemuth, T. R. Powers, R. E. Goldstein, Phys. Rev. Lett.
84, 1623–1626 (2000).
33. W. Huang, M. Khalid Jawed, Comput. Fluids 228, 105038
(2021).
34. W. Gilpin, S. Uppaluri, C. P. Brangwynne, Biophys. J. 108,
1887–1898 (2015).
35. A. Kudrolli, B. Ramirez, Proc. Natl. Acad. Sci. U.S.A. 116,
25569–25574 (2019).
36. V. P. Patil, vppatil28/tangle_code: Tangle simulation code,
Zenodo (2023); doi: https://doi.org/10.5281/zenodo.7519605.
37. V. P. Patil et al., Zenodo (2023); https://doi.org/10.5281/
zenodo.7519337.
AC KNOWLED GME NTS
We thank S. Emelianov (Georgia Tech) for sharing instrument
facilities for ultrasound imaging. J. Dunkel thanks the World
Premier International Research Center Initiative International
Institute for Sustainability with Knotted Chiral Meta Matter at
Hiroshima University for hospitality and support while parts of this
work were completed. Funding: This work was supported by a
MathWorks Fellowship (V.P.P.), a Stanford Science Fellowship (V.P.P.),
Patil et al., Science 380, 392–398 (2023)
28 April 2023
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
the NSF Graduate Research Fellowship Program (H.T.), a Georgia
Institute of Technology (Georgia Tech) President’s Fellowship (H.T.
and D.Q.), the Georgia Tech President’s Undergraduate Research
Award (E.K.), the MIT Mathematics Robert E. Collins Distinguished
Scholar Fund (J.D.), and Sloan Foundation Grant G-2021-16758 (J.D.).
M.S.B. acknowledges funding support from NIH Grant R35GM142588;
NSF Grants MCB-1817334; CMMI-2218382; and CAREER IOS-1941933,
and the Open Philanthropy Project. Author contributions: V.P.P.,
H.T., J.D., and M.S.B. conceptualized the research. V.P.P. and J.D.
developed theory. V.P.P. performed simulations and analytical
calculations. H.T. and M.S.B. designed the experiments. H.T., E.K.,
T.C., and D.Q. conducted the ultrasound experiments, for which T.C.
and V.P.P. performed the analysis. H.T. and E.K. conducted the worm
tangling and untangling experiments. J.D. and M.S.B. supervised
the research. V.P.P., H.T., J.D., and M.S.B. contributed to writing the
manuscript. All authors discussed and revised the manuscript.
Competing interests: The authors declare that they have no
competing interests. Data and materials availability: The code used
for numerical simulations is available at Zenodo (36). Additional
datasets are available at Zenodo (37). License information:
Copyright © 2023 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science. No
claim to original US government works. https://www.science.org/
about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade7759
Materials and Methods
Supplementary Text
Figs. S1 to S15
Movies S1 to S4
References (38–58)
View/request a protocol for this paper from Bio-protocol.
Submitted 7 September 2022; accepted 1 March 2023
10.1126/science.ade7759
Patil et al., Science 380, 392–398 (2023)
28 April 2023
7 of 7
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.