Add files using upload-large-folder tool

.gitattributes

@@ -96,3 +96,15 @@ d133fe86da3f5e7fc0805a9516ff43084c5bcb580a23b834688486bfdb132260/preprint/images
 d133fe86da3f5e7fc0805a9516ff43084c5bcb580a23b834688486bfdb132260/preprint/images/Figure_1.tif filter=lfs diff=lfs merge=lfs -text
 d133fe86da3f5e7fc0805a9516ff43084c5bcb580a23b834688486bfdb132260/preprint/images/Figure_4.tif filter=lfs diff=lfs merge=lfs -text
 d133fe86da3f5e7fc0805a9516ff43084c5bcb580a23b834688486bfdb132260/preprint/images/Figure_5.tif filter=lfs diff=lfs merge=lfs -text
+8c2073492d425364bbbe1b837523a84f3d3cb228d38041789f767035e5ccd7e4/preprint/images/Figure_5.tif filter=lfs diff=lfs merge=lfs -text
+8c2073492d425364bbbe1b837523a84f3d3cb228d38041789f767035e5ccd7e4/preprint/images/Figure_4.tif filter=lfs diff=lfs merge=lfs -text
+8c2073492d425364bbbe1b837523a84f3d3cb228d38041789f767035e5ccd7e4/preprint/images/Figure_2.tif filter=lfs diff=lfs merge=lfs -text
+8c2073492d425364bbbe1b837523a84f3d3cb228d38041789f767035e5ccd7e4/preprint/images/Figure_7.tif filter=lfs diff=lfs merge=lfs -text
+8c2073492d425364bbbe1b837523a84f3d3cb228d38041789f767035e5ccd7e4/preprint/images/Figure_1.tif filter=lfs diff=lfs merge=lfs -text
+8c2073492d425364bbbe1b837523a84f3d3cb228d38041789f767035e5ccd7e4/preprint/images/Figure_6.tif filter=lfs diff=lfs merge=lfs -text
+8c2073492d425364bbbe1b837523a84f3d3cb228d38041789f767035e5ccd7e4/preprint/images/Figure_3.tif filter=lfs diff=lfs merge=lfs -text
+b23f24d7cb67e73af08721dd80f6c2b96c31be47afbac7a1d20ec1dd7d4fbd1d/preprint/images/Figure_6.tif filter=lfs diff=lfs merge=lfs -text
+b23f24d7cb67e73af08721dd80f6c2b96c31be47afbac7a1d20ec1dd7d4fbd1d/preprint/images/Figure_2.tif filter=lfs diff=lfs merge=lfs -text
+b23f24d7cb67e73af08721dd80f6c2b96c31be47afbac7a1d20ec1dd7d4fbd1d/preprint/images/Figure_1.tif filter=lfs diff=lfs merge=lfs -text
+b23f24d7cb67e73af08721dd80f6c2b96c31be47afbac7a1d20ec1dd7d4fbd1d/preprint/images/Figure_3.tif filter=lfs diff=lfs merge=lfs -text
+b23f24d7cb67e73af08721dd80f6c2b96c31be47afbac7a1d20ec1dd7d4fbd1d/preprint/images/Figure_4.tif filter=lfs diff=lfs merge=lfs -text

0072c2609668ba1c13914173f2a943ea6ee628a9fdbc1b605b78b41def605213/preprint/images_list.json

@@ -0,0 +1,18 @@
+[
+    {
+        "type": "image",
+        "img_path": "images/Figure_1.png",
+        "caption": "Overview of sampling sites and recovered biodiversity for ESB samples from terrestrial, limnic and marine ecosystems spanning a time window from 1985 to 2022. A) ESB sampling sites of tree leaves, zebra mussels, bladderwrack and blue mussels across Germany. B) Total OTU and order level richness (log scale) of different taxonomic groups across the tree of life associated with tree leaves, zebra mussels, bladderwrack and blue mussels. Taxonomic groups are represented by different colors. Icons refer to natural DNA sampler species.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_2.png",
+        "caption": "Multidecadal trends of \u03b1-diversity, as well as temporal- and spatial \ud835\udefd-diversity in communities across the tree of life natural DNA sampler organisms from the ESB. A) Trends in OTU richness (\u03b1-diversity) of the associated communities between 1991 and 2021. B) Temporal OTU turnover (temporal \ud835\udefd-diversity) of the associated communities calculated from the turnover component of Jaccard distance is shown as a function of the distance in years between samples of the same sampling site. C) Spatial OTU turnover (spatial \ud835\udefd-diversity) of the associated communities between 1991 and 2021. Trends describe the degree of dissimilarity in community composition between different sampling locations over time. OTU richness and turnover values were summarized as mean with standard error bars across sampling locations and time windows. D)-F) Diversity trends from A)-C) reduced to their respective slope. Filled circles indicate significant departures from the null expectations generated through the dynamic model for community assembly, suggesting an out-of-equilibrium dynamic.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    }
+]

0072c2609668ba1c13914173f2a943ea6ee628a9fdbc1b605b78b41def605213/preprint/preprint.md

@@ -0,0 +1,179 @@
+# Abstract
+
+Detecting the imprints of global environmental change on biological communities is a paramount task for ecological research¹,²,³,⁴. But due to a lack of standardized long-term biomonitoring data, community assembly in the Anthropocene remains poorly understood⁵,⁶,⁷. Novel sources of data for analyzing biodiversity change across time and space are urgently needed⁸,⁹. By metabarcoding highly standardized biota samples from a long-term pollution monitoring archive in Germany, we here analyze four decades of community diversity for tens of thousands of species across the tree of life. The samples – tree leaves, marine macroalgae, and marine and limnic mussels – represent natural community DNA samplers¹⁰,¹¹, preserving a taxonomically diverse imprint of their associated biodiversity at the time of collection⁹,¹². We find no evidence for localized diversity declines¹, but a strong, gradual compositional turnover as a universal pattern of temporal biodiversity change in Germany’s terrestrial and aquatic ecosystems⁵,¹³,¹⁴. This turnover results in biotic homogenization in terrestrial ecosystems, indicative of country-wide diversity decline. Our work uncovers a massive cryptic biodiversity loss across ecosystems, resulting from out-of-equilibrium ecological dynamics. This highlights the immense promise of alternative sample sources to provide standardized time series data of biodiversity change in the Anthropocene.
+
+[Biological sciences/Ecology/Community ecology](/browse?subjectArea=Biological%20sciences%2FEcology%2FCommunity%20ecology) [Biological sciences/Ecology/Biodiversity](/browse?subjectArea=Biological%20sciences%2FEcology%2FBiodiversity) [Earth and environmental sciences/Ecology/Ecological genetics](/browse?subjectArea=Earth%20and%20environmental%20sciences%2FEcology%2FEcological%20genetics)
+
+# Main
+
+Global ecosystems have undergone unprecedented change in the past decades. To understand the consequences of this transformation for biodiversity, standardized, long-term and taxonomically broad biomonitoring data are essential. Such data, however, are lacking for most taxa and ecosystems<sup>15,16</sup>. Available time series are often short or incomplete and limited to a few target taxa and study locations<sup>7,17</sup>.
+
+A promising solution to this problem is offered by environmental specimen banks (ESBs), long-term pollution monitoring archives, which collect indicator organisms from various terrestrial and aquatic ecosystems. ESB samples are collected according to highly standardized protocols and stored at low temperatures<sup>18</sup>, ensuring an excellent preservation of the sample-associated nucleic acids. The indicator species collected by ESBs comprise metaorganisms representing diverse communities of interacting taxa, each of which has left a trace of its DNA in the sample. Recent work has shown that such metaorganisms are excellently suited “natural samplers” for studying the surrounding biota via DNA metabarcoding<sup>9,10</sup>. Examples include sponges and mussels, which filter DNA particles from the water column, hence providing a snapshot of their aquatic communities, or plant material which retains DNA traces of interacting arthropods<sup>11,12</sup>. The long-term archives of ESBs can thus serve as community DNA samplers, providing the standardized time series data so urgently needed to understand biodiversity change.
+
+Here, we use metabarcoding of samples from the German ESB, one of the largest, technologically advanced and longest-operating ESBs, to reconstruct biodiversity change across broad taxonomic, spatial and temporal scales. We focus on four sample types from terrestrial, limnic and marine ecosystems and recover sample associated communities from across the tree of life. Using archived leaves from tree canopies, we characterize communities of canopy-associated fungi, bacteria and arthropods. Samples of a dominant marine macroalgal species along the European coastline reveal coastal bacterial and animal communities associated with the alga. Finally, marine and limnic mussels provide an imprint of the surrounding bacterial and eukaryotic communities in oceans and rivers. Our analysis yields highly standardized community diversity data for tens of thousands of species in parallel, many of which represent cryptic and largely understudied taxonomic groups.
+
+Using these data, we explore common patterns of biodiversity change across the tree of life in terrestrial, marine and limnic habitats in Germany over the past four decades. Recent work has highlighted different responses of biota to changing environmental conditions in the Anthropocene. We explore the generality of these responses by testing three hypotheses for the temporal variation of biodiversity: 1) Stressful conditions have led to extinctions of species at individual sites, i.e. local declines of α-diversity<sup>1</sup>. 2) The losses of species are countered by the immigration of novel taxa, leading to a pattern of biotic turnover (𝛽-diversity) without an overall loss of ɑ-diversity. This turnover could occur 2a) gradually with the changing environment<sup>5</sup>, or 2b) rapidly, when the community reaches a tipping point<sup>2,20</sup>. 3) The immigration of species across broad geographic scales, for example of widespread invasive taxa, leads to a pattern of biotic homogenization, i.e. a loss of 𝛽-diversity across space<sup>14</sup>. To test these hypotheses, we developed a dynamic model for community assembly, based on the equilibrium theory of island biogeography<sup>19</sup>, which generates null expectations of diversity trends in the absence of disturbance (Extended Data Figure 3A).
+
+Figure 1. Overview of sampling sites and recovered biodiversity for ESB samples from terrestrial, limnic and marine ecosystems spanning a time window from 1985 to 2022. A) ESB sampling sites of tree leaves, zebra mussels, bladderwrack and blue mussels across Germany. B) Total OTU and order level richness (log scale) of different taxonomic groups across the tree of life associated with tree leaves, zebra mussels, bladderwrack and blue mussels. Taxonomic groups are represented by different colors. Icons refer to natural DNA sampler species.
+
+We metabarcoded 550 samples of archived natural sampler organisms from three marine, nine limnic and nine terrestrial sites in Germany (Figure 1B; Supplementary Table 1). Using time series of leaves from trees (1985-2022), marine macroalgae (1985-2021), and marine (1985-2020) as well as limnic mussels (1994-2020), we reconstructed communities associated with these organisms at the time of sampling. Rarefaction and bootstrapping analyses indicated sufficient sequencing depth and sampling size for biodiversity estimations at both local and regional scales (Extended Data Figure 1 + Extended Data Table 1). Our analysis recovered highly diverse prokaryote and eukaryote communities (Figure 1B), a total of 66,184 operational taxonomic units (OTUs) in 751 orders and 102 phyla. Tree leaves recovered 5,183 OTUs of bacteria in 94 orders, 6,250 fungal OTUs in 113 orders and 3,271 metazoan (mainly arthropod) OTUs in 24 orders. We found 5,474 bacterial OTUs in 101 orders and 787 metazoan OTUs in 78 orders in marine macroalgae. 21,266 OTUs of bacteria in 180 orders and 3,551 OTUs of microeukaryotes (mainly algae and protozoa) in 160 orders were found in marine blue mussels. In limnic zebra mussels, we found 14,292 bacterial OTUs in 184 orders, 5,587 microeukaryote (mainly algae and protozoa) OTUs in 173 orders and 523 metazoan OTUs in 71 orders (Figure 1B, Supplementary Table 2).
+
+The detected species accurately represented their respective ecosystems and natural sampler organisms (Extended Data Figure 2; Supplementary Table 3). For example, various typical coastal metazoans were found in bladderwrack samples and numerous species of eukaryotic algae reflect the phytoplankton community surrounding mussels. Typical canopy-dwelling arthropods and leaf-associated fungi and bacteria were recovered from leaves (Supplementary Table 3). The species detected in tree canopies showed a high specificity for their respective host tree, with each tree species harboring a unique community of leaf-associated taxa (Extended Data Figure 2). Typical monophagous arthropods or host-specific microorganisms were exclusively found on their respective host tree (Supplementary Table 3). The communities recovered from all sample types were also site-specific, with many taxa limited to single sites, likely indicating their specific ecological requirements or biogeographic affinities. For example, highly disparate communities were associated with bladderwrack and blue mussels from the Baltic versus the Northern Sea; these two seas are distinguished by pronounced salinity differences. Additionally, our two Northern Sea sampling sites, which are separated by about 200 km, harbored different sets of taxa. The same held true for the leaf-associated communities at different forest sites across Germany and the zebra mussel-associated communities in different river systems (Extended Data Figure 2), which have different biogeographic affinities. These results underline the accuracy and efficiency of natural sampler time series for the targeted enrichment of biological communities.
+
+Figure 2. Multidecadal trends of α-diversity, as well as temporal- and spatial 𝛽-diversity in communities across the tree of life natural DNA sampler organisms from the ESB. A) Trends in OTU richness (α-diversity) of the associated communities between 1991 and 2021. B) Temporal OTU turnover (temporal 𝛽-diversity) of the associated communities calculated from the turnover component of Jaccard distance is shown as a function of the distance in years between samples of the same sampling site. C) Spatial OTU turnover (spatial 𝛽-diversity) of the associated communities between 1991 and 2021. Trends describe the degree of dissimilarity in community composition between different sampling locations over time. OTU richness and turnover values were summarized as mean with standard error bars across sampling locations and time windows. D)-F) Diversity trends from A)-C) reduced to their respective slope. Filled circles indicate significant departures from the null expectations generated through the dynamic model for community assembly, suggesting an out-of-equilibrium dynamic.
+
+The communities recovered by the natural DNA samplers provide highly standardized and well resolved temporal biodiversity data across ecosystems and the tree of life. This offers an unprecedented opportunity to test for general patterns of biodiversity change. We analyzed temporal patterns of α-diversity (OTU richness), compositional OTU turnover (temporal 𝛽-diversity) and biotic homogenization (spatial 𝛽-diversity). To evaluate the significance of the observed trends, we compared them to null expectations of community changes in the absence of anthropogenic disturbances. These expectations were based on a new dynamic model of community ecology we developed, built upon the equilibrium theory of island biogeography<sup>5,19</sup> (Extended Data Figure 3A; see Methods).
+
+Recent and often prominently featured work on insect decline has highlighted losses of α-diversity (richness) as a driving feature of biodiversity change in the Anthropocene<sup>1,17,21</sup>. Our data did not support this assumption, refuting our first hypothesis. No universal trend for α-diversity was identified. Irrespective of ecosystems and taxonomic groups, richness either remained stable, increased, or declined (Figure 2A+D, Extended Data Figure 4A). A particularly pronounced drop in richness was found in marine prokaryotes. In contrast, richness strongly increased in limnic prokaryotes (Figure 2A+D) across the studied river systems. Interestingly, a reverse pattern was found for limnic and marine micro-eukaryotes, which showed α-diversity losses of in limnic habitats, but increases in marine sites (Figure 2A+D). A more common pattern was found in terrestrial ecosystems, where most studied taxa showed a slight, increase of α-diversity at individual. Surprisingly, we even found a slight albeit significant increase in richness for terrestrial arthropod communities, supporting recent work suggesting that forest canopy arthropods are not affected by pronounced insect decline<sup>9</sup>.
+
+In contrast to α-diversity, a clear universal trend was found for compositional turnover (temporal 𝛽-diversity). This trend significantly deviated from the null expectations in all studied communities, supporting our second hypothesis<sup>5</sup>. We found considerable losses of OTUs in all communities, which however, were countered by the immigration of novel, likely better adapted taxa (Figure 2B+E, Extended Data Figure 4B). This led to an out-of-equilibrium dynamic with a stronger-than-expected temporal turnover of community composition (Extended Data Table 2). Slightly different temporal dynamics of compositional turnover were observed across taxa (Extended Data Figure 3B). The higher rates of local extinction and immigration observed in metazoans and algae led to OTU-poor, rapidly-changing communities. Bacterial communities were also characterized by higher immigration rates, but combined with lower local extinction rates. This resulted in more diverse, dynamic communities. In contrast, lower extinction and immigration rates observed in fungi generated slower taxonomic turnover compared with other communities, but still faster than expected. The observed turnovers were gradual in all communities: no abrupt breaks of community composition were detected (Figure 2B+E). The observation of a gradual compositional turnover as a predominant pattern of biodiversity change in the Anthropocene is well supported by other recent work<sup>4,5,22</sup>. We found no indication of rapid state shifts in communities reaching tipping points<sup>2</sup>. This is also supported by the patterns at higher taxonomic levels. While we observed a significant compositional OTU turnover in all the ecosystems and taxonomic groups, the turnover did not affect higher taxonomic ranks: the temporal composition of phyla, classes or orders remained remarkably stable in all the natural samplers (Extended Data Figure 5). Hence, those taxa that disappeared were primarily replaced by closely related and likely functionally similar ones, which are better adapted to changing environmental conditions.
+
+The observed temporal change of 𝛽-diversity, i.e. a compositional turnover of communities, can also be seen at the level of thousands of individual OTUs within our data. Each sampler type and data set showed replacements of various OTUs, with gains and losses approximately balanced (Extended Data Figure 6). Various novel colonizers were detected, among them typical invasive species. For example, the occurrence of the invasive Pacific oyster (<em>Crassostrea gigas</em>) in coastal bladderwrack in our 2004 data reflects the timing of its actual introduction into the German coastline<sup>23</sup>. Similarly, the colonization of various plant pathogens can be traced over time. For example, Norway spruce and European beech at the national park Harz have been colonized by a <em>Taphrina</em> species and spruces at Lake Belau were infested by <em>Coniothyrium</em> (Supplementary Table 4). At the same time, declines in the occurrence of several taxa are evident, for example the common periwinkle (<em>Littorina littorea</em>) in marine ecosystems and the green silver-lines (<em>Pseudoips prasinana</em>) in forest ecosystems. Both are considered common species and their decline was previously overlooked. ESB samples thus can serve as an important early warning system for both the decline of local species and the emergence of problematic invaders<sup>9,10</sup>.
+
+We last explored patterns of spatial 𝛽-diversity, e.g. biotic homogenization. Similar to α-diversity, no universal trend across ecosystems was evident from our data (Figure 2C+F). Most aquatic sampler organisms did not show a clear trend of spatial homogenization over time. Some of the communities became even more heterogeneous across space, for example in marine microeukaryotes. The increasing spatial heterogeneity was especially evident in prokaryotes and microeukaryotes in limnic sites (Figure 2C+F). Different river systems in Germany have become more heterogenous over time, possible due to colonization by novel taxa from different biogeographic regions, for example between Danube and Rhine system. Interestingly, a general pattern of homogenization across space was evident in terrestrial canopy ecosystems. All the taxonomic groups, from prokaryotes to fungi and arthropods, showed a significant spatial homogenization of communities over time in the terrestrial samples (Figure 2C+F, Extended Data Figure 4C). As in the temporal 𝛽-diversity, the observed change was gradual rather than abrupt. This finding is in line with recent observations that widespread generalist species benefit from environmental change and replace more specialized ones<sup>14,24,25,26</sup>. Terrestrial ecosystems did not lose diversity at the level of individual sites, in fact many sites even showed an increasing α-diversity. Despite the local increase of diversity, a clear loss of diversity was evident at larger spatial scales. This indicates the immigration of widespread taxa across the tree of life in Germany’s canopy ecosystems.
+
+In summary, our work closes critical gaps in understanding of temporal and spatial biodiversity change in the Anthropocene. Archived natural samplers provide time series data of unprecedented taxonomic resolution and standardization and enable the study of patterns of biodiversity across the tree of life<sup>18</sup>, from prokaryotes to multicellular eukaryotes. We report an out-of-equilibrium dynamic in both aquatic and terrestrial ecosystems in Germany, potentially driven by anthropogenic disturbances<sup>27</sup>. The widespread taxonomic replacement and biotic homogenization we observe across deeply divergent taxonomic groups constitute serious environmental issues, which are hitherto insufficiently characterized. Our data indicate that thousands of species have disappeared from Germany's aquatic and terrestrial ecosystems in the past decades. The majority of the increasing or declining taxa we observed are highly cryptic and rarely detected by monitoring programs and biodiversity research, which focus on prominent taxa like plants<sup>28</sup>, vertebrates<sup>29</sup> or insects<sup>1</sup>. At the same time, many of the taxa disappearing from our dataset represent critical elements in food webs and are essential to ecosystem stability, for example phytoplankton in aquatic ecosystems or arthropods in tree canopies<sup>30</sup>. Our data highlight a hitherto poorly understood facet of environmental change: a massive cryptic loss of biodiversity by turnover and biotic homogenization.
+
+# References
+
+1. Seibold, S. et al. Arthropod decline in grasslands and forests is associated with landscape-level drivers. *Nature*, **574**, 671-674 (2019). https://doi.org/10.1038/s41586-019-1684-3.
+2. Barnosky, A. D. et al. Approaching a state shift in Earth’s biosphere. *Nature*, **486**, 52-58 (2012). https://doi.org/10.1038/nature11018.
+3. Jandt, U. et al. More losses than gains during one century of plant biodiversity change in Germany. *Nature*, **611**, 512-518 (2022). https://doi.org/10.1038/s41586-022-05320-w.
+4. Blowes, S. A. et al. The geography of biodiversity change in marine and terrestrial assemblages. *Science*, **366**, 339-345 (2019). DOI: 10.1126/science.aaw1620
+5. Dornelas, M. et al. Assemblage time series reveal biodiversity change but not systematic loss. *Science*, **344**, 296-299 (2014). DOI: 10.1126/science.1248484.
+6. Johnson, T. F. et al. Revealing uncertainty in the status of biodiversity change. *Nature*, **628**, 788-794 (2024). https://doi.org/10.1038/s41586-024-07236-z
+7. Crossley, M. S. et al. No net insect abundance and diversity declines across US Long Term Ecological Research sites. *Nature Ecology & Evolution*, **4**, 1368-1376 (2020). https://doi.org/10.1038/s41559-020-1269-4.
+8. Outhwaite, C. L., Gregory, R. D., Chandler, R. E., Collen, B. & Isaac, N. J. B. Complex long-term biodiversity change among invertebrates, bryophytes and lichens. *Nature Ecology & Evolution*, **4**, 384-392 (2020). https://doi.org/10.1038/s41559-020-1111-z.
+9. Krehenwinkel, H. et al. Environmental DNA from archived leaves reveals widespread temporal turnover and biotic homogenization in forest arthropod communities. *eLife*, **11**, e78521 (2022a). https://doi.org/10.7554/eLife.78521.
+10. Junk, I., Schmitt, N. & Krehenwinkel, H. Tracking climate-change-induced biological invasions by metabarcoding archived natural eDNA samplers. *Current Biology*, **33**, 943-944 (2023). https://doi.org/10.1016/j.cub.2023.07.035.
+11. Mariani, S., Baillie, C., Colosimo, G. & Riesgo, A. Sponges as natural environmental DNA samplers. *Current Biology*, **29**, 401-402 (2019). https://doi.org/10.1016/j.cub.2019.04.031.
+12. Weber, S. et al. Molecular diet analysis in mussels and other metazoan filter feeders and an assessment of their utility as natural eDNA samplers. *Molecular Ecology Resources*, **23**, 471-485 (2023). https://doi.org/10.1111/1755-0998.13710.
+13. Pilotto, F. et al. Meta-analysis of multidecadal biodiversity trends in Europe. *Nature Communications*, **11**, 3486 (2020). https://doi.org/10.1038/s41467-020-17171-y.
+14. Blowes, S. A. et al. Synthesis reveals approximately balanced biotic differentiation and homogenization. *Science Advances*, **10**, eadj9395. (2024). DOI: 10.1126/sciadv.adj9395.
+15. Bálint, M. et al. Environmental DNA Time Series in Ecology. *Trends in Ecology & Evolution*, **33**, 945-957 (2018). DOI: 10.1016/j.tree.2018.09.003.
+16. Magurran, A. E. et al. Long-term datasets in biodiversity research and monitoring: assessing change in ecological communities through time. *Trends in Ecology & Evolution*, **25**, 574-582 (2010). https://doi.org/10.1016/j.tree.2010.06.016.
+17. Wagner, D., Grames, E. M., Forister, M. L., Stopak, D. Insect Decline in the Anthropocene: Death by a thousand cuts. *PNAS*, **118**, e2023989118 (2021a). https://doi.org/10.1073/pnas.2023989118.
+18. Fliedner, A. et al. Environmental specimen banks and the European Green Deal. *Science of the Total Environment*, **852**, 158430 (2022). https://doi.org/10.1016/j.scitotenv.2022.158430.
+19. MacArthur, R.H., & Wilson, E.O. An equilibrium theory of insular zoogeography. *Evolution*, **17**, 373-387 (1963). https://doi.org/10.2307/2407089.
+20. Maucieri, D.G., Starko, S. & Baum, J.K. Tipping points and interactive effects of chronic human disturbance and acute heat stress on coral diversity. *Proceedings of the Royal Society B*, **290**, 20230209 (2023). https://doi.org/10.1098/rspb.2023.0209.
+21. Wagner, D., Fox, R, Salcido, D. M. & Dyer, L. A. A window to the world of global insect declines: Moth biodiversity trends are complex and heterogeneous. *PNAS*, **118**, e2002549117 (2021b). https://doi.org/10.1073/pnas.2002549117.
+22. Magurran, A. E. et al. Divergent biodiversity change within ecosystems. *PNAS*, **11**, 1843-1847 (2018). https://doi.org/10.1073/pnas.1712594115.
+23. Brandt, G., Wehrmann, A. & Wirtz, K.W. Rapid invasion of *Crassostrea gigas* into the German Wadden Sea by larval supply. *Journal of Sea Research*, **59**, 279-296 (2008). https://doi.org/10.1016/j.seares.2008.03.004.
+24. Stuart-Smith, R. D., Mellin, C., Bates, A. E. & Edgar, E. J. Habitat loss and range shifts contribute to ecological generalization among reef fishes. *Nature Ecology & Evolution*, **5**, 656–662 (2021). https://doi.org/10.1038/s41559-020-01342-7.
+25. Dharmarajam, G., Gupta, P., Vishnudas, C. K. & Robin, V.V. Anthropogenic disturbance favours generalist over specialist parasites in bird communities: Implications for risk of disease emergence. *Ecology Letters*, **24**, 1859-1868 (2021). https://doi.org/10.1111/ele.13818.
+26. Xu, W. B. et al. Regional occupancy increases for widespread species but decreases for narrowly distributed species in metacommunity time series. *Nature Communications*, **14**, 1463 (2023). https://doi.org/10.1038/s41467-023-37127-2.
+27. Marta, S., Brunetti, M., Manenti, R., Provenzale, A. & Ficetola, G.F. Climate and land-use changes drive biodiversity turnover in arthropod assemblages over 150 years. *Nature Ecology & Evolution*, **5**, 1291-1300 (2021). https://doi.org/10.1038/s41559-021-01513-0.
+28. Jandt, U. et al. More losses than gains during one century of plant biodiversity change in Germany. *Nature*, **611**, 512-518 (2022). https://doi.org/10.1038/s41586-022-05320-w.
+29. Semenchuk, P. et al. Relative effects of land conversion and land-use intensity on terrestrial vertebrate diversity. *Nature Communications*, **13**, 615 (2022). https://doi.org/10.1038/s41467-022-28245-4.
+30. Field, C. B., Behrenfeld, M. J., Randerson, J. T. & Falkowski, P. Primary production of the biosphere: Integrating terrestrial and oceanic components. *Science*, **281**, 237-240 (1998). DOI: 10.1126/science.281.5374.237.
+
+# Methods
+
+## Specimen Bank Data
+
+The German Environmental Specimen Bank (ESB) has been in operation since the early 1980s. The ESB collects samples of indicator species from various terrestrial and aquatic ecosystems. These species serve as accumulators of environmental chemicals and provide a detailed image of pollution and ecosystem health. The long operational history of ESBs results in exceptionally long time series of environmental pollution. To make pollution analysis comparable between years, ESB samples are collected according to highly standardized protocols. Samples are taken at the same time of the year, at identical sites and using identical protocols. ESB collection is done using sterile equipment to avoid carryover of even trace amounts of pollutants between samples. To ensure preservation of unstable chemical compounds, the samples are stored over liquid nitrogen after collection and for the long term, halting all chemical and biological degradation. To acquire an integrative view of pollution in an ecosystem, ESB samples are large, each one including hundreds to thousands of specimens or tissue compartments (in case of trees) <sup>18,31-36</sup>. Each sample is cryo-milled to a fine powder of a grain size of 200 µm, thoroughly homogenizing all traces of chemicals <sup>37</sup>.
+
+Recent work shows that ESBs, being set up as chemical pollution archives, also makes them ideal for studies on biodiversity change. ESB indicator species can be considered metaorganisms or natural community DNA samplers, which preserve an imprint of the surrounding biological community at the time of sampling. The highly standardized and contamination-free sampling and sample processing conditions, coupled with storage at ultra-low temperatures, make ESB samples perfectly suited for metabarcoding. The cryomilling of large ESB samples also guarantees an even distribution of community DNA traces among the sample and breaks open cell walls of various microorganisms, whose DNA is then uniformly released. Former studies have already extensively tested and highlighted the suitability of different ESB samples for retrospective biodiversity monitoring <sup>9,10,12</sup>. Here, we use four different types of ESB samples from terrestrial, limnic and marine habitats as natural community DNA samplers to measure biodiversity change across four decades.
+
+1) **Tree leaves:** The ESB collects leaves from three common tree species in Germany, namely European beech (<em>Fagus sylvatica</em>), Norway spruce (<em>Picea abies</em>) and Lombardy poplar (<em>Populus nigra</em>). The leaves are collected once annually or biannually and serve as samplers for aerial pollutants deposited on the leaves <sup>31,32,33</sup>. ESB leaf samples are sampled from different forest ecosystem types, spanning a land use gradient from core zones of national parks to timber forests, forests neighboring agricultural sites, and urban parks. Sample series from nine sites were included in this study, starting from 1985. Each sample contains hundreds of leaves from at least 15 individual trees, milled to a fine powder. These samples contain DNA traces of all organisms that interacted with the tree canopy at the time of collection <sup>9</sup>. Here we characterize communities of canopy-associated arthropods, fungi and bacteria.
+
+2) **Bladderwrack** (<em>Fucus vesiculosus</em>): This macroalgae is widespread along the European coastline and makes up a significant part of the biomass in Europe’s coastal ecosystems. Marine pollutants are enriched in the tissue of the algae, making it an ideal sentinel species for pollution <sup>34</sup>. Three sites have been sampled annually or biannually for bladderwrack thalli beginning in 1985. ESB samples from two North Sea sites are collected at intervals of two months, six times a year, and then merged into a pooled annual sample. The third site at the Baltic Sea is sampled twice a year. Bladderwrack is a critical species in coastal ecosystems, providing a habitat for countless taxa. All these taxa leave detectable DNA traces in the ESB sample <sup>12</sup>. Here we characterize communities of animals and bacteria that interacted with the bladderwrack.
+
+3) **Blue mussels** (<em>Mytilus edulis</em>): This is the most common mussel in coastal ecosystems of Northern and Central Europe. Blue mussels constantly filter the water column for planktonic organisms and organic particles. In doing so, they enrich pollutants in their tissue, making them an excellent sentinel species for pollution monitoring <sup>35</sup>. The ESB has collected blue mussels at three coastal sites in Germany since 1985. The mussel’s entire soft tissue including respiratory water is used for the sample. Annual or biannual samples of hundreds of mussels are compiled from six sampling events at the North Sea and two at the Baltic Sea. With each mussel filtering roughly 1 liter of water per hour, these samples contain a comprehensive imprint of the annual planktonic biodiversity at the sampling site <sup>12</sup>. Here we characterize communities of eukaryotic plankton and mussel-associated bacteria.
+
+4) **Zebra mussels** (<em>Dreissena polymorpha</em>): The limnic zebra mussel is an invasive species from the Black Sea region, which has colonized nearly all major rivers of Germany since the 1960s. Like the blue mussel, zebra mussels are highly efficient filter feeders. Since the 1990s, zebra mussels are reared by the ESB on special plate stacks, which are then placed in four major German rivers for about one year, allowing the mussels to accumulate pollutants in their tissue. The mussels are then collected from the plate stacks, immediately deep-frozen and a sample of soft tissue including respiratory water is compiled from thousands of mussels <sup>36</sup>. The samples from nine sites used here provide an overview of limnic biodiversity from major rivers of Germany <sup>12</sup>. Here we characterize communities of animals, eukaryotic plankton and mussel-associated bacteria.
+
+## DNA extraction, library preparation, sequencing and sequence processing
+
+Samples were processed as described in Krehenwinkel et al. (2022a) and Junk et al. (2023). Work steps were performed under clean benches to avoid carryover and cross-contamination. We isolated DNA from 200 mg of homogenate from each sample, which was shown to be sufficient to recover the sample-associated diversity <sup>9</sup>. DNA was extracted in one or two replicates depending on the sample type (Supplementary Table 5) using a CTAB protocol (OPS Diagnostics, New Jersey, USA), which proved best suited to extract high purity DNA from these sample types. The DNA extracts were then amplified for different DNA barcode markers to enrich various taxonomic groups from the samples (for a list of barcode markers and PCR conditions see Supplementary Table 5). PCR was performed using the Qiagen Multiplex PCR kit (Hilden, Germany) in 10-µl volumes according to the manufacturer’s protocols. Primers were chosen to amplify the associated community, but not the ESB indicator species itself, whose DNA dominates the extract. To characterize bacterial communities (all sample types), we amplified the V1 or V5-7 region of 16SrDNA <sup>39,40,41</sup>. For terrestrial arthropods (tree leaf samples), we used a mitochondrial Cytochrome Oxidase I (COI) marker <sup>38</sup>. For fungi (tree leaf samples) we used the ITS1 region of the nuclear ribosomal cluster <sup>42,43</sup>. The variable V9 region of nuclear 18SrDNA was targeted to characterize communities of aquatic animals and eukaryotic plankton (bladderwrack and mussel samples <sup>12</sup>; for primer details, see Supplementary Table 5). PCR success was checked on 1.5 % agarose gels, and the PCR products were then amplified in another round of PCR to add Illumina Truseq adapters and unique combinations of dual indexes <sup>44</sup> (Supplementary Table 5). All final libraries were pooled in approximately equimolar amounts, cleaned of leftover primers using 1X AMPure beads XP (Beckman-Coulter, California, USA), and then sequenced on an Illumina Miseq (California, USA) using paired-end sequencing with 500-cycle V2 and 600-cycle V3 kits. To ensure reproducibility of our data and to recover rare species, we ran several PCR replicates for every sample, which were indexed and sequenced separately. The number of PCR replicates was adapted based on sample type and marker and varied between three and six (Supplementary Table 5). Blank DNA extractions were included in every batch of extractions, and non-template control PCRs were run alongside all PCR reactions. All controls were sequenced along with the samples to provide a baseline for carryover or cross-contamination during processing.
+
+Forward and reverse reads were merged using PEAR <sup>45</sup> with a minimum quality of 20 and a minimum overlap of 50 bp. The merged reads were then quality-filtered by limiting the number of expected errors in a sequence to 1 <sup>46</sup> and transformed to fasta format using USEARCH <sup>47</sup>. Primer sequences were trimmed off using UNIX scripts. Long 18SrDNA amplicons (~350 bp) of aquatic metazoan and microeukaryotic plankton generated from zebra mussels were trimmed to match the corresponding short amplicon of ~150 bp. Since both metazoan and phytoplankton amplicons span exactly the same nuclear 18SrDNA region, all sequences were combined into one file. Likewise, reads generated from the three tree species were saved to one file for each marker. After trimming, the resulting file for each marker and sample type was dereplicated and clustered into zero radius OTUs (hereafter OTUs) using the USEARCH pipeline. OTU tables were built for each sample type and marker, also using USEARCH. Taxonomy was annotated using blast2taxonomy script v1.4.2 <sup>48</sup> after BLAST searching <sup>49</sup> against the entire NCBI Genbank database for 18SrDNA and COI with a maximum number of 10 target sequences. The Silva database <sup>50</sup> was used for annotating 16SrDNA sequences, and the UNITE database <sup>51</sup> for fungal ITS1. The FungalTraits database <sup>52</sup> was used for the functional annotation of fungi. Taxonomic assignments based on BLAST hits with a base pair length of less than 80 % of the amplicon length and/or less than 85 % sequence identity were removed. We excluded all taxa except Bacteria, Algae/Protozoa, Metazoa or Fungi from the respective datasets.
+
+Only OTUs with a minimum of 3 reads per sample replicate were retained. The OTU tables were checked for contamination using negative controls. PCR replicates were merged and all datasets were checked for sufficient sequencing depth and sampling size (Extended Data Figure 1; for resulting sampling and OTU count as well as number of phyla and orders see Supplementary Table 2 “cleaned dataset”).
+
+## Statistical model and analyses of community diversity
+
+For each type of sample, we only (i) selected sites that were sampled for at least 5 years, (ii) only kept sampling years represented by at least 50 % of the sites and (iii) removed years that were isolated from the others (>2 years). We also removed samples with low read coverage (less than 50 % of the median number of reads). Finally, because OTU read abundances from metabarcoding datasets are subject to many biases <sup>53</sup>, we converted OTU abundances into binary presence/absence data and only analyzed trends in terms of OTU occurrence. To limit cross-contamination, we considered an OTU as present in a sample if it represents at least 0.01 % of the total reads (for resulting sampling and OTU count as well as number of phyla and orders see Supplementary Table 2 “filtered dataset (model)”; Extended Data Table 1).
+
+We measured community diversity trends in three different ways. First, in each community, in each year, we computed the ɑ-diversity using the OTU richness as a measure of local diversity at a given time. Second, we computed 𝛽-diversities between pairs of communities (1) sampled at the same site in different years or (2) sampled at different sites in the same year. (1) gives an idea of temporal turnover in community composition (temporal 𝛽-diversity), whereas (2) indicates the spatial turnover of the communities (spatial 𝛽-diversity). We measured 𝛽-diversities using the turnover component of the Jaccard distances (R-package betapart <sup>54</sup>).
+
+To identify temporal trends in these diversity indices, temporal models were fitted using mixed linear models accounting for the temporal autocorrelation between sampling years and the effect of the different sampling sites. We used the lme function (R-package nlme <sup>55</sup>) with the corAR1 temporal correlation and the different sites as random effects. We fitted these temporal models with either the ɑ-diversities or the temporal and spatial 𝛽-diversities as response variables.
+
+The significance of the observed trends was evaluated by comparing them to null expectations of community changes. Changes in community composition occur as a result of immigration and extinction events, which are influenced by neutral and/or niche factors. This generates a dynamic equilibrium with ever-changing communities, even in the absence of any kind of disturbance. Following Dornelas et al. (2014), we built upon the equilibrium theory of island biogeography (ETIB) <sup>19</sup> to set up a dynamic model for community assembly that generates null expectations of diversity trends in the absence of disturbance (Extended Data Figure 3A). The ETIB is a lineage-based model of species colonization of a local community (the island) from a metacommunity (the continent). In its simplest form <sup>19</sup>, at each time step, it assumes that each OTU has a probability to migrate from the metacommunity to the local community, and once settled in the community, each OTU has a probability to go extinct. The number of <em>new</em> immigration events (i.e. of species not already in the community) per time step is given by where is the total number of OTUs in the metacommunity and is the number of OTUs already present in community; it declines as the number of OTUs in the community increases. The number of extinction events per time step, given by , increases with the number of OTUs in the community. An equilibrium is reached when the number of immigration events per time step equals the number of extinction events, i.e. . The equilibrium number of OTUs in the community is given by . This simple form of ETIB implies a linear decrease or increase of the number of new immigration events (resp. extinction events) per unit of time with the number of settled species in the local community. It assumes that all species have the same probabilities to migrate or go extinct (neutrality), and that these probabilities do not depend on the number of species in the community, implying that there is a negligible influence of interspecific competition on immigration and extinction. It thus applies best to communities that are far from carrying capacity. This model has the advantage of being very straightforward to simulate using a simple discrete-time Markov chain.
+
+Given the incomplete sampling of communities typically achieved with metabarcoding techniques, we assumed that each OTU present in community is observed at each time step with a fixed probability . This extra parameter can be handled using hidden Markov models. We assumed that the rates , , and vary from one community to the other due to various extrinsic and intrinsic factors (e.g. distance to the metacommunity, community size, and environmental factors).
+
+Assuming neutrality, i.e. that all OTUs have the same immigration and extinction rates, is a strong assumption often proven to be unrealistic <sup>56</sup>. We therefore relaxed the assumption of neutrality. In our non-neutral model, we assumed that immigration (resp. extinction) rates for each OTU are sampled from a beta distribution with parameters and (resp. ). The rates are therefore sampled around (resp. ) with a variance that is inversely proportional to : a large corresponds to scenarios of neutrality whereas closer to 0 indicates that immigration and extinction rates are very different across OTUs. While each OTU has specific immigration and extinction rates in each community, we assumed that the ranks of the OTUs in terms of immigration and extinction rates are conserved across communities (i.e. an OTU with a low extinction rate in community compared with other OTUs will also have a low extinction rate in community ). We thus obtained a non-neutral model derived from the equilibrium theory of island biogeography, which assumes that the presence of an OTU in a community results from the balance between immigration and extinction, and that each OTU is characterized by specific rates of immigration and extinction centered around the average rates. At equilibrium, some OTUs are more likely to be present in the community (e.g. OTUs with higher immigration and lower extinction rates).
+
+Instead of testing different parameters values chosen <em>a priori</em> <sup>5</sup>, we implemented a new inference strategy to adjust the model parameters to the empirical data using a sequential technique. First, in each site, the sampling fraction was inferred using ACE (R-package vegan <sup>57</sup>). Second, we estimated the average rates and by fitting the neutral model to each community using a hidden Markov model (R-package seqHMM <sup>58</sup>). We used these estimates as community-specific estimates of the average rates of immigration and extinction of the non-neutral model. Third, given , , and , we used simulation-based inferences using artificial neural networks to estimate the parameters and . We generated a large number of simulated datasets by sampling and from uniform prior distributions and simulating the corresponding non-neutral model of community assembly, and for each of these simulations, we recorded ɑ-, spatial and temporal 𝛽-, and γ-diversities through time across all the sampled sites. We specifically incorporated subsampling into the simulations (with probability ), such that not all OTUs present in the local communities are observed, mimicking the detection bias of metabarcoding: simulated ɑ-, spatial and temporal 𝛽-, and γ-diversities are therefore directly comparable to empirical diversities. For , we used a uniform prior distribution between the number of observed OTUs and three times the ACE estimate of γ-diversity; for , we used a uniform prior between 1 and 5. We started the simulations at year 1500 with a random community composition at each site (each OTU has a probability to be initially present in the community, where is the theoretical number of species at equilibrium in the neutral model, given , and : ). Next, we simulated community composition over time in each site until 2023, sampled the communities according to , and recorded community composition through time for the years of sampling. We then computed for each simulation the ɑ- and 𝛽-diversities across all the sampled sites. We used the same methods and sampling scheme as for empirical data (the number of sampling sites varied through time) to obtain measures of ɑ-, spatial and temporal 𝛽-diversities. We trained an artificial neural network to estimate and from time series of ɑ-, 𝛽-, and γ-diversities using the Python library Keras <sup>59</sup>, with 100,000 simulations per dataset until reaching a sufficient predictive power. Once trained, the artificial neural network takes as input ɑ- and 𝛽-diversities through time and outputs estimates of and . We used a neural network with 3 intermediate layers, containing 132, 64, and 32 neurons respectively, and with ELU activation functions. We prevented overfitting by using a dropout of 0.5 at each intermediate layer. Input and output data were scaled between 0 and 1 before fitting and the simulations were split between the training set (90 %) and the test set (10 %). Once validated (Extended Data Figure 7), we finally applied the trained neural network to our empirical data, and obtained corresponding and values.
+
+Given the estimated and values, we simulated our non-neutral model 1000 times with these parameters to generate time series of ɑ- and 𝛽-diversities under an equilibrium model of community assembly. We then compared empirical and simulated temporal trends and considered an empirical trend to be significant if it was above or below 95 % of the simulated trends. We interpreted significant deviations from the simulated equilibrium model of community assembly as indicative of out-of-equilibrium dynamics in the empirical data, potentially driven by anthropogenic disturbances.
+
+We then investigated the temporal trends of OTU occurrences, as some OTUs may be recent invaders and others may have gone extinct at the regional scale. In each ecosystem, we only looked at OTUs present in at least 10 % of the samples and across at least 2 sites. For each OTU, we first tested whether its occurrence in the communities tended to vary through time. To do so, we fitted a generalized linear mixed model with a binomial response (presence/absence of the OTU in a community) and considered the sampling site as random effects.
+
+## References Methods
+
+31. Tarricone, K., Klein, R., Paulus, M. & Teubner, D. Guideline for Sampling and Sample Processing Red beech (<em>Fagus sylvatica</em>). Umweltbundesamt (2018b). https://www.umweltprobenbank.de/upb_static/fck/download/SOP_ESB_Red_Beech_V2.0.3_2018_en.pdf.
+
+32. Klein, R., Tarricone, K., Teubner, D. & Paulus, M. Guideline for Sampling and Sample Processing Norway Spruce (<em>Picea abies</em>)/ Scots Pine (<em>Pinus sylvestris</em>). Umweltbundesamt (2018). https://www.umweltprobenbank.de/upb_static/fck/download/SOP_ESB_Spruce_Pine_V2.0.2_2018_en.pdf.
+
+33. Tarricone, K., Klein, R., Paulus, M. & Teubner, D. Guideline for Sampling and Sample Processing Lombardy Poplar (<em>Populus nigra</em> ‘Italica’). Umweltbundesamt (2018a). https://www.umweltprobenbank.de/upb_static/fck/download/SOP_ESB_Lombardy_Poplar_V2.0.3_2018_en.pdf.
+
+34. Quack, M. et al. Guideline for Sampling and Sample Processing Bladderwrack (<em>Fucus vesiculosus</em>). Umweltbundesamt (2010). https://www.umweltprobenbank.de/upb_static/fck/download/SOP_Blasentang.pdf.
+
+35. Paulus, M., Klein., R. & Teubner, D. Guideline for Sampling and Sample Processing Blue Mussel (<em>Mytilus edulis</em> <em>complex</em>). Umweltbundesamt (2018). https://www.umweltprobenbank.de/upb_static/fck/download/SOP_ESB_Blue_Mussel_V2.1.0_2018_en.pdf.
+
+36. Teubner, D., Klein, R., Tarricone, K. & Paulus, M. Guideline for Sampling and Sample Processing Zebra Mussel (<em>Dreissena polymorpha</em>). Umweltbundesamt (2018). https://www.umweltprobenbank.de/upb_static/fck/download/SOP_ESB_Zebra_Mussel_V2.1.1_2018_en.pdf.
+
+37. Rüdel, H., Uhlig, S. & Weingärtner, M. Zerkleinerung und Homogenisierung von Umweltproben durch Cryomahlung. Fraunhofer-Institut für Molekularbiologie und Angewandte Oekologie (2008). https://www.umweltprobenbank.de/upb_static/fck/download/IME_SOP_Probenvorbereitung_Dez2008_V200.pdf
+
+38. Krehenwinkel, H., Weber, S., Künzel, S. & Kennedy, S. R. The bug in a teacup-monitoring arthropod-plant associations with environmental DNA from dried plant material. <em>Biology Letters</em>, <strong>18</strong>, 20220091 (2022b). https://doi.org/10.1098/rsbl.2022.0091.
+
+39. Donia, M. S. Fricke, W. F., Partensky, F. & Schmidt, E. W. Complex microbiome underlying secondary and primary metabolism in the tunicate Prochloron symbiosis. <em>PNAS</em>, <strong>108</strong>, 1423-1432 (2011). https://doi.org/10.1073/pnas.1111712108.
+
+40. Bodenhausen, N., Horton, M. W. & Bergelson, J. Bacterial communities associated with the leaves and the roots of Arabidopsis thaliana. <em>PloS One</em>, <strong>8</strong>, e56329 (2013). https://doi.org/10.1371/journal.pone.0056329.
+
+41. Chelius, M. K. & Triplett, E. W. The Diversity of Archaea and Bacteria in Association with the Roots of Zea mays L. <em>Microbial Ecology</em>, <strong>41</strong>, 252-263 (2001). DOI: 10.1007/s002480000087.
+
+42. Gardes, M. & Bruns, T. D. ITS primers with enhanced specificity for basidiomycetes--application to the identification of mycorrhizae and rusts. <em>Molecular Ecology,</em> <strong>2</strong>, 113-118 (1993). DOI: 10.1111/j.1365-294x.1993.tb00005.x.
+
+43. White, T. J., Bruns, T., Lee, S.& Taylor, J. Amplification and direct sequencing of fungal ribosomal RNA genes for phylogenetics. In : T. J. White, coord.: PCR Protocols: Elsevier, 315-322 (1990). https://doi.org/10.1016/B978-0-12-372180-8.50042-1.
+
+44. Lange, V. et al. Data from: Cost-efficient high-throughput HLA typing by MiSeq amplicon sequencing. <em>BMC Genomics</em> <strong>15</strong>, 63 (2014). https://doi.org/10.1186/1471-2164-15-63.
+
+45. Zhang, J., Kobert, K., Flouri, T. & Stamatakis, A. PEAR: a fast and accurate Illumina Paired-End reAd mergeR. <em>Bioinformatics</em>, <strong>30</strong>, 614-620 (2014). https://doi.org/10.1093/bioinformatics/btt593.
+
+46. Edgar, R. C. & Flyvbjerg, H. Error filtering, pair assembly and error correction for next-generation sequencing reads. <em>Bioinformatics</em>, <strong>31</strong>, 3476-3482 (2015). https://doi.org/10.1093/bioinformatics/btv401.
+
+47. Edgar, R. Usearch. Lawrence Berkeley National Lab (LBNL), Berkeley, CA (United States) (2010).
+
+48. Schöneberg, Y. „yschoeneberg/blast2taxonomy“. (2023). Zenodo. doi: 10.5281/zenodo.10009721.
+
+49. Altschul S. F., Gish W., Miller W., Myers E. W. & Lipman D. J. Basic local alignment search tool. <em>Journal of Molecular Biology</em>, <strong>215</strong>, 403-410 (1990). https://doi.org/10.1016/S0022-2836(05)80360-2.
+
+50. Quast, C. et al. The SILVA ribosomal RNA gene database project: improved data processing and web-based tools. <em>Nucleic Acids Research</em>, <strong>41</strong>, 590-596 (2012). https://doi.org/10.1093/nar/gks1219.
+
+51. Abarenkov, K. et al. The UNITE database for molecular identification and taxonomic communication of fungi and other eukaryotes: sequences, taxa and classifications reconsidered. <em>Nucleic Acids Research</em>, <strong>52</strong>, 791-797 (2024). https://doi.org/10.1093/nar/gkad1039.
+
+52. Põlme, S. et al. FungalTraits: a user-friendly traits database of fungi and fungus-like stramenopiles. <em>Fungal Diversity</em> <strong>105</strong>, 1-16 (2020). https://doi.org/10.1007/s13225-020-00466-2.
+
+53. Piñol J., Mir G., Gomez-Polo P. & Agustí N. Universal and blocking primer mismatches limit the use of high-throughput DNA sequencing for the quantitative metabarcoding of arthropods. <em>Molecular Ecology Resources</em>, <strong>15</strong>, 819-830 (2015). https://doi.org/10.1111/1755-0998.12355.
+
+54. Baselga, A. et al. Package ‘betapart’: Partitioning Beta Diversity into Turnover and Nestedness Components (2023). https://cran.r-project.org/web/packages/betapart/index.html.
+
+55. Pinheiro, J. et al. Package ‘nlme’: Linear and Nonlinear Mixed Effects Models (2024). https://cran.r-project.org/web/packages/nlme/index.html.
+
+56. Lomolino M. V. A call for a new paradigm of island biogeography. <em>Global Ecology and Biogeography</em>, <strong>9</strong>, 1-6 (2001). https://doi.org/10.1046/j.1365-2699.2000.00185.x.
+
+57. Oksanen, J. et al. Package ‘vegan’: Community Ecology Package (2024). https://cran.r-project.org/web/packages/vegan/vegan.pdf.
+
+58. Helske, S. & Helske, J. Mixture Hidden Markov Models for Sequence Data : The seqHMM Package in R. <em>Journal of Statistical Software</em>, <strong>88</strong>, Article 3 (2019). https://doi.org/10.18637/jss.v088.i03.
+
+59. Chollet, F. et al. Keras. (2015). https://github.com/fchollet/keras.
+
+# Supplementary Files
+
+- [TableS1S5.xlsx](https://assets-eu.researchsquare.com/files/rs-5139547/v1/27da36f4c65799b78a89133d.xlsx)  
+  Supplementary Table 1-5
+
+- [ExtendedDataFigures.docx](https://assets-eu.researchsquare.com/files/rs-5139547/v1/2544e2e5b0f84679d6c1359b.docx)

02e1743cee69bc1a7db10cf4c0bd148b57b0c52861ec56f8c2eba9309e3dd556/preprint/images_list.json

@@ -0,0 +1,42 @@
+[
+    {
+        "type": "image",
+        "img_path": "images/Figure_1.png",
+        "caption": "a) Reaction scheme of the in vitroreconstructed part of the nucleotide salvage pathway. The network consists of 6 enzymes and 15 substrates/products, resulting in a set of ODEs containing 42 kinetic parameters. b) Positive and negative correlations between the forward sensitivities between the kinetic parameters with respect to the measured output.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_2.png",
+        "caption": "a) Schematic of the experimental workflow. Enzymes are immobilized on gel beads and placed in CSTR with 6 inlets containing different substrates The output is measured offline on an Ion paired HPLC, 8 species can be observed over time indicated by the magnifying glass icon. b) Computational workflow to design an information dense dataset and train a kinetic model. In step one the OED algorithm evolves control inputs (i.e. inflow rates of the 6 inlets) to be maximally informative. In step two this data is added to a training dataset which is subsequently used to fit a model in step three. In step four we use the previous iteration of the model to predict the outcome of the latest experiment, utilizing this round as test data.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_3.png",
+        "caption": "a) Input flow rates for each of the 6 inputs substrates evolved by the OED algorithm. b) Data as measured on HPLC (black triangles) and the fit of the model to the data (solid lines).",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_4.png",
+        "caption": "a) Distribution of fits of the parameters including either only the first (number of datapoints N = 211) or the latest iteration (number of datapoints N = 166), the parameter set is included if the fit score deviates no less than 15% from the best fit. We note that after new rounds are added the distributions of the parameters decreases. b) Prediction of the last experiment (black triangles) in the dataset using the model trained on the dataset obtained after the first (shaded blue) and second (shaded orange) iteration of the cycle.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_5.png",
+        "caption": "a) shows the inverse design of different ratios, we opted to screen a large space of experimental and select 7 ratios to test along a range of summed substrate input concentrations (numbered spots). Each color is a simulated ratio, the standard deviation is the simulated deviation around the predicted mean (y-axis). The conversion efficiency is the predicted fraction of nucleobases that is converted to a triphosphate. To calculate the efficiency of the adenine conversion we first subtract the ATP concentration input from the measured ATP output. b) shows the experiments, labeled 1-7, with both the simulated concentrations including confidence interval (N = 20) and the HPLC measurement. The ratio between ATP (blue), UTP (grey) and GTP (green) are shown on top. c) Shows the prediction error defined as the percentage the simulated mean deviates from the HPLC data on the y-axis (averaging the error for the three triphosphates) and the total concentrations of the input substrates on the x-axis.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    }
+]

02e1743cee69bc1a7db10cf4c0bd148b57b0c52861ec56f8c2eba9309e3dd556/preprint/preprint.md

@@ -0,0 +1,143 @@
+# Abstract
+
+Kinetic modelling of *in vitro* constructed enzymatic reaction works is vital to understand and control the complex behaviours emerging from the abundant nonlinear interactions inside. However, modelling is severely hampered by the lack of training data. Here, we introduce a methodology that combines an active learning-like approach and flow chemistry to efficiently create optimized datasets for a highly interconnected enzymatic reactions network with multiple inputs and multiple outputs. The optimal experimental design (OED) algorithm designed a sequence of out-of-equilibrium perturbations to maximise the information about the reaction kinetics, yielding a descriptive model that allowed inverse design of the output of the network towards any cost function. We experimentally validated the model by forcing the network to produce different product ratios while maintaining a minimum level of overall conversion efficiency. Our workflow scales with the complexity of the system and enables the inverse design of previously unobtainable network outputs.
+
+[Biological sciences/Biochemistry/Biocatalysis](/browse?subjectArea=Biological%20sciences%2FBiochemistry%2FBiocatalysis)  
+[Physical sciences/Chemistry/Analytical chemistry/Microfluidics](/browse?subjectArea=Physical%20sciences%2FChemistry%2FAnalytical%20chemistry%2FMicrofluidics)  
+[Biological sciences/Systems biology/Biochemical networks](/browse?subjectArea=Biological%20sciences%2FSystems%20biology%2FBiochemical%20networks)  
+[Biological sciences/Systems biology/Control theory](/browse?subjectArea=Biological%20sciences%2FSystems%20biology%2FControl%20theory)
+
+# Introduction
+
+Living cells rely on enzymatic reaction networks (ERNs) to produce energy and building blocks to support cellular processes. Evolution has shaped these ERNs into interconnected sub-pathways to generate multiple outputs from multiple inputs, driving product formation across complex kinetic landscapes. Recently, significant progress has been made in reconstituting ERNs *in vitro* with the aim of building a cell from the bottom up<sup>1–4</sup>, or to produce value-added chemicals from sustainable substrates as an advanced biotechnology<sup>5–8</sup>. However, these networks typically do not capture one of the essential features of biological ERNs, which produce multiple outputs from multiple inputs. The design of such networks remains challenging due to the lack of sufficiently informative experimental datasets that can be utilized to train kinetic models which provide qualitative and quantitative insights into the dynamic properties of large ERNs<sup>9,10</sup>.
+
+Typically, the design of ERNs towards specific outcomes, like increasing the overall efficiency, is achievable by searching a large combinatorial space of inputs and measure the product formation of the ERN to generate a large experimental dataset. Experimentally, this is prohibitively time-, labour- and cost-intensive<sup>11</sup>. Recently Pandi et al. has shown such screening process could be significantly improved by an AI based active learning protocol<sup>12</sup>. Additionally, promising advances have been published recently, utilizing machine learning to derive qualitative understanding of reaction networks and individual reaction mechanisms<sup>9,10,13−15</sup>. Yet, these black box approaches are limited in their ability to provide both qualitative and quantitative insights into large ERNs. This gap stems from a lack of informative data, that usually consist of measured steady state outcomes which are kinetically non-informative and not sufficient to approximate the kinetic landscape of complex ERNs. To address this, time-course data tracking the responses of ERNs to controlled perturbations are needed. This is demonstrated by both Shen et al. and Hold et al. in both batch and flow respectively, who characterized networks by adding the enzymes sequentially and measuring the change in product formation<sup>16–18</sup>. However, as the complexity and scale of an ERN increases (substrate competition, allosteric interactions, feedback loops, futile cycles, etc.) choosing a set of perturbations ‘intuitively’ such that we obtain relevant information about the kinetic landscape becomes increasingly difficult.
+
+Here, we present a scalable method that trains a kinetic model iteratively, by adding new and more informative experiments to a training dataset in each optimization cycle (akin to active learning). It incorporates an optimal experimental design (OED) algorithm that evolves a sequence of out-of-equilibrium perturbations to be maximally informative. We subsequently test the utility of the model by using the experimental outcomes of these perturbation experiments as test data for the previous iteration of the model. Using this approach, we demonstrate that a limited number of design iterations is sufficient to obtain data of sufficient quality to map the dynamic landscape of the ERN and obtain a measure of control over it as a multi-input multi-output (MIMO) system *in vitro*.
+
+# Results And Discussion
+
+The *in vitro* constructed ERN in this work derives from the nucleotide salvage pathway (Fig. 1a), which regenerates nucleotides for cellular processes by recovering bases and nucleosides from the degradation of RNA and DNA. This pathway has been reconstructed *in vitro* to synthesize labelled nucleotides. <sup>19–21</sup> The network consists of six enzymes and starts with phosphoribosyl pyrophosphate (PRPP), which can be converted from glucose via the pentose phosphate pathway and is coupled by the enzyme APRT and UPRT to nucleobases adenine and uracil, respectively, to form the monophosphate nucleotides AMP and UMP. For solubility reasons we did not include guanine as a nucleobase and started from GMP. AMP, UMP and GMP are subsequently converted to their corresponding diphosphate nucleotides (NDPs) by enzymes AK, GMPK and UMPK, respectively, using ATP as cofactor. Finally, NDPs are converted to NTPs by a single enzyme, PK. In total, this system consists of six enzymes catalysing eight reversible reactions, where PK is shared between three substrates, and resource competition for ATP, PEP and PRPP throughout the network.
+
+Translating the reactions of the ERN into a fully generalized ODE (ordinary differential equations) model results in a system of 15 equations with 42 kinetic rates (assuming reversible bi-substrate reaction kinetics, see Supplementary information 1.2 Eq. 2). Large model descriptions of non-linear ERNs are needed to quantitatively predict their behaviour, but fitting these models such that they maintain their predictive power is challenging. This ‘parameter problem’ is present in all models but in ODEs it can be viewed from the perspective of a parameter’s forwards sensitivity to the observed species (Fig. 1b). <sup>22</sup> These sensitivities map onto the contribution a parameter has to the observed rates of change over time (Supplementary information 1.2 Eq. 3). When these sensitivities correlate with one another, the observations can be approximated by the model by modifying both rates simultaneously. A positive correlation between the forward sensitivities of kinetic rates implies a similar effect on the rate of change of the observed species, thus the model can fit the data by increasing the value of one rate whilst decreasing the value of its partner, a negative correlation implies an opposing effect on the rate of change, thus, to fit the data the kinetic rates need both to either increase or decrease.
+
+This unidentifiability means many combinations of kinetic rates can approximate the data (not just the ‘true’ rates), which in turn leads to prediction errors as the experimental conditions change from those used to generate the initial training data. <sup>22–24</sup> Thus, experimental data can be deemed uninformative if the inability to discern which reactions contribute most to the flux of a species at a specific time results in prediction errors as conditions change. Generally, it is easier to completely identify rates in simplified models, but their quantitative predictive power will be limited as mechanistic assumptions are readily broken (Supplementary information 1.6.2, Fig. S6). Conversely, detailed mechanistic models are more descriptive but it is harder to identify kinetic rates.
+
+However, from a broadly practical perspective, precisely identifying individual rates is not needed to control the behaviour of an ERN, a model just needs to approximate the kinetic landscape adequately and the remaining uncertainty needs to be manageable. To address this efficiently, we adapted an active learning approach commonly applied in machine learning with the singular goal of controlling ERNs. We utilized optimal experimental design (OED) to design experiments that maximize information about the ERN in the data, and subsequently fitted a kinetic model and tested its predictive power. This cycle was repeated until the uncertainty around the predictions was reduced and they matched the experimental outcome.
+
+Figure 2a shows the experimental workflow. First, all enzymes were individually immobilised on microfluidically produced hydrogel beads with a diameter of 50 µm. <sup>25</sup> The activity of each enzyme after immobilisation was measured separately. Next, enzyme-loaded beads were loaded into a microfluidic continuous stirred-tank reactor (CSTR), for experimental details see Method. The CSTR chamber itself has a volume of 100 µl and has six inlets for each of the input substrates uracil, GMP, adenine, ATP, PEP and PRPP and a single outlet. Samples were collected from the output at different intervals depending on the total flow rates by a fraction collector and analysed offline by ion-pair HPLC. <sup>26</sup> The analysis of the chromatographic peaks provides a compositional pattern of eight input substrate, intermediates, and product molecules (uracil, adenine, UMP, GMP, ADP, ATP, UTP and GTP), each changing at every input combination.
+
+Figure 2b shows the optimal experimental design workflow. In step one a hybrid swarm/evolutionary algorithm evolves an input flow profile for each of the six inputs at three different flowrates. <sup>27,28</sup> This algorithm scores input patterns by minimizing the D-Fisher information criterion (Supplementary information 1.2 Eq. 4). <sup>29</sup> This criterion is obtained by computing the determinant of the Fisher information matrix which is derived from the parameter sensitivities (Supplementary information 1.2 Eq. 3). This metric maps onto the volume of the parameter space where the ODE model can fit to the experimental data. <sup>29,30</sup> This means the algorithm is driven to find a combination of input sequences that breaks the correlation between parameter sensitivities (if only temporary). The transition between different total flow rates results in different output compositions and serves as another control parameter that increases the information content about substrate conversion fluxes in the data. At high flow rates input molecules and monophosphates are detected (as only a fraction of substrate has been converted); at low flow rates increased NTP formation is observed (Supplementary information 1.5 Fig. S4). In step two this data is added to a training dataset, the model is fit to this data in step three. In step four the predictive power of the model is assessed by using the previous iteration of the model to predict the current experiment (test data), if the predictive power is not sufficient or no longer improves, the cycle is terminated; if not, the cycle continues, and the latest iteration of the model and database is used to design a new experiment in step one. <sup>31</sup>
+
+A total of three iterations of the optimization cycle were performed (excluding a calibration), each time exchanging the microfluidic chip, altering the enzyme concentrations (Supplementary information 2.2 Fig. S9-S12). The lower and upper boundary of the concentration ranges for the substrates were based on the enzyme activity assays and substrate solubility (Supplementary information 3.4 Fig. S14-S25). The initial experiment (not part of the cycle) is manually designed and ‘calibrates’ the model (Supplementary information 2.2 Fig. S9). This allows for the subsequent OED of an informative input sequence since more knowledge about the system equates to better OED outcomes. <sup>25</sup> Fig. 3a shows the substrate inputs for final experiment of the optimization cycle, illustrating the non-intuitive character of the evolved input sequence. Figure 3b shows the complexity of the time-course data and the convergence of the model fit to the data.
+
+Figure 4 places these data in the context of the optimization cycle. Figure 4a shows parameter distributions of the model trained in the first iteration (left) and the parameter distribution of the model trained in the second iteration (right). We note a significant decline in the distribution width of most kinetic parameters. Figure 4b shows the prediction (shaded area) made by the model from iteration one and the model from iteration two for the experiment shown in Fig. 3. The latter shows a drastic reduction in the variance around the prediction and highlights that the model can approximate the behaviour of the ERN quantitatively. This presents us with new opportunities, beyond traditional optimization schemes that focus on maximizing the yield of a single product. Here, we demonstrate how we can use the final iteration of the model for inverse design to control MIMO system towards any cost function. <sup>28,32</sup> We opted to tune the ATP/GTP/UTP output ratios whilst maintaining a minimal conversion efficiency - defined as the percentage of nucleobases converted to triphosphates - of 60%. Figure 5a shows the outcome of a sampling process, we randomly generated 10<sup>6</sup> substrate input combinations, each simulated twenty times using different combinations of fitted kinetic rates (keeping ratios that could be approximated by integers). Each dot represents a different condition, the color indicates the ratio between ATP/UTP/GTP. The y-axis shows the average standard deviation of these simulations to the simulated mean of ATP, UTP and GTP (the predicted uncertainty and thus possible prediction error due to poorly resolved kinetics). The x-axis shows the conversion efficiency. We semi-randomly selected seven experimental conditions manually representing seven ATP/UTP/GTP ratios in Fig. 5a, including one repeated ratio (experiment 1 & 7) and one experiment with a lower conversion efficiency (experiment 3). To ensure we selected dissimilar experiments we aimed for a range of summed substrate concentrations.
+
+Figure 5b shows the simulated confidence interval of the predicted final yield and the yield as measured on the HPLC. For experiments 1–5 uncertainties and total output concentrations vary but predictions still match. For very low input concentrations of UMP, guanine, adenine, and ATP in experiments 6 and 7, the predictions error increases drastically even though the simulated standard deviation is low. This relation between the prediction error, quantified as the percentage the simulated mean deviates from the HPLC measurement and the summed input concentration of the nucleobases is shown in Fig. 5c. It highlights that the model is quantitatively accurate from a total substrate input concentration of 0.3 mM. The cause can be two-fold: most likely, the signal to noise ratio for the HPLC measurement decreases, leading to larger variations in the experimental data, or, low substrate concentrations might require different model assumptions. Either way, this observation marks the practical boundary of the current model, allowing us to efficiently probe conditions in the identified operable space. In Supplementary information 1.6.1 Fig. S5-S7 we investigated the relationship between the size of the training dataset and the prediction error further. Additionally, we explicitly tested model assumptions and found that reversibility is needed to obtain quantitatively accurate predictions even though these models can fit the data.
+
+# Conclusion
+
+We have presented a methodology to design informative training data and map the kinetic landscape of an ERN as efficiently as possible. By designing sufficiently complex experiments we were able to restrict the combinations of potential kinetic rates such that they map onto real product formation fluxes sufficiently well to obtain quantitative control, balancing identifiability, mechanistic detail, and predictive power. The number of optimization iterations required depends on a combination of the complexity of the network and the quality and resolution of experimental data. If the system is highly non-linear, more certainty about the rates will be needed as smaller deviations from the true value are punished more severely. In contrast, very linear and orthogonal networks will likely require significantly fewer optimization cycles to enable a form of inverse design. The full potential of this pipeline lies in the inverse design of network outputs for which there is no design ‘blueprint’. We have demonstrated this in our current work by forcing the network to produce a certain product ratio while maintaining a certain conversion efficiency. In the future, ever more complex cost functions can be defined, for which our pipeline will then find optimal solutions. This cost function could also explicitly include methods to investigate the validity of proposed reaction mechanisms by explicitly breaking mechanistic assumptions of different models during the experimental design/training data generation process. Overall, our pipeline will benefit optimization of enzymatic pathways or integration of multiple modules into complex networks used in (cell-free) synthetic biology.
+
+# Methods
+
+## Materials
+
+Enzymes adenylate kinase (AK) and pyruvate kinase (PK) and all chemicals were purchased from Sigma and directly used without further processes. Enzymes adenine phosphoribosyl transferase (APRT) and, uracil phosphoribosyl transferase (UPRT) were expressed and purified as described by Arthur et al. <sup>33</sup>, Genes for guanosine monophosphate kinase (GMPK) and, uridine monophosphate kinase (UMPK) were PCR amplified from E. coli K12 using gene specific primers, cloned into pET15b, expressed overnight at 30°C (GMPK) and 18°C (UMPK) in E. coli BL21(DE3) and purified according to protocols modified from Oeschger et al <sup>34</sup> (GMPK) and Serina et al <sup>35</sup> (UMPK) to accommodate Ni<sup>2+-</sup> sepharose purification. Purified enzymes were dialysed against 20 mM potassium phosphate buffer (pH 7.2) prior to immobilization. All the enzymes were immobilised on microfluidic produced hydrogel beads, as reported. <sup>25</sup> After immobilisation, all the enzyme-beads were freeze dried and stored in -20<sup>o</sup> C. While using, 1 mg of each enzyme beads was dissolved in 31 ul IVTT buffer (pH 7.3, 9 mM magnesium acetate, 5 mM potassium phosphate, 95 mM potassium glutamate, 5 mM ammonium chloride, 0.5 mM calcium chloride, 1 mM spermidine, 8 mM putrescine, 1 mM dithiothreitol, 10 mM creatine phosphate).
+
+## Flow experiments setup
+
+Cetoni Nemesys syringe pumps with Hamilton syringes were used to control input and the flow profile was programmed using the Cetoni neMESYS software. <sup>25,36</sup> Before performing the designed flow profile, the whole system was equilibrium with buffer for two hours. The outflow of the CSTR was collected using a fraction collector, collecting for either 30 or 15 minutes or three droplets per fraction. The ion-pair HPLC analysis was adapted from ref 26 and performed on Shimadzu Nexera X3 HPLC system with an Inertsil ODS-4 column (3 μm, 150×4.6 mm; GL Science) and a guard column (3 μm; 10 × 4.6 mm) at 40 °C. The elution gradient was as follows: 100% buffer A (100 mM potassium phosphate buffer (pH 6.4) with 8 mM ion-pair reagent tetrabutylammonium bisulfate, filtered before use) for 13 min; 0-77% linear gradient of buffer B for 22 min; 77-100% buffer B (70% buffer A with 30% acetonitrile) for 1 min; and 100% buffer B for 14 min. The flow rate was maintained at 1 ml/min. Peaks were identified by comparison with standard samples. The concentration was obtained from the integrated peak areas with the calibration curve of each standard.
+
+## Modelling, fitting, data processing, parameterisation
+
+An overview for the software that performs the optimizations can be found in Supplementary Information 1. The package is written in Python 3.8 (python software foundation, Delaware US). A generated text-based model object <sup>25</sup> is translated to an SBML and AMICI object modified from ref 27 and ref 37. AMICI is an ODE compilation package to C++ which is continuously updated. <sup>38-41</sup> Code can be found at huckgroup github at http://github.com/huckgroup/OED. For more information on OED and parameter estimation see Supplementary Information.
+
+# References
+
+1. Berhanu, S., Ueda, T. & Kuruma, Y. Artificial photosynthetic cell producing energy for protein synthesis. *Nat Commun* **10**, 1325 (2019).
+
+2. Bhattacharya, A., Brea, R. J., Niederholtmeyer, H. & Devaraj, N. K. A minimal biochemical route towards de novo formation of synthetic phospholipid membranes. *Nat Commun* **10**, 300 (2019).
+
+3. Lee, K. Y., Park, S.-J., Lee, K. A. *et al.* Photosynthetic artificial organelles sustain and control ATP-dependent reactions in a protocellular system. *Nat Biotechnol* **36**, 530–535 (2018).
+
+4. Pols, T., Sikkema, H. R., Gaastra, B. F. *et al.* A synthetic metabolic network for physicochemical homeostasis. *Nat Commun* **10**, 4239 (2019).
+
+5. Burgener, S., Luo, S., McLean, R. *et al.* A roadmap towards integrated catalytic systems of the future. *Nat Catal* **3**, 186–192 (2020).
+
+6. Valliere, M. A., Korman, T. P., Arbing, M. A., & Bowie, J. U., A bio-inspired cell free system for cannabinoid production from inexpensive inputs. *Nat. Chem. Biol.* **16**, 1427–1433 (2020).
+
+7. Rasor, B. J., Vögeli, B., Landwehr, G. M., Bogart, J. W. *et al.* Toward sustainable, cell-free biomanufacturing. *Curr Opin Biotechnol* **69**, 136–144 (2021).
+
+8. Miller, T. E., Beneyton, T., Schwander, T., Diehl, C *et al.* Light-powered CO2 fixation in a chloroplast mimic with natural and synthetic parts. *Science* **368**, 649–654 (2020).
+
+9. Yu, T., Boob, A. G., Volk, M. J., Liu, X., Cui, H., & Zhao, H., Machine learning-enabled retrobiosynthesis of molecules. *Nat. Catal.* **6**, 137–151 (2023).
+
+10. Margraf, J. T., Jung, H., Scheurer, C., & Reuter, K., Exploring catalytic reaction networks with machine learning. *Nat. Catal.* **6**, 112–121 (2023).
+
+11. Morgado, G., Gerngross, D., Roberts, T. M. & Panke, S., Synthetic biology for cell-free biosynthesis: fundamentals of designing novel in vitro multi-enzyme reaction networks. *Adv. Biochem. Eng. Biotechnol.* **162**, 117–146 (2018).
+
+12. Pandi, A., Diehl, C., Kharrazi, A. Y., Scholz, S. A. *et al.* A versatile active learning workflow for optimization of genetic and metabolic networks. *Nat Commun* **13**, 3876 (2022).
+
+13. Wen, M., Spotte-Smith, E. W. C., Blau, S. M., McDermott, M. J., Krishnapriyan, A. S., & Persson, K. A., Chemical reaction networks and opportunities for machine learning. *Nat Comput Sci* **3**, 12–24 (2023).
+
+14. Burés, J., & Larrosa, I., Organic reaction mechanism classification using machine learning. *Nature* **613**, 689–695 (2023).
+
+15. Faulon, J. L., & Faure, L., In silico, in vitro, and in vivo machine learning in synthetic biology and metabolic engineering. *Curr Opin Chem Biol* **65**, 85–92 (2021).
+
+16. Shen, L., Kohlhaas, M., Enoki, J., Meier, R., Schönenberger, B. *et al.* A combined experimental and modelling approach for the Weimberg pathway optimisation. *Nat Commun* **11**, 1098 (2020).
+
+17. Bujara, M., Schumperli, M., Pellaux, R., Heinemann, M. & Panke, S., Optimization of a blueprint for in vitro glycolysis by metabolic real-time analysis. *Nat. Chem. Biol.* **7**, 271–277 (2011).
+
+18. Hold, C., Billerbeck, S. & Panke, S., Forward design of a complex enzyme cascade reaction. *Nat Commun* **7**, 12971 (2016).
+
+19. Parkin, D. W., Leung, H. B., & Schramm, V. L., Synthesis of nucleotides with specific radiolabels in ribose. Primary $^{14}$C and secondary $^{3}$H kinetic isotope effects on acid-catalyzed glycosidic bond hydrolysis of AMP, dAMP, and inosine, *J Biol Chem.* **259**, 9411–9417 (1984).
+
+20. Tolbert, T. J., & Williamson, J. R., Preparation of specifically deuterated and $^{13}$C-labeled RNA for NMR studies using enzymatic synthesis, *J. Am. Chem. Soc.* **119**, 12100–12108 (1997).
+
+21. Nelissen, F. H. T., Girard, F. C., Tessari, M., Heus, H. A., & Wijmenga, S. S., Preparation of selective and segmentally labelled single-stranded DNA for NMR by self-primed PCR and asymmetrical endonuclease double digestion, *Nucleic Acids Res.* **37**, e114 (2009).
+
+22. Gábor, A., Villaverde., A. F., & Banga, J. R., Parameter identifiability analysis and visualization in large-scale kinetic models of biosystems, *BMC Syst. Biol.* **11**, 54 (2017).
+
+23. Kreutz, C., Raue, A., Kaschek, D., & Timmer, J., Profile likelihood in systems biology. *FEBS Journal* **280**, 2564–2571 (2013).
+
+24. Raue, A., Kreutz, C., Maiwald, T., Bachmann, J. *et al.* Structural and practical identifiability analysis of partially observed dynamical models by exploiting the profile likelihood. *Bioinformatics* **25**, 1923–1929 (2009).
+
+25. Baltussen, M., Wiel, J., Regueiro, C., Jakštaitė M., & Huck, W. T. S., A Bayesian approach to extracting kinetic information from artificial enzymatic networks, *Anal. Chem.* **94**, 7311 (2022).
+
+26. Nakajima, K., Kitazume, S., Angata, T., Fujinawa, R. *et al.* Simultaneous determination of nucleotide sugars with ion-pair reversed-phase HPLC, *Glycobiology* **20**, 865–871 (2010).
+
+27. Sluijs, B., Maas, R. J. M., Linden, A. J., Greef, T. F. A., & Huck, W. T. S., A microfluidic optimal experimental design platform for forward design of cell-free genetic networks, *Nat Commun* **13**, 3626, (2022).
+
+28. Smith, R. W., Sluijs, B., & Fleck, C., Designing synthetic networks in silico: a generalised evolutionary algorithm approach, *BMC Syst. Biol.* **11**, 118 (2017).
+
+29. Sinkoe, A., & Hahn, J., Optimal experimental design for parameter estimation of an IL-6 signaling model, *Processes* **5**, 49 (2017).
+
+30. Aguiar, P. F., Bourguignon, B., Khots, M. S., Massart, D. L., & Phan-Than-Luu, R., D-optimal designs. *Chemometr Intell Lab Syst* **30**, 199–210 (1995).
+
+31. Ruess, J., Parise, F., Milias-Argeitis, A., & Lygeros, J., Iterative experiment design guides the characterization of a light-inducible gene expression circuit, *Proc. Natl. Acad. Sci. U.S.A.* **112**, 8148–8153 (2015).
+
+32. Otero-Muras, I., & Carbonell, P., Automated engineering of synthetic metabolic pathways for efficient biomanufacturing. *Metab. Eng.* **63**, 61–80 (2021).
+
+33. Arthur, P. K., Alvarado, L. J., & Dayie, T. K., Expression, purification and analysis of the activity of enzymes from the pentose phosphate pathway, *Protein Expression and Purification* **76**, 229–237 (2011).
+
+34. Oeschger, M. P., & Bessman, M. J., purification and properties of guanylate kinase from escherichia coli, *J. Biol. Chem* **241**, 5452–5460 (1966).
+
+35. Serina, L., Blondin, C., Krin, E., Sismeiro, O. *et al.* Escherichia coli UMP-kinase, a member of the aspartokinase family, is a hexamer regulated by guanine nucleotides and UTP. *Biochemistry* **34**, 5066–5074 (1995).
+
+36. Helwig, B., Sluijs, B., Pogodaev, A. A., Postma, S. G. J., & Huck, W. T. S., Bottom-Up Construction of an Adaptive Enzymatic Reaction Network. *Angew. Chem.* **130**, 14261–14265 (2018).
+
+37. Choi, K., Medley, J. K., König, M., Stocking, K. *et al.* Tellurium: An extensible python-based modeling environment for systems and synthetic biology. *Biosystems* **171**, 74–79 (2018).
+
+38. Fröhlich, F., Weindl, D., Schalte, Y., Pathirana, D. *et al.* AMICI: high-performance sensitivity analysis for large ordinary differential equation models. *Bioinformatics* **37**, 3676–3677 (2021).
+
+39. Lakrisenko, P., Stapor, P., Grein, S., Paszkowski, Ł. *et al.* Efficient computation of adjoint sensitivities at steady-state in ODE models of biochemical reaction networks. *PLoS Comput Biol* **19**, e1010783 (2023).
+
+40. Schälte, Y., Fröhlich, F., Stapor, P., Vanhoefer, J. *et al.* pyPESTO - parameter estimation toolbox for python (v0.2.16). Zenodo (2023).
+
+41. Schmiester, L., Weindl, D., & Hasenauer, J., Efficient gradient-based parameter estimation for dynamic models using qualitative data. *Bioinformatics* **37**, 4493–4500 (2021).
+
+# Supplementary Files
+
+- [SupplementaryInformationInverseDesignofEnzymaticReactionStates.docx](https://assets-eu.researchsquare.com/files/rs-2646906/v1/ed65ebc516183f50bf95360f.docx)  
+  Supplementary information is available online.

03958f873bda5dcf088e0d766c000075ed7adb61f8e7f76a9a5c0ea0d50554df/preprint/images_list.json

@@ -0,0 +1,58 @@
+[
+    {
+        "type": "image",
+        "img_path": "images/Figure_1.png",
+        "caption": "RPCs identification and their developmental trajectories during human retinal development. A) UMAP plot of integrated scRNA-Seq fetal retina cells. Each cluster was identified based on expression of retinal specific cell markers. Highly expressed markers for each clusters are shown in Table S2. B) RPCs and transient neurogenic progenitors named T1, T2 and T3 were identified. Highly expressed markers for each cluster along the pseudotime trajectory are shown in Table S2. C, D) Pseudotime analysis demonstrating transition from early to late RPCs, and to T1 progenitors, which further commit to either T2 or T3 transient neurogenic progenitors. E) Gene expression heatmap showing similarities in gene expression patterns between early and late RPCs, but distinct gene expression signatures in T1, T2 and T3 neurogenic progenitors.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_2.png",
+        "caption": "ST analysis of 8 PCW human eye sections reveals the location of early RPCs in the CMZ. A) Representative histological staining of the 8 PCW fresh frozen human eye section. B, C) Spatial localisation of the 12 clusters identified from the ST analysis. Highly expressed markers for each cluster are shown in Table S3. D) UMAP of spatial transcriptomics scRNA-Seq data. E) Subclustering of ciliary margin zone (cluster 4) reveals the presence of two subclusters namely RPCs and ciliary body, and iris pigmented epithelial cells. G and I) Expression violin plots showing the expression of early RPCs in the peripheral retina (CMZ) and late RPCs in the central retina, respectively. H and J) Spatial localisation of early and late RPCs in the CMZ and central retina, respectively.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_3.png",
+        "caption": "ST analyses demonstrate decreased presence of early RPCs in the ciliary margin zone as development proceeds from 10 to 13 PCW. A-C) Spatial localisation of all cell clusters identified from the ST analyses (left hand side). The CMZ, early and late RPCs cluster localisations are shown on the right-hand side. Highly expressed markers for each cluster are shown in Table S3. Note the scarce presence of early RPCs in the CMZ of 13 PCW human eye. D) Early RPCs reach a peak at 7.5-8 PCW and then decline from 10 PCW onwards in the human developing retina. The ratio of early to late RPCs is inferred from the scRNA-Seq data.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_4.png",
+        "caption": "Single cell ATAC-Seq analysis of developing retina samples reveals cell type specific chromatin accessibility profiles. A) The number and type of chromatin accessibility profiles for each cell type. B) Heatmap showing differentially accessible of chromatin accessibility peaks (columns) for each cell type (rows). C) Representative examples of chromatin accessibility peaks for retinal cell specific marker genes. Each track represents the aggregate scATAC signal of all cells from the given cell type normalized by the total number of reads in TSS regions.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_5.png",
+        "caption": "Motif analysis of accessible DNA peaks predicts cell type specific TFs in the developing human retina. A) Heatmap of transcription factor binding motifs enriched in each cell type. More significant enrichment is indicated by the darker colours. B) Footprinting analysis of selected TFs predicted to show a significant enrichment in RPCs. C) Footprinting analysis of selected TFs predicted to show a significant enrichment in transient neurogenic progenitors and retinal neurons. Additional abbreviations to those mentioned in the main text: Gly ACs- glycinergic amacrine cells, GABA ACs \u2013 gabaergic amacrine cells, ST ACs \u2013 starburst amacrine cells, HCs \u2013 horizontal cells, MG- Muller glia cells.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_6.png",
+        "caption": "Representative gene regulatory networks in RPCs depicting activated upstream regulators (A, B) and inhibited upstream regulators (C, D) and their target genes. Upstream regulatory networks were generated with IPA using differentially expressed genes from the scRNA-Seq data and differential accessibility analysis in the scATAC-Seq data. The networks show predictions of upstream regulators which might be activated or inhibited to explain observed upregulation/downregulations in the data. The barplots next to each molecule represent the relative expression in the sRNA-Seq (column 1) and scATAC-Seq datasets (column 2). The colours for the network nodes/barplots indicate observed upregulation/ increased chromatin accessibility (red), predicted upregulation/increased chromatin accessibility (orange), observed downregulation (green) and predicted downregulation/ decreased chromatin accessibility (blue). The colour of the edges represents the relationships between the molecules; orange = prediction and observation are consistent with activation; blue = prediction and observation are consistent with downregulation; yellow = prediction and observation are inconsistent; and grey relationship between the molecules is available in the IPA knowledge database. *- indicates duplicates in scATAC-Seq dataset.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_7.png",
+        "caption": "TEAD binding plays a significant role in retinal lamination and RPCs, RGCs and photoreceptor specification. A) Footprinting analysis of TEAD2 and TEAD4\u00a0 showing a significant enrichment in RPCs. Additional abbreviations to those mentioned in the main text: Gly ACs- glycinergic amacrine cells, GABA ACs \u2013 gabaergic amacrine cells, ST ACs \u2013 starburst amacrine cells, HCs \u2013 horizontal cells, MG- Muller glia cells. B-J) Quantitative immunofluorescence analyses for the presence of VSX2+ RPCs, Ki67+ proliferating cells, SCNG+ RGCs and Recoverin+ and CRX+ photoreceptor precursors reveal loss of RPCs, disturbed retinal lamination, and attenuation of photoreceptor and RGCs specification. White arrowheads show the presence of rosettes comprised of RPCs or Ki67+ proliferating cells. Scale bars 100 \u00b5M for B-E and 50 \u00b5M for the bottom panel insets.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    }
+]

03958f873bda5dcf088e0d766c000075ed7adb61f8e7f76a9a5c0ea0d50554df/preprint/preprint.md

@@ -0,0 +1,226 @@
+# Abstract
+
+The emergence of retinal progenitor cells (RPCs) and differentiation to various retinal cell types represent fundamental processes during retinal development. Herein, we provide a comprehensive single cell characterisation of transcriptional and chromatin accessibility changes that underline RPC specification and differentiation over the course of human retinal development up to midgestation. Our lineage trajectory data demonstrate the presence of early RPCs, which transit to late RPCs, and further to transient neurogenic progenitors, that give rise to all the retinal neurons. Combining single cell RNA-Seq with spatial transcriptomics of early eye samples, we demonstrate for the first time the transient presence of early RPCs in the ciliary margin zone with decreasing occurrence from 8 PCW of human development. In RPCs, we identified a significant enrichment for TEAD transcription factor binding motifs, which when inhibited led to loss of RPCs and retinal lamination, and inhibition of photoreceptor and retinal ganglion cell differentiation.
+
+**Biological sciences/Developmental biology**  
+**Health sciences/Molecular medicine**  
+**developing human retina**  
+**single cell**  
+**RNA-Seq**  
+**ATAC-Seq**  
+**spatial transcriptomics**  
+**ciliary margin**  
+**GRNs**  
+**TEADs**
+
+# Introduction
+
+Retina is the innermost, light-sensitive tissue that lines the back of the eye and is vital for light sensing and image processing. The retina is derived from a germinal zone in the optic vesicle in which neuroepithelial cells proliferate to give rise to the six types of retinal neuronal and one glial cell type, organized within three different layers. All these cell types derive from retinal progenitor cells (RPCs) in an orderly spatio-temporal manner that has been well studied in vertebrates<sup>1</sup>. Great progress has been achieved in the last five years providing both gene expression and epigenetic profiles of the developing human retina<sup>2, 3, 4</sup>. Single cell (sc) RNA-Seq studies have enabled molecular characterisation of all the retinal cell types and have provided the developmental trajectories that lead to retinal cell specification from RPCs during retinal development<sup>5, 6, 7, 8, 9</sup>. The transcriptomic studies have been complemented with sc assay for transposase-accessible chromatin sequencing (scATAC-seq), enabling identification of key transcription factors relevant to specific retinal cell fates and the gene regulatory networks (GRNs) that underline the cell-state changes<sup>7, 10, 11, 12</sup>. Notwithstanding, these studies have not provided as yet any information on the spatial resolution of RPCs or retinal neurons.
+
+The retina of fish and amphibians continuously integrates new neurons generated from RPCs residing in the distal tip of the retina known as the ciliary margin zone (CMZ). In chicks the CMZ also contributes to retinal neurogenesis during development, but the adult chick CMZ cells have a more restricted potential as they only contribute to a small fraction of the retina and moreover do not participate in retinal regeneration following injury<sup>13</sup>. In mice, the ciliary body which arises from the CMZ was postulated to contain a population of pigmented cells that were able to form clonogenic spheres and could be differentiated to express marker genes found in photoreceptors, bipolar and Muller glia cells<sup>14</sup>. However subsequent studies demonstrated that these pigmented ciliary epithelial cells fail to differentiate into retinal neurones <em>in vitro</em> or <em>in vivo</em><sup>15</sup>. Instead, progenitors distinct from the classical RPCs, characterised by Msx1 expression and located in the proximal CMZ, were shown to give rise to non-pigmented ciliary epithelial cells and multipotent neural RPCs<sup>16</sup>. Whether the human CMZ displays comparable properties during retinal development remains unclear.
+
+Herein we generated a scRNA-Seq atlas of 24 samples spanning 13 time points from 7.5-21 PCW of human development, enabling identification of early and late RPCs and their transition to neurogenic progenitors and retinal neurons. Complementing scRNA-Seq with spatial transcriptomics (ST) we demonstrate the transient presence of early RPCs in the CMZ of developing human retina. Complementary scATAC-Seq identified key transcription factors and signalling pathways that underline RPCs proliferation and differentiation. Together, these comprehensive single cell analyses shed light on the molecular mechanisms governing human retinal development and provide guidance on the generation of RPCs and differentiated retinal cell types from human pluripotent stem cells (PSCs).
+
+# Results
+
+scRNA-Seq atlas of human developing eyes and retinas  
+To better understand the emergence of RPCs, their heterogeneity and differentiation to retinal cell types, we performed scRNA-Seq of 11 embryonic and fetal eyes spanning 7.5-15 PCW and 13 retinal samples from 10-21 PCW (Table S1). Each sample was embedded using Uniform Manifold Approximation and Projection (UMAP) and clustered using Seurat graph-based clustering (Figure S1, Table S1). In our earlier study of human retinal development, we reported expression of RPC markers (e.g. VSX2) at the peripheral margin of developing human retina as early as 6.5 PCW, and at the outer neuroblastic zone of central retina at 7.8 PCW, marking this time period an important developmental window of human RPC divergence. To better understand this process at the single cell level, we first focused our scRNA-Seq analysis in the 7.5-8.5 PCW embryonic eyes, revealing the presence of retinal and non-retinal cell clusters (Figure S1A). At this stage of human development, the neuroepithelium-derived optic vesicle is surrounded by periocular mesenchyme (POM), which is derived from neural crest and mesoderm and contributes to anterior and non-neural ocular tissue development. In accordance, we identified neural crest cell clusters adjacent to POM and corneal endothelial and corneal stromal cells (Figure S1A). The POM cell clusters of 7.5-8.5 PCW eyes displayed high expression of characteristic markers described in other species such as collagen chains (COL3A1, COL5A1, COL1A1), proteoglycan decorin (DCN) and latent Transforming Growth Factor Beta Binding Protein 1 (LTBP1). Other non-retinal cell clusters including extraocular muscle, ocular surface epithelium, red blood cells, monocytes and macrophages were also identified in most eye samples of this developmental window (Figure S1A, Table S1).
+
+Amongst the retinal cell clusters, RPCs and retinal ganglion cells (RGCs) were identified in all 7.5 - 8.5 PCW eyes, consistent with our previous findings of RGC presence at the basal side of the inner neuroblastic zone at 8 PCW. A recent scRNA-Seq study has documented the presence of early and late RPCs and their transcriptional signature. In accordance, we found that RPCs clusters of 7.5 and 7.7 PCW eyes displayed high expression of characteristic early RPC markers such as SFRP2, RAX, PAX6, VSX2, ZIC2, ZIC1, SIX3, SIX6 etc. (Table S1). While the expression of these markers was maintained in RPCs found in the 8 and 8.5 PCW eye specimens, expression of late RPC markers such as CCND1, ASCL1 and HES6 became prominent, suggesting the co-existence of early and late RPCs. A proliferating cone photoreceptor cluster marked by high expression of proliferation (TOP2A, PCNA), neurogenic progenitors (HES6) and cone markers (RXRG, GNB3) was detected in two eye specimens obtained at 7.7 and 8 PCW (Figure S1A) as well as one of the 10 PCW samples (Figure S1B). In the latter sample, the proliferating cone cluster was distinct to the mature cones. Notably a small cluster of horizontal cells was detected in two specimens of 8 and 8.5 PCW of development (Figure S1A), corroborating with the first reported emergence of immature horizontal cells at day 59 of human development. A retinal pigment epithelium (RPE) cell cluster with high expression of characteristics markers namely PMEL, TYRP1, MITF and GJA1 was identified in all eyes of this stage (Figure S1A, Table S1).
+
+A recent scRNA-Seq study provided evidence of the existence of neurogenic RPCs in addition to the early and late RPCs, based on gene expression profiling. Sridhar and colleagues further defined the neurogenic into intermediate transitional T1, T2 and T3 cell populations with the capacity to give rise to defined types of retinal neurons at specific developmental windows. Our analysis of 10-14 PCW samples demonstrated the presence of some of these neurogenic transitional progenitors; however, a better definition was obtained from the analysis of retinal samples (> 10 PCW) due to a higher number of analysed cells within the retina per se (Figure S1B, Table S1). From 14 PCW, we were able to detect all types of retinal neurons apart from bipolar and Muller glia cells, which were detected from 14 and 16 PCW respectively (Figure S1B,C, Table S1). Published topographical studies report the highest density of microglia in the retinal periphery and retinal margin at ~ 8 PCW. Their presence in the central retina is not observed until 12 PCW (Diaz-Araya et al., 1995). In accordance with these studies, presence of cell clusters with high expression of microglia markers were observed from 14 PCW retinal samples (Figure S1B, C, Table S1).
+
+## Pseudo-temporal analyses reveal transition of RPCs to T1-T3 neurogenic cell clusters and retinal neurons
+
+To identify gene regulatory networks (GRNs) that control RPCs specification and differentiation we integrated the transcriptomes of all retinal cells from 7.5-16 PCW human embryonic/fetal eyes with those of 10-21 PCW retinas. 48,856 cells were embedded using UMAP and clustered with Seurat (Figure 1A). Forty-three clusters were identified: of these, thirty-eight were composed of RPCs and retinal neurons, one of microglia, one of fibroblasts and two expressed markers of more than one cell type (Table S2). These last four clusters and an amacrine cell cluster with high expression of mitochondrial markers (cluster 41) were removed prior to further analysis. For illustrative purposes, all cell clusters defined as the same cell type, are annotated with the same color (Figure 1A).
+
+Six RPC clusters were identified with enriched expression of typical markers including SFRP2, DAPL1, HES1, HMGA1, PAX6, RAX etc. Adjacent to the RPCs, a Muller glia cell cluster was found with high expression of RLBP1, SLC1A3, APOE, VIM, CLU, SOX9 (Table S2). Eight clusters named proliferating RPCs were also found next to RPCs and were characterized by high expression of proliferation markers (e.g., MKi67, TOP2A) and RPCs markers (SFRPP2, SOX2, HES1, ID3 etc.) (Figure 1A). All differentiated retinal cell types were easily identified based on the expression of characteristic cell markers (Table S2). Between the RPCs and retinal neurons, we identified a progenitor cell cluster with high expression of ATOH7, HES6 and DLL3, which was defined as the transitional T1 cluster. Two other cell clusters emanating from T1 were identified leading to the horizontal and amacrine cells clusters, and to photoreceptors and bipolar cells. These were defined as T2 and T3 based on the high expression of PTF1A and PRDM13 and FABP7 and OTX2 respectively (Table S2).
+
+To fully define the transitions from RPCs to T1, T2 and T3 neurogenic progenitors, we performed pseudo-temporal analysis of gene expression changes following cell cycle regression (Figure 1B). This analysis revealed bimodal densities of RPCs, mirroring the transition between early and late RPCs described by Lu and colleagues, followed by the transitional T1 and then T2 and T3 neurogenic progenitors (Figure 1C, D). Early and late RPCs showed overall similar expression patterns of markers, but also some differences in the expression level. For example, the inhibitor of Wnt pathway (SFRP2) was more highly expressed in the early RPCs, while Muller glia cell markers (VIM, FOS) were more prominent in the late RPCs (Figure 1E, Table S2). High expression of well described neurogenic markers such as ATOH7, HES6 and OLIG2 was characteristic of the T1 and to a lesser extent of the T3 progenitor cluster. In accordance with the position of T2 between horizontal and amacrine cells, high expression of key transcription factors (ONECUT2, TFAP2C) required for differentiation to these two lineages was observed (Table S2). The transitional cluster T3 shared the expression of highly expressed T1 markers, and additionally displayed ethe expression of typical photoreceptor precursors (RXRG, NRL) and bipolar cell markers (VSX1), consistent with its positioning between the T1, and photoreceptor and bipolar cell clusters (Figure 1A, C, Table S2). To tease out the transcriptional machinery that controls the specification of each differentiated retinal cell type and their transition, separate pseudo-temporal analyses of RGCs, horizontal and amacrine, and photoreceptor and bipolar cells were conducted (Figures S2, Table S2). These analyses show that RGCs go through a T1 transition (data not shown), horizontal and amacrine cells through a T1-T2 (Figure S2A) and photoreceptor and bipolar cells through a T1-T3 transition state (Figure S2B), corroborating recently published scRNA-Seq data on few stages of human fetal retinal development.
+
+During the course of these analyses many transcription factors highly expressed on defined progenitor subtypes or retinal neurons were identified. Human horizontal cells express the well-known markers such as ONECUT1, ONECUT3, LHX1, PROX1 and show considerable overlap in gene expression with amacrine cells which are characterised by high expression of TFAP2A, LHX9 and MEIS2 (Figure S2A, Table S2). The cone and rod precursors were characterised by high expression of THRB, and NRL and NR2E3 respectively, but interestingly also shared high expression of RGC and horizontal and amacrine cell markers (Figure S2B, Table S2). For example, high expression of RGC marker SNCG was found in the cone precursor cluster, while rod precursors displayed high expression of PROX1, a marker of retinal progenitors, horizontal and AII amacrine cells. These findings were further corroborated by immunofluorescence analysis showing co-expression of cone photoreceptor marker RXRG with RGC marker SNCG in the 8 and 11 PCW, which tailed off at 12 PCW (Figure S3). Notably, these co-expressing cells were in a proliferative state at 8 PCW, but not in the later stages of development.
+
+## ST analyses reveal the location of early RPCs in the CMZ of 8-13 PCW developing human eyes
+
+Our pseudo-temporal analysis above and a recent scRNA-seq analysis of human retinal development identified clear transcriptional signatures of early and late RPCs, however their spatial location in the developing retina has not been defined to date. To investigate this in detail, we performed ST analysis on an 8 PCW eye sample (Figure 2A-D) revealing the presence of 12 spatially organised cell clusters (Table S3) including POM, lens, vitreous, corneal stroma, extraocular muscle and RPE (Figure 2C). Notably, the optic stalk (cluster 9, Figure 2C) was clearly defined by the ST analysis and characterised by high expression of PAX2, ZIC1, HES1, SOX2, LHX2, THY1, PAX5 and ID3 (Table S3). Histologically, a single cell layer was present at the peripheral retina, but outer and inner neuroblastic layers were present in the central retina (Figure 2A, C). The cells residing in the inner neuroblastic layer of central retina (cluster 5) were characterised by high expression of RGC markers such as GAP43, PRPH, SNCG, INA, NEFL and NEFM, corroborating our previous immunofluorescence findings.
+
+High expression of genes typically marking the ciliary margin (e.g., WNT2B), eye field (RAX and PAX6), earlyRPCs (ID3, HMGA1, MDK, SFRP2) and pigmented cell (PMEL, TYR and TYRP1) markers (Table S3) were characteristic of the cluster 4, which was defined as CMZ. Subclustering of cluster 4 resulted in two further cell subclusters (Figure 2E, F) with subcluster 0 expressing at high level marker genes associated with early RPCs (FGF19, SFRP2, DAPL1, ZIC1, ID1, ID3, HMGA1, EEF1A1, TPT1, TMSB4X), and subcluster 1 expressing at high level ciliary body (KCNJ8, TPM2, ADGRA2A) and iris pigment epithelium signature markers (TYR, TRYP1, DCT, SILV, MLANA, PMEL). In contrast, the progenitor cells residing in the outer neuroblastic layer of the central retina (cluster 10, Figure 2C) were characterised by high expression of late RPC markers such as RORB, CKB, HES1, HES5, ASCL1, HES6 and NEUROD1 (Table S3). This led us to hypothesise that the early and late RPCs could be spatially located in different regions of the developing human retina. To investigate this further, we generated a gene expression signature for early and late RPCs from our pseudo-temporal analysis. Violin plots (Figure 2G, I) indicated the highest expression of early RPCs markers in the subcluster 0 of CMZ (cluster 4), while the highest expression of the late RPCs markers was observed in the outer neuroblastic layer of central retina (cluster 10). These findings were further corroborated by the spatial mapping, showing the predominant localisation of early RPCs in the CMZ, the late RPCs in the central retina (Figure 2H, J), and the transitional neurogenic clusters T1, T2 and T3 predominantly in the outer neuroblastic layer of the central retina (cluster 10, Figure S4). Interestingly, an early RPCs expression signature was also observed in and around the optic stalk (Figure 2H), albeit less intense than in the CMZ.
+
+Similar cell clusters with enriched expression of RPCs and pigmented epithelial cell markers were consistently found in the ciliary margin zone of 10, 11 and 13 PCW eye specimens analysed with the ST approach (Figure 3A-C, Table S3); but the frequency of early RPCs was very much reduced compared to the 8 PCW sample. We were unable to fit larger size eyes to the current Visium ST slides, thus it was not possible to extend the ST analyses to fetal eyes older than 13 PCW. However, we utilised the scRNA-Seq data and plotted the ratio of early to late RPCs, showing a significant reduction in fraction of early RPCs from 8 PCW onwards (Figure 3D).
+
+To complement the ST analyses and obtain more insights into early and late RPC localisation into the CMZ during retinal development, we performed RNA-Scope investigations using a marker of early RPCs (ZIC1), iris pigmented epithelium (TFPI2), late RPCs/neurogenic progenitors marker (HES6) and ciliary body (OPTC) in eye samples from 6.3-16 PCW. Consistently with data obtained above, we observed ZIC1 expression in the CMZ as well as rest of retinal neuroepithelium of 6.6 PCW retinas (Figure S5A, B). In contrast, the expression of HES6 was first seen in the central retina (Figure S5B) spreading to the periphery, up to CMZ, from 8 PCW until the last 16 PCW specimens examined (Figure S5D, G, Figure S6C, D, G). Although ZIC1 expression was still present in the CMZ of 8 PCW retinas (Figures S5C, D) it was reduced and by 10 PCW, it was localised in small patches at the posterior end of CMZ of less than 50% of eye sections assessed (Figure S5E, F, G). In 13 and 16 PCW, ZIC1 expression was completely absent from the CMZ (Figure S6A-F), however its expression in the rest of retinal neuroepithelium persisted in a co-expressing pattern with HES6 (Figure S6C, D, G). Together these data corroborate localisation of early RPCs in the CMZ of developing human retinas with decreasing frequency from 8 PCW of development and reveal the propagation of neural retina differentiation from the centre to the periphery.
+
+## Characterisation of chromatin accessibility during human retinal development
+
+To investigate chromatin accessibility during human retinal development, we performed scATAC-Seq analyses of two developing human eyes (8.5 PCW), and ten retinal samples dissected from 10-21 PCW (Table S4). 118,883 cells were captured using the 10XGenomics Chromium Single Cell ATAC Library and Gel Bead Kit. The corresponding scRNA-Seq datasets were used as reference maps to identify ATAC-Seq clusters for each sample (Figure S7). Similarly, to scRNA-Seq analysis, defined clusters of corneal stroma, epithelium, endothelium, and periocular mesenchyme, CMZ, extraocular muscle, red blood cells, microglia and optic nerve were found in the 8 PCW human eye samples alongside retinal clusters comprised of RPCs, RGCs, horizontal and amacrine cells and transient neurogenic cluster T1 (Figure S7A). All the retinal cell types were present between 10-14 PCW, albeit some at the precursor stage (for example rod precursors at 10 PCW), in addition to RPCs and the three types of transient neurogenic clusters T1, T2, T3 (Figure S7A). From 16 PCW, small clusters of microglia and the last-born cell type, Muller glia cells were evident (Figure S7B), corroborating the scRNA-Seq data and previously published sequential order of retinal cell emergence.
+
+54,045 retinal cells from 12 samples and 9 developmental timepoints encompassing 8-21 PCW with a 3694,28 average median fragments per cell were integrated. The cells were clustered using the chromatin accessibility peaks near previously known marker genes resulting in 22 clusters (Figure S7C, Table S4). These included the abundant cell types (for example rods, cones) as well as the rarer cell types (e.g., microglia) and several types of amacrine cells (gabaergic, glycinergic and starburst). The DNA accessibility peaks were classified using annotation from cellranger and associated with promoters (if found within -1000 to +100 bp of the transcription start sites), exons, introns, distal (if found within 200 kb of the closest transcription start site), or intergenic regions (if not mapped to any genes) (Figure 4A, Table S5). This analysis enabled us to identify cell type specific regions of accessibility in RPCs, transient neurogenic progenitors T1, T2, T3 and retinal neurons (Figure 4B). All clusters displayed scATAC-Seq marker peak enrichment of cell type specific marker genes (Figure 4C).
+
+We then went on to predict transcription factor (TF) binding motifs within the scATAC peaks using Signac (Figure 5A, Table S6), followed by foot printing validation analysis (Figure 5B, C). In RPCs, we identified binding motifs for TFs expressed in the optic stalk (VAX1, VAX2), eye field (RAX) and RPCs (LHX6, VSX2, SOX2) as well as TF binding motifs that were shared with other cell types (for example FOS, NFI family members and SOX6 binding motifs were shared with Muller glia cells, Figure 5A). Notably we identified binding motifs for TFs not previously associated with RPCs, for example EN1, GBX2, LBX2, SHOX2 and GSX1 (Figure 5A,B, Table S6).
+
+In accordance with sequential emergence of T2 and T3 neurogenic clusters from T1, we observed a set of shared TF binding motifs for NEUROG2, NEUROD1, HAND2, TAL1:TCF3, PITF1A, ATOH7, BHLHA15, MYF5 and ATOH1 (Figures 5A). The earliest born retinal cell types, RGCs displayed the POU4F and EBF family members binding motifs, whilst horizontal and amacrine cells were characterised by ONECUT and CUX, and TFAP2 and MEIS family members binding motifs respectively (Figure 5A, C). Rod and cone photoreceptors are derived from the T3 neurogenic progenitors; hence shared binding motifs were identified for TFs such as OTX2, CRX and DMBX1 (Figure 5A, C), which are well described in the literature for their role in photoreceptor specification. Importantly, shared binding motifs in T3 and photoreceptors were also discovered for PITX1 and GSC TFs not associated previously with a role in rods or cones (Figure 5A). Bipolar cells are also derived from the T3 progenitors; hence enrichment of T3 TF binding motifs (such as OTX2) was observed in addition to less well characterised TFs such ZBTB18 (Figure 5C). Enrichment of NFI binding motif family members was observed for glia cells (Muller glia and microglia) in accordance with their role in regulating specification of the late-born cell types in the retina.
+
+## scATAC-Seq enables prediction of gene regulatory networks and novel TFs that govern retinal development
+
+To identify GNRs governing the RPCs differentiation to transient neurogenic progenitors (T1, T2, T3) and the retinal neurons, the upstream regulator tool in IPA was used, combined with overlay analysis of DA peaks. A large number of upstream regulators including TGFβ, IGF-1, FGF-2, Sonic hedgehog were predicted to be activated in RPCs (Table S7). Their main common characteristic was predicted activation of key genes encoding proteins important for RPCs proliferation (HES1, CCND1, ID3) and inhibition of transcription factors (ATOH7, NEUROD4, NEUROD1, PTF1A, HES6) (Figures 6A, B) that define the RPC competence to RGCs, horizontal and amacrine cell fates, and photoreceptor differentiation. Amongst the predicted inhibited upstream regulators, we found key transcription factors such as PAX6 and ASCL1, shown to control the timing and specificity of retinal neurogenesis (Figures 6C, D). Together these findings suggest the presence of a finely tuned balance between activation of proliferation regulators and inhibition of neurogenic cues to maintain RPC self-renewal. Conversely in the transient neurogenic progenitors, we predicted the activation of upstream neurogenic regulators (e.g., ASCL1 in T1, FOXA2 in T2, GTF2IRD1 in T3), which control the expression of TF necessary for retinal cell type specific differentiation (Figure S8A-C) and inhibition of regulators (e.g. BMP4, LIN28A, IGF-1) that govern RPCs fate (Figure S8D-F).
+
+Performing the same analysis on the differentiated retinal cells revealed some interesting insights and predicted novel upstream regulators (Table S7). For example, Eomesodermin (EOMES), a target gene of Pou4f2, required for RGC and optic nerve development in mouse, was predicted to be activated in horizontal cells (Figure S9B), resulting in activation of LHX1, a transcription factor that specifies horizontal cells, while suppressing LHX9, a transcription factor required for amacrine cell subtype specification. We identified a novel putative upstream regulator in RGCs, namely KLF2 (Figure S9A), which is predicted to inactivate Notch signalling, an important event required for RGC differentiation. Our scRNA-Seq data have shown that the basic-helix-loop-helix PTF1A is highly expressed in the T2 progenitors, which give rise to amacrine and horizontal cells. A similar role for PTF1A has been demonstrated during mouse retinal development, albeit transient expression in all types of amacrine cells has been demonstrated in the zebrafish. In accordance with these findings, the PTF1A upstream regulator was identified in amacrine cells, resulting in activation of TFs that promote amacrine cell specification and function (Figure S9C).
+
+Our analysis was also able to predict more complex upstream regulators, which are ligand-activated and able to play a role in the control of gene expression in a cell type specific manner. For example, the peroxisome proliferator-activated receptor γ (PPARγ), was identified as a putative novel upstream regulator in cone photoreceptors (Figure S9D) and predicted to activate amongst others PRDM1, which has been demonstrated to stabilise photoreceptor cell fate in OTX2+ progenitors by preventing bipolar cell induction. In contrast, the upstream regulator TCF7 in bipolar cells is predicted to suppress PRDM1 and activate expression of key genes important for bipolar cell function such as GNAO1 and TRPM1 (Figure S9F). G-protein coupled receptors (GPCRs) play a significant role in many tissues by transducing complex signalling networks that coordinate gene expression. In accordance, one of the putative activated upstream regulators in rods, was the GPCR Rhodopsin (RHO), the most abundant protein in rods which functions as the primary photoreceptor molecule of vision (Figure S9E). Our data suggests that RHO may regulate the transcription of rod specific phosphodiesterases (e.g., PDE6G, PDE6B) and transducins (GNAT1, GNGT1).
+
+## scATAC-seq trajectory analysis reveals a role for TEAD transcription factors in retinal lamination and RPC, photoreceptor and RGCs specification
+
+The TF binding motif analysis in RPCs (Figure 5A) revealed an enrichment for TEA domain (TEAD) TFs in RPCs (Figure 7A, Table S6). TEADs play an essential role in mediating YAP-dependent gene expression resulting in transcription of target genes responsible for cell proliferation, inhibition of apoptosis or retinal neurogenesis. During retinal development, YAP’s expression is restricted to the outer neuroblastic layer where RPCs reside. The TEAD TFs display different expression patterns in the retina, with Tead2 being highly expressed in the proliferating cells located at the basal side of the outer neuroblastic layer of mouse retina and Tead3 being highly enriched in the inner neuroblastic layer (genepaint.org).
+
+To assess if TEADs binding is essential for RPC development and differentiation, we treated PSCs-derived retinal organoids with TEAD palmitoylation inhibitor MGH-CP1, that blocks TEAD2/4-YAP interaction and suppresses the expression of their target genes in cancer cell lines. Treatment was started at day 25, coinciding with immunofluorescence detection of RPCs in the retinal organoids. Three different doses (2.5, 5 and 10 µM) of MGH-CP1 were added every two days to the culture media for 14 days. Under control conditions (treated with vehicle) retinal organoids develop a bright phase retinal neuroepithelium by day 28 (day 3 of treatment) which expanded over time. By the end of treatment the majority of control retinal organoids (> 75%) had full coverage with bright phase retinal neuroepithelium, with a minority of organoids displayed partial coverage (15%) or no coverage (7%) at all (Figure S10A, B). In contrast, the great majority of retinal organoids treated with MGH-CP1, displayed either partial or full loss of bright phase neuroepithelium in a dose dependent manner. These findings were fully corroborated by immunofluorescence analyses, which showed the presence of partial retinal neuroepithelium harbouring VSX2+ RPCs in organoids treated with 5 µM MGH-CP1 and much reduced and mislocalised RPCs in the organoids treated with 10 µM MGH-CP1 (Figure 7B, F). The latter were also often characterised by the presence of SNCG+ RGCs in the apical layer of the organoids and internal rosettes with VSX2+ RPCs or Ki67 proliferating cells (Figure 7C-E). Quantitative analysis demonstrated attenuated RPCs, photoreceptor, and RGCs specification in a dose-dependent manner in retinal organoids treated with MGH-CP1 (Figure 7f, H-J). In accordance with TEADs role in activating transcription of genes important for cell proliferation, we noted a significant reduction in the fraction of Ki67+ cells in the retinal organoids treated with the highest dose of MGH-CP1 (Figure 7G), which also showed a small but significant increase in the percentage of apoptotic cells (Figure S10C).
+
+# Discussion
+
+In this study we analyzed the transcriptional and DNA accessibility profiles of over 277,000 cells from 35 human eyes and retinal samples collected from 7.5-21 PCW of human development to delineate the diversification of RPCs and the GRNs that govern their differentiation. Using spatiotemporal single cell RNA-Seq analyses we demonstrate the transient localisation of early RPCs in the CMZ of developing human retinas and propagation of neural retina differentiation from the centre to the periphery. Single cell ATAC-Seq analysis revealed a significant enrichment of TEAD transcription factor binding motifs in RPCs, which when inhibited led to loss of retinal lamination, and attenuated RPCs, photoreceptor and RGCs differentiation.
+
+The human eye is a heterogenous entity, comprised of diverse tissues derived from neuroectoderm, neural crest and mesenchymal cells. Single cell RNA-Seq studies have focused on the transcriptome of developing and/or adult cornea<sup>51</sup>, iris, ciliary body<sup>52</sup>, neural retina<sup>5, 53</sup><sup>6</sup>, RPE and choroid<sup>54</sup><sup>55</sup>. However, a comprehensive single cell spatiotemporal profiling of human developing eyes has not been attempted before. Herein, we undertook scRNA-Seq analysis of 10 developing human eyes encompassing 7.5-10 PCW of development, revealing for the first-time transcriptional signature of neural crest, extraocular muscle and POM cells in addition to ciliary body, iris pigmented epithelium, ocular surface epithelium, stroma and endothelium, neural retina and RPE cells. In all cases, the neural crest cell clusters were closely associated with cell types derived therefrom including corneal stroma and endothelium, POM, periocular connective tissue and melanocytes, or those that require early signals from neural crest for their development such as extraocular muscle<sup>56</sup> and lens fibers<sup>57</sup>. Defects in neural crest formation are at the heart of several severe craniofacial and ocular anomalies including ocular coloboma, glaucoma etc.<sup>17</sup>, hence a comprehensive understanding of transcriptome of neural crest cells and its derivatives as described herein, is of great importance for understanding the complexities underlying congenital eye diseases.
+
+Taking advantage of ST, we were able to spatially locate various structures in the developing human eyes of 8-13 PCW, revealing for the first time the spatial and single cell transcriptome of the optic stalk, which forms the primitive connection between the retina and the brain. Importantly, the ST approach enabled us to analyse at single cell resolution the CMZ in developing human eyes of 8, 10, 11 and 13 PCW. Using PSCs-derived retinal organoids, Kuwahara and colleagues provided evidence of a putative CMZ, containing sphere-forming cells, able to generate<em>de novo</em> retinal cells<sup>59</sup>. However, whether the human CMZ is a source of RPCs during embryogenesis has been a long-standing question in the field. Using a combination of single cell RNA-Seq and ST, we show the presence and localisation of early RPCs in the CMZ of 8 PCW human eyes with decreasing frequency as development proceeds from 8 PCW onwards. Moreover, the CMZ located early RPCs were characterised by distinct expression signature compared to the ciliary body and iris pigmented epithelium, suggesting that they are a distinct entity to the putative stem cell ciliary body cells suggested by Gautam and colleagues<sup>52</sup>. Together our data provide for the first-time evidence on the transitional presence of early RPCs in the CMZ at a particular stage of human retinal development that needs to be functionally validated both<em>in vivo</em> using a larger number of early eye specimens and<em>in vitro</em> in retinal organoids by combining cell barcoding with single cell analyses. We were also able to locate some of the early RPCs in and around the optic stalk. The region around the optic stalk is initially continuous with the optic fissure margins. In the human eye fissure closure begins in the middle and most of the posterior part eventually forms part of the optic nerve head. Like the CMZ, the fissure margins are a zone of transition between the neural retina and RPE, thus it is likely that the optic stalk region may retain characteristics in common with the CMZ. In accordance, high expression of early RPC markers (<em>SOX2, HES1, ZIC1, NR2F1, LHX2</em>) and typical optic stalk marker PAX2, was observed in the optic stalk cluster (<strong>Table S3</strong>). The relationship between the early RPCs and optic stalk and how these contribute to retinogenesis, remains to be further investigated.
+
+In addition to early and late RPCs, we were able to demonstrate the presence of transient neurogenic progenitors, named T1, T2, and T3 by both scRNA- and ATAC-seq, corroborating recently published data<sup>6</sup>. The pseudotime analyses demonstrated that RGCs transit through the T1 neurogenic progenitors, horizontal and amacrine cells via T1, T2 progenitors and bipolar cells and photoreceptors via T1, T3 progenitors. We were also able to demonstrate that both cone and rod photoreceptors develop from a precursor stage which precedes their maturation. Notably, we also observed at the early stages of development some “plasticity” regarding lineage transcription factor expression such as<em>SNCG</em>, a marker of RGCs was expressed in cone precursors and<em>PROX1</em>, a marker of horizontal and amacrine cells expressed in rod photoreceptors. Strikingly, Prox1 expression has been noted in rod progenitors as well as the Müller glia cells in the fish<sup>60</sup>, while high co-expression of RGC markers was reported recently in the proliferating cones of retinoblastoma pluripotent stem cell derived organoids as well as patient retinoblastoma samples<sup>61, 62</sup>. Together these data suggest that although the expression of certain TFs maps to individual retinal cell trajectories, some of these may be reused temporary during the development of other retinal cell types or redeployed upon re-entry into the cell cycle and conversion to malignancy.
+
+We complemented the scRNA-Seq with ATAC-Seq analysis of similarly staged developing human eyes and retinas. We found the scATAC-Seq to be more informative when it came to subclustering of amacrine cells into the three main subtypes namely gabaergic, glycinergic and starburst amacrine cells. Notably we also found scATAC-Seq to predict cell type clusters at earlier stages than scRNA-Seq analysis. For example, horizontal and amacrine cell precursors were identified at 8 PCW in the scATAC-Seq but 10 PCW in the scRNA-Seq. This could be due to DNA chromatin accessibility preceding key TF expression and cell fate determination. Similarly to recently published studies of scATAC-Seq in the developing and adult retina<sup>7, 10, 11, 12</sup>, we were able to reveal the accessible chromatin regions and putative TF binding motifs for the RPCs, transient neurogenic progenitors and all the retinal cells as they emerged during the course of development. This identified both well characterised TFs (e.g., POU4F in RGCs, ONECUT2 in horizontal cells, RAX in RPCs) as well as novel TFs (e.g. EN1, GBX2, LBX2, SHOX2, TEAD1-3 and GSX1 in RPCs), which deserve further functional validation in animal models and retinal organoids.
+
+The scATAC-seq analyses enabled us to predict GRNs and novel TFs that govern retinal neurogenesis. In particular, enrichment of TEAD binding in RPCs was highlighted, suggesting a role for the Yap-Hippo signalling pathway during human retinal development. Evidence demonstrating the role of Hippo-Yap signalling during retinogenesis in mice has shown that developing retinas devoid of Yap display disrupted apical junctions, rosette formation and loss of laminar arrangements<sup>49</sup>. Our data fully corroborate these findings, showing the presence of rosettes within retinal organoids and disrupted laminar organisation when TEAD binding is disrupted. In addition to playing a role in retinal lamination, activated Yap has been shown to simultaneously interact with TEADs and Rx1 during zebrafish embryogenesis to drive the expression of proliferation-related genes and attenuate the trans-activation of photoreceptor genes respectively<sup>63</sup>. In our organoid model, we observed attenuated photoreceptor differentiation, which suggests that upon inhibition of TEAD binding, Yap may interact with the human ortholog of Rx1 (RAX) to suppress photoreceptor differentiation. Importantly we observed a significant reduction in RPCs and cellular proliferation at the highest dose of the inhibitor, but reduced RGCs and photoreceptor specification at both 5- and 10 µM doses, suggesting that the latter two events may be more sensitive to TEAD inhibition. While activation of Hippo-Yap signalling has been studies in the context of Muller glia cell activation<sup>64, 65</sup>, our data provide for the first-time important insights into the function into Hippo-Yap and related TEAD signalling during the very early stages of human retinogenesis.
+
+The data generated herein have been submitted to open access online resources adding valuable information to the currently existing scRNA and ATAC-Seq to increase the sample and read size, but most importantly the novel use of ST to reveal the localisation of RPCs and neurogenic progenitors provides important insights into the long-debated source of retinal progenitors during human development.
+
+# Methods
+
+**scRNA- and -ATAC-Seq**
+
+11 samples of developing human eyes and 14 samples of neural retina from 7.5-21 post-conception weeks (Tables S1, S4) were obtained from the Human Developmental Biology Resource under ethics permission 08/H0906/21+5 issued by the NorthEast Newcastle and North Tyneside 1 Research Ethics Committee. All samples were isolated and dissociated to single cells using a neurosphere dissociation kit (Miltenyi Biotech). Approximately 10,000 cells from each sample were captured, and sequencing libraries generated using the Chromium Single Cell 3’ Library & Gel Bead Kit (version 3, 10x Genomics). 10,000 of the subsequent nuclei were captured, and sequencing libraries generated using the Chromium Single Cell ATAC Library & Gel Bead Kit (version 1, 10x Genomics). Single cell RNA-Seq libraries were sequenced to 50,000 reads per cell and scATAC-Seq libraries were sequenced to 25,000 reads per nucleus on an Illumina NovaSeq 6000.
+
+**scRNA-Seq analysis**
+
+The BCL files were de-multiplexed using CellRanger mkfastq version 3.01 and then aligned and quantify them against the human reference genome GRCh38 using Cellranger count. We performed quality control checks for each sample in R for each sample and removed any cells with fewer than 1000 reads or 500 genes or greater than 10% mitochondrial reads. Any cells which expressed haemoglobin genes were also removed from the analysis. Doublets were predicted using DoubletFinder and filtered from the data. The Seurat R package (version 4.3.0) was then used to process the data prior to integration. Firstly, the raw data was normalised using the standard parameters. The FindVariableFeatures function used to select 2000 highly variable genes. The data was then scaled using ScaleData and the following variables were regressed out "percent.mt", "nCount_RNA", "nFeature_RNA". Principle component reduction with the 2000 highly variable genes selected, was applied to the scaled data using FindPCA function.
+
+Harmony (version 0.1.1) was used to remove sample batch effects from the data. A Uniform Manifold Approximation and Projection (UMAP) reduction was than applied to the first 10 harmony corrected components. The Seurat graph-based method was used to cluster the data. Resolutions from 0.2 to 2.2 were tested. Differential expression analysis using the standard settings in the FindMarkers function from Seurat were used to identify markers genes within each cluster. Cell types were then assigned to these clusters (Table S1). We then performed pseudotime on 4 branches of the UMAP, namely: RPC-T1-T2-T3, RPC, RPC-T1-RGC-T2-HC-AC, RPC-T1-T2-HC-AC, RPC-T1-T2-RGCs-HCs-ACs and T1-T3-Cones-Rods-BC. The data was subset by these cell type groups we re-clustered the data using the method described in the previous section. The cell types were re-annotated to ensure robust assignment of the cell types. Monocle 3 was used to the order the cells and infer a pseudotime trajectory within the separate branches. We used Seurat FindAllMarkers to identify differentially expressed genes in the different cell types within each branch. The top 10 genes for each cell type, with the cells ordered by pseudotime order, were visualised in heatmaps.
+
+**scATAC-ATAC analysis**
+
+Peaks were detected using Cellranger ATAC software (version 1.2) in each of the samples. A set of shared peaks was then defined using Bedtools merge (version 2.30) and the Cellranger ATAC reanalyse function was then used to call peaks using the shared peak set. The datasets were imported using Signac and quality control steps were performed to remove cells low quality cells. We excluded cells with fewer than 20% of reads in peak region fragments, or less than 3000 peak region fragments. Cells with a TSS enrichment score of less than 2 and Blacklist ratio greater than 0.05 or a nucleosome signal of less than 4 were also removed from downstream analysis.
+
+Signac was then used to perform term frequency-inverse document frequency (TF-IDF) normalisation and singular value decomposition (SVD) dimension reduction for each individual sample. This was followed by UMAP reduction using components 2 to 30, and cluster analysis using Seurat. Signac was then used to generate a gene activity matrix based on open regions for each cell and FindAllMarkers from Seurat was used to predict upregulated genes for each cluster. These gene lists were used to assign cell type identity to the clusters. Retinal cell types were selected from each sample for integration and the normalisation, dimension reduction, and clustering steps, and cell type annotation steps described for the individual samples were applied to the combined dataset. We then identified differentially accessible peaks for each of the annotated cell types using the logistic regression (LR) test from the FindAllMarkers function. The average peak value for each cell type was calculated and differentially accessible peaks were plotted using the ComplexHeatmap package.
+
+Chromvar was used to compute per cell motif activity scores for each cell and FindAllMarkers was used to compute enriched motifs for each cell type. The top enriched motifs ordered by average difference in z-score between each cell type are shown in a heatmap generated with the ComplexHeatmap package. Motif plots and Footprint plots were generated using Signac. The differentially accessible peaks were analysed using Qiagen Ingenuity Pathway Analysis (IPA). The lists of promoters were used to predict upstream regulators and motif data was overlaid onto the prediction to look for consensus in the predictions.
+
+**ST**
+
+Fresh frozen sections of four fetal retina samples of 8, 10, 11 and 13PCW were used for the spatial transcriptome analyses performed with the Visium Spatial Gene Expression kit from 10XGenomics. First the tissue optimisation was performed defining 30 minutes as the most optimal permeabilization time window. The ST procedure was performed according to manufacturer’s instructions. Four tissue sections from each sample were carefully placed into the four capture areas, fixed, haematoxylin and eosin stained and images in order to preserve histological information. This makes it possible to overlay the cell tissue image and the gene expression data in a later step. After permeabilization, reverse transcription reagents were added on top of the tissues. The tissues were subsequently removed, leaving the cDNA coupled to the arrayed oligonucleotides on the slide. The cDNA-RNA hybrids were cleaved off the chip and the sequencing libraries were prepared. The sequencing depth varied between 25,000,000 and 200,000,000 million reads.
+
+**ST analysis**
+
+Spaceranger version 1.0 was used to demulitplex the data, generate gene expression matrices by aligning the FASTQs to GRCH38. The pipeline was also used to calculate spot co-ordinates using fiducial detection and to identify the spots covered by the tissue. The results from Spaceranger were imported into R using Spaniel (version 1.12). The data from each sample, which was derived from 4 consecutive sections within the tissue, was normalized and clustered, with a resolution of 0.5, using Seurat as described in the scRNA analysis section. FindAllMarkers was used to identify differentially expressed genes clusters. Spaniel was used to visualize the clusters overlaid on the tissue and generate spatial expression plots.
+
+**RNAscope**
+
+RNAscope in situ hybridization assay was used to determine the expression profile of ZIC1, TFPI2, OPTC and HES6 during the development of the human retina. Formalin-fixed, paraffin embedded staged human fetal eyes were provided by the MRC/Wellcome Trust funded Human Developmental Biology Resource; www.hdbr.org). Tissue sections were taken on a microtome at 8μm intervals to SuperFrost microscope slides and baked for 1 h at 60°C before the paraffin was removed in xylene. The sections were first dehydrated in two changes of 100% ethanol before a target retrieval was performed by heating the sections for 8 min at 95°C and incubating with a protease enzyme cocktail (ACD- Cat. No. 322381) for 15 min at 40°C. RNAscope probes Hs-TFPI2 (ACD- Cat No. 470361), Hs-OPTC-C3 (ACD- Cat No. 1165211-C3), Hs-ZIC1-C2 (ACD- Cat No. 542991-C2), Hs-PAX2 (ACD- Cat No. 442541), Hs-HES6-C3 (ACD- Cat No. 521301-C3) were hybridised to the tissue for 2 h at 40°C followed by multiple rounds of signal amplification. Positive (ACD- Cat No. 320861) and negative (ACD- Cat No. 320871) control probes were used to confirm specificity. The annealed probes were detected using Opal fluorophores OPAL 570 (c1) OPAL 650 (C2) and OPAL 520 (C3) and imaged using a Zeiss microscope and ZEN software.
+
+**Immunofluorescence analyses (IF)**
+
+Human fetal retinal tissue was fixed and IF performed on cryostat sections as previously described (Mellough et al., 2019). Sections were reacted against the following primary antibodies: CRX (1:200, Abnova, H00001406-MO2), Ki67 (1:200, Abcam, ab15580), Recoverin (1:1000, Millipore-Merck, ab5585), RXRγ (1:200, Santa Cruz Biotechnology, sc-555), SNCG (1:500, Antibodies.com, A121664) and VSX2 (1:100, Santa Cruz, sc-365519). Secondary antibodies were conjugated to Alexa488 (Jackson Immuno Research Laboratories), Cy3 (Jackson Immuno Research Laboratories) and Alexa 647 (Thermo Fisher). Antibody specificity was assessed by omitting the primary antibodies. Images were obtained using a Zeiss Axio Imager.Z1 microscope with ApoTome.2 accessory equipment and AxioVision or Zen software. Between 5 and 10 images were collected from each IF analyses. Images are displayed as a maximum projection and adjusted for brightness and contrast in Adobe Photoshop CS6 (Adobe Systems).
+
+**Retinal organoid differentiation**
+
+WT2 hiPSCs derived and characterised in our lab<sup>66</sup> were expanded in mTESR™1 (StemCell Technologies, 05850) on growth factor reduced Matrigel (BD Biosciences, San Jose, CA) coated plates at 37°C and 5% CO<sub>2</sub>. Retinal organoids were generated as followed: hiPSCs were dissociated into single cells using Accutase (Gibco, A1110501) and seeded at a density of 7,000 cells/well onto U-bottom 96-well plates (Helena, 92697T) prior coated with Lipidure (AMSbio, AMS.52000011GB1G) in mTeSR™1 with 10 μM Y-27632 ROCK inhibitor (Chemdea). After 2 days, 200 μl of differentiation medium as described in Dorgau et al.<sup>67</sup> was added. Half of the differentiation medium was changed every 2 days until day 18 of differentiation. Then, the media was supplemented with 10% Fetal Calf Serum (FCS; Life Technologies, UK), Taurine (Sigma-Aldrich) and T3 (Sigma-Aldrich) and retinal organoids were transferred to 6-well low attachment plates (Corning, 3471). Retinoic Acid (RA; 0.5 µM; Sigma-Aldrich) was added from day 90 to day 120 of differentiation. Media was changed every 2-3 days. MGH-CP1 (Sigma-Aldrich) at different concentrations (2.5µM, 5µM and 10µM) or the vehicle control (DMSO) were added to the culture media and refreshed at every media change. Retinal organoids were incubated with MGH-CP1 or DMSO for 2 weeks, starting at day 25 of differentiation and were collected after the incubation period at day39 of differentiation.
+
+# References
+
+1. Marquardt T, Gruss P. Generating neuronal diversity in the retina: one for nearly all. *Trends Neurosci* **25**, 32-38 (2002).
+2. Mellough CB, et al. An integrated transcriptional analysis of the developing human retina. *Development* **146**, (2019).
+3. Aldiri I, et al. The Dynamic Epigenetic Landscape of the Retina During Development, Reprogramming, and Tumorigenesis. *Neuron* **94**, 550-568 e510 (2017).
+4. Hoshino A, et al. Molecular Anatomy of the Developing Human Retina. *Dev Cell* **43**, 763-779 e764 (2017).
+5. Lu Y, et al. Single-Cell Analysis of Human Retina Identifies Evolutionarily Conserved and Species-Specific Mechanisms Controlling Development. *Dev Cell* **53**, 473-491 e479 (2020).
+6. Sridhar A, et al. Single-Cell Transcriptomic Comparison of Human Fetal Retina, hPSC-Derived Retinal Organoids, and Long-Term Retinal Cultures. *Cell Rep* **30**, 1644-1659 e1644 (2020).
+7. Wang SK, et al. Single-cell multiome of the human retina and deep learning nominate causal variants in complex eye diseases. *Cell Genom* **2**, (2022).
+8. Hu Y, et al. Dissecting the transcriptome landscape of the human fetal neural retina and retinal pigment epithelium by single-cell RNA-seq analysis. *PLoS Biol* **17**, e3000365 (2019).
+9. Singh RK, Winkler PA, Binette F, Petersen-Jones SM, Nasonkin IO. Comparison of Developmental Dynamics in Human Fetal Retina and Human Pluripotent Stem Cell-Derived Retinal Tissue. *Stem Cells Dev* **30**, 399-417 (2021).
+10. Lyu P, et al. Gene regulatory networks controlling temporal patterning, neurogenesis, and cell-fate specification in mammalian retina. *Cell Rep* **37**, 109994 (2021).
+11. Xie HH, et al. Chromatin accessibility analysis reveals regulatory dynamics of developing human retina and hiPSC-derived retinal organoids. *Sci Adv* **6**, (2020).
+12. Finkbeiner C, Ortuno-Lizaran I, Sridhar A, Hooper M, Petter S, Reh TA. Single-cell ATAC-seq of fetal human retina and stem-cell-derived retinal organoids shows changing chromatin landscapes during cell fate acquisition. *Cell Reports* **38**, (2022).
+13. Fischer AJ, Reh TA. Identification of a proliferating marginal zone of retinal progenitors in postnatal chickens. *Dev Biol* **220**, 197-210 (2000).
+14. Tropepe V, et al. Retinal stem cells in the adult mammalian eye. *Science* **287**, 2032-2036 (2000).
+15. Gualdoni S, et al. Adult Ciliary Epithelial Cells, Previously Identified as Retinal Stem Cells with Potential for Retinal Repair, Fail to Differentiate into New Rod Photoreceptors. *Stem Cells* **28**, 1048-1059 (2010).
+16. Belanger MC, Robert B, Cayouette M. Msx1-Positive Progenitors in the Retinal Ciliary Margin Give Rise to Both Neural and Non-neural Progenies in Mammals. *Developmental Cell* **40**, 137-150 (2017).
+17. Williams AL, Bohnsack BL. The Ocular Neural Crest: Specification, Migration, and Then What? *Front Cell Dev Biol* **8**, (2020).
+18. Doane KJ, Ting WH, McLaughlin JS, Birk DE. Spatial and temporal variations in extracellular matrix of periocular and corneal regions during corneal stromal development. *Experimental Eye Research* **62**, 271-283 (1996).
+19. Bi L, Lwigale P. Transcriptomic analysis of differential gene expression during chick periocular neural crest differentiation into corneal cells. *Dev Dyn* **248**, 583-602 (2019).
+20. Bassett EA, Wallace VA. Cell fate determination in the vertebrate retina. *Trends Neurosci* **35**, 565-573 (2012).
+21. Dyer MA, Livesey FJ, Cepko CL, Oliver G. Prox1 function controls progenitor cell proliferation and horizontal cell genesis in the mammalian retina. *Nat Genet* **34**, 53-58 (2003).
+22. Perez de Sevilla Muller L, Azar SS, de Los Santos J, Brecha NC. Prox1 Is a Marker for AII Amacrine Cells in the Mouse Retina. *Front Neuroanat* **11**, 39 (2017).
+23. Kubo F, Takeichi M, Nakagawa S. Wnt2b controls retinal cell differentiation at the ciliary marginal zone. *Development* **130**, 587-598 (2003).
+24. Mui SH, Kim JW, Lemke G, Bertuzzi S. Vax genes ventralize the embryonic eye. *Genes Dev* **19**, 1249-1259 (2005).
+25. Furukawa T, Kozak CA, Cepko CL. rax, a novel paired-type homeobox gene, shows expression in the anterior neural fold and developing retina. *P Natl Acad Sci USA* **94**, 3088-3093 (1997).
+26. Balasubramanian R, Bui A, Ding Q, Gan L. Expression of LIM-homeodomain transcription factors in the developing and mature mouse retina. *Gene Expr Patterns* **14**, 1-8 (2014).
+27. Zou CJ, Levine EM. Vsx2 Controls Eye Organogenesis and Retinal Progenitor Identity Via Homeodomain and Non-Homeodomain Residues Required for High Affinity DNA Binding. *Plos Genet* **8**, (2012).
+28. Taranova OV, et al. SOX2 is a dose-dependent regulator of retinal neural progenitor competence. *Gene Dev* **20**, 1187-1202 (2006).
+29. Nishida A, et al. Otx2 homeobox gene controls retinal photoreceptor cell fate and pineal gland development. *Nat Neurosci* **6**, 1255-1263 (2003).
+30. Furukawa T, Morrow EM, Cepko CL. Crx, a novel otx-like homeobox gene, shows photoreceptor-specific expression and regulates photoreceptor differentiation. *Cell* **91**, 531-541 (1997).
+31. Wong L, Weadick CJ, Kuo C, Chang BSW, Tropepe V. Duplicate dmbx1 genes regulate progenitor cell cycle and differentiation during zebrafish midbrain and retinal development. *Bmc Dev Biol* **10**, (2010).
+32. Clark BS, et al. Single-Cell RNA-Seq Analysis of Retinal Development Identifies NFI Factors as Regulating Mitotic Exit and Late-Born Cell Specification. *Neuron* **102**, 1111-+ (2019).
+33. Wall DS, et al. Progenitor cell proliferation in the retina is dependent on Notch-independent Sonic hedgehog/Hes1 activity. *Journal of Cell Biology* **184**, 101-112 (2009).
+34. Das G, Choi Y, Sicinski P, Levine EM. Cyclin D1 fine-tunes the neurogenic output of embryonic retinal progenitor cells. *Neural Dev* **4**, 15 (2009).
+35. Du Y, Yip HK. The Expression and Roles of Inhibitor of DNA Binding Helix-Loop-Helix Proteins in the Developing and Adult Mouse Retina. *Neuroscience* **175**, 367-379 (2011).
+36. Wu FG, et al. Single cell transcriptomics reveals lineage trajectory of retinal ganglion cells in wild-type and Atoh7-null retinas. *Nat Commun* **12**, (2021).
+37. Cherry TJ, Wang S, Bormuth I, Schwab M, Olson J, Cepko CL. NeuroD Factors Regulate Cell Fate and Neurite Stratification in the Developing Retina. *J Neurosci* **31**, 7365-7379 (2011).
+38. Fujitani Y, et al. Ptf1a determines horizontal and amacrine cell fates during mouse retinal development. *Development* **133**, 4439-4450 (2006).
+39. Bae S, Bessho Y, Hojo M, Kageyama R. The bHLH gene Hes6, an inhibitor of Hes1, promotes neuronal differentiation. *Development* **127**, 2933-2943 (2000).
+40. Philips GT, et al. Precocious retinal neurons: Pax6 controls timing of differentiation and determination of cell type. *Developmental Biology* **279**, 308-321 (2005).
+41. Nelson BR, Hartman BH, Ray CA, Hayashi T, Bermingham-McDonogh O, Reh TA. Acheate-scute like 1 (Ascl1) is required for normal delta-like (Dll) gene expression and notch signaling during retinal development. *Dev Dyn* **238**, 2163-2178 (2009).
+42. Mao CA, Kiyama T, Pan P, Furuta Y, Hadjantonakis AK, Klein WH. Eomesodermin, a target gene of Pou4f2, is required for retinal ganglion cell and optic nerve development in the mouse. *Development* **135**, 271-280 (2008).
+43. Poche RA, Kwan KM, Raven MA, Furuta Y, Reese BE, Behringer RR. Lim1 is essential for the correct Laminar positioning of retinal horizontal cells. *J Neurosci* **27**, 14099-14107 (2007).
+44. Nelson BR, Gumuscu B, Hartman BH, Reh TA. Notch activity is downregulated just prior to retinal ganglion cell differentiation. *Dev Neurosci-Basel* **28**, 128-141 (2006).
+45. Jusuf PR, Harris WA. Ptf1a is expressed transiently in all types of amacrine cells in the embryonic zebrafish retina. *Neural Development* **4**, (2009).
+46. Brzezinski JA, Lamba DA, Reh TA. Blimp1 controls photoreceptor versus bipolar cell fate choice during retinal development. *Development* **137**, 619-629 (2010).
+47. Morgans CW, et al. TRPM1 is required for the depolarizing light response in retinal ON-bipolar cells. *P Natl Acad Sci USA* **106**, 19174-19178 (2009).
+48. Lee M, Goraya N, Kim S, Cho SH. Hippo-yap signaling in ocular development and disease. *Dev Dynam* **247**, 794-806 (2018).
+49. Kim JY, et al. Yap is essential for retinal progenitor cell cycle progression and RPE cell fate acquisition in the developing mouse eye. *Developmental Biology* **419**, 336-347 (2016).
+50. Sun Y, et al. Pharmacological blockade of TEAD-YAP reveals its therapeutic limitation in cancer cells. *Nat Commun* **13**, (2022).
+51. Collin J, et al. A single cell atlas of human cornea that defines its development, limbal progenitor cells and their interactions with the immune cells. *Ocul Surf* **21**, 279-298 (2021).
+52. Gautam P, et al. Multi-species single-cell transcriptomic analysis of ocular compartment regulons. *Nat Commun* **12**, 5675 (2021).
+53. Lukowski SW, et al. A single-cell transcriptome atlas of the adult human retina. *EMBO J* **38**, e100811 (2019).
+54. Collin J, et al. Single cell RNA sequencing reveals transcriptional changes of human choroidal and retinal pigment epithelium cells during fetal development, in healthy adult and intermediate age-related macular degeneration. *Hum Mol Genet*, (2023).
+55. Voigt AP, et al. Single-cell transcriptomics of the human retinal pigment epithelium and choroid in health and macular degeneration. *Proc Natl Acad Sci U S A* **116**, 24100-24107 (2019).
+56. Bohnsack BL, et al. Development of Extraocular Muscles Requires Early Signals From Periocular Neural Crest and the Developing Eye. *Arch Ophthalmol-Chic* **129**, 1030-1041 (2011).
+57. Grocott T, Johnson S, Bailey AP, Streit A. Neural crest cells organize the eye via TGF-beta and canonical Wnt signalling. *Nat Commun* **2**, 265 (2011).
+58. Harris WA, Perron M. Molecular recapitulation: the growth of the vertebrate retina. *Int J Dev Biol* **42**, 299-304 (1998).
+59. Kuwahara A, Ozone C, Nakano T, Saito K, Eiraku M, Sasai Y. Generation of a ciliary margin-like stem cell niche from self-organizing human retinal tissue. *Nat Commun* **6**, (2015).
+60. Cid E, et al. Prox1 expression in rod precursors and Muller cells. *Exp Eye Res* **90**, 267-276 (2010).
+61. Rozanska A, et al. pRB-Depleted Pluripotent Stem Cell Retinal Organoids Recapitulate Cell State Transitions of Retinoblastoma Development and Suggest an Important Role for pRB in Retinal Cell Differentiation. *Stem Cells Transl Med* **11**, 415-433 (2022).
+62. Liu J, et al. A high-risk retinoblastoma subtype with stemness features, dedifferentiated cone states and neuronal/ganglion cell gene expression. *Nat Commun* **12**, 5578 (2021).
+63. Asaoka Y, Hata S, Namae M, Furutani-Seiki M, Nishina H. The Hippo Pathway Controls a Switch between Retinal Progenitor Cell Proliferation and Photoreceptor Cell Differentiation in Zebrafish. *Plos One* **9**, (2014).
+64. Kastan N, et al. Small-molecule inhibition of Lats kinases may promote Yap-dependent proliferation in postmitotic mammalian tissues. *Nat Commun* **12**, (2021).
+65. Kastan NR, et al. Development of an improved inhibitor of Lats kinases to promote regeneration of mammalian organs. *P Natl Acad Sci USA* **119**, (2022).
+66. Melguizo-Sanchis D, et al. iPSC modeling of severe aplastic anemia reveals impaired differentiation and telomere shortening in blood progenitors. *Cell Death Dis* **9**, (2018).
+67. Dorgau B, et al. Laminin gamma3 plays an important role in retinal lamination, photoreceptor organisation and ganglion cell differentiation. *Cell Death Dis* **9**, 615 (2018).
+
+# Supplementary Files
+
+- [TableS1.xlsx](https://assets-eu.researchsquare.com/files/rs-3160527/v1/c6ecb269230124530fb09a05.xlsx)  
+  Table S1: Single cell RNA-Seq analysis of whole human eyes or retina from 7.5-21 PCW. Highly expressed markers for each cell type are shown on separate spreadsheets for each sample together with HDBR sample accession number.
+
+- [TableS2.xlsx](https://assets-eu.researchsquare.com/files/rs-3160527/v1/41bfc91c47dbc1ac58537d57.xlsx)  
+  Table S2: Single cell RNA-Seq analysis of integrated data set of retinal cells from 7.5- 21 PCW. Highly expressed markers for each cell type in the integrated analysis are shown in the first spreadsheet. Highly expressed markers characterising lineage transitions from RPCs to T1, T2, T3 are shown in the second spreadsheet, those characterising lineage transitions from RPCs to T1, T2 and RGCs, amacrine and horizontal cells are shown in the second and third spreadsheet, and those characterising lineage transitions from RPCs to T1, T2 and amacrine and horizontal cells are shown in the fourth spreadsheet.
+
+- [TableS3.xlsx](https://assets-eu.researchsquare.com/files/rs-3160527/v1/87f61de74f4486a48797d731.xlsx)  
+  Table S3: Spatial transcriptomic analysis of 8, 10, 11 and 13 PCW human eyes. Highly expressed markers for each cell type are shown on separate spreadsheets for each sample together with HDBR sample accession number.
+
+- [TableS4.xlsx](https://assets-eu.researchsquare.com/files/rs-3160527/v1/b08e817aa9cb2af22d6e89a6.xlsx)  
+  Table S4: Single cell ATAC-Seq analysis of whole eyes or retina samples from 8- 21 PCW. Gene activity estimates were generated using Signac. Genes with high activity scores for each cell type are shown on separate spreadsheets for each sample. The final sheet shows the gene activity scores for the integrated datasets. The retina cell type from each sample were integrated using harmony and the data was then re-clustered to produce these results.
+
+- [TableS5.xlsx](https://assets-eu.researchsquare.com/files/rs-3160527/v1/324fcb55fe7cb88d87726e24.xlsx)  
+  Table S5: Differential accessibility peaks for each retinal cell type. Peaks were linked to genes using Cellranger and classified as either promoter, distal or intergenic. The distance from the gene is shown for each peak.
+
+- [TableS6.xlsx](https://assets-eu.researchsquare.com/files/rs-3160527/v1/d9c066d5d4ff4c44c8917916.xlsx)  
+  Table S6: Predicted transcription factor binding for each cell type of the developing human retina. ChromVAR was used to characterise transcription factor motifs for each cell type.
+
+- [TableS7.xlsx](https://assets-eu.researchsquare.com/files/rs-3160527/v1/210a6a4572cbbe60da886032.xlsx)  
+  Table S7: List of significant regulators of gene expression in RPCs, transient neurogenic progenitors and retinal neurons.
+
+- [SupplementcombinedNatCom.pdf](https://assets-eu.researchsquare.com/files/rs-3160527/v1/467d61dbdac62c93b6e11085.pdf)