Single-cell analysis reveals cellular heterogeneity and limits of marker-based assessment in retinal ganglion cell-enriched organoid cultures

13 Human pluripotent stem cell (hPSC)-derived retinal organoids provide an in vitro system for 14 generating retinal ganglion cells (RGCs), yet the cellular composition and developmental 15 fidelity of RGC-enriched cultures remain insufficiently characterised. Here, we tested an 16 RGC-enriched approach involving dissociation of hPSC-derived retinal organoids at day 40, 17 corresponding to peak expression of RGC markers, followed by two-dimensional culture 18 conditions intended to enrich for RGC survival. Flow cytometry was used to assess the 19 expression of RGC markers, including POU4F, ISL1, SNCG, and THY1. Across four 20 samples, POU4F expression ranged from 79-95%, ISL1 from 18-58%, SNCG from 22%-91% 21 and THY1 from 3%-29%, indicating substantial variability between markers and samples. 22 Single-cell RNA sequencing analysis of 73,642 cells identified multiple retinal lineages, 23 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 2 including retinal progenitors, RGCs, photoreceptor-committed cells, amacrine and horizontal 24 cells, and retinal pigment epithelium (RPE), as well as off-target populations comprising 25 HOX-enriched posterior neural cells and other cell types. Cellular composition varied across 26 samples. Transcriptomically defined RGCs accounted for 19-45% of cells across samples, 27 with different subtypes identified. These findings indicate that marker-based assessments 28 alone may overestimate RGC identity and provide a detailed single-cell characterisation of 29 cellular heterogeneity in RGC-enriched retinal organoid cultures. 30 31

32 Retinal ganglion cells, single-cell RNA sequencing, differentiation, retinal organoids, hPSC 33

34 The human retina offers a unique and accessible window into the organisation of the central 35 nervous system. Retinal ganglion cells (RGCs), the sole projection neurons of the retina, 36 convey visual information to central targets via the optic nerve. Loss of RGCs is a defining 37 pathological feature of optic neuropathies such as glaucoma, which remains a leading cause 38 of irreversible blindness worldwide [1]. Current therapies slow disease progression by 39 lowering intraocular pressure but cannot restore RGCs once degenerated [2] , highlighting a 40 critical need for physiologically relevant human models capable of recapitulating RGC 41 development, maturation, and vulnerability to support disease modelling, drug screening, and 42 potential cell-based replacement therapies. However, achieving this goal requires reliable 43 access to human RGCs, which are difficult to obtain and maintain in vitro. Indeed, RGCs 44 constitute less than 1% of all retinal cells in the adult eye [3,4], making the prospect of 45 isolating large numbers of human RGCs from donor tissue challenging. Human pluripotent 46 stem cells (hPSCs) including both embryonic stem cells (hESCs) and induced pluripotent 47 stem cells (hiPSCs), represent a promising alternative source for generating human RGCs. 48 For RGCs, two challenges remain: (i) generating sufficiently pure and high yield of RGCs, 49 and (ii) maintaining them in culture in a manner that preserves their developmental 50 trajectories and diverse subtypes. Previous studies identified 40-46 RGC subtypes in the 51 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 3 mouse retina [5,6] and 18 subtypes in primates [7], although the exact number in humans 52 remains unknown. This diversity is biologically important: RGC subtypes differ in their 53 morphology, function, and potentially disease vulnerability [8]. Hence, heterogeneity must be 54 captured to allow the reliable study of disease mechanisms and therapeutic responses. 55 Many RGC differentiation protocols rely on using small molecules to induce two-56 dimensional (2D) differentiation with positive selection based on a limited set of surface 57 markers [9–12], yet such strategies may bias cultures toward particular molecular subtypes. 58 Additionally, these cultures lack the three-dimensional (3D) organisation and signalling 59 gradients that drive retinogenesis in vivo , hence miss the spatial cues required for 60 physiologically appropriate specification and maturation. Furthermore, flow-cytometric 61 marker-based enrichment is rarely complemented by transcriptome-level validation, leaving 62 uncertainties about subtype composition and cellular heterogeneity. Retinal organoids offer 63 an opportunity to overcome these limitations by providing an in vivo -like environment that 64 supports the emergence of laminated retinal architecture and intrinsic patterning cues that 65 shape early RGC identity. Further, retinal organoids recapitulate the developmental hierarchy 66 of retinogenesis, in which RGCs are the first neuronal population to emerge [13]. This 67 temporal advantage allows the generation of comparatively enriched RGC populations when 68 organoids are harvested at early stages. Building on this principle, we applied an RGC-69 enriched approach to examine RGC differentiation and overall cellular composition in hPSC-70 derived retinal organoids. Organoids were dissociated at day 40, corresponding to peak 71 expression of POU4F (BRN3) transcription factors [14], which are key regulators of RGC 72 differentiation, survival and function [15–17], and replated for an additional 14 days in 73 medium formulated to support RGC survival and maturation. We used flow cytometry and 74 single-cell RNA sequencing (scRNA-seq) to assess enrichment efficiency, sample-to-sample 75 variability, and the spectrum of retinal and off-target cell types present in the resulting 76 cultures. Our findings demonstrate that protein marker-based assessment alone can 77 overestimate RGC identity and highlight the necessity of single-cell transcriptomic validation 78 for accurate evaluation of hPSC-derived RGC differentiation. 79 Methodology 80 Ethics 81 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 4 This work was approved by the office of Research Ethics and Integrity of the University of 82 Melbourne (2026-32991-75214-7), as per the requirements of the NHMRC, in accordance 83 with the Declarations of Helsinki. 84 hPSC maintenance 85 HESC H9 (WiCell) and hiPSC WAB-0222 [18] lines were maintained on Matrigel-coated 86 plates (Corning, #354230) in mTeSR™ Plus medium (STEMCELL Technologies, #100-87 0276). Medium was changed every other day, and cells were passaged weekly for routine 88 maintenance. 89 Retinal organoid differentiation 90 Retinal organoids were generated following a previously published protocol [14], with some 91 modifications. On Day-1, hPSC cultures at 70% confluency were dissociated using ReLeSR 92 (STEMCELL Technologies, #100-0484), and 4,000-5,000 cells were seeded into each well of 93 a 96-well low adhesion U-bottom plate (Corning, #7007) in mTeSR™ Plus containing 20 µM 94 Y-27632 (STEMCELL Technologies, #72304). Plates were centrifuged at 300 × g for 3 95 minutes to promote aggregate formation. On Day 0, each well received 50 µL mTeSR™ Plus 96 and 50 µL Neural Induction Medium (NIM; DMEM/F12 [1:1; Thermo Fisher Scientific, 97 #11320033] supplemented with 1% N2 [Thermo Fisher Scientific, #17502001], 1% MEM 98 non-essential amino acids [Sigma, #M7145], 1× penicillin-streptomycin [Gibco, 99 #15140122], and 2 µg/mL heparin [STEMCELL Technologies, #7980]), supplemented 100 with 20 µM Y-27632. On Day 1, 40-45 aggregates were transferred directly into 10-101 centimeter polystyrene dishes (Corning, #430591) and fed with 6 mL fresh mTeSR™ Plus 102 and 6 mL NIM. Medium was refreshed with NIM on Days 2 and 3. On Day 6, cultures were 103 changed to NIM containing 50 ng/mL BMP4 (R&D systems, #314-BP-010). On Day 8, wells 104 of a 6-well plate were pre-coated with 500 µL fetal bovine serum (FBS; Cytiva HyClone, 105 #SH30084.03), and 20-30 aggregates were plated per well. On Day 9, half of the BMP4-106 containing medium was replaced with NIM to achieve a final concentration of 10% FBS, 107 followed by half-medium changes on Days 12 and 15. On Day 16, organoids were lifted and 108 transferred into 10-centimeter dishes containing Retinal Differentiation Medium (RDM; 109 DMEM/F12 [3:1], 2% B27 supplement [Life Technologies, #17504044], 1% MEM non-110 essential amino acids, and 1× penicillin-streptomycin) containing 1% FBS. RDM was 111 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 5 replaced on Days 18 and 20 using media supplemented with 3% and 5% FBS, respectively. 112 On Day 22, organoids were transitioned to Advanced RDM (ARDM; DMEM/F12 [3:1], 113 supplemented with 2% B27, 1% MEM non-essential amino acids, 1× penicillin-streptomycin, 114 1× GlutaMAX, 10% FBS, and 100 µM taurine). Half-medium changes were performed every 115 2-3 days until Day 40. 116 On Day 40, retinal organoids were dissociated using the Papain Dissociation Kit 117 (Worthington Biochemical Corporation, #LK003153). Following a 30-minute incubation at 118 37 °C, organoids were triturated 20 times with a 1-mL pipette to generate a suspension 119 enriched for small aggregates. The suspension was transferred to a 15-mL tube containing an 120 equal volume of 10 mg/mL ovomucoid protease inhibitor and centrifuged at 300 × g for 5 121 minutes. The resulting pellet was resuspended in 1 mL Neurobasal-based neuronal 122 differentiation medium (NDM; Neurobasal Plus [Thermo Fisher Scientific, #A3582901], 123 supplemented with 1% MEM non-essential amino acids, 1% GlutaMAX [Thermo Fisher 124 Scientific, #35050061], 1% 45% glucose [Merck , #G8769], 1× penicillin-streptomycin, 1× 125 B27, 1× N2, 1× CultureOne [Thermo Fisher Scientific, #A3320201], and 1× Normocin 126 [InvivoGen, #ant-nr-2]). All supplements were added fresh immediately before use. Viable 127 cells were counted using trypan blue exclusion and plated onto 12-well plates coated with 128 poly-D-lysine (2 µg/cm²; Sigma-Aldrich, #P0899-10MG) and laminin (1 µg/cm²; Sigma-129 Aldrich, #L2020-1MG) at a density of 200,000 cells/cm². Following dissociation, cells were 130 cultured in NDM containing 20 µM Y-27632, 10 ng/mL CNTF (PeproTech; #450-13-50UG), 131 40 ng/mL BDNF (PeproTech, #450-02-50UG), 10 µM forskolin (Biogem/Lonza, #6652995), 132 and 3 µM DAPT (Abcam, #AB120633), applied as a full medium change. Y-27632 was 133 added immediately after plating. On Day 41, a half-medium change with NDM containing 134 CNTF, BDNF, forskolin, and DAPT was performed, with medium added gently along the 135 well perimeter to avoid dislodging adherent cells. A full medium change was performed on 136 Day 43 to remove debris and eliminate residual Y-27632. From Days 44-47, half-medium 137 changes were performed every 2-3 days while maintaining CNTF, BDNF, forskolin, DAPT, 138 and CultureOne. From Days 48-54, medium changes continued as above, with forskolin 139 increased to 25 µM; Day 48 involved a full medium change, followed by half-medium 140 changes every 2-3 days. 141 142 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 6 Flow cytometry 143 hPSC-RGCs cultured for 14-16 days on PDL/laminin-coated 24-well plates were dissociated 144 using either TrypLE Express (Thermo Fisher Scientific; 12604-021, 10 minutes, 37 °C) or the 145 Papain Dissociation Kit (Worthington Biochemical Corporation; LK003153, 20 minutes, 37 146 °C), depending on cell numbers. The papain kit was more effective at preserving cell viability 147 in cultures with lower cell yields. Cells were collected by centrifugation at 340 × g for 5 148 minutes at 4 °C, and the resulting pellets were resuspended in DPBS- 2% BSA (FACS 149 buffer). Live/dead staining was performed using the Fixable Violet 405 dye kit (Thermo 150 Fisher Scientific, #L34964, 30 minutes, room temperature). Cells were then washed in the 151 FACS buffer, and centrifuged at 340 × g for 5 minutes at 4 °C. For extracellular staining, 152 cells were resuspended in FACS buffer and incubated with Anti-CD90 BV510 (BD Horizon, 153 #563070) for 45 minutes on ice. For intracellular staining, cells were first resuspended in 154 DPBS, fixed with 4% paraformaldehyde for 10 minutes at room temperature, and 155 permeabilized in 0.1% BSA in DPBS-0.05% Triton X-100-0.1% Tween-20 for 15 minutes at 156 room temperature. Cells were subsequently resuspended in 0.1% BSA- DPBS- 0.1% Tween-157 20 and incubated with the following antibodies for 45 minutes at room temperature: Anti-158 SNCG Alexa Fluor 488 (Santa Cruz, #sc-65979), Anti-ISL1 PE (BD Pharmingen, #562547), 159 Anti-GFAP Alexa Fluor 647 (BD Pharmingen, #560298), and Anti-POU4F Alexa Fluor 647 160 (Santa Cruz, catalog #sc-390780). Antibody selection was based on markers that had been 161 detected in hPSC-derived RGCs. The pan-RGC marker RBPMS was therefore not included 162 due to its variable and often low expression in these cultures [19]. Stained cells were analysed 163 using a CytoFLEX LX flow cytometer, and data were processed with FlowJo v10.10 164 software. A negative control using hPSC-derived RPE cells is shown in Fig. S1. 165 Single cell preparation of iPSC-RGCs 166 hPSC-RGCs cultured for 14-16 days on PDL/laminin-coated 24-well plates were dissociated 167 using the Papain Dissociation kit according to the manufacturer’s protocol. Briefly, cells were 168 incubated with papain/DNase solution at 37 °C for 20 minutes, gently triturated, and the 169 enzymatic reaction was quenched using the albumin-ovomucoid inhibitor solution. Following 170 dissociation, cells were centrifuged and washed in 1% BSA, then sequentially filtered 171 through 30 µm pre-separation filters (Miltenyi Biotec; 130-041-407) and kept on ice. Cell 172 viability and concentration were determined by Trypan Blue exclusion using a Countess 3 FL 173 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 7 Automated Cell Counter (Thermo Fisher; AMQAF2000). Pelleted cells were fixed for long-174 term storage using the GEM-X Flex Sample Preparation v2 Kit (10x Genomics; 1000781) 175 according to the manufacturer’s CG000782 protocol for GEM-X Flex Gene Expression. 176 Generation of single cell GEMs and sequencing libraries 177 Single-cell suspensions were processed by the Australian Genome Research Facility using 178 the 10x Genomics Chromium GEM-X Flex Gene Expression Human 4-plex assay according 179 to the manufacturer’s protocol. For each sample, 3 × 10 /i1 fixed cells were hybridised with 180 uniquely barcoded probe sets (BC001-BC004) for 20 hours at 42 °C, washed, and pooled at 181 equal concentrations. Pooled cells were loaded onto a Chromium X instrument with GEM-X 182 FX Chips, combining barcoded Gel Beads, master mix, and Partitioning Oil B to generate 183 single-cell Gel Beads-in-Emulsion (GEMs) targeting recovery of ~80, 000 cells. GEMs were 184 transferred to a thermal cycler to ligate the left-hand and right-hand probes that remained 185 hybridised to their target RNA, hybridise Gel Bead primers to the capture sequence of the 186 ligated probe pairs, and extend barcode sequences. Following emulsion breaking with 187 Recovery Reagent, the ligated and extended products were PCR-amplified, cleaned, and 188 indexed. Libraries were quality-assessed on a TapeStation D1000, quantified by qPCR, and 189 sequenced on an Illumina NovaSeq X Plus (10B flow cell; 150 bp paired-end + 10 % PhiX). 190 Mapping of reads to transcripts and cells 191 Base-call files were processed using Cell Ranger v7.1.0 (10x Genomics) configured for the 192 Chromium Single Cell 3 ′ v3.1 chemistry. Reads were aligned to the Homo sapiens reference 193 genome (GRCh38, Ensembl release 109). Cell Ranger performed default barcode and UMI 194 correction to generate unfiltered gene-by-cell count matrices. No library aggregation was 195 performed. 196 197 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 8 scRNA-seq processing, integration, and visualisation 198 Cell recovery and ambient RNA correction 199 Raw gene-barcode count matrices were processed to remove empty droplets and correct 200 ambient RNA contamination. Ambient RNA contamination was corrected using the SoupX 201 package [20]. For each sample, the ambient RNA profile was estimated from barcodes not 202 present in the cell-containing matrix (raw-only barcodes) with total UMI counts in the range 203 1-100 UMIs. Non-empty putative cells were pre-clustered in Seurat (v5) using LogNormalize 204 → FindVariableFeatures → ScaleData → Run principal component analysis (PCA; 30 PCs) 205 → FindNeighbours → FindClusters (Leiden, resolution = 0.4). These quick cluster labels 206 were supplied to SoupX, contamination fraction ( ρ ) was estimated using autoEstCont, and 207 corrected counts were generated with adjustCounts. 208 Post-processing and doublet detection 209 Based on the pre-filtering quality-control metrics (Fig. S2), Seurat objects were reconstructed 210 from SoupX-corrected counts and filtered using thresholds of nFeature_RNA 1000-8000 and 211 mitochondrial RNA percentage < 30%. Doublets were detected on raw counts using 212 scDblFinder (SingleCellExperiment backend; serial execution) [21], and only singlets were 213 retained. 214 Normalisation and Harmony Integration 215 Each sample was normalised independently using SCTransform v2 (glmGamPoi backend) 216 [22], regressing out mitochondrial transcript fraction. After merging samples, principal 217 component analysis was performed on SCT residual features, and sample-associated 218 technical variation was mitigated using Harmony. Harmony aligns transcriptionally similar 219 cell states across samples without assuming technical replicates or enforcing alignment of 220 non-overlapping populations. Harmony embeddings were used for UMAP visualisation and 221 construction of the shared nearest-neighbour graph (dims 1-30). Clustering was performed 222 using the Leiden algorithm across a range of resolutions (0.1-1.0) to assess cluster stability 223 and granularity. 224 Dimensionality reduction and clustering 225 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 9 Principal component analysis (PCA; 50 components) was performed on the Harmony-226 integrated object. To determine the number of biologically informative dimensions, we 227 primarily examined the Elbow plot, which revealed a clear inflection point at approximately 228 20-25 principal components ( Fig. S3A). This observation was supported by the cumulative 229 variance plot, which showed that the majority of structured variation was captured within the 230 first ~20-25 PCs, with progressively smaller gains thereafter ( Fig. S3B). To ensure retention 231 of potentially informative higher-order components while avoiding excessive noise, a 232 conservative cutoff of the first 30 PCs was selected for all subsequent neighbour graph 233 construction, UMAP embedding, and clustering. 234 Cluster stability assessment and resolution selection 235 UMAP embeddings were generated using Harmony-corrected embeddings (min.dist = 0.3, 236 spread = 1.0) with fixed random seeds to ensure reproducibility [23]. Graph-based clustering 237 was performed using the Leiden algorithm (algorithm = 4) across a range of resolutions from 238 0.1 to 1.0 in increments of 0.1 [24]. Cluster stability across resolutions was assessed by 239 calculating the Adjusted Rand Index (ARI) between clustering solutions at consecutive 240 resolutions. ARI values increased rapidly between low resolutions (0.2-0.4) and reached a 241 high, stable range from resolution 0.4 onwards, indicating robust preservation of cluster 242 structure across increasing granularity ( Fig. S4A). Consistent with this, cluster visualisation 243 demonstrated coherent cluster propagation with minimal fragmentation across intermediate 244 resolutions, forming a stable plateau between resolutions 0.4 and 0.5 ( Fig. S4B). Based on 245 the convergence of ARI stability and clustree topology, a resolution of 0.5 was selected for 246 all downstream analyses. 247 Differential gene expression and marker identification 248 Differential expression analysis was performed on the Harmony-integrated object using 249 Seurat v5. Prior to differential expression analysis, clusters containing fewer than 100 cells 250 were excluded to ensure sufficient statistical power. The object was prepared for DE using 251 PrepSCTFindMarkers(). Cluster-specific markers were identified using FindAllMarkers 252 (test.use = "wilcox", only.pos = TRUE, min.pct = 0.10, log2FC threshold = 0.25), comparing 253 each cluster against all other cells. Gene identifiers were standardised to HGNC symbols 254 using Ensembl v109 via biomaRt. 255 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 10 Cell type annotation and marker validation 256 Gene symbols were standardised to HGNC conventions using HGNChelper [25] and 257 biomaRt [26]. For each Seurat cluster, average gene expression (SCT assay) was z-scored 258 and visualised with ComplexHeatmap to confirm marker specificity and to guide manual 259 annotation. Clusters containing fewer than 100 cells were excluded. The remaining clusters 260 were assigned to retinal progenitor cells, RGCs, amacrine cells, horizontal cells, 261 photoreceptors, retinal pigment epithelium (RPE), Multilineage (stressed), Other ( HOX-262 enriched), or Other based on the expression of established retinal marker genes from human 263 datasets (Table 1) [27–29]. 264 265 Table 1: Established retinal marker genes from human datasets. 266 Cell Identity Marker genes Retinal progenitor cell SOX2, PAX6, VSX2, LHX2, MKI67, ATOH7, FABP7, HES6, DLL3, PTF1A RGC POU4F2 (BRN3B), ISL1/2, EOMES, RBPMS, SNCG, THY1, NEFL, NEFM, GAP43, SLC17A6, POU6F2, ELAVL3, ELAVL4 Amacrine cell GAD1, TFAP2A, PRDM13, CHAT, SLC6A9 Horizontal cell PROX1, LHX1, ONECUT1, ONECUT2, GJA10, RORB Photoreceptor OTX2, CRX, THRB, NRL, RHO, GNAT1, NR2E3, ARR3, GNAT2, PDE6H RPE RPE65, MITF, TTR, TYR, TYRP1, RDH5 Multilineage (stressed) FOS, JUN, ATF3, DDIT3, CDKN1A, HSPA1B Other (HOX-enriched) HOXB5-8 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 11 267 Projection of Multilineage (stressed) cells 268 To infer the lineage affiliation of “Multilineage (stressed)” cells, we applied Seurat’s label-269 transfer framework using non-stressed cells as the reference population. Cells in cluster 19, 270 defined based on Harmony-integrated clustering, were treated as the query set. Transfer 271 anchors were computed using PCA-based reduction on normalised expression profiles, and 272 lineage predictions for each stressed cell were generated using TransferData. 273 Trajectory inference, lineage reconstruction, and pseudotime estimation 274 The developmental root was defined by proliferative activity within retinal progenitor cell 275 clusters. Cell-cycle phase was assigned using Seurat’s CellCycleScoring with canonical S-276 phase and G2/M gene sets. Clusters 15 and 17 exhibited high proliferative activity; however, 277 cell-cycle phase analysis revealed distinct cycling states. Cluster 15 comprised a mixed 278 cycling population (50.9% S phase, 49.1% G2/M), whereas cluster 17 was almost exclusively 279 G2/M (98.9%), indicating a highly synchronised mitotic progenitor population. Cluster 17 280 was therefore selected as the trajectory root, as it represents the most actively cycling 281 progenitor state. In contrast, differentiated neuronal populations, including RGCs (61-95.8% 282 G1), interneurons (75.8-84.6% G1), and photoreceptor-committed cells (53.8-71.2% G1), 283 were predominantly in G1, consistent with post-mitotic states. 284 Cell differentiation trajectories were inferred using Slingshot (v2.16.0) applied to the 285 PCA embeddings derived from the Harmony-integrated object (R v4.5.1) [30]. Trajectory 286 inference was performed in a family-wise manner, with separate analyses conducted for 287 major retinal lineages without cross-lineage interference. For each lineage family, Slingshot 288 inferred smooth principal curves through the cluster topology, generating continuous 289 pseudotime values for all included cells. Pseudotime distributions for downstream 290 comparison and visualisation were summarised using ridge plots (ggridges). 291 RGC Subclustering and Subtype Annotation 292 Clusters 2, 3, 11, and 14 were identified as RGCs and extracted for downstream analysis. All 293 analyses were performed in R (v4.5.1) using Seurat (v5) together with harmony, mclust, 294 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 12 clustree, dplyr, and ggplot2. This RGC population was reprocessed independently. The RNA 295 assay was re-normalised using SCTransform, and PCA was performed on the resulting SCT 296 assay, computing 50 PCs. Dimensionality selection was guided by PCA elbow plots as well 297 as variance explained and cumulative variance plots ( Fig. S5). Based on these metrics, the 298 first 20 PCs were retained for downstream neighbour graph construction and dimensionality 299 reduction. A shared nearest-neighbour graph was constructed using the Harmony embeddings 300 (dimensions 1-20), and Leiden clustering was applied across a resolution grid ranging from 301 0.1 to 1.0 (step size = 0.1). Clustering stability was assessed by calculating the ARI between 302 adjacent resolutions using mclust and by visual inspection of cluster relationships using 303 clustree. Based on ARI profiles and clustree topology, a final clustering resolution of 0.3 was 304 selected (Fig. S6). 305 RGC subtype markers were selected based on previously reported human RGC 306 marker genes associated with major RGC subtypes (Table 2), including α /Parasol, Midget, 307 direction-selective ganglion cells (DSGC), orientation-selective ganglion cells (OSGC; J-308 RGC-like), large sparse RGCs, and intrinsically photosensitive RGCs (ipRGCs) [7,29,31–309 33]. 310 Table 2: Established marker genes for human RGC subtypes. 311 Subtypes Markers α /Parasol SPP1 CAV2 CHRNA2 POU6F2 FABP4 Midget TPBG TBR1 GUCY1A1 RBPMS2 DSGC DCX CDH6 FSTL4 OSGC (J-RGC) JAM2 Large sparse SATB2 ipRGC OPN4 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 13

312 Generation of RO and incorporation of RGC enrichment strategies 313 314 Figure 1. Generation of retinal organoids and subsequent RGC enrichment. (A) 315 Schematic diagram illustrating the generation of retinal organoids, and ( B) the subsequent 316 enrichment of RGCs following dissociation of retinal organoids at day 40 (D40). Created 317 with BioRender.com (C-F) Brightfield images showing retinal organoids at D40 (left panel) 318 and the corresponding RGC-enriched enriched cell populations at D54 (right panel) for ( C) 319 H9_RGC1, (D) H9_RGC2, (E) WAB_RGC2, (F) WAB_RGC1. Scale bar: 100 µm. 320 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 14 321 The RGC enrichment approach was developed by integrating elements from prior studies 322 (Fig. 1A,B). Retinal organoids were first generated using a modified version of a previously 323 established protocol [14], then dissociated and plated for further differentiation, a step that 324 has been shown to favour enrichment of neuronal populations expressing CD90 [10]. Retinal 325 organoids were dissociated at day 40, a developmental stage reported to exhibit peak POU4F 326 expression, a canonical RGC marker [14]. Finally, cultures were maintained under 327 conditions previously shown to support the survival and maturation of emerging RGC-like 328 cells [9](Fig. 1B). 329 Retinal organoids derived from two hPSC lines were produced in duplicate 330 independent cultures (H9_RGC1, H9_RGC2, WAB_RGC1, and WAB_RGC2). All samples 331 underwent the same differentiation procedure from the hPSC stage to retinal organoids 332 formation. Typically, lifted retinal organoids after day 16 acquire a smooth, rounded 333 morphology with a distinct bright halo along the periphery, indicating the establishment of 334 laminated retinal layers. This characteristic morphology was observed in H9_RGC1, 335 H9_RGC2, and WAB_RGC2 ( Fig. 1C-E, left panels). In contrast, WAB_RGC1 organoids 336 became flattened during the plating-down period, likely as a result of disruption to the 337 embryoid body surface during manual transfer. Despite the flattening, these organoids were 338 lifted and retained for downstream culture and analysis to assess whether such morphological 339 changes could serve as a potential morphological selection criterion in future protocols. Upon 340 lifting, WAB_RGC1 organoids developed irregular, clumped edges around a central dense 341 core, and lacked the characteristic bright peripheral halo, suggesting disrupted lamination 342 (Fig. 1F, left panel). In addition, regions of folded or thickened tissue were observed on the 343 surface of WAB_RGC1 organoids, indicating a failure to establish proper neuroepithelial 344 polarity ( Fig. 1F , left panel). We therefore included WAB_RGC1 as a suboptimal 345 differentiated control to assess whether altered organoid morphology would influence RGC 346 yield. 347 After 14 days of RGC enrichment following organoid dissociation, cells from 348 H9_RGC1, H9_RGC2, and WAB_RGC2 formed multiple well-defined neuronal clusters 349 interconnected by thin, long dendritic processes, typical of neuronal networks ( Fig. 1C-E , 350 right panels). By contrast, WAB_RGC1-derived cultures displayed multiple dense aggregates 351 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 15 containing several large soma-like centres with multiple extended multiple neurites ( Fig. 1F, 352 right panel; large soma-like centres indicated by arrowheads). 353 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 16 Flow cytometric assessment of RGC marker expression 354 355 356 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 17 Figure 2. Flow cytometry assessment of RGC-associated marker expression. (A) 357 Representative flow cytometry dot plots showing expression of POU4F, ISL1, SNCG, and 358 THY1 across four samples (H9_RGC1, H9_RGC2, WAB_RGC2, and WAB_RGC1). Rows 359 correspond to individual samples and columns to individual markers. Gates were defined 360 using unstained populations. Percentages denote the fraction of marker-positive cells within 361 the live singlet population. (B) Violin plots showing log-normalised mRNA expression of 362 canonical RGC markers POU4F family members ( POU4F1, POU4F2, POU4F3), ISL1, 363 SNCG, and THY1 across H9_RGC1, H9_RGC2, WAB_RGC2, WAB_RGC1. Each violin 364 represents the distribution of transcript abundance across individual cells. 365 366 To quantify the abundance of RGC-like cells within each enriched culture, flow cytometry 367 analysis was performed using four markers, which in combination are indicative of RGC 368 identity: POU4F (recognising POU4F family proteins), ISL1, SNCG, and THY1 ( Fig. 2A; 369 Fig. S7). All samples demonstrated high proportions of POU4F+ cells (79.0 to 95.1%), 370 consistent with efficient induction of sensory neurons/ RGC lineage identity. SNCG, a small 371 cytosolic protein highly enriched in RGCs [34], was used as an additional marker of RGC 372 differentiation. ISL1, a transcription factor expressed in RGCs and other inner retinal neurons 373 [35,36], showed greater variability across samples (18.0-58.0%), reflecting the known stage-374 dependent temporal dynamic expression of ISL1 during RGC differentiation [35]. SNCG 375 expression was similarly high in H9_RGC1, H9_RGC2 and WAB_RGC2 (81.3-91.0%), but 376 was markedly reduced in WAB_RGC1 (21.5%), consistent with its less organised neurite 377 architecture. THY1, a surface protein historically used as an RGC antigen in the retina [37], 378 displayed the greatest variability, ranging from 3.7% in WAB_RGC2 to 29.3% in H9_RGC1. 379 This variability aligns with the recognised instability of THY1 as a surface marker as it labels 380 only a subset of RGCs and is also detectable in non-RGC retinal cell types [38]. Together, 381 these data indicate the presence of RGC-associated marker expression across all samples, 382 with some variability among individual markers. 383 All RGC markers used for flow-cytometry validation were also detected at the 384 transcript level in the scRNA-seq dataset; however, the mRNA expression of these individual 385 markers varied substantially across all samples ( Fig. 2B; Fig. S8 ). POU4F1, POU4F2, and 386 POU4F3 transcripts were detected in only a small fraction of cells across all samples, with 387 the majority of cells exhibiting zero or near-zero expression. ISL1 mRNA was detected in a 388 small subset of cells primarily in H9_RGC1, while H9_RGC2, WAB_RGC2, and 389 W AB_RGC1 showed little to no detectable expression. SNCG transcripts were detected in 390 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 18 subsets of cells in H9_RGC1, H9_RGC2, and WAB_RGC2, with a distinct population 391 exhibiting elevated expression, whereas SNCG expression was minimal in WAB_RGC1, 392 where most cells showed near-zero transcript levels. THY1 transcripts were present in subsets 393 of cells in H9_RGC1 and H9_RGC2 but were largely absent in WAB_RGC samples, with 394 only rare cells showing detectable expression. These findings indicate that reliance on flow 395 cytometry, through detection of stable or residual protein, may overestimate RGC abundance 396 relative to scRNA-seq. 397 398 399 400 401 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 19 402 403 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 20 Figure 3. Single-cell transcriptomic profiling of RGC-enriched dissociated retinal 404 organoid cells. (A) UMAP visualisation showing the distribution of the 22 transcriptionally 405 defined clusters identified in the integrated dataset. (B) Heatmap displaying z-scored average 406 log-normalised RNA expression of different marker genes across clusters. Rows correspond 407 to marker genes grouped by major retinal cell classes, and columns represent the 22 clusters. 408 Left annotation indicates the retinal cell type associated with each marker set, and the top 409 annotation reflects the assigned cell-type identity for each cluster. ( C) Stacked bar charts 410 showing the proportional composition of annotated retinal cell types in each sample. ( D) 411 Ridge plots illustrating pseudotime distributions for the major developmental lineages. Each 412 ridge reflects the relative density of cells and the progression of cells along pseudotime 413 within a given lineage. ( E) Lineage tree reconstructed from pseudotime analysis, showing 414 hierarchical branching of retinal progenitor populations into downstream neuronal and 415 epithelial fates. Each node is labelled with its cluster identifier, with bracketed numbers 416 indicating the corresponding cluster IDs. RPC: retinal progenitor cell, RGC: retinal ganglion 417 cell, AC: amacrine cell, HC: horizontal cell, PR: photoreceptor, RPE: retinal pigment 418 epithelium. 419 Identification and characterisation of 22 subpopulations 420 Across the four samples, a total of 73,642 high-quality single cells were retained for 421 downstream integrated analysis (Table S1 ). At a clustering resolution of 0.5, a total of 22 422 transcriptionally distinct clusters were identified ( Fig. 3A). However, cluster 22 contained 423 fewer than 100 cells (n = 91) and was therefore excluded from downstream analyses. Based 424 on marker expression, clusters were annotated into different cell types ( Fig. 3B ). Sample-425 level comparisons revealed notable differences in differentiation bias across the four datasets 426 (Fig. 3C). In H9_RGC1, RGCs constituted the largest population (45%), followed by 427 amacrine cells (15%), retinal progenitor cells (14%), HOX-enriched cells (13%), and 428 horizontal cells (7%). In H9_RGC2, HOX-enriched cells were the most abundant (30%), with 429 RGCs and amacrine cells each accounting for 19%, followed by retinal progenitor cells 430 (11%), and photoreceptor-committed cells (10%). WAB_RGC2 displayed 38% of cells 431 classified as photoreceptor-committed cells and 27% as RGCs, alongside horizontal cells 432 (12%), other minor populations (8%), retinal progenitor cells (7%), and amacrine cells (6%). 433 WAB_RGC1 exhibited the highest proportion of retinal progenitor cells (47%), with RGCs 434 (10%), amacrine cells (10%), RPE (9%), photoreceptor-committed cells (7%), HOX-enriched 435 cluster (5%), and other populations (5%) (Table S2). 436 Six retinal progenitor cells clusters (4, 10, 15, 17, 18, 21) were annotated based on the 437 expression of canonical retinal progenitor cell markers such as VSX2, along with key 438 transcription factors regulating neuronal fate, including PAX6 and SOX2 [39–41]. They also 439 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 21 expressed LHX2, which is essential for maintaining open chromatin during retinogenesis and 440 for gliogenesis [42,43]. Cell cycle-related genes were unevenly distributed among progenitor 441 subpopulations, with the G2/M phase marker MKI67 showing predominant expression in 442 clusters 15 and 17. 443 RGC clusters were annotated based on the expression of canonical markers including 444 POU4F2, ISL1, ISL2, EOMES, SNCG, THY1, NEFL, NEFM, and RBPMS [5,44–47]. Using 445 these markers, four transcriptionally distinct RGC subpopulations were identified ( Fig. 3B), 446 All of these clusters (2, 3, 11, and 14) expressed POU4F2 . Cluster 3 represented the earliest 447 RGC-fated state, characterised by high expression of ATOH7, HES6, and DLL3. Clusters 2 448 and 14 showed strong co-expression of POU4F2, ISL1, RBPMS, NEFL, NEFM, SNCG, 449 THY1, SLC17A6, and GAP43, indicative of maturing RGCs. Clusters 11 and 14 exhibited 450 partial expression of horizontal cell-associated markers such as ONECUT1 and ONECUT2 , 451 indicating transcriptional heterogeneity within the enriched RGC population. 452 Two amacrine cell clusters (5 and 7) were enriched with an inhibitory neuron marker 453 GAD1 [27], while a horizontal cell cluster (6) expressed PROX1, ONECUT1, ONECUT2 and 454 GJA10 [48–50]. Photoreceptor-committed cells identity was defined using established lineage 455 mark ers, including OTX2, CRX, THRB, and NRL, together with rod-specific markers ( RHO, 456 GNAT1, NR2E3) and cone-specific markers ( ARR3, GNAT2, PDE6H) [51–55]. Four clusters 457 (8, 9, 13 and 20) showed strong expression of OTX2 and CRX , confirming photoreceptor 458 lineage commitment. Among these, cluster 9 displayed the most advanced phototransduction 459 programme, co-expressing cone-specific markers ( ARR3, GNAT2, PDE6H) together with 460 rod-associated genes ( NRL, PDE6B, GNAT1), indicating a population of mixed rod-cone 461 photoreceptor precursors. Additionally, cluster 16 showed the expression of RPE 65, MITF, 462 TYR, TYR and RDH5, classical markers of RPE [56]. 463 Cluster 19 exhibited a stress-associated transcriptional profile characterised by high 464 expression of FOS, JUN, ATF3, DDIT3, CDKN1A, and HSPA1B [57–61]. This cluster was 465 annotated as a Multilineage (stressed) population, and it was less than 5% of the total cells 466 across all samples respectively. Differential expression analysis revealed induction of genes 467 associated with apoptosis initiation, cellular stress signalling and early cell-death pathways, 468 although no coherent set of canonical lineage markers was detected. To determine the 469 underlying lineage composition masked by stress-induced transcriptional reprogramming, 470 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 22 label-transfer analysis was performed (Table 4 ). This analysis revealed that cluster 19 471 comprised mainly of retinal progenitor cells, followed by “Other” cell types, RPE, HOX-472 enriched cells, and amacrine cells (Table 4). 473 Table 4: Lineage composition of the Multilineage (stressed) cluster identified by label-474 transfer analysis. 475 Family # of cells Percentage of cells (%) Retinal progenitor cell 483 35.3 RGC 43 3.1 Amacrine cell 125 9.1 Horizontal cell 14 1.0 Photoreceptor-committed cell 28 2.0 RPE 144 10.5 Other (HOX-enriched) 136 9.9 Other 397 29.0 476 Cluster 1 was annotated as HOX-gene-enriched cells, marked by high expression of 477 HOX-related genes such as HOXB5-8. These cells also expressed neuronal structural genes 478 such as NEFL, NEFM and ELAVL4 but lacked retinal identity markers, indicating the 479 formation of off-target posterior neural cell types not belonging to anterior retinal tissue. 480 HOX-enriched populations were most abundant in the two H9-derived samples. In addition, 481 cluster 12 was annotated as “Other” as it did not exhibit retinal cell markers. 482 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 23 Pseudotime analysis reveals progressive retinal lineage trajectories 483 To reconstruct differentiation pathways across retinal lineages, we performed pseudotime and 484 lineage trajectory analysis on the Harmony-integrated dataset comprising all four samples. 485 Clusters annotated as “Multilineage (stressed)” and “Other” were excluded as they exhibited 486 mixed lineage marker expression and/or off-target transcriptional programs, rather than a 487 single coherent retinal identity. Inclusion of these cells could distort transcriptional similarity 488 relationships and lead to artificial trajectory splits that do not reflect genuine developmental 489 lineage decisions. Cells were ordered along developmental trajectories based on 490 transcriptional similarity. Ten major lineages were identified, corresponding to retinal 491 progenitor, RGC, amacrine, horizontal, photoreceptor, and RPE fates, each branching from 492 the retinal progenitor cell cluster 17, which showed the highest expression of the proliferation 493 marker MKI67 (Fig. 3D). The earliest pseudotime positions were occupied by the RPC1-494 RPC3 populations, which together formed the central developmental trunk. RPC4 495 predominantly gave rise to two RGC lineages and to RPC5, which forms the root of all other 496 major retinal cell lineages, including an additional retinal progenitor cell lineage, an amacrine 497 cell lineage, one horizontal lineage, two photoreceptor lineages, and one RPE lineage ( Fig. 498 3E). 499 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 24 500 Figure 4. Subclustering and molecular characterisation of RGC subpopulations. ( A) 501 UMAP showing the clusters (in purple) selected for subsequent reclustering of the RGC 502 subpopulation. ( B) UMAP showing the seven distinct clusters identified. ( C) Heatmap 503 showing z-scored average log-normalised RNA expression of RGC subtypes markers. Rows 504 represent marker genes grouped by major RGC subtypes, and columns represent the seven 505 clusters. The left annotation indicates the subtype associated with each marker set. (D ) 506 UMAP visualisation of RGC subtype marker expression using log-normalised RNA 507 expression values. For each marker, cells with detectable expression (log-normalised 508 expression > 0) are highlighted in purple, while non-expressing cells are shown in grey. Each 509 panel represents a single marker gene, with gene symbols displayed at the top of each plot. 510 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 25 511 Transcriptional characterisation of RGC subtypes 512 To further resolve the molecular diversity within the RGC lineage, we re-clustered cells from 513 RGC-associated clusters (2, 3, 11 and 14) at a resolution of 0.3, resulting in seven 514 transcriptionally distinct RGC subclusters ( Fig. 4A , B). Canonical human RGC markers 515 [7,29,31–33] (Table 2) were detected across these subclusters, including markers for parasol, 516 midget, DSGC, OSGC, and large sparse RGCs (Fig. 4C ), suggesting that the cultures 517 generate multiple RGC subtypes. Most subclusters showed overlapping expression of several 518 RGC subtype marker genes, suggesting incomplete transcriptional segregation. Although 519 clusters 1 and 4 showed elevated OPN4 signal (Fig. 4C), it was detected in only 49 cells for 520 cluster 1, and 11 cells in cluster 4 in the log-normalised RNA matrix. As this proportion fell 521 below the minimum detection threshold (≥ 5% of cells expressing the marker), OPN4 was not 522 identified in the differential expression list (Fig. 4C, D). For each selected marker gene, cells 523 with normalised expression greater than zero were considered positive, and Harmony-derived 524 UMAP embeddings were used to visualise the spatial distribution of marker-positive cells 525 (Fig. 4D). 526 Given that OPN4 marks a rare subtype of RGC, we assessed whether additional rare 527 melanopsin-expressing cells were missed due to normalisation and subsequent subclustering. 528 Examination of the unnormalised, SoupX-adjusted RNA count matrix, where low-abundance 529 transcripts remain easier to detect than in SCTransform-normalised data [62], identified 128 530 OPN4+ cells, comprising 120 cells with low transcript abundance and 8 cells with higher 531 abundance (Fig. S9 and Table S3 ). Mapping these OPN4+ cells onto the Harmony-derived 532 UMAP embedding revealed that approximately half were located within the horizontal 533 cluster (cluster 6 of the full integrated RGC-enriched retinal organoid dataset; Fig. 3), while 534 51 cells were located within the RGC cluster (cluster 2; Fig. 3). 535

536 This work provides an in-depth analysis of RGC differentiation within hPSC-derived retinal 537 organoid cultures using complementary flow cytometry and single-cell transcriptomic 538 approaches. Discrepancies between protein- and transcript-based readouts highlight the 539

of marker-based assessments and underscore the importance of transcriptomic 540 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 26 benchmarks for defining RGC identity in heterogeneous cultures. Across samples, protein-541 based assays suggested robust RGC enrichment, whereas transcriptomic profiling revealed a 542 smaller RGC population alongside significant contributions from non-RGC neuronal 543 populations. For instance, over 80% of cells were deemed POU4F+ and SNCG+ in 544 H9_RGC2, yet scRNA-seq identified more than one-third of cells belonging to HOX-545 enriched cell clusters. These HOX-enriched clusters may represent posterior CNS-like 546 neurons, such as spinal interneurons or hindbrain sensory populations, which are known to 547 transiently express POU4F family proteins during development [63]. Similarly, WAB_RGC2 548 showed high proportions of POU4F+ cells and SNCG+ cells by flow cytometry, whilst 549 scRNA-seq showed that the major cell types within this differentiation were from the 550 photoreceptor lineage. These can also transiently express SNCG [64], which is widely 551 expressed across multiple peripheral and central nervous system neuronal populations [65]. 552 ISL1 expression further illustrates the limitations of protein-based classification, as it 553 varies with cellular maturation and is not restricted to RGCs, being detectable in other 554 neuronal populations [66]. This likely contributes to the variable proportion of ISL1+ cells 555 observed across samples. In contrast, THY1+ cells consistently represented the smallest 556 fraction, reflecting the dynamic nature of THY1 expression, its downregulation in stressed 557 RGCs [67], and its presence in subsets of amacrine cells [68]. In H9_RGC1 and 558 WAB_RGC2, scRNA-seq identified a greater proportion of RGCs than suggested by flow 559 cytometry, highlighting the limited reliability of THY1 as a commonly used standalone 560 marker for estimating RGC content [10]. These findings indicate that RGC-associated 561 proteins can be detected outside the canonical RGC lineage, whereas scRNA-seq more 562 accurately excludes non-RGC cells. This discrepancy reflects the broad and often transient 563 expression of commonly used RGC markers during early neurogenesis, together with their 564 persistence in off-target neuronal lineages. As a result, protein marker positivity alone can 565 overestimate RGC identity when lineage resolution is incomplete. Future flow cytometry 566 approaches could improve specificity by combining POU4F and SNCG with exclusion 567 markers for photoreceptors and/or GAD1 for amac rine cells, thereby ensuring that marker-568 positive populations more faithfully represent RGCs. 569 Methodological differences between flow cytometry and scRNA-seq further 570 contribute to these contrasting readouts. Flow cytometry enriches larger and healthier 571 dissociated cells through gating and is highly sensitive to stable or residual protein, 572 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 27 potentially enriching early neurons that transiently express RGC-associated markers, as 573 discussed above. In contrast, scRNA-seq captures a broader and less selectively filtered 574 population, including fragile or transcriptionally immature cells, and classifies identity based 575 on integrated gene expression programs rather than individual markers. Although transcript 576 dropout is an inherent limitation of scRNA-seq with typical protocols capturing only ~10-577 20% of transcripts [69–71], reliance on multi-gene signatures reduces the impact of false 578 negatives [72] and provides a more conservative and lineage-resolved definition of RGC 579 identity. 580 Single-cell analysis resolved multiple RGC transcriptional states, alongside relatively 581 higher proportions of retinal progenitor cells, off-target posterior neural populations, and 582 photoreceptor-committed cells. Rather than forming a uniform population, RGCs in enriched 583 cultures occupied a continuum of transcriptional states. Pseudotime reconstruction supported 584 the coexistence of early, intermediate, and maturing RGC populations. Multiple RGC 585 subtypes were also identified. Rare subtypes, including melanopsin-expressing RGCs, were 586 detected at low abundance and were localised to specific transcriptional clusters, consistent 587 with genuine low-level expression. However, these populations were particularly sensitive to 588 analytical thresholds and normalisation strategies, underscoring the limitations of single-cell 589 approaches for resolving rare neuronal populations and highlighting the importance of 590 cautious interpretation when assessing low-abundance transcripts in developing systems. 591 In the presence of multiple developmental states within the culture, the observed 592 overlap between RGC transcriptional programs and markers associated with other early-born 593 retinal neurons can be understood in the context of retinal development. RGCs, amacrine and 594 horizontal cells arise from shared ATOH7 + progenitors, and fate commitment during early 595 retinogenesis occurs gradually rather than through abrupt transitions [73]. Transitional cells 596 may therefore transiently activate transcriptional programs associated with multiple retina 597 lineages before terminal differentiation. Consistent with this, developing RGCs or RGC-like 598 neurons have been found to express markers associated with amacine and horizontal cells 599 [74,75]. Hence, when interpreting the scRNAseq analysis, low-level or partial expression of 600 lineage-associated markers should not be interpreted as definitive evidence of fate switching, 601 but rather as a reflection of developmental immaturity or incomplete fate resolution. 602 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 28 This developmental heterogeneity complicates marker-based assessments of RGC 603 enrichment. As enriched cultures contain early RGCs, maturing neurons, and diverse subtype 604 identities, no single surface or intracellular marker can reliably distinguish true RGCs from 605 transient intermediates or off-target neurons with overlapping antigen expression. This 606

is particularly relevant for THY1-based enrichment strategies, which are 607 commonly used for immunopanning and positive selection [9–12]. In our study, THY1 608 expression was variable and lacked specificity, being detected across multiple retinal and 609 non-retinal neuronal populations. These findings suggest that THY1-based approaches are 610 likely to isolate a heterogeneous mixture rather than a well-defined RGC population and that 611 more stringent multi-marker or transcriptomically informed selection strategies will be 612 necessary to obtain high-fidelity RGC enrichment. However, identifying surface markers that 613 are both specific and stable across RGC subtypes remains an important and unresolved 614 challenge. 615 Another feature of our dataset is the presence of HOX-enriched cell populations, 616 particularly in hESC-derived organoids. From a developmental perspective, this is 617 unexpected, as the retina originates from the anterior neuroectoderm, whereas HOX genes 618 pattern posterior regions of the CNS [76,77] . However, our findings are consistent with 619 previous studies reporting upregulation of HOX gene programs in hPSC-derived retinal 620 organoids [18,78,79]. Although the abundance and composition of these clusters vary across 621 studies, the recurrent detection of posterior HOX -expressing cells suggests that off-target 622 neural identities represent a common feature of current retinal organoid differentiation 623 protocols. Modulating retinoid acid availability during early differentiation such as the use of 624 B27 supplement without vitamin A can shift neural identity toward anterior domains [80,81], 625 and could potentially reduce the proportion of HOX-enriched cells. 626 While this protocol yielded variable proportions of RGCs across samples, the 627 resulting RGC content (19-45%), excluding the poorly differentiated retinal organoids from 628 WAB_RGC1, is comparable to previously reported yields from hPSC-to-RGC differentiation 629 strategies validated by scRNA-seq. Earlier studies reported RGC proportions of around 12% 630 using a 2D differentiation protocol and 17% in RGC-enriched retinal organoid systems 631 [18,83], indicating that the overall efficiency observed here is above the range of 632 transcriptomically validated approaches. Nevertheless, sample-to-sample variability indicates 633 opportunities for further optimisation, as early exclusion of poorly patterned organoids and 634 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 29 refinement of dissociation timing may reduce retention of retinal progenitor and 635 photoreceptor-committed cells [14,84]. 636

637 A limitation of this study is that RGC subtype classification is based on transcriptional 638 profiles rather than functional properties. Classical RGC subtypes are defined by 639 morphology, connectivity and electrophysiological behaviour [33], which cannot be resolved 640 by scRNA-seq alone. As a result, transcriptionally defined clusters may not map directly onto 641 established functional RGC classes and may instead reflect developmental states, stress 642 responses or culture-induced divergence. Future studies integrating transcriptomics with 643 electrophysiology, connectivity mapping and functional assays will be required to directly 644 molecular identity with RGC function. 645 646 Data availability 647 All data have been deposited in the ArrayExpress database 648 (https://www.ebi.ac.uk/arrayexpress). 649 650 Supplementary Information 651 Nine supplementary figures, three supplementary tables, and an additional Excel file 652 containing differential gene expression results with an adjusted p-value < 0.05. 653 654 Acknowledgments 655 The authors acknowledge Dr. Magdaline Sakkas of the Melbourne Cytometry Platform, The 656 University of Melbourne, for her instruction on using the Beckman Coulter CytoFLEX LX 657 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 30 flow cytometer, and Dr. Peter Lau of the Australian Genome Research Facility for his 658 assistance in processing samples for scRNA-seq analysis. 659 660 Authors’ Contributions 661 Conceptualisation: all authors; Methodology: all authors; Investigation: J.Y.W.M.; Data 662 analysis: J.Y.W.M.; Writing original draft: J.Y.W.M., A.P.; Writing review & editing, all 663 authors; Funding Acquisition: all authors; Supervision and project administration: A.P. 664 665 Funding 666 This research was supported by funding from PYC therapeutics, a University of Melbourne 667 Department of Anatomy and Physiology Early Career Seeding Grants (JYWM) and a Dame 668 Kate Campbell Fellowship (AP). 669 670

& Correspondence. Requests for materials and correspondence should be 671 addressed to Jessica Ma and Alice Pébay. 672 673 Competing Interests 674 A.P. is a scientific advisor to PYC Therapeutics, and a director and shareholder of CellTellus 675 Pty Ltd. 676 677 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 31

678 1. Vision Loss Expert Group of the Global Burden of Disease Study, GBD 2019 Blindness 679 and Vision Impairment Collaborators. Global estimates on the number of people blind or 680 visually impaired by glaucoma: A meta-analysis from 2000 to 2020. Eye (Lond). 681 2024;38:2036–46. 682 2. Luo X, Salgueiro Y, Beckerman SR, Lemmon VP, Tsoulfas P, Park KK. Three-683 dimensional evaluation of retinal ganglion cell axon regeneration and pathfinding in whole 684 mouse tissue after injury. Exp Neurol. 2013;247:653–62. 685 3. Curcio CA, Allen KA. Topography of ganglion cells in human retina. J Comp Neurol. 686 1990;300:5–25. 687 4. Curcio CA, Millican CL, Allen KA, Kalina RE. Aging of the human photoreceptor mosaic: 688 evidence for selective vulnerability of rods in central retina. Invest Ophthalmol Vis Sci. 689 1993;34:3278–96. 690 5. Rheaume BA, Jereen A, Bolisetty M, Sajid MS, Yang Y, Renna K, et al. Single cell 691 transcriptome profiling of retinal ganglion cells identifies cellular subtypes. Nat Commun. 692 2018;9:2759. 693 6. Tran NM, Shekhar K, Whitney IE, Jacobi A, Benhar I, Hong G, et al. Single-Cell Profiles 694 of Retinal Ganglion Cells Differing in Resilience to Injury Reveal Neuroprotective Genes. 695 Neuron. 2019;104:1039–55.e12. 696 7. Peng Y-R, Shekhar K, Yan W, Herrmann D, Sappington A, Bryman GS, et al. Molecular 697 Classification and Comparative Taxonomics of Foveal and Peripheral Cells in Primate 698 Retina. Cell. 2019;176:1222–37.e22. 699 8. Morgan JE, Bevan RJ, Cimaglia G. Retinal Ganglion Cell Subtypes and Their 700 Vulnerability in Glaucoma. Methods Mol Biol. 2025;2858:191–205. 701 9. Chavali VRM, Haider N, Rathi S, Vrathasha V, Alapati T, He J, et al. Dual SMAD 702 inhibition and Wnt inhibition enable efficient and reproducible differentiations of induced 703 pluripotent stem cells into retinal ganglion cells. Sci Rep. 2020;10:11828. 704 10. Rabesandratana O, Chaffiol A, Mialot A, Slembrouck-Brec A, Joffrois C, Nanteau C, et 705 al. Generation of a Transplantable Population of Human iPSC-Derived Retinal Ganglion 706 Cells. Front Cell Dev Biol. 2020;8:585675. 707 11. VanderWall KB, Huang K-C, Pan Y, Lavekar SS, Fligor CM, Allsop AR, et al. Retinal 708 Ganglion Cells With a Glaucoma OPTN(E50K) Mutation Exhibit Neurodegenerative 709 Phenotypes when Derived from Three-Dimensional Retinal Organoids. Stem Cell Reports. 710 2020;15:52–66. 711 12. Edo A, Sugita S, Futatsugi Y, Sho J, Onishi A, Kiuchi Y, et al. Capacity of Retinal 712 Ganglion Cells Derived from Human Induced Pluripotent Stem Cells to Suppress T-Cells. Int 713 J Mol Sci [Internet]. 2020;21. Available from: http://dx.doi.org/10.3390/ijms21217831 714 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 32 13. Eiraku M, Takata N, Ishibashi H, Kawada M, Sakakura E, Okuda S, et al. Self-organizing 715 optic-cup morphogenesis in three-dimensional culture. Nature. 2011;472:51–6. 716 14. Harkin J, Peña KH, Gomes C, Hernandez M, Lavekar SS, So K, et al. A highly 717 reproducible and efficient method for retinal organoid differentiation from human pluripotent 718 stem cells. Proc Natl Acad Sci U S A. 2024;121:e2317285121. 719 15. Wagstaff EL, Ten Asbroek ALMA, Ten Brink JB, Jansonius NM, Bergen AAB. An 720 alternative approach to produce versatile retinal organoids with accelerated ganglion cell 721 development. Sci Rep. 2021;11:1101. 722 16. Mu X, Klein WH. A gene regulatory hierarchy for retinal ganglion cell specification and 723 differentiation. Semin Cell Dev Biol. 2004;15:115–23. 724 17. Sridhar A, Hoshino A, Finkbeiner CR, Chitsazan A, Dai L, Haugan AK, et al. Single-Cell 725 Transcriptomic Comparison of Human Fetal Retina, hPSC-Derived Retinal Organoids, and 726 Long-Term Retinal Cultures. Cell Rep. 2020;30:1644–59.e4. 727 18. Daniszewski M, Senabouth A, Liang HH, Han X, Lidgerwood GE, Hernández D, et al. 728 Retinal ganglion cell-specific genetic regulation in primary open-angle glaucoma. Cell 729 Genom. 2022;2:100142. 730 19. Gudiseva HV, Vrathasha V, He J, Bungatavula D, O’Brien JM, Chavali VRM. Single 731 Cell Sequencing of Induced Pluripotent Stem Cell Derived Retinal Ganglion Cells (iPSC-732 RGC) Reveals Distinct Molecular Signatures and RGC Subtypes. Genes (Basel) [Internet]. 733 2021;12. Available from: http://dx.doi.org/10.3390/genes12122015 734 20. Young MD, Behjati S. SoupX removes ambient RNA contamination from droplet-based 735 single-cell RNA sequencing data. Gigascience [Internet]. 2020;9. Available from: 736 http://dx.doi.org/10.1093/gigascience/giaa151 737 21. Germain P-L, Lun A, Garcia Meixide C, Macnair W, Robinson MD. Doublet 738 identification in single-cell sequencing data using. F1000Res. 2021;10:979. 739 22. Hafemeister C, Satija R. Normalization and variance stabilization of single-cell RNA-seq 740 data using regularized negative binomial regression. Genome Biol. 2019;20:296. 741 23. McInnes L, Healy J, Saul N, Großberger L. UMAP: Uniform Manifold Approximation 742 and Projection. J Open Source Softw. 2018;3:861. 743 24. Traag VA, Waltman L, van Eck NJ. From Louvain to Leiden: guaranteeing well-744 connected communities. Scientific Reports. 2019;9:1–12. 745 25. Oh S, Abdelnabi J, Al-Dulaimi R, Aggarwal A, Ramos M, Davis S, et al. HGNChelper: 746 identification and correction of invalid gene symbols for human and mouse. F1000Res. 747 2020;9:1493. 748 26. Durinck S, Spellman PT, Birney E, Huber W. Mapping identifiers for the integration of 749 genomic datasets with the R/Bioconductor package biomaRt. Nat Protoc. 2009;4:1184–91. 750 27. Wang SK, Nair S, Li R, Kraft K, Pampari A, Patel A, et al. Single-cell multiome of the 751 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 33 human retina and deep learning nominate causal variants in complex eye diseases. Cell 752 Genom [Internet]. 2022;2. Available from: http://dx.doi.org/10.1016/j.xgen.2022.100164 753 28. Xu Z, Liao X, Li N, Zhou H, Li H, Zhang Q, et al. A Single-Cell Transcriptome Atlas of 754 the Human Retinal Pigment Epithelium. Front Cell Dev Biol. 2021;9:802457. 755 29. Li J, Wang J, Ibarra IL, Cheng X, Luecken MD, Lu J, et al. Integrated multi-omics single 756 cell atlas of the human retina [Internet]. Res Sq. 2023. Available from: 757 http://dx.doi.org/10.21203/rs.3.rs-3471275/v1 758 30. Street K, Risso D, Fletcher RB, Das D, Ngai J, Yosef N, et al. Slingshot: cell lineage and 759 pseudotime inference for single-cell transcriptomics. BMC Genomics. 2018;19:477. 760 31. Langer KB, Ohlemacher SK, Phillips MJ, Fligor CM, Jiang P, Gamm DM, et al. Retinal 761 Ganglion Cell Diversity and Subtype Specification from Human Pluripotent Stem Cells. Stem 762 Cell Reports. 2018;10:1282–93. 763 32. Yan W, Peng Y-R, van Zyl T, Regev A, Shekhar K, Juric D, et al. Cell Atlas of The 764 Human Fovea and Peripheral Retina. Scientific Reports. 2020;10:1–17. 765 33. Kim US, Mahroo OA, Mollon JD, Yu-Wai-Man P. Retinal Ganglion Cells-Diversity of 766 Cell Types and Clinical Relevance. Front Neurol. 2021;12:661938. 767 34. Surgucheva I, Weisman AD, Goldberg JL, Shnyra A, Surguchov A. Gamma-synuclein as 768 a marker of retinal ganglion cells. Mol Vis. 2008;14:1540–8. 769 35. Pan L, Deng M, Xie X, Gan L. ISL1 and BRN3B co-regulate the differentiation of 770 murine retinal ganglion cells. Development. 2008;135:1981–90. 771 36. Wu F, Kaczynski TJ, Sethuramanujam S, Li R, Jain V, Slaughter M, et al. Two 772 transcription factors, Pou4f2 and Isl1, are sufficient to specify the retinal ganglion cell fate. 773 Proc Natl Acad Sci U S A. 2015;112:E1559–68. 774 37. Martersteck EM, Hirokawa KE, Evarts M, Bernard A, Duan X, Li Y, et al. Diverse 775 Central Projection Patterns of Retinal Ganglion Cells. Cell Rep. 2017;18:2058–72. 776 38. Rodriguez AR, de Sevilla Müller LP, Brecha NC. The RNA binding protein RBPMS is a 777 selective marker of ganglion cells in the mammalian retina. J Comp Neurol. 2014;522:1411–778 43. 779 39. Zou C, Levine EM. Vsx2 controls eye organogenesis and retinal progenitor identity via 780 homeodomain and non-homeodomain residues required for high affinity DNA binding. PLoS 781 Genet. 2012;8:e1002924. 782 40. Sun J, Zhao Y, McGreal R, Cohen-Tayar Y, Rockowitz S, Wilczek C, et al. Pax6 783 associates with H3K4-specific histone methyltransferases Mll1, Mll2, and Set1a and 784 regulates H3K4 methylation at promoters and enhancers. Epigenetics Chromatin. 2016;9:37. 785 41. Taranova OV, Magness ST, Fagan BM, Wu Y, Surzenko N, Hutton SR, et al. SOX2 is a 786 dose-dependent regulator of retinal neural progenitor competence. Genes Dev. 787 2006;20:1187–202. 788 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 34 42. de Melo J, Zibetti C, Clark BS, Hwang W, Miranda-Angulo AL, Qian J, et al. Lhx2 Is an 789 Essential Factor for Retinal Gliogenesis and Notch Signaling. J Neurosci. 2016;36:2391–405. 790 43. Zibetti C, Liu S, Wan J, Qian J, Blackshaw S. Epigenomic profiling of retinal progenitors 791 reveals LHX2 is required for developmental regulation of open chromatin. Commun Biol. 792 2019;2:142. 793 44. Clark BS, Stein-O’Brien GL, Shiau F, Cannon GH, Davis-Marcisak E, Sherman T, et al. 794 Single-Cell RNA-Seq Analysis of Retinal Development Identifies NFI Factors as Regulating 795 Mitotic Exit and Late-Born Cell Specification. Neuron. 2019;102:1111–26.e5. 796 45. Freeman NE, Templeton JP, Orr WE, Lu L, Williams RW, Geisert EE. Genetic networks 797 in the mouse retina: growth associated protein 43 and phosphatase tensin homolog network. 798 Mol Vis. 2011;17:1355–72. 799 46. Hu Y, Wang X, Hu B, Mao Y, Chen Y, Yan L, et al. Dissecting the transcriptome 800 landscape of the human fetal neural retina and retinal pigment epithelium by single-cell 801 RNA-seq analysis. PLoS Biol. 2019;17:e3000365. 802 47. Peirson SN, Oster H, Jones SL, Leitges M, Hankins MW, Foster RG. Microarray analysis 803 and functional genomics identify novel components of melanopsin signaling. Curr Biol. 804 2007;17:1363–72. 805 48. Dyer MA, Livesey FJ, Cepko CL, Oliver G. Prox1 function controls progenitor cell 806 proliferation and horizontal cell genesis in the mammalian retina. Nat Genet. 2003;34:53–8. 807 49. Hahn J, Monavarfeshani A, Qiao M, Kao AH, Kölsch Y, Kumar A, et al. Evolution of 808 neuronal cell classes and types in the vertebrate retina. Nature. 2023;624:415–24. 809 50. Shen B, Fang T, Dai M, Jones G, Zhang S. Independent losses of visual perception genes 810 Gja10 and Rbp3 in echolocating bats (Order: Chiroptera). PLoS One. 2013;8:e68867. 811 51. Milam AH, Rose L, Cideciyan AV, Barakat MR, Tang W-X, Gupta N, et al. The nuclear 812 receptor NR2E3 plays a role in human retinal photoreceptor differentiation and degeneration. 813 Proc Natl Acad Sci U S A. 2002;99:473–8. 814 52. Miyamoto M, Aoki M, Hirai K, Sugimoto S, Kawasaki K, Imai R. A nonsense mutation 815 in Gnat1, encoding the alpha subunit of rod transducin, in spontaneous mouse models of 816 retinal dysfunction. Exp Eye Res. 2010;90:63–9. 817 53. Renninger SL, Gesemann M, Neuhauss SCF. Cone arrestin confers cone vision of high 818 temporal resolution in zebrafish larvae. Eur J Neurosci. 2011;33:658–67. 819 54. Ronning KE, Allina GP, Miller EB, Zawadzki RJ, Pugh EN Jr, Herrmann R, et al. Loss of 820 cone function without degeneration in a novel Gnat2 knock-out mouse. Exp Eye Res. 821 2018;171:111–8. 822 55. Shimizu-Matsumoto A, Itoh K, Inazawa J, Nishida K, Matsumoto Y, Kinoshita S, et al. 823 Isolation and chromosomal localization of the human cone cGMP phosphodiesterase gamma 824 cDNA (PDE6H). Genomics. 1996;32:121–4. 825 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 35 56. Plaza Reyes A, Petrus-Reurer S, Padrell Sánchez S, Kumar P, Douagi I, Bartuma H, et al. 826 Identification of cell surface markers and establishment of monolayer differentiation to 827 retinal pigment epithelial cells. Nat Commun. 2020;11:1609. 828 57. Kueng-Hitz N, Grimm C, Lansel N, Hafezi F, He L, Fox DA, et al. The retina of c-fos-/- 829 mice: electrophysiologic, morphologic and biochemical aspects. Invest Ophthalmol Vis Sci. 830 2000;41:909–16. 831 58. Fernandes KA, Harder JM, Kim J, Libby RT. JUN regulates early transcriptional 832 responses to axonal injury in retinal ganglion cells. Exp Eye Res. 2013;112:106–17. 833 59. Syc-Mazurek SB, Fernandes KA, Wilson MP, Shrager P, Libby RT. Together JUN and 834 DDIT3 (CHOP) control retinal ganglion cell death after axonal injury. Mol Neurodegener. 835 2017;12:71. 836 60. Piri N, Kwong JMK, Gu L, Caprioli J. Heat shock proteins in the retina: Focus on HSP70 837 and alpha crystallins in ganglion cell survival. Prog Retin Eye Res. 2016;52:22–46. 838 61. Lazzarini R, Nicolai M, Pirani V, Mariotti C, Di Primio R. Effects of senescent secretory 839 phenotype acquisition on human retinal pigment epithelial stem cells. Aging (Albany NY). 840 2018;10:3173–84. 841 62. Lu S, Yang J, Yan L, Liu J, Wang JJ, Jain R, et al. Transcriptome size matters for single-842 cell RNA-seq normalization and bulk deconvolution. Nat Commun. 2025;16:1246. 843 63. Nishida K, Matsumura S, Uchida H, Abe M, Sakimura K, Badea TC, et al. Brn3a controls 844 the soma localization and axonal extension patterns of developing spinal dorsal horn neurons. 845 PLoS One. 2023;18:e0285295. 846 64. Dorgau B, Collin J, Rozanska A, Zerti D, Unsworth A, Crosier M, et al. Single-cell 847 analyses reveal transient retinal progenitor cells in the ciliary margin of developing human 848 retina. Nat Commun. 2024;15:3567. 849 65. Morato Torres CA, Wassouf Z, Zafar F, Sastre D, Outeiro TF, Schüle B. The Role of 850 Alpha-Synuclein and Other Parkinson’s Genes in Neurodevelopmental and 851 Neurodegenerative Disorders. Int J Mol Sci [Internet]. 2020;21. Available from: 852 http://dx.doi.org/10.3390/ijms21165724 853 66. Elshatory Y, Deng M, Xie X, Gan L. Expression of the LIM-homeodomain protein Isl1 in 854 the developing and mature mouse retina. J Comp Neurol. 2007;503:182–97. 855 67. Huang W, Fileta J, Guo Y, Grosskreutz CL. Downregulation of Thy1 in retinal ganglion 856 cells in experimental glaucoma. Curr Eye Res. 2006;31:265–71. 857 68. Raymond ID, Vila A, Huynh U-CN, Brecha NC. Cyan fluorescent protein expression in 858 ganglion and amacrine cells in a thy1-CFP transgenic mouse retina. Mol Vis. 2008;14:1559–859 74. 860 69. Riemondy KA, Ransom M, Alderman C, Gillen AE, Fu R, Finlay-Schultz J, et al. 861 Recovery and analysis of transcriptome subsets from pooled single-cell RNA-seq libraries. 862 Nucleic Acids Res. 2019;47:e20. 863 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint 36 70. Bacher R, Kendziorski C. Design and computational analysis of single-cell RNA-864 sequencing experiments. Genome Biol. 2016;17:63. 865 71. Hwang B, Lee JH, Bang D. Single-cell RNA sequencing technologies and bioinformatics 866 pipelines. Exp Mol Med. 2018;50:1–14. 867 72. Amezquita RA, Lun ATL, Becht E, Carey VJ, Carpp LN, Geistlinger L, et al. 868 Orchestrating single-cell analysis with Bioconductor. Nat Methods. 2020;17:137–45. 869 73. Yang Z, Ding K, Pan L, Deng M, Gan L. Math5 determines the competence state of 870 retinal ganglion cell progenitors. Dev Biol. 2003;264:240–54. 871 74. Sapkota D, Chintala H, Wu F, Fliesler SJ, Hu Z, Mu X. Onecut1 and Onecut2 872 redundantly regulate early retinal cell fates during development. Proc Natl Acad Sci U S A. 873 2014;111:E4086–95. 874 75. Agarwal D, Dash N, Mazo KW, Chopra M, Avila MP, Patel A, et al. Human retinal 875 ganglion cell neurons generated by synchronous BMP inhibition and transcription factor 876 mediated reprogramming. NPJ Regen Med. 2023;8:55. 877 76. Fuhrmann S. Eye morphogenesis and patterning of the optic vesicle. Curr Top Dev Biol. 878 2010;93:61–84. 879 77. Krumlauf R. Hox genes in vertebrate development. Cell. 1994;78:191–201. 880 78. Swamy V, McGaughey D. Eye in a Disk: eyeIntegration Human Pan-Eye and Body 881 Transcriptome Database Version 1.0. Invest Ophthalmol Vis Sci. 2019;60:3236–46. 882 79. Wahlin KJ, Cheng J, Jurlina SL, Jones MK, Dash NR, Ogata A, et al. CRISPR Generated 883 SIX6 and POU4F2 Reporters Allow Identification of Brain and Optic Transcriptional 884 Differences in Human PSC-Derived Organoids. Front Cell Dev Biol. 2021;9:764725. 885 80. Brown G. Retinoic acid receptor regulation of decision-making for cell differentiation. 886 Front Cell Dev Biol. 2023;11:1182204. 887 81. Yelon D, Stainier DYR. Pattern formation: swimming in retinoic acid. Curr Biol. 888 2002;12:R707–9. 889 82. Sanjurjo-Soriano C, Erkilic N, Damodar K, Boukhaddaoui H, Diakatou M, Garita-890 Hernandez M, et al. Retinoic acid delays initial photoreceptor differentiation and results in a 891 highly structured mature retinal organoid. Stem Cell Res Ther. 2022;13:478. 892 83. Teotia P, Niu M, Ahmad I. Mapping developmental trajectories and subtype diversity of 893 normal and glaucomatous human retinal ganglion cells by single-cell transcriptome analysis. 894 Stem Cells. 2020;38:1279–91. 895 84. Matsushita T, Onishi A, Matsuyama T, Masuda T, Ogino Y, Kageyama M, et al. Rapid 896 and efficient generation of mature retinal organoids derived from human pluripotent stem 897 cells via optimized pharmacological modulation of Sonic hedgehog, activin A, and retinoic 898 acid signal transduction. PLoS One. 2024;19:e0308743. 899 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 10, 2026. ; https://doi.org/10.64898/2026.02.08.704697doi: bioRxiv preprint