To interrogate transcriptomic changes during human photoreceptor development, dissociated RPCs and photoreceptor precursors were FACS-enriched from 18 retinae, ages FW13-19 (Figure 1—figure supplement 1A), and isolated using microfluidics or direct sorting into microliter droplets, followed by full-length cDNA synthesis, paired-end sequencing, and alignment to Ensembl transcript isoforms (Figure 1A). The FACS enrichment was based on a prior cone isolation method (Xu et al., 2014) but with wider gating to include rods and RPCs. After sequencing, we excluded all cells with <100,000 read counts and 18 cells expressing one or more markers of retinal ganglion, amacrine, and/or horizontal cells (POU4F1, POU4F2, POU4F3, TFAP2A, TFAP2B, ISL1) and concurrently lacking photoreceptor lineage marker OTX2. This yielded 794 single cells with averages of 3,750,417 uniquely aligned reads, 8278 genes detected, and 20,343 Ensembl transcripts inferred (Figure 1—figure supplement 1A–C). Sequencing batches were normalized and transcriptomes clustered and visualized in uniform manifold approximation and projection (UMAP) plots that integrated cells across different retinae, ages, isolation methods, and sequencing runs (Figure 1B and C and Figure 1—figure supplement 1D–F).

Low-resolution Louvain clustering (level 0.4) generated six clusters that segregated into mostly distinct UMAP domains (Figure 1B). One cluster was comprised of RPCs and Mϋller glia (MG), with specific expression of LHX2, VSX2, SOX2, and SLC1A3, while five clusters were comprised of cells with photoreceptor features, with wide expression of OTX2 and CRX and cluster-specific rod- and cone gene expression (Figure 1D and Figure 1—figure supplement 2). In UMAP space, the cluster designated immature photoreceptor precursors (iPRPs) intermixed with the RPC/MG population, extended towards early L/M cone precursors, and was predominantly comprised of cells expressing the L/M cone determinant THRB (Figure 1B and D).

Two clusters highly expressing the rod determinant NR2E3 were designated early-maturing rod (ER) and late-maturing rod (LR) (Figure 1B), based on the latter’s increased expression of rod phototransduction genes GNAT1, CNGB1, PDE6G, GNGT1, and RHO (Figure 1D and Figure 1—figure supplement 2C). Few rods were detected at FW13, whereas both early and late rods were detected from FW15 to FW19 (Figure 1C), corroborating prior reports (Lu et al., 2020; Zuo et al., 2024). Differential expression analysis revealed other genes upregulated in late rods (GNB1, SAMD7, NT5E) (Omori et al., 2017; Gagliardi et al., 2018) as well as the downregulated CRABP2, DCT, and FABP7 (Figure 1—figure supplement 3A and Supplementary file 1A). Genes upregulated in the LR cluster were enriched for photoreceptor and light sensing gene ontologies (Figure 1—figure supplement 3B) including spectrin binding, likely relating to proteins that control photoreceptor polarity and synapse formation (Chen et al., 2009; Burger et al., 2021), and purine biosynthesis and ribonucleotide metabolic processes, potentially related to the developing photoreceptors’ high NAD⁺ requirements (Sokolov et al., 2021).

Cones segregated into distinct S- and L/M-cone clusters, with differential expression of cone subtype markers (OPN1SW, THRB), previously identified S-cone enriched genes (CCDC136, UPB1; Lukowski et al., 2019; Peng et al., 2019; Kallman et al., 2020), novel S-cone enriched genes (MEGF10, NRXN3, ACKR3), and the L/M cone transcription factor ISL2 (Lu et al., 2020), among others (Figure 1B and D, Figure 1—figure supplement 3C, Supplementary file 1B). Gene ontology analysis did not reveal relevant terms enriched in S cones, whereas L/M cones were enriched for protein translation related ontologies (Figure 1—figure supplement 3D) due to increased expression of ribosomal protein genes RPL23A, RPLP0, RPS19, RPS27, RPS27A, RPS29, RPS3A (Supplementary file 1B).

In UMAP space, the THRB⁺ L/M-cone cluster segregated into a large 366 cell proximal population and a five-cell distal population inferred to represent early-maturing and late-maturing stages, respectively, based on the latter’s increased expression of cone phototransduction genes OPN1LW (encoding L-opsin), PDE6H, and GUCA1C (Figure 1E and Figure 1—figure supplement 3E and F), analogous to RHO, PDE6G, and GNGT1 upregulation in late rods. L/M cone precursors from different age retinae occupied different UMAP regions, suggesting age-related differences in L/M cone precursor maturation (Figure 1C). Compared to the early L/M population, late L/M cones had upregulation of three M-opsin genes, TTR, encoding the retinol-binding protein transthyretin, PCP4, a small protein that binds calmodulin previously noted in foveal cones (Voigt et al., 2019), and MYL4, a myosin light chain gene upregulated in retinal organoid L/M cones (Kallman et al., 2020), among others (Figure 1—figure supplement 3E and F, Supplementary file 1C). The low proportions of OPN1MW⁺ and OPN1LW⁺ late-maturing L/M cones are consistent with a prior analysis of similar-age retinae and with the further upregulation of these proteins in later maturation (Lu et al., 2020).

We next asked whether similar distinctions between early-maturing and late-maturing L/M cone and rod precursors were observed in prior studies. Indeed, a 3' single nucleus (sn) RNA-seq analysis of ~220,000 retinal cells from post-conception week (pcw) 8–23 distinguished the cone precursor versus ML cone and the rod precursor versus rod subclasses (Zuo et al., 2024). (N.b., we retain the ‘pcw’ and ‘ML cone’ terms of Zuo et al. when describing their data and the synonymous ‘FW’ and ‘L/M cone’ for our data to maintain continuity with past publications.) The Zuo et al. cone and rod precursor versus cone and rod photoreceptor comparisons were not strictly analogous to our early-maturing versus late-maturing precursor comparisons in that our early-maturing precursors excluded immature cone and rod precursors. Still, the comparisons revealed many of the same differentially expressed genes (Figure 1—figure supplement 3A, E, G and H; Supplementary file 1A, C, D and E).

To further interrogate cell identities, we used SCENIC to identify cluster-specific transcription factor regulons, which represent the overall expression of single transcription factors and their likely coregulated target genes (Van de Sande et al., 2020). The highest specificity regulons defined major cell populations including RPC/MG-specific E2F2, E2F3, VSX2, and PAX6; pan-photoreceptor NEUROD1, OTX2, and CRX; rod-specific NRL; and L/M cone-specific THRB and ISL2 (Figure 2 and Supplementary file 1F). SCENIC also distinguished ER and LR states via the latter’s increased NRL, CRX, ATF4, and LHX3 and decreased HMX1 and RAX activities (p<0.0005 for each, Dunn test). RAX activity also decreased in the 5 cell late-maturing L/M cone group (Figure 2B), supporting its distinct transcriptomic identity and suggesting a similar mode of late cone and late rod maturation. Additionally, iPRPs expressed their most specific regulons, LHX9 and OLIG2, at levels similar to RPC/MGs along with photoreceptor-related regulons, consistent with the transitional nature of this population (Figure 2A and Supplementary file 1F). However, SCENIC did not identify S cone-specific regulons, in keeping with the notion that S cones represent a default photoreceptor state induced by pan-photoreceptor factors such as OTX2, CRX, and NEUROD1 in the absence of NRL and THRB (Swaroop et al., 2010). Thus, deep, full-length scRNA-seq enabled identification of regulons underlying RPC and developing photoreceptor states at the single-cell level.

Although cone and rod precursors segregated into distinct clusters, mRNAs encoding rod-determining factor NRL, L/M cone-determining factor TRβ, and cone marker RXRγ were co-expressed in both rod and cone precursor populations, with mean NRL expression only 4.3-fold higher in the ER vs LM cluster, mean THRB expression 5.0-fold higher in LM vs LR, and mean RXRG expression 4.3-fold higher in LM vs ER (Figure 1—figure supplement 2C and D). Cone NRL expression was unexpected given NRL’s role in rod fate determination (Kallman et al., 2020) and rod-specific NRL regulon activity (Figure 2A and D). Similarly, rod RXRG and THRB expression were unexpected given their roles in cone gene expression and fate determination (Ng et al., 2001; Roberts et al., 2005) and L/M cone-specific THRB regulon activity (Figure 2A and D). Accordingly, we used full-length scRNA-seq data to determine if cone and rod precursors differentially express NRL, THRB, and RXRG transcript isoforms.

For NRL, three assigned transcript isoforms (ENST00000397002, ENST00000561028, and ENST00000558280) are predicted to encode the canonical full-length NRL protein (FL-NRL) (RefSeq NP_001341697.1), while two others (ENST00000560550 and ENST00000396995) are previously uncharacterized transcripts predicted to use an alternative ‘P2’ promoter and first exon, here termed exon 1T (Figure 3A–C). The novel transcripts are predicted to encode an N-terminally truncated NRL protein (Tr-NRL) retaining the leucine zipper DNA binding domain but lacking the minimal transactivation domain (Friedman et al., 2004; Figure 3C and Figure 3—figure supplement 1A and B). While all transcript isoforms were inferred to be more highly expressed in the ER rod cluster versus the LM cone cluster, the ratio of all FL-NRL:Tr-NRL transcripts was 2.9:1 in early rod precursors and 2.2:1 in late rod precursors, in contrast to 0.67:1 in L/M cone precursors (Figure 3B). Consistent with the assigned isoform ratios, mean read coverage of the Tr-NRL-specific exon 1T was higher in S and LM cones, while coverage of FL-NRL-specific exon 1 was higher in rods (Figure 3C, red vs. black arrowheads). Comparing the reads mapped to each first exon relative to total reads further confirmed that the Tr-NRL exon 1T predominated in individual cones whereas the FL-NRL exon 1 predominated in rods (Figure 3D). The cone cells’ higher proportional expression of Tr-NRL first exon sequences was validated by RNA fluorescence in situ hybridization (FISH) of FW16 fetal retina in which NRL immunofluorescence was used to identify rod precursors, RXRγ immunofluorescence was used to identify cone precursors, and FISH probes specific to Tr-NRL exon 1T or to FL-NRL exons 1 and 2 were used to assess Tr-NRL and FL-NRL expression (Figure 3E and F).

While the Tr-NRL-encoding NRL isoforms were not to our knowledge previously described, another NRL isoform that initiated within exon 2 and lacked the NRL transactivation domain due to alternative exon 2 splicing, termed DD10 (Figure 3C), was previously identified in adult retina (Swaroop et al., 1992). Concordantly, we detected reads spanning the unique DD10 splice junction, yet at lower levels than the unique Tr-NRL junction (5,942 vs 57,048).

As transcript isoforms inferred from short-read sequencing do not necessarily reflect the original transcript structures, we further defined the NRL isoforms expressed in early cone and rod precursors by performing 5' rapid amplification of cDNA ends (RACE) on the already generated single cell cDNA libraries from 23 early rod (ER) and 21 early L/M cone (LM) cells, followed by nanopore long-read sequencing of the pooled RACE products. The long-read sequencing revealed isoforms consistent with FL-NRL, Tr-NRL, DD10, and several other NRL isoforms with alternative transcription initiation and alternative splicing within exon 2 as well as within the Tr-NRL exon 1T (Figure 3G and Figure 3—figure supplement 1C). In keeping with the computationally assigned isoforms and differential exon usage, ER libraries had a higher proportion of FL-NRL exon 1 and exon 2 reads, and LM libraries had a higher proportion of Tr-NRL exon 1T reads (Figure 3G and Figure 3—figure supplement 1C). Moreover, alternative splicing within NRL exon 2 was more prevalent in LM libraries, affecting 6788 of 29,177 (23 %) of exon 2 reads compared to 5378 of 85,860 (6 %) in ER cells (p<0.0001; Chi-square with Yates correction), resulting in rarer full-length (exon 1-2-3) transcripts than inferred from short-read sequencing. Thus, long-read sequencing revealed cell type-biased NRL isoform expression with a preponderance of FL-NRL transcripts in early rods and disrupted FL-NRL and Tr-NRL transcript isoforms in L/M cones.

Despite our detection of L/M cone RNAs encoding Tr-NRL and FL-NRL, cone expression of NRL protein has not been reported. To assess endogenous Tr-NRL expression, we performed immunoblot analysis of CHLA-VC-RB31 retinoblastoma cells (Stachelek et al., 2023), which were predicted to express FL-NRL and Tr-NRL transcripts in a cone-like 0.73:1 ratio, with an antibody that recognizes both FL-NRL and Tr-NRL proteins (Figure 3—figure supplement 2A and B). As in other retinoblastoma cells (Khanna et al., 2006), FL-NRL increased in response to retinoic acid and proteasome inhibition, whereas Tr-NRL was not detected (Figure 3—figure supplement 2C). Similarly, Tr-NRL was not detected in EGFP⁺, RXRγ⁺ cones following lentiviral transduction of an explanted fetal retina with an EGFP-P2A-Tr-NRL cassette (Figure 3—figure supplement 2D–G). These findings suggest that Tr-NRL is poorly translated or unstable in most cone precursors.

As Tr-NRL might be expressed in contexts that were not examined in our analyses, we explored the function of the putative Tr-NRL protein. As both Tr-NRL and the previously described NRL DD10 lack the NRL transactivation domain and as DD10 interferes with FL-NRL transactivation (Rehemtulla et al., 1996), we examined if Tr-NRL similarly opposes FL-NRL transcriptional activity. Indeed, in luciferase reporter assays, Tr-NRL suppressed FL-NRL activation of a PDE6B promoter (Figure 3H).

For THRB, the most highly assigned transcript isoforms encoded the L/M cone-specific TRβ2 (ENST00000280696) and the more widely expressed TRβ1 (ENST00000396671 and others; Figure 4A-C, Figure 4—figure supplement 1). While both TRβ1 and TRβ2 promote L/M cone fate determination (Eldred et al., 2018), the isoform encoding TRβ2 predominated in L/M cones while isoforms encoding TRβ1 predominated in early rods (Figure 4B). Moreover, late rods preferentially expressed THRB exons 1–6 and the THRB 3' untranslated region (UTR), implying that RNAs encoding full-length TRβ proteins were rare (Figure 4B and C). Notably, a higher percentage of reads extended from the exon 4 and exon 6 splice donor sequences into the subsequent introns in LR versus LM cells (Figure 4D and E; p≤0.001 for both, two-tailed Chi square test), suggesting that premature transcription termination (PTT) in introns 4 and 6 preferentially limits full-length TRβ expression in the LR population. The inferred PTT events are consistent with structures of the assigned Ensembl isoforms (Figure 4B, Figure 4—figure supplement 1).

To further evaluate PTT events, we used THRB 3' RACE and long-read sequencing on single-cell cDNA libraries from 23 L/M cone cells and five LR cells selected for high THRB expression. RACE reactions were performed separately with primers complementary to the TRβ1-specific exon 4 and to the TRβ2-specific first exon (Figure 4F). Sequencing of LM and LR RACE products corroborated pronounced PTT in introns 4 and 6, with greater intron 6 PTT in LR versus LM cells (Figure 4F and Figure 4—figure supplement 2). However, we did not corroborate the late rods’ proportionately higher intron 4 PTT, likely due to the small number of LR cells examined and heterogeneity in PTT frequency in individual cells. Long read sequencing also revealed PTT following the TRβ2-specific exon and a novel transcription-terminating exon following the canonical exon 5 observed solely in TRβ2 transcripts (Figure 4F and Figure 4—figure supplement 2). 3' RACE transcripts rarely extended into the 3' UTR in LM or LR cells, suggesting that reads mapping to this region do not reflect protein-coding RNAs and confound the assessment of protein-coding THRB mRNA expression. These analyses demonstrate that THRB is regulated by multiple PTT events in rod and cone precursors as well as by cell type-specific promoter utilization and independent 3' UTR RNA expression.

For RXRG, short read sequencing reads were assigned to several isoforms that differed in their 5' promoter position and exon utilization (Figure 4—figure supplement 3A–C). However, 5' RACE and long-read sequencing did not support differential isoform expression, and quantitative imaging revealed an average ~3.5-fold higher RXRγ protein in cones compared to rods (Figure 4—figure supplement 3D and E). Thus, long-read sequencing clarified that RXRG expression is moderately higher in human cone versus rod precursors without evidence of cell-type-specific isoforms.

To further define photoreceptor developmental states, we subdivided the initial six cell populations with higher resolution clustering (level 1.6), identified high resolution cluster-associated genes and regulons, and inferred each cell’s rate and direction of transcriptomic change using RNA velocity (La Manno et al., 2018; Figure 5A and B). Increased clustering resolution divided the RPC/MG cluster into separate RPC and Müller glia (MG) groups, divided L/M cones into four subgroups (LM1 – LM4), which partially overlapped in UMAP space, and divided the iPRP cluster into two clusters here designated immature cone precursors (iCPs) and immature rod precursors (iRPs), the latter also drawing cells from the low-resolution ER cluster (Figure 5A and C). Similar clusters were observed at reduced k.param values that define nearest neighbors.

The distinction between RPCs and MG was corroborated by the expression of known marker genes, with the RPC cluster having increased expression of cell cycle markers (CCNE2, CCNA2, CCNB2) and the cell cycle-associated PBK and E2F7 (Figure 5D, Figure 5—figure supplement 1A, and Supplementary file 1G). The MG cluster was spatially segregated, lacked expression of most cell cycle genes, and expressed genes known to be expressed in both populations (CCND1, SLC1A3, PAX6, LHX2, VSX2) as well as the MG marker RLBP1 (Roesch et al., 2008; Pereiro et al., 2020; Blackshaw et al., 2004; Figure 5C and D and Figure 5—figure supplement 1A). Differential gene expression analyses revealed upregulation of cell-cycle-related genes and G2/M checkpoint and E2F target ontologies in the RPC cluster but no significantly upregulated ontologies in the MG population (Figure 5—figure supplement 1B and C), consistent with early MGs resembling quiescent RPCs (Walcott and Provis, 2003). In contrast to other studies, we did not distinguish primary RPCs (PRPCs) from neurogenic RPCs (NRPCs; Clark et al., 2019; Lu et al., 2020; Zuo et al., 2024), likely due to the underrepresentation of RPCs in our dataset.

The high-resolution iCP and iRP clusters included ‘RPC-localized’ cells positioned adjacent to RPCs and ‘non-RPC-localized’ cells adjacent to the LM and ER clusters in UMAP space (Figure 5C). RNA velocity suggested that the RPC-localized cell groups flowed towards the larger early maturing L/M cones and rods, respectively (Figure 5B and C). Compared to RPCs and MG, RPC-localized iCP and iRP cells had minimal expression of cyclin RNAs or RPC markers PAX6, LHX2, and VSX2, and had increased expression of photoreceptor determinants OTX2 and CRX (Figure 5D), consistent with their being immediately post-mitotic photoreceptor precursors. iCP cells in this region upregulated the early cone cell fate determinant ONECUT1 (Emerson et al., 2013; Finkbeiner et al., 2022), the L/M cone determinant THRB, and the cone differentiation marker GNAT2, whereas iRP cells trended towards higher expression of the rod determinant NR2E3 (Figure 5D and Figure 5—figure supplement 2A); however, such RNAs may be more highly expressed in cone- and rod-fated NRPCs than in the unspecified RPCs in our study. Notably, THRB and GNAT2 expression did not significantly change while ONECUT1 declined in the subsequent non-RPC-localized iCP and LM1 stages, whereas NR2E3 and NRL dramatically increased on transitioning to the ER state (Figure 5—figure supplement 2A).

In the RPC-localized region, iCPs had higher ONECUT1, THRB, and GNAT2, whereas iRPs trended towards higher NRL and NR2E3 (p=0.19, p=0.054, respectively; Figure 5D). While we detected differential expression when selectively examining these genes of interest, no genes were differentially expressed (Padj <0.05) in a transcriptome-wide comparison, likely due to the lack of statistical power given the small number of cells examined.

To our knowledge, past studies have not distinguished immature cone and rod precursors (i.e. iCPs and iRPs) from the subsequent maturing cone and rod precursor states. To explore whether early cone and rod precursors with similar properties are present in droplet-based scRNA-seq studies, we examined gene expression in spatiotemporally segregated cone and rod precursor populations as well as in cone and rod NRPCs in the Zuo et al., 2024 3' snRNA-seq dataset. This revealed that cone NRPCs and the earliest (pcw 8–10) cone precursors had high ONECUT1 and THRB yet minimal NR2E3 (Figure 5—figure supplement 2B). Similarly, rod NRPCs had high NRL and NR2E3, yet it was not possible to examine the earliest (pcw 10–13) rod precursors, as many had cone-like gene expression profiles, suggesting they may have been mis-assigned (Figure 5—figure supplement 2B, gray boxes). Further evaluation suggested misassignment of a smaller proportion of rod and cone precursors at other ages (Figure 5—figure supplement 3). In particular, cells potentially misassigned as rod precursors were located in the cone precursor UMAP region and highly expressed THRB (Figure 5—figure supplement 3D, **), while cells potentially misassigned as cone precursors were located in the rod precursor UMAP region and highly expressed NR2E3 (Figure 5—figure supplement 3E, *). When considering only the main cone and rod precursor UMAP regions, early (pcw 8–13) cone precursors expressed THRB and lacked NR2E3 (Figure 5—figure supplement 3D and E, blue arrows), while early (pcw 10–15) rod precursors expressed NR2E3 and lacked THRB (Figure 5—figure supplement 3D and E, red arrows), similar to RPC-localized iCPs and iRPs in our study (Figure 5D). We conclude that Louvain clustering of deep full-length scRNA-seq distinguished immediately post-mitotic iCP and iRP populations with features similar to early cone and rod precursors in a large 3' snRNA-seq dataset.

In a further comparison, we noted lower detection of GNAT2 in L/M cone precursors and lower detection of CCNE2, CCNA2, and CCNB2 in RPCs in 3' snRNA-seq versus full-length scRNA-seq datasets (Figure 5—figure supplement 2A and B). This suggests that deep full-length scRNA-seq enabled more sensitive detection of certain cell-type-specific RNAs along with improved discrimination of immature cone versus rod precursor states.

SCENIC regulons further clarified the identities of the four RPC-localized clusters. The MG cluster was best specified (i.e., had highest regulon specificity scores) by PAX6 and SOX9, both previously described in RPCs and MGs (Roesch et al., 2008; Marquardt et al., 2001; Poché et al., 2008), whereas RPCs were best specified by E2F2 and E2F3 (Figure 5E). These RPC and MG regulons were low or absent in the early post-mitotic iCPs and iRPs, suggesting they undergo an abrupt cell state change. Moreover, RPC-localized iCPs had greater THRB regulon signal than RPC-localized iRPs, while RPC-localized iRPs had higher NRL regulon signal (p<0.05 for both; Figure 5E). Interestingly, RPC-localized and/or non-localized iCPs also had higher activity of the pan-photoreceptor regulons LHX3, OTX2, and NEUROD1 compared to their iRP counterparts (Figure 5—figure supplement 2C and D). Finally, activities of the cone-specific THRB and ISL2 regulons, the rod-specific NRL regulon, and the pan-photoreceptor LHX3, OTX2, CRX, and NEUROD1 regulons increased to varying extents on transitioning from immature iCP or iRP states to the early-maturing LM1 or ER states (Figure 5—figure supplement 2C and D).

In addition to the immediately post-mitotic RPC-localized iCPs, the iCP cluster included cells bridging the UMAP region between early maturing cones and early maturing rods (Figure 6A). Many iCP and ER cells in this bridge region expressed RNAs encoding cone markers (GNAT2, THRB), rod markers (GNAT1, NR2E3), or, in many cases, both (Figure 6B), suggestive of a proposed hybrid cone/rod precursor state more extensive than implied by the co-expression of different THRB and NRL isoforms (Ng et al., 2011).

To determine whether the intermixed cone and rod gene expression reflects a hybrid cone-rod precursor state, we examined bridge region NRL and THRB regulon activities, which embody the overall transcriptomic effects of these factors. Indeed, in the bridge region, cells with NRL and THRB regulon signals intermixed, and some cells showed signals for both (Figure 6C, arrows). However, iCP cells had low NRL regulon z-scores similar to L/M cones and most had THRB z-scores above that of adjacent ER cells (Figure 6D). The difference between regulon signals increased in rod and cone clusters outside the bridge region as NRL or THRB regulons became more dominant (Figure 6D).

To identify cells with early cone- and rod-related RNA co-expression in the developing retina, we performed multiplex RNA FISH for early cone marker GNAT2 and rod marker NR2E3 combined with immunofluorescence (IF) staining for RXRγ, which has high expression in outermost neuroblastic layer (NBL) cone precursors and low expression in middle NBL rod precursors (as earlier shown in Figure 4—figure supplement 3E), and for NR2E3, which is detected solely in rods. To infer the spatiotemporal pattern of such expression across human retinal development, we examined GNAT2 and NR2E3 RNA co-expression in RXRγ⁺ cone precursors in the outermost NBL and in RXRγ⁺ rod precursors in the middle NBL across 13 regions of a FW14 retina section (Figure 7; see Figure 7—figure supplement 1 for IF + FISH images). Limiting our analysis to the outer and middle NBL allowed us to disregard RXRγ⁺ retinal ganglion cells in the retinal ganglion cell layer or inner NBL.

The analyses revealed that most of the far peripheral (hence, nascent) outer NBL RXRγ⁺ cone precursors (Figure 7A and B, regions 2 and 13) lacked detectable GNAT2 and NR2E3 RNA (green bars in Figure 7C), whereas those in the more mature central retina (regions 3–12) were uniformly GNAT2⁺ (yellow and red bars in Figure 7C). However, starting with regions 3 and 11, some outermost NBL GNAT2⁺ cones were also NR2E3⁺ (red bars in Figure 7C). The proportion of outermost NBL cells that co-expressed GNAT2 and NR2E3 RNA increased in the more central retina (regions 5–10), yet all of these cells had strong RXRγ staining and lacked NR2E3 protein, consistent with their having a cone identity (illustrated in Figure 7E). In contrast, most middle NBL RXRγ⁺ cells had either a low number of NR2E3 RNA puncta without NR2E3 protein (NR2E3⁺/NR2E3^–, magenta bars in Figure 7D), likely in nascent rods, or prominent NR2E3 RNA with NR2E3 protein (NR2E3⁺/NR2E3⁺, purple bars in Figure 7D), in maturing rods (Figure 7D and E). However, we did not detect GNAT2 puncta in middle NBL RXRγ⁺, NR2E3⁺ cells, including the most peripheral RXRγ⁺ cells with low-level NR2E3 RNA (Figure 7E). In summary, most photoreceptor precursors with GNAT2 and NR2E3 RNA co-expression had high RXRγ, no detectable NR2E3 protein, and outermost NBL positions expected of cone precursors. This supports the notion that, in our scRNA-seq analyses, bridge region iCP cells with combined GNAT2 and NR2E3 RNA expression are likely cone-directed.

Cone and rod precursor populations that co-express cone and rod marker genes were also evident in the Zuo et al. 3' snRNA-seq dataset. While the earliest cone precursors were NR2E3-negative and the earliest rod precursors were THRB-negative (as described above), starting at pcw 15, later cone precursors expressed NR2E3 (Figure 5—figure supplement 3E, green arrows) and later rod precursors expressed THRB (Figure 5—figure supplement 3D, purple arrows). THRB was highest in the most distal rod precursors (yet still lower than in cone precursors), corroborating our findings in Figures 1D and 4B–C. In contrast, in the 3' snRNA-seq analysis, GNAT2 expression was sparse in rod precursors of all ages (Figure 5—figure supplement 3F), consistent with the lack of GNAT2 in rod precursors in our RNA FISH analyses and implying that different cone- and rod-related genes have different tendencies to be expressed in the other photoreceptor precursor type.

Thus, a 3' snRNA-seq analysis confirmed the initial production of immature photoreceptor precursors with either L/M cone-precursor-specific THRB or rod-precursor-specific NR2E3 expression, followed by lower-level co-expression of their counterparts, NR2E3 in cone precursors and THRB in rod precursors. However, in the Zuo et al. analyses, the co-expression was first observed in well-separated UMAP regions, as opposed to a region that bridges the early cone and early rod populations in our UMAP plots. These findings are consistent with the notion that cone- and rod-related RNA co-expression begins in already fate-determined cone and rod precursors, and that such precursors aberrantly intermixed in our UMAP ‘bridge region’ due to their insufficient representation in our dataset.

To further assess relationships between iCP subpopulations, we examined the expression and UMAP distributions of the four iCP marker genes identified in Figure 5—figure supplement 1A. Among these genes, CHRNA1 was mainly expressed in bridge region iCP and ER cells, whereas ONECUT1 was biased to cone-directed iCPs, lncRNA CTC-378H22.2 was expressed in a narrow zone of THRB⁺, GNAT2⁺ iCPs, and S100A6 was more widely expressed (Figure 8A). iCP cells also expressed ATOH7 and DLL3, which were previously proposed to define transitional photoreceptor precursors and promote cone fate determination (Clark et al., 2019; Lu et al., 2020; Sridhar et al., 2020). ATOH7 was largely restricted to RPC-localized and bridge region iCPs, whereas DLL3 was broadly expressed similar to S100A6 (Figure 8A).

A similar pattern was seen in the Zuo et al., 2024 3' snRNA-seq dataset, where CHRNA1, ONECUT1, S100A, ATOH7, and DLL3 were most highly expressed in the youngest and developmentally earliest ML cone and rod precursors and persisted to various extents during cone and rod maturation, whereas CTC-378H22.2 (ENSG00000259436) had more restricted and cone-specific expression (Figure 8—figure supplement 1A). However, the early cone and rod precursor UMAP regions did not adjoin one another (in contrast to our UMAP bridge region), suggesting that the cone and rod precursors expressing the above markers had begun their distinct trajectories and that their juxtaposition in our UMAP analyses is spurious. Also, different iCP markers had different spatiotemporal expression: CHRNA1 and ATOH7 were most prominent in peripheral retina with ATOH7 strongest at pcw 10 and CHRNA1 strongest at pcw 13; CTC-378H22.2 was prominently expressed from pcw 10–13 in both the macula and the periphery; and DLL3 and ONECUT1 showed the earliest, strongest, and broadest expression (Figure 8—figure supplement 1B). The distinct patterns suggest that these factors have spatiotemporally distinct roles in cone precursor differentiation.

As CHRNA1 appeared to be the most specific marker of both early cone and early rod precursors, we evaluated whether CHRNA1 RNA marked specific cone and rod populations in the developing retina using RNA-FISH combined with RXRγ and NRL immunofluorescent staining. In a FW12 retina, we detected the highest CHRNA1 in the earliest, most peripheral NRL⁺ rods and RXRγ⁺ cones and fewer CHRNA1⁺ cells in more mature, central regions (p<0.005 in cones; Figure 8B–D), consistent with its expression in early cone and rod precursors.

To identify factors that regulate transcriptomic states during cone cell fate determination, we examined the UMAP distributions of the most iCP-specific transcription factor regulons, OLIG2 and LHX9 (Figure 5—figure supplement 2D, Supplementary file 1H). As Olig2 was previously detected in mouse RPCs in which Onecut1 enabled Thrb expression and cone fate determination (Emerson et al., 2013; Hafler et al., 2012), the OLIG2 regulon was expected to be most active in RPCs and in early, immediately post-mitotic iCPs. While the OLIG2 and LHX9 regulons were indeed active in RPCs, both were also active in a narrow zone of ONECUT1⁺ iCPs positioned farthest from S cones and immediately preceding the upregulation of the THRB regulon, a location with sparse OLIG2 RNA expression (Figure 8—figure supplement 2). These data indicate that the OLIG2 and LHX9 regulons are active in and potentially relevant to human L/M cone fate determination in a post-mitotic iCP subpopulation.

After fate commitment, L/M cones initiate a maturation process with upregulation of RNAs and proteins related to phototransduction, axonogenesis, synaptogenesis, and outer segment morphogenesis (Hendrickson et al., 2012; Hoshino et al., 2017). To evaluate whether early L/M cone maturation is comprised of distinct transcriptomic cell states, we assessed marker gene and regulon differences between high-resolution clusters LM1, LM2, LM3, and LM4. While clusters LM1-4 were distinguished by sequential increasing expression of ACOT7, RTN1, PDE6H, OLAH, and NPFF, they showed only subtle differences among the three regulons with highest LM1-4 specificity scores, THRB, ISL2, and LHX3 (Figure 9A and B). The lack of cluster-specific marker genes and regulons suggests that LM1-LM4 represent different stages of a graded maturation process.

We further evaluated gene expression changes in maturing cones by pseudotemporally ordering iCP and L/M cone precursors (Figure 9C). This identified 967 pseudotime-correlated genes (q-value <0.05, expression >0.05 in >5% of cells; Supplementary file 1I) in seven gene modules (Figure 8—figure supplement 1). Among the top 20 pseudotime-correlated genes in each module, we identified four lncRNAs that were sequentially expressed (Figure 9D). To determine if these lncRNAs distinguish developmentally distinct maturing cones in vivo, we probed their expression via multiplex RNA FISH in co-stained RXRγ⁺ cone precursors across a FW16 retina (Figure 9E and F). Quantitation of FISH puncta defined four cone maturation zones based on expression peaks and significant count differences for each lncRNA as color-coded in Figure 9G: the most peripheral cones with high CTC-378H22.2 and HOTAIRM1 (blue), peripheral cones with high HOTAIRM1 only (green), cones with low expression of all four lncRNAs (red), and parafoveal/foveal cones with upregulated RP13143G15.4, CTD-2034I21.2, and CTC-378H22.2 (purple) (Figure 9G). The detection of foveal CTC-378H22 by ISH but not in late LM transcriptomes may relate to a lack of the most mature cones in our scRNA-seq analyses. These data support the concept that maturing L/M cones sequentially express specific lncRNAs as they develop.

We next assessed whether early-maturing cone transcriptomic features are conducive to the proliferative response to pRB loss (Xu et al., 2014; Singh et al., 2018). Given the high transcriptomic similarity across the L/M cone population, we compared gene expression between all early-maturing L/M cones and early-maturing rods. This identified 422 genes upregulated and 119 downregulated in cones (p<0.05, log2FC>|0.4|) (Figure 10—figure supplement 1A, Supplementary file 1J). Among cone-enriched genes, the top three enriched ontologies related to translation initiation, protein localization to membrane, and MYC targets (Figure 10—figure supplement 1B). The upregulation of MYC target genes was of interest given that many MYC target genes are also targets of MYCN, that MYCN protein is highly expressed in maturing (ARR3⁺) cone precursors but not in NRL⁺ rods (Figure 10A), and that MYCN is critical to the cone precursor proliferative response to pRB loss (Cobrinik, 2024; Xu et al., 2009; Xu et al., 2014). Indeed, whereas MYC RNA was not detected, the LM cone cluster had increased MYCN RNA (log₂FC = 0.54) and MYCN regulon activity, representing the seventh highest LM cluster regulon specificity score (Figure 10B and C, Supplementary file 1F).

Among other differentially expressed genes, we noted the L/M cone-specific upregulation of SYK (Figure 10D, Supplementary file 1J), which encodes a non-receptor tyrosine kinase. Whereas SYK was previously implicated in retinoblastoma genesis and proposed to be induced in response to pRB loss (Zhang et al., 2012), its expression was not previously reported in developing fetal retina. Indeed, our scRNA-seq analyses revealed that SYK RNA expression increased from the iCP stage through cluster LM4, in contrast to its minimal expression in rods (Figure 10E). Moreover, SYK expression was abolished in the five-cell group with properties of late-maturing cones (characterized in Figure 1E), here displayed separately from the other LM4 cells and designated LM5 (Figure 10E). Similarly, immunohistochemical staining revealed high SYK protein expression in immature (ARR3^-) and early-maturing (ARR3⁺) cones from the retinal periphery to the maturing foveal cones at FW16, while SYK was not detected in the most mature foveal cones at FW18 (Figure 10F). The loss of SYK protein and RNA expression with cone maturation is consistent with the lack of SYK in cones of normal retina adjacent to retinoblastoma tumors (Zhang et al., 2012) and implies that SYK is a defining feature of the human early cone precursor state. While MYCN RNA was also preferentially expressed in early cones, it did not increase as much relative to RPCs nor decline as much in late-maturing cone precursors when compared to SYK RNA dynamics (Figure 10E).

To determine if SYK might contribute to retinoblastoma initiation, dissociated fetal retinal cells were transduced with an RB1-directed shRNA (shRB1-733) known to induce cone precursor proliferation (Xu et al., 2014), treated with the selective SYK inhibitor GS-9876 (Blomgren et al., 2020) for 12 days, and examined for cone precursor cell cycle entry by co-staining for RXRγ and Ki67. GS-9876 treatment suppressed the proliferative response to pRB knockdown at all concentrations from 1.0 to 5.0 µM (Figure 10G), consistent with the notion that cone precursor intrinsic SYK activity contributes to the proliferative response to pRB loss and is retained in retinoblastoma cells (Figure 10H). However, given potential SYK inhibitor off-target effects, validation of the role of SYK in retinoblastoma initiation will require gene ablation studies.

Human fetal samples were provided by Advanced Bioscience Resources, Inc, Alameda, CA, or collected under Institutional Review Board approval at the University of Southern California (protocol HS-13–00399), and Children’s Hospital Los Angeles (protocol CHLA-14–00122). Following the patient decision for pregnancy termination, patients were offered the option of donation of the products of conception for research purposes, and those that agreed signed an informed consent. This did not alter the choice of termination procedure, and the products of conception from those who declined participation were disposed of in a standard fashion. Fetal age was determined according to the American College of Obstetrics and Gynecology guidelines (Pettker et al., 2017). Each retina subjected to sequencing was from a different donor.

CHLA-VC-RB31 was produced as in Stachelek et al., and HEK-293T and NIH-3T3 cells obtained from the American Type Culture Collection (ATCC). CHLA-VC-RB31 and HEK-293T were authenticated by short tandem repeat (STR) analysis and all cell lines tested (negative) for mycoplasma by the University of Arizona Genetics Core, which has the original CHLA-VC-RB31 and HEK-293T STR profile records. NIH-3T3 cells were authenticated by STR analysis by the ATCC Mouse Cell Authentication Service.

Marker features for each cluster were identified using the Wilcoxon rank sum tests and specificity scores were computed using the genesorteR (Ibrahim and Kramann, 2019) R package. Only genes with an adjusted p-value (pAdj) <0.05 were considered for further analyses. Differential expression analysis of full-length scRNA-seq was performed using the Seurat FindMarkers function with the Wilcoxon rank sum test on log-normalized counts. Results were displayed as volcano plots made with EnhancedVolcano (Blighe et al., 2018). Differential expression analysis of a 3' snRNA-seq was performed using the Seurat FindMarkers function with the Wilcoxon rank sum test on log-normalized counts. Results were displayed as volcano plots made with EnhancedVolcano (Blighe et al., 2018).

Overrepresentation analyses were performed with WebGestalt (Liao et al., 2019). Analyses were run with the default settings (5–2000 genes per category, Benjamini-Hochberg multiple-testing correction), gene list enrichment was compared to the genome reference set and ontologies were displayed with a false discovery rate (FDR)<0.05 and weighted set cover redundancy reduction. All gene sets evaluated were provided within WebGestalt and the same three were used unless otherwise indicated: GO – Biological Process, KEGG, and Hallmark50.

Transcription factor regulons were identified using pySCENIC, the python-based implementation of Single-Cell Regulatory Network Inference and Clustering (SCENIC; Van de Sande et al., 2020; Aibar et al., 2017). The tool was run using the basic settings as shown (https://pyscenic.readthedocs.io/en/latest/tutorial.html). Initial filtering required a gene to have a minimum of 3 raw counts in 1% of cells to be considered for inclusion in a regulon. AUC scores were z-score normalized for comparison between regulons.

RNA velocity analysis was performed with R package velocyto (La Manno et al., 2018). Spliced and unspliced count matrices were assembled using the run_smartseq2 subcommand with repeat-masked ENSEMBL build 87. RNA velocity was calculated using cell kNN pooling with a gamma fit based on a 95% quantile, and velocities were overlaid on integrated UMAP visualizations using velocyto. Pseudotemporal trajectories were calculated using Monocle 3 (Cao et al., 2019). The SeuratWrappers package provided integration between Seurat’s final integrated dimensional reduction and Monocle. The data was subset to include only cone-directed iCPs and L/M cones, then a principal graph was fit across the UMAP plot and a root cell chosen to represent the latest maturation point in the OPN1LW⁺ late-maturing cone group. After assigning pseudotime values to each cell, genes that significantly changed as a function of pseudotime were identified and those with a correlation q-value of <0.05 with expression >0.5 in at least 5% of the cells present in the pseudotime were grouped into modules of co-regulation by performing UMAP and Louvain community analysis.

FW18.5 retina was partially dissociated with papain, cultured overnight in retinal culture media at 37 °C in a six-well plate, then frozen in 10% DMSO solution and stored in liquid nitrogen. Samples were thawed and revived in retinal culture media overnight and infected with lentivirus carrying shRNAs targeting RB1 (shRB1-733; Xu et al., 2014) or a scrambled control (shSCR) and diluted β-mercaptoethanol, insulin, glutamine, penicillin, and streptomycin to the final retina culture medium concentrations along with 5 µg/ml Polybrene, for 24 hr. After replacing 2/3 of media in each well with fresh media, cells were treated with GS-9876 (MedChemExpress, HY-109091) in DMSO at or with the same volume of DMSO vehicle control. After 12 days, cells were attached to poly-L-Lysine-coated coverslips for 3 hr, fixed in 4% PFA for 10 min, washed in 1 x PBS three times, and stored at –20 °C until immunocytological staining with antibodies to Ki67, RXRγ, and GFP (for YFP staining) as described (Xu et al., 2014).

Statistical methods and packages are cited in the Methods text and figure legends. Statistical comparison between more than two groups of cells for gene expression, regulon activity, or exon count proportions used the non-parametric Kruskal-Wallis test followed by pairwise post-hoc Dunn tests with Benjamini-Hochberg correction. Mean expression levels of NRL and RXRG isoforms across cell clusters were evaluated by Welch’s t-tests for groups with unequal variance; p-values for these tests were estimated using bootstrapping with 1,000,000 replications per comparison. Changes in FISH puncta counts were examined within each channel individually using zero-inflated negative binomial regression models with robust standard error terms. Distance across the sample tissue was used to model degenerate zero inflation. P values reported in the manuscript text and legends are Wald Chi-Squared tests that evaluate differences in count between inflection points. Analyses were performed using Microsoft Excel, the R statistical language in RStudio, or Stata SE v14.2.