Distinct epithelial gene expression patterns define cell fate differences of skin keratinocytes and cornea limbal stem cells

To characterize gene expression patterns that define the cell fate difference between human cornea LSCs and human skin KCs (Fig 1A), we used cultured LSCs established from post mortem limbal biopsies and basal KCs from skin donors. Both cultured cells have the capacity to regenerate stratified epithelial tissues in vitro [41,42] and have high p63 expression (S1 Fig), thus exhibiting the progenitor cell state. We performed comparative RNA-seq analyses from bulk and pseudobulk RNA-seq data (aggregated from scRNA-seq using cultured LSCs and KCs) experiments and incorporated our data with publicly available RNA-seq data for robustness (S1 Table) (S1 Fig) [16,32]. Single-cell data was aggregated because no measurable heterogeneity was detected in these cultured cells, except for cell cycle differences (S1E Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. RNA-seq analysis of LSCs and KCs.

(A) Schematic picture of the epidermis and the limbus. (B) Heatmap of normalized expression of differentially expressed genes between LSCs and KCs (adjusted pval < 0.01, log2 FC > 1.5). Differentially expressed genes are clustered using k-means clustering with 2 clusters. (C) GO term enrichment of KC-high genes. (D) GO term enrichment of LSC-high genes. (E) PROGENy pathway activity analysis, with scores sorted based on the LSC/KC ratio. Pathways depicted in red are differential, the color is gray if non-differential, orange if higher in LSC, and purple if higher in KCs. (F) GSEA of differentially expressed genes identified in aniridia patient LSCs, as compared to controls, up- and down-regulated genes (aniridia high and low, respectively) were tested for enrichments against the KC-LSC fold change. For all underlying data, see S4 Table, GEO GSE206922, GSE206923, and GSE242995. GO, Gene Ontology; GSEA, gene set enrichment analysis; KC, keratinocyte; LSC, limbal stem cell.

https://doi.org/10.1371/journal.pbio.3002336.g001

Through pairwise comparison, we identified 1,251 differential expressed genes between LSCs and KCs. Among them, 793 genes had higher expression in LSCs (referred to as LSC-high genes), while 459 differential genes were more highly expressed in KCs (referred to as KC-high genes). This analysis resulted in typical genes for both epithelial cell types: LSC-high genes with limbal and corneal epithelial genes including KRT19 and the eye master regulator PAX6 (Figs 1B and S1F), whereas KC-high genes contained epidermal markers such as KRT1, KRT10, LCE3D, and LCE3C. Although some of these detected genes are associated with epithelial stratification, e.g., KRT3 and KRT12 for the cornea and KRT1 and KRT10 for the epidermis, their expression was significantly lower (5- to 20-fold) than their expression in stratified epithelial cells [16,39] (S1G Fig), confirming the progenitor states of the cultured LSCs and KCs. It should be noted that TP63 is highly expressed in both LSCs and KCs (S1F Fig), and therefore is not identified as differential.

Gene Ontology (GO) enrichment analysis [43,44] of KC-high genes identified enrichment of GO terms related to the “epidermis” and “skin development” (Fig 1C). GO terms associated with “response to virus” were enriched due to detected immune and interferon-related genes among KC-high genes, consistent with the gene set enrichment analysis (GSEA) using the hallmark gene set of the MsigDB collection [45,46] that identified enrichment of the interferon-alpha and gamma response (S4 Table). Finally, PROGENy pathway target gene analysis [47] identified higher VEGF signaling in KCs (Fig 1E), indicating that VEGF vascularization-related genes are lowly expressed or completely repressed in LSCs. This is in line with the avascularized state of the cornea. LSC-high genes were enriched for terms such as “wound healing” and “positive regulation of body fluid levels” as well as “visual system development” and “regulation of nervous system development” (Fig 1D). PROGENy analysis identified the TNF-α and NFKβ pathways associated with LSCs (Fig 1E), such as CXCL1,3,5,6 and TNF-α. Consistently, TNF-α and NF-kB signaling was also identified by KEGG pathway [48], GO, and GSEA analyses using the hallmark gene set of the MsigDB collection (S2 Fig). Finally, GSEA enrichment using the C8 dataset of the MsigDB collection that contains single-cell dataset gene lists also identified enrichment for genes in the “Descartes fetal eye corneal and conjunctival epithelial cells” (S4 Table) with more signal in the LSCs.

Next, we asked the question of whether LSCD-associated LSCs acquire a KC-like cell fate. To test this, we examined the cell fate of LSCs from patients with aniridia, a disease mostly caused by PAX6 haploinsufficiency. We performed RNA-seq analysis of primary LSCs of 2 aniridia patients and postmortem cornea-extracted LSC controls. Aniridia and control LSCs were both expanded in keratinocyte serum-free medium (KSFM) conditions and were on passage 3 when processed. These data were integrated with other aniridia RNA-seq data published previously using the same procedures and culture conditions [49] (S1 Table). We obtained 73 differentially down-regulated genes (aniridia low genes) and 22 up-regulated genes (aniridia high genes) in aniridia patient LSCs, as compared to LSCs from healthy controls (S3 Fig). Aniridia low genes included the PAX6 target gene KRT12 and other corneal and epithelial genes such as TGFBI, CLND1, GJB6, IL36G, LAYN, NMU, and TMEM47. Many of these epithelial genes are potential PAX6 target genes reported in an immortalized LSC model where 1 allele of PAX6 was deleted [50]. We then applied GSEA to compare them to LSC and KC gene expression signatures, to investigate whether these deregulated genes due to PAX6 haploinsufficiency represent the changed cell fate of these aniridia LSCs. Indeed, as expected, aniridia low genes were enriched among genes expressed highly in LSCs (P-value 0.000028) (Fig 1F), indicating a loss of the LSC cell fate in aniridia LSCs. Among aniridia high genes, we found GATA3 present in genes expressed highly in KCs (S3 Fig). Nevertheless, there was no detected enrichment of aniridia high genes among genes expressed highly in KCs representing the KC cell fate, arguing against the postulated model that PAX6-deficient LSCs acquire a KC-like cell fate at the transcriptome level. These data suggest that additional mechanisms, such as TFs other than PAX6, contribute to the cell fate difference between KCs and LSCs.

The epigenetic states of cis-regulatory elements correlate with gene expression patterns

To understand the mechanisms underlying different cell fate controls of LSCs and KCs, we identified CREs and their epigenetic states that drive gene expression differences. We generated an extensive multi-omics dataset of LSCs and KCs and integrated these with other published data (S4A Fig and S2 and S3 Tables).

The complete dataset included ATAC-seq for open chromatin regions representing CREs [32] and ChIP-seq of histone modifications, H3K27ac and H3K4me3 marking active CREs, and H3K27me3 that marks repressed chromatin regions (Fig 2A) [17,32]. Using ATAC-seq analysis, we identified 124,062 CREs in the 2 cell types. Approximately 80% of these CREs were accessible in both cell types (S5 Fig). The public ATAC-seq data was compared to the data generated using Pearson correlation (S5F Fig). This comparison revealed a distinct biological difference between KCs and LSCs, which was more prominent than technical variation across different laboratories and techniques.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. CRE analysis.

(A) Schematic overview of CRE identification and quantification. Signals of each analysis were quantified by different window sizes covering the ATAC-seq peak summit. (B) Heatmap of the Z-scores of the quantile normalized ATAC-seq and histone mark signals near LSC- and KC-high genes. For promoter CREs, it corresponds to the closest CRE within 20 kb to the TSS. For enhancer CREs, the signals of all CREs within a 100-kb window near a TSS were quantified, distance weighted, and summed. (C) GO term enrichment of LSC- and KC-high genes and genes close (within 20 kb) to differential CREs. (D) PAX6 and GATA4 TSS loci with signals of RNA-seq, ATAC-seq, ChIP-seq of H3K27ac, H3K4me3, and H3K27me3 in KCs and LSCs. For the underlying data, see S5 Table, GEO GSE206918, GSE206920, and the trackhub in the Zenodo entry [51]. CRE, cis-regulatory element; KC, keratinocyte; LSC, limbal stem cell; TSS, transcription start site.

https://doi.org/10.1371/journal.pbio.3002336.g002

To examine the differential epigenetic states of these CREs in LSCs and KCs, we quantified ATAC-seq and histone modification signals in windows covering these CREs (Fig 2A). This resulted in 35,348 CREs with differential epigenetic signals, about 28.5% of CREs (S5 Fig). To assess the correlation between these differential CREs and the expression of their nearby genes, we considered both CREs at the promoter regions (promoter CREs) and enhancer CREs located within 50-kb distance from the genes (enhancer CREs) (Figs 2B and S4B). As expected, high ATAC, H3K4me3, and H3K27ac signals correlated with high gene expression, while the strong signals of repressive H3K27me3 correlated well with lowly expressed genes or genes with undetectable expression in the corresponding cell types, e.g., the loci of PAX6, GATA3, HOXA9, and TNF-α (Figs 2B, 2D and S4D). The strong repression signals marked by H3K27me3 in KCs at genes that are key for the LSC fate such as PAX6 suggest a repression mechanism in KCs to prevent inappropriate gene expression that defines the LSC fate. All annotated CRE regions are available on a publicly accessible trackhub [51] (see Data availability).

Next, we performed GO analysis on genes that are close to differential CREs (Fig 2C). For H3K27ac and H3K4me3 that mark active CREs, GO terms such as “epidermis” and “skin development” were identified for CREs with strong signals in KCs. Furthermore, “positive regulation of cell adhesion” and “extracellular matrix organization” terms were found for CREs with strong H3K27ac and H3K4me3 signals in LSCs. This is consistent with identified differentially expressed genes in each corresponding cell type. In contrast, the repressive mark H3K27me3 is anti-correlated with gene expression (Fig 2C).

Furthermore, in line with the enrichment of TNF-α and NF-kB signaling pathways in LSCs identified by PROGENy analysis of differentially expressed genes, higher H3K27ac and H3K4me3 signals were present in the loci of TNF-α and NF-kB target genes in LSCs, as compared to in KCs, while these loci in KCs were repressed by H3K27me3 (S4C Fig).

Gene regulatory network analysis identifies transcription factors controlling distinct epithelial cell fates and their hierarchy

Using the identified differential CREs, we set out to identify key TFs driving the cell fate differences between LSCs and KCs. TF-binding motif enrichment was performed using Gimme Motifs [52] in all differential CREs marked by ATAC, H3K4me3, H3K27ac, and/or H3K27me3 signals. In general, TF motifs enriched in CREs with active marks, ATAC, H3K4me3, or H3K27ac, in 1 cell type, were also enriched within regions with the repressive H3K27me3 mark in the other cell type (Figs 3A and S6A). For example, TF motifs that are linked to FOXC1, TEAD1, JUN, PAX6, FOS, RUNX2, OTX1, ELF3, SOX9, and REL were detected in differential CREs marked by high active mark signals in LSCs but also marked by high H3K27me3 in KCs. Consistent with our expectation, the enrichment of the PAX6 motif in differential CREs with higher active mark signals was detected in LSCs, as PAX6 is specific for LSCs but not for KCs. As REL is a TF involved in TNF-α and NF-kB pathways, the detection of the REL motif is consistent with the enrichment of TNF-α and NFKβ signaling genes among LSC-high genes. Notably, the FOS motif that is associated with FOS, FOSL1, FOSL2, and JUN (S6A Fig) was present in approximately 10% of all variable CREs in LSCs, the highest among all motifs (Fig 3A). Motifs enriched in KC active CREs included those linked to KLF6, GRHL1, HOXC10, GATA3, NFIA, CEBPA, and CTCF. These motifs were also enriched in CREs marked by high H3K27me3 signals in LSCs. Enriched motifs could mostly be linked to TFs with high expression differences between LSCs and KCs, e.g., FOXC1, PAX6, and FOS are highly expressed in LSCs, while HOXC10, GATA3, and CEBPA are highly expressed in KCs (Fig 3A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. TFs and TF hierarchy controlling distinct epithelial cell identity.

(A) Heatmap of motif enrichment Z-scores detected in variable CREs and the corresponding TFs. The percentage of CREs containing the motifs and the expression ratio of TFs in LSCs and KCs are indicated. (B) ANANSE influence score plot of TFs identified in ESC-KC (x-axis) and ESC-LSC (y-axis) comparison. Circle size represents the maximum number of target genes of a TF. The color represents log2FC between LSC/KC (orange LSC high; purple KC high). (C) ANANSE influence score plot of TFs identified in KC-LSC comparison. (D) ANANSE influence score plot of TFs identified in LSC-KC comparison. (E) TF hierarchy is indicated by the binding score of a TF to its target TF locus, and the cell type-specific regulation is indicated by the binding score difference of the TF at the target TF locus between cell types. When a binding score difference in KC-LSC comparison is greater than the mean of the difference in ESC-KC and ESC-LSC comparison, this TF regulation of the target TF is annotated as either KC- (purple arrows) or LSC-specific (orange arrows) regulation. Otherwise, the regulation is annotated as “shared regulation” for both cell types (gray arrows). The degree of binding score difference is indicated by the thickness of the arrows. Outdegree node size represents the number of target genes. Fold change of TF gene expression in LSC and KCs is represented by orange (LSC-high) and purple (KC-high) colors. For the underlying data, see the Zenodo entry [51]. CRE, cis-regulatory element; ESC, embryonic stem cell; KC, keratinocyte; LSC, limbal stem cell; TF, transcription factor.

https://doi.org/10.1371/journal.pbio.3002336.g003

To predict gene regulation by considering the expression of both TFs and their targets, we applied ANANSE [53], a gene regulatory network method to identify key TFs for cell identities. ANANSE integrates CRE activities and TF motif predictions with the expression of TFs and their target genes to generate a gene regulatory network of the specific cell type. Subsequently, a pairwise comparison of gene regulatory networks from 2 cell types is performed to identify the most influential TFs that differentiate between the 2 cell types. The overall importance of identified TFs is represented by their influence score, scaled from 0 to 1 (Materials and methods) [53].

Because many key TFs for LSC and KC fates that are shared in these 2 cell types such as p63 had similar gene expression levels (S1F Fig), these TFs could not be identified through a direct comparison of the differential gene regulatory network between these 2 cell types using ANANSE. Therefore, we decided to utilize embryonic stem cells (ESCs), a completely different cell type as compared to either LSCs or KCs, as the reference point for this differential network analysis. This enabled us to identify not only distinct but also shared TFs for LSC and KC fates. Using RNA-seq, ATAC-seq, and H3K27ac data from ESCs [54], we performed a pairwise comparison of gene regulatory networks between ESCs versus KCs or ESCs versus LSCs. When predicting the TFs driving the LSC or KC fates from ESCs, ANANSE resulted in 70 epithelial TFs that had influence scores above 0.5 in both ESC-LSC and ESC-KC pairwise differential network analysis. Many TFs are known to be important for epithelial cell function, such as TP63, EHF, TFAP2A, TFAP2C, FOSL2, the KLF family protein 3,4,5,6,7, JUNB, CEBPD, CEBPB, and RUNX1. We classified these TFs as shared epithelial TFs (Figs 3B and S7). Intriguingly, when conducting a network outdegree analysis that quantifies the number of targets of a TF within the top 20 TFs, FOSL2, JUN, TP63, and TFAP2A were identified as regulating most other TFs. This implies their importance in driving other TF expression (S7B and S7C Fig). The highest outdegree of FOSL2 is in line with the high percentage of detected FOS motif (Fig 3A).

Importantly, in the ESC-LSC and ESC-KC differential network analysis, TFs with high influence scores in LSCs but with undetectable (PAX6, ELF3, OTX1, PPARD) or low (FOSL1 and SMAD3) influence scores in KCs were considered as LSC-specific TFs (Fig 3B). Direct pairwise differential network analysis between LSCs and KCs (Fig 3C) was largely consistent with these findings. Interestingly, FOXC1 was annotated as a shared epithelial TF, whereas in the direct KC-LSC comparison, it was identified as a specific TF for the LSC fate. This is probably due to the higher expression of FOXC1 in LSCs. For KC-specific TFs, we only identified HOXA9 in the ESC-LSC and ESC-KC differential network analysis (Fig 3B). The direct pairwise differential network analysis between KCs and LSCs (Fig 3D) confirmed this. Next to HOXA9, the direct comparison identified other TFs such as GATA3, IRX4, and CEBPA with high influence scores in KCs, indicating that KC-LSC pairwise comparison is more sensitive for detecting KC-specific TFs.

Finally, we set out to dissect the TF regulatory hierarchy for the cell identity differences between LSCs and KCs, by identifying potential target TFs of the shared and specific TFs. For this analysis, we did not consider the expression level of TFs themselves, to ensure that the prediction is mainly driven by the potential binding of a TF to its target loci, represented by the binding score. If a target TF is regulated by a TF with similar binding scores in both LSCs and KCs compared to ESCs, this regulation is annotated as a “shared regulation”; if the binding score is significantly higher in one cell type than in the other (Materials and methods), the regulation of the TF-target TF pair is annotated as “cell type-specific regulation.” We included the top shared and specific TFs, 15 shared epithelial TFs, 6 LSC specific TFs (PAX6, ELF3, OTX1, PPARD, SMAD3, and FOSL1), and the KC-specific TF HOXA9 (Figs 3B, 3E, S6B, S6C and S6D). As expected, many shared TFs are regulating each other via “shared regulation” (Fig 3E, gray arrows). Consistently, cell type-specific TFs regulate their target TFs via “cell type-specific regulation” (Fig 3E, orange arrows), e.g., PAX6 is predicted to be regulated by SMAD3 and PPARD. Furthermore, many autoregulation loops were also detected, e.g., PAX6 in LSCs and HOXA9 in KCs. We also found that shared TFs may regulate cell type-specific TFs via “cell type-specific regulation.” For example, p63, FOXC1, and TFAP2A were identified as shared TFs between KCs and LSCs, but they were predicted to regulate PAX6 in LSCs.

In summary, our molecular characterization using KCs and LSCs cultured in vitro identified shared and cell type-specific TFs for the LSC and KC fates. p63, FOSL2, EHF, TFAP2A, KLF5, RUNX1, CEBPD, and FOXC1 are among the shared epithelial TFs for both LSCs and KCs. PAX6, SMAD3, OTX1, ELF3, and PPARD are LSC specific TFs for the LSC fate, and HOXA9, IRX4, CEBPA, and GATA3 were identified as KC-specific TFs. Furthermore, LSC and KC fates are defined by cooperative regulation of both shared and cell type-specific TFs.

Single-cell RNA-seq analysis of the cornea and the epidermis validates the expression of key transcription factors controlling cell fates

Since our multi-omics analysis was performed on LSCs and KCs cultured in vitro, we assessed scRNA-seq datasets derived from the cornea and the epidermis to confirm that the molecular signatures of LSCs and KCs in our study indeed represent those of epithelial stem cells maintaining the corneal limbus and the epidermis [55,56]. By clustering single cells according to marker gene expression (S8 Fig), we selected the cell clusters corresponding to the stem cells as pseudobulk for further differential gene expression analysis. For the epidermis, we selected cells with high KRT14, KRT5, and low KRT1 and KRT10 expression as basal KCs, and for the cornea, cells with high S100A2, PAX6, and TP63 expression and without CPVL expression as LSCs, because CPVL has been proposed as a marker with neural crest origin [56].

Consistent with the in vitro findings, the in vivo LSCs expressed high levels of PAX6, ELF3, FOXC1, FOSL1, OTX1, and SMAD3, whereas the in vivo KCs expressed high levels of HOXA9, CEBPA, and GATA3 (Fig 4A). GO analysis identified similar functions of differentially expressed genes between the in vivo LSCs and KCs, as compared to those from in vitro cultured cells (S8E and S8F Fig). Furthermore, PROGENy analysis of differentially expressed genes between in vivo LSCs and KCs showed that TNF-α and NF-kB pathway genes are significantly enriched in in vivo LSCs, e.g., CXCL1,2,3,8,20 and NFKB1 (Fig 4B). GSEA analysis using the hallmark gene set of the MsigDB collection also identified enrichment for TNF-α signaling genes (S5 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Validation of key TF expression using in vivo scRNA-seq.

(A) Fold change comparison of identified TFs using in vivo and in vitro data. (B) PROGENy pathway analysis of in vivo LSCs and KCs. (C) ANANSE influence score plot of in vivo basal KCs to LSCs. (D) ANANSE influence score plot of in vivo LSC to basal KCs. (E) Summary of the identified shared and specific TFs. For the underlying data, see S6 Table, GEO GSE155683, GSE147482 [55,56], and the Zenodo entry [51]. KC, keratinocyte; LSC, limbal stem cell; scRNA-seq, single-cell RNA-seq; TF, transcription factor.

https://doi.org/10.1371/journal.pbio.3002336.g004

As the data of CREs of in vivo tissues were not available, we utilized the in vitro ATAC and H3K27ac datasets for differential gene regulatory network analysis, together with the in vivo pseudobulk data used for gene expression analysis above. Here, we assumed that, since the GRN analysis is largely driven by gene expression data, this analysis assessed the influence of TFs on in vivo LSC and KC fate differences. Overall, the in vivo data identified similar cell type-specific TFs, as compared to in vitro cultured cells (Fig 4C and 4D). In in vivo LSCs, PPARD however was not detected, PAX6, ELF3 FOXC1, and FOSL1 exhibited the highest influence scores, and TFs with the highest influence scores in in vivo basal KCs were HOX TFs and a few others such as CEBPA, GATA3, and IRX4.

Taken together, our analyses showed clear consistency between in vivo and in vitro derived data and identified key TFs driving the cell fate of between LSCs and KCs (Fig 4E).

FOSL2 is a novel transcription factor controlling the LSC fate and associated with corneal opacity

As several key TFs defining the LSC fate, either shared or cell type specific, such as p63 and PAX6, respectively, are associated with corneal opacity, we explored the potential involvement of LSC TFs in the pathomechanism of corneal diseases. For this, we leveraged the whole genome sequencing data in the 100,000 Genomes Project at Genomics England UK to identify variants of uncertain significance that may have functional consequences. For establishing a suitable cohort, we identified a total number of 33 unsolved participants with human phenotype ontology (HPO) terms associated with corneal opacity (S8 Table). Next, we screened for variants in the coding regions of the top 15 shared epithelial TFs as well as the 6 LSC-specific TFs identified in our study. In a proband with band keratopathy (HP:0000585), we identified a rare de novo heterozygous missense variant in FOSL2 (2:28412095:C:T, genome build GRCh38/hg38, NM_005253.4:c.628C>T) (genomAD allele frequency 0.0000922) (S9A Fig), giving rise to a predicted damaging amino acid change (NP_005244.1:p.(Arg210Cys)), based on most major prediction tools (S9A Fig and S9 Table).

As the role of FOSL2 in LSCs is completely unknown, we first investigated the protein expression of FOSL2 within the cornea. The immunostaining results showed that FOSL2 protein is expressed in both limbus and central cornea (Figs 5A and S9B), co-localizing in the nucleus with other LSC TFs, e.g., p63, PAX6, and its closely related AP1 complex factor FOSL1 (S9C–S9F Fig). The same was also visible in LSCs (and KCs, with the exception of PAX6) cultured in vitro (S9G–S9J Fig). The identified variant associated with corneal opacity, together with the expression pattern, indicated an important function of FOSL2 in the cornea.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Transcription factors regulation of corneal disease-associated genes.

(A) FOSL2 and p63 staining of the peripheral cornea. (B) TFs that bind to gene loci associated with corneal abnormalities with significantly higher occurrence (FDR), as compared to binding to all genes in the whole genome. FDR was calculated with Fisher exact testing. (C) Dot plot showing the TF binding intensity (color bar) and the number of binding peaks (npeaks, dot size) near the disease genes that contain a significant number of potential TF binding. The number of peaks is within a 100 kb region of the TSS; TF binding intensity score is the weighted z score of the quantile log normalized intensities distance weighted. FECD, Fuchs Endothelial Corneal Dystrophy. For underlying data see S7 Table, GEO GSE206920, GSE236440, and GSE156272. TF, transcription factor; TSS, transcription start site.

https://doi.org/10.1371/journal.pbio.3002336.g005

To further explore the association of FOSL2 with corneal diseases, we questioned whether the target genes of FOSL2 are enriched for corneal disease genes, similar to other known LSC TFs such as PAX6, FOXC1, and p63. The rationale behind this question is based on the concept that many disease-associated TFs could contribute to the diseases not only because of their mutations but also due to abnormal regulation of their target genes [29]. Here, our analyses investigated a broader scope of corneal diseases, because many of these known disease-associated TFs can influence multiple tissues of the eye. For this analysis, we employed 2 types of data, DNA-binding profiles of key LSC TFs to predict their target genes and a comprehensive list of corneal disease genes. For TF DNA-binding profiles, we generated cleavage under targets and release using nuclease (CUT&RUN) of FOSL2 and ChIP-seq of the p63 protein in LSCs and additionally incorporated publicly available ChIP-seq data of PAX6, FOXC1, RUNX1, and SMAD3 in LSCs [30,32]. Except for FOSL2, all other TFs are known to regulate the LSC fate. To collect genes that are associated with corneal phenotypes, we used 2 approaches. First, we performed a literature search and constructed a list of 161 genes associated with LSCD, inherited corneal opacification diseases, and ocular as well as systemic syndromes with known corneal manifestations (S7 Table). In addition to this curated disease gene list, we used disease genes assembled in the EyeDiseases database [57]. To assess whether TF binding of the key TFs to corneal disease gene loci is more probable than random, we mapped TF binding sites to the nearest genes in order to detect a statistically significant enrichment of disease genes bound by TFs (S10A Fig). With this method, we identified that the genes associated to our “curated corneal disease list” were significantly more likely to be bound by PAX6, FOXC1, RUNX1, and FOSL2 (Figs 5B and S10C). The binding signal of these TFs within the transcription start site (TSS) regions of the top cornea disease-related genes is summarized by the quantification of binding signals (Fig 5C). In parallel, we investigated the TF-disease association by combining TF binding signals with the distance between the TF binding peaks and the disease genes and employed a Mann–Whitney U statistical test to identify a potential binding increase (S10B Fig). This gave rise to similar TF-disease association enrichment results (S10C Fig).

Overall, our analysis highlighted that FOSL2 and several other key LSC TFs binding is enriched in the TSS regions of cornea disease genes, suggesting that they directly regulate cornea disease genes.

FOSL2 is a direct PAX6 target gene and regulates angiogenesis and tight junction genes in LSCs

To further assess the function of FOSL2 in LSCs and corneal opacity, we performed siRNA knockdown of FOSL2 in primary LSCs. Knockdown of FOSL2 was validated using quantitative polymerase chain reaction (qPCR) and RNA-seq and displayed an efficiency of over 80%, as compared to control siCTR (0.12-fold ± 0.068, n = 4) (Figs 6A, S11A and S11B), and gave rise to 212 differentially expressed genes. Because PAX6 is an undisputed regulator in LSCs and defects are associated with LSCD and corneal opacity, we also performed PAX6 knockdown (siPAX6). Intriguingly, FOSL2 expression was significantly down-regulated upon PAX6 knockdown, as compared to siCTR (0.5-fold) (Fig 6A). Furthermore, several high-confidence PAX6 binding peaks were identified surrounding the FOSL2 locus (Fig 6B), demonstrating that FOSL2 is a direct target gene of PAX6 in LSCs.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. FOSL2 siRNA knockdown results in deregulation of angiogenesis and tight junction genes.

(A) Normalized transcriptome difference of FOSL2 and PAX6 upon knockdown of FOSL2 (siFOSL2) and PAX6 (siPAX6) (* pval<0.05, ** pval<0.01, *** pval<0.001, paired DESEQ2 differential expression testing). (B) FOSL2 TSS locus with PAX6 binding signal, ATAC seq, and H3K27ac ChIP-seq in LSCs. (C) Heatmap of differentially expressed genes in siFOSL2, with Zscore plotted. (D) GO term enrichment of genes up-regulated in siFOSL2 (E) GSEA enrichment analysis based on the expression foldchange (siFOSL2/siCTR) of genes with nearby FOSL2 binding signal. (F) Expression of JAG1, TGFBI, CLDN1, and CLDN4 were measured in control LSCs (siCTL) and FOSL2 siRNA-knock down (siFOSL2) samples (n = 4). Values represent the fold change difference of siFOSL2/siCTR and were normalized to internal housekeepers GAPDH and ACTB (* pval<0.05, ** pval<0.01, *** pval<0.001, unpaired t test analysis). The data underlying this figure can be found at S10 Table. (G) JAG1, TGFBI, and CLDN1 TSS loci with FOSL2 CUT&RUN signal, ATAC-seq, and H3K27ac ChIP-seq in LSCs. For the underlying data see S10 Table, GEO GSE242990, GSE236440, and the Zenodo entry [51]. CUT&RUN, cleavage under targets and release using nuclease; GO, Gene Ontology; GSEA, gene set enrichment analysis; LSC, limbal stem cell; TSS, transcription start site.

https://doi.org/10.1371/journal.pbio.3002336.g006

Among the 212 differentially regulated genes detected in siFOSL2 (Fig 6C), 108 were up-regulated, with significant enrichment of genes involved in “negative regulation of cell population proliferation,” “blood vessel morphogenesis,” “angiogenesis,” and “endothelial cell differentiation” (Fig 6D). Furthermore, predicted FOSL2 target genes based on the nearby FOSL2 binding sites were enriched among up-regulated genes by siFOSL2 (Fig 6E). This indicates that FOSL2 may function primarily as a repressive factor in LSCs. Relevant up-regulated genes included TGFBI and JAG1 which have known roles in angiogenesis and CLDN1 which is involved in tight junctions (Fig 6F). Furthermore, FOSL2 binding signals were identified near these genes (Fig 6G), indicating that these genes are direct targets of FOSL2.

Interestingly, 103 down-regulated genes upon siFOSL2 did not give any enrichment in GO analysis, suggesting a less prominent role of FOSL2 in gene activation. Nevertheless, CLDN4 and CLDN7, other tight junction genes, are significantly down-regulated (Figs 6C, 6F and S11G), suggesting an altered barrier function in FOSL2-deficient LSCs. Other down-regulated genes include TGM1 and ABCA12, linked to epidermal permeability barrier disorders (S11G Fig).

To summarize, our results demonstrated that, among TFs that define the LSC fate, FOSL2 is a novel LSC TF associated with cornea opacity.

Ethical statement

All procedures for establishing and maintaining human primary keratinocytes were approved by the ethical committee of the Radboud University Medical Center (“Commissie Mensgebonden Onderzoek Arnhem-Nijmegen”) (CMO-nr:2004/132). All donors have given informed consent in written forms.

All procedures for establishing and maintaining LSCs were conducted according to the principles expressed in the Declaration of Helsinki. The use of LSCs extracted from corneal scleral donor rims was approved by the Ethics Committee of the Saarland (Number 226/15 and 110/17). Aniridia patients consented in written form to research limbal biopsy during ocular surgery.

Permission to use the donor corneas for research purposes was given by the donor’s next of kin by signing a written consent form released by the Regional Transplant Service and the National Transplant Service, without ethics committee approvals. According to Article 4 of the Italian Law 91 of 1st April 1999, donor corneas that cannot be used for transplantation (i.e., unsuitable for biological reasons or anamnestic reasons) can be used for research purposes if the aim is to ameliorate corneal transplantation or progress towards a cure of corneal diseases. The human corneas used in this study were unsuitable for transplantation and obtained by Fondazione Banca degli Occhi del Veneto (www.fbov.org; Venice, Italy) for research purposes.

KC and LSC cell culture in vitro

KCs were isolated and cultured as previously described [16]. Briefly, after isolation, primary KCs were cultured in Keratinocyte Basal Medium supplemented with 100 U/ml Penicillin/Streptomycin, 0.1 mM ethanolamine, 0.1 mM O-phosphoethanolamine, 0.4% (vol/vol) bovine pituitary extract, 0.5 μg/ml hydrocortisone, 5 μg/ml insulin, and 10 ng/ml epidermal growth factor. Medium was refreshed every other day until the cells were 90% confluent.

Limbal tissues were acquired as previously described [91]. Two aniridia limbal tissue single biopsies were obtained from the superior limbus during penetrating keratoplasty from 2 patients with congenital aniridia as previously described [92]. Genetics of the aniridia patients were identified to be c.33delC p.Gly12Valfs*19 (NM_000280.2) for AN55 and c.990_993dup p.Met332Alafs*10 for AN40 (S1 Table). Both patients their corneas were in Lagali AAK-Stage 4 during the Keratoplasty [31].

Cell isolation was performed as previously described [91]. Briefly, limbal tissue was digested in collagenase A solution (4 mg/ml) in KSFM (Thermo Fisher Scientific; Waltham, Massachusetts, United States of America) for 20 h at 37°C. Cell suspensions were filtered through a use of Flowmi micro strainer (SP Bel-Art; Wayne, New Jersey, USA). LSC clusters were dissociated with trypsin-EDTA (0.05%) solution and cultivated in KSFM. Medium was refreshed every other day. Subconfluent (80% to 90%) limbal epithelial cells were harvested at passage 3. Aniridia LSCs were harvested at passage 3.

Next to this approach, other LSC samples (LSC-Aberdam, S1 Table) were isolated from postmortem donated peripheral corneal epithelium and cultured as previously described [50]. Briefly after isolation, they were expanded and cultured in KSFM (Gibco, Life Technologies) supplemented with 25 μg/ml bovine pituitary extract (BPE; Gibco, Life Technologies), 0.2 ng/ml epidermal growth factor (EGF, Peprotech, Neuilly-sur-Seine, France), 0.4 mM CaCl₂, 2 mM Glutamine (Gibco, Life Technologies), and 100 U/ml Penicillin/Streptomicin (Gibco, Life Technologies). Medium was refreshed every other day until the cells were 90% confluent.

Finally for CUT&RUN of LSCs, LSCs were isolated and cultured as described previously [93,94]. Briefly, corneal limbal tissue was dissected from human donor corneas, cut into small pieces, and incubated with 10 ml trypsin (0.05% trypsin—0.01% EDTA solution, Life Technologies, 25300–062) for 30 min at 37°C for 4 consecutive times. Every time, the supernatants containing the cells were collected, neutralized, centrifuged at 1,000 rpm for 5 min and cells co-cultured with a feeder layer (consisting of lethally irradiated 3T3-J2 cells) in LSC growth medium. The LSC growth medium contains 2:1 Dulbecco’s Modified Eagle Medium (Life Technologies, 21969035) and F12 (Life Technologies, 21765029), 10% Fetal Bovine Serum (Life Technologies, 10099–141), 50 mg/ml penicillin-streptomycin (Life Technologies, 15140122), 4 mM glutamine (Life Technologies, 25030081), 0.18 mM adenine grade I (Pharma Waldhof GMBH, 4010-21-2), 0.4 mg/ml hydrocortisone (Flebocortid Richter, Sanofi, AIC013986029), insulin (Humulin R, Lilly, Canada, HI0210), 2 nM triiodothyronine (Liotir, IBSA, AIC036906016), and 8.1 mg/ml cholera toxin QD (List Biological Laboratories, 9100B). Cultures were incubated at 37°C with 5% CO₂. Medium was changed every 2 days with further addition of 10 ng/ml EGF (Cell Genix GmbH, Germany 1416–050). Cultures were switched to KSFM and used after 3 passages.

siRNA treatment

Primary human limbal epithelial stem cells were isolated from healthy donors and cultivated in 6-well plates in KSFM medium (Cat. Nr. 17005042, Thermo Fisher, Gibco, Life Technologies, Paisley, United Kingdom) supplemented with 500 ng EGF, 12.5 mg BPE (Cat. Nr. 37000015, Thermo Fisher, Gibco, Life Technologies, Paisley, UK) and 0.1% P/S. After reaching 70% to 80% confluence, cells were transfected using Lipofectamin 2000 transfection reagent. For this purpose, 5 μl Lipofectamin was added to 250 μl Optimem and incubated for 5 min, at room temperature (RT). In a separate tube, siRNA of interest (information and concentrations used are summarized in Table 1) was diluted in 250 μl Optimem. The Lipofectamin was added to the siRNA mixture and incubated for 20 min, RT. After the incubation time, transfection mixture was added dropwise to the cells. The cells were incubated at 37°C and 5% CO2 and medium were changed after 24 h. Cells were collected 48 h after transfection for further analysis.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. siRNA oligonucleotide sequences for FOSL2 and PAX6 knockdown.

https://doi.org/10.1371/journal.pbio.3002336.t001

Bulk RNA-seq

Total RNA was isolated using the Quick–RNA MicroPrep kit (Zymo Research), according to the manufacturer’s protocol. RNA concentrations were measured using the the DeNovix DS-11FX spectrometer, and 500 ng of RNA was prepared for sequencing using the KAPA RNA HyperPrep Kit with RiboErase (Kapa Biosystems). Libraries were sequenced on the NextSeq 500 (Illumina), generating an average of 15 to 20 million reads per sample.

Single-cell RNA-seq library preparation and sequencing

A single-cell suspension was made using trypsin, after which cells were filtered using a 40 μm filter to remove cell clumps. Cells were stained with 7-AAD. The live cells were selected for and FACS-sorted onto 384-well plates containing primers with unique molecular identifiers, according to the SORT-Seq protocol [95]. Plates were spun down (1,200 × g, 1 min, 4°C) and ERCC spike-in mix (1:50,000) was dispensed by a Nanodrop (BioNex) into each well, and 150 nl of the Reverse Transcription (RT) mix was dispensed into each well. Thermal cycling conditions were set at 4°C 5 min; 25°C 10 min; 42°C 1 h; 70°C 10 min. The library of each plate was pooled together and the cDNA was purified using AmpureXP (New England BioLabs) beads. Overnight in vitro transcription (Ambion MEGA-Script) was carried out at 16°C, with the lid set at 70°C. An exonuclease digestion step was performed thereafter for 20 min at 37°C, followed by fragmentation of the RNA samples. After a beads cleanup, the samples were subjected to library RT and amplification to tag the RNA molecules with specific and unique sample indexes (Illumina), followed by a final beads cleanup (1:0.8, reaction mix: beads) and the sample cDNA libraries were eluted with DNAse free water. Libraries were quantified using the KAPPA quantification kit following manufacturer’s protocol after which the plates were sequenced on the NextSeq 500 (Illumina) for 25 million reads per plate.

ChIP-seq

Chromatin for ChIP was prepared as previously described [96] with minor modifications. On average, 0.5 M cells were used in each ChIP. Antibodies against H3K27ac (Diagenode #C15410174, 1.2 μg), H3K4me3 (Diagenode #C15410003, 1 μg), H3K27me3 (Diagenode #C15410069, 1.5 μg), p63 (Santa Cruz #H129, 1 μg, recognizing the C-terminal α tail of p63) were used in ChIP assay. Afterwards, 5 ng DNA fragments were pooled and proceeded on with library construction using KAPA Hyper Prep Kit (Kapa Biosystems #KK8504) according to the standard protocol. The prepared libraries were then sequenced using the NextSeq 500 (Illumina) according to standard Illumina protocols.

Immunofluorescence stainings

Human donor corneas were fixed with 4% paraformaldehyde overnight at 4°C, soaked in increasing gradients of sucrose solutions (7.5%, 15%, and 30% in PBS, at least 30 min each at 4°C, the last one overnight) and then embedded in OCT compound, frozen, and cut into 5- to 7-mm sections. Sections were permeabilized in a solution containing 0.5% Triton X-100 for 10 min, incubated in 5% BSA (Sigma-Aldrich A7906) for 1 h at RT and then incubated overnight at 4°C with primary antibodies. The following primary antibodies (all diluted at 1:100) were tested: PAX6 (rabbit, Biolegend Poly19013); FOSL1 aka Fra-1 (C-12) (mouse, Santa Cruz Biotechnology, sc-28310); Anti-FOSL2 (rabbit, Sigma, HPA004817); p63 (mouse, DAKO, M7317). Sections were extensively washed in 1× PBS and incubated with secondary antibodies (Alexa Fluor 488 anti-rabbit A11008, Rhodamine Red-X goat anti-mouse IgG R6393, Rhodamine Red-X goat anti-rabbit IgG R6394, Alexa Fluor 488 goat anti-mouse A11001, all from Invitrogen, all diluted at 1:500) for 45 min at RT. Sections were eventually mounted with medium containing DAPI Fluoromont-G (EMS, Società Italiana Chimici, Rome, Italy, 17984–24) and pictures taken and evaluated through an Eclipse Ti Nikon microscope (Nikon, Amstelveen, the Netherlands).

LSCs and primary keratinocytes were grown in cell culture chamber slides (Corning, New York, USA), and, when confluent, fixed using 4% PFA solution. Immunostainings were performed as previously described [97]. Briefly, cells were washed, permeabilized for 15 min with 0.1% Triton X-100, washed again, and blocked for 2 h in 1% FBS. Primary antibody incubation was done overnight at 4°C with the following antibodies diluted in blocking buffer: PAX6 (1:300, 901301 Biolegend), P63 (1:100, sc8344 SCBT), FOSL1 (1:100, sc20310, SCBT), and FOSL2 (1:200, HPA004817 Sigma Aldrich). Cells were washed thoroughly and incubated with secondary antibodies AlexaFlour 488 (1:400, A11008 Invitrogen) and AlexaFluor 647 (1:400, A31571 Invitrogen, California, USA). Slides were mounted with Vectashield (VectorLabs, California, USA) and analyzed with Zeiss AI Sample Finder (Zeiss, Jena, Germany). Fiji/ImageJ software (National Institutes of Health, Maryland, USA) was used for overlay images preparation.

CUT&RUN

CUT&RUN was performed using the CUT&RUN Assay Kit (Cell Signaling Technology #86652) and according to the manufacturer instructions. Briefly, 200,000 live primary LSCs were used per immunoprecipitation (IP). For FOSL2 IPs, 10 ml of antibody (D2F1E, Rabbit mAb #19967) was used per IPs, 2 IPs were performed per replica to ensure a sufficient amount of DNA. The negative control IgG included in the kit was used to check for enrichment by quantitative PCR.

For high-throughput sequencing, libraries were prepared using KAPA HyperPrep Kit (Kapa Biosystems #KK8504) using 5 ng of DNA. Libraries concentration and fragment repartition were checked by automated electrophoresis (Bioanalyzer Instrument, Agilent). Libraries were sequenced using the NextSeq 500 (Illumina) and according to their protocol.

Reverse transcription quantitative PCR (RT-qPCR) validation

Approximately 250 ng RNA were used for cDNA synthesis using the iScript cDNA synthesis kit (BioRad, California, USA). Quantitative PCR (qPCR) was carried out using 4 μl cDNA (diluted 1:20) with iQ SYBR Green Supermix (BioRad, California, USA) and specific primers (Biolegio, Nijmegen, NL):

GAPDH Forward (Fw): AAG GAG TAA GAC CCC TGG ACC A

GAPDH Reverse (Rv): GCA ACT GTG AGG AGG GGA GATT

ACTB Fw: TTC TAC AAT GAG CTG CGT G

ACTB Rv: GGG GTG TTG AAG GTC TCA AA

JAG1 Fw: GAT CGC CTG CTC AAA GGT CT

JAG1 Rv: GAC TGG AAG ACC GAC ACT CG

TGFBI Fw: CCT TTG AGA CCC TTC GGG CTG

TGFBI Rv: TCG AAG GCC TCA TTG GTC GG

CLDN1 Fw: CCC CAG TCA ATG CCA GGT ACG

CLDN1 Rv: TCG GGG ACA GGA ACA GCA AA

OCLN Fw: TCT AGG ACG CAG CAG ATT GGT

OCLN Rv: TCA GGC CTG TAA GGA GGT GG

All disease gene qPCRs were performed using a BioRad CFX96 system following standard cycling conditions. Data were analyzed by the ΔΔCT method, normalized to housekeepers GAPDH and ACTB.

FOSL2 and PAX6 siRNA RT-qPCR

RNA extraction were performed using RNA isolation kit (RNA Purification Plus Micro Kit, Cat. Nr. 47700, (Norgen Biotek CORP. Canada)) according to the manufacturer’s instructions. Optional DNA column digestion was performed. The eluted RNA samples were stored at −80°C until further use. NEB One Taq RT-PCR kit (One Taq RT-PCR Kit, New England Biolabs, Frankfurt, Germany) was used for cDNA synthesis, following the manufacturer’s instructions. Samples were stored at −20°C. The qPCR mix contained 1 μl of the specific primer solution, 5 μl SYBR Green Mix (Qiagen N.V., Venlo, the Netherlands), and 3 μl nuclease-free water (total volume: 9 μl). To investigate the mRNA expression levels, 1 μl cDNA of interest was added to the qPCR mix. The “QuantStudio 5 real-time PCR system” (Thermo Fisher Scientific, Waltham, Massachusetts, USA) was used for the qPCR reactions. The amplification conditions were 95°C for 10 s, 60°C for 30 s, and 95°C for 15 s (40 cycles). Samples were measured in duplicates and experiments were repeated for at least 5 donors. Values were normalized to the expression value of TATA-Box binding protein as an endogenous control gene using the ΔΔCT method and the fold change (2ΔΔCT-value) was used for statistical analysis. All primers were purchased from Quantitect Primer assay (Qiagen N.V., Venlo, the Netherlands; PAX6: QT00071169, FOSL2: QT01000881, and TBP: QT00000721).

RNA-seq, ATAC-seq ChIP-seq, and CUT&RUN data preprocessing

Preprocessing of reads was done automatically with workflow tool seq2science v0.7.1 [98]. Paired-end reads were trimmed with fastp v0.20.1 [99] with default options. Genome assembly GRCh38.p13 was downloaded with genomepy 0.11.1 [100]. Public samples were downloaded from the Sequence Read Archive [101] with help of the NCBI e-utilities and pysradb [102]. The effective genome size was estimated per sample by khmer v2.0 [103] by calculating the number of unique kmers with k being the average read length per sample. scATAC fastq files were merged to pseudobulk by combining all fastq files from each plate using the bash command cat.

Reads of ChIP-seq, CUT&RUN, and ATACseq were aligned with bwa-mem v0.7.17 [104] with options ’-M’.

Reads of RNAseq samples were aligned with [105] with default options. Afterwards, duplicate reads were marked with Picard MarkDuplicates v2.23.8 [106]. General alignment statistics were collected by samtools stats v1.14 [107]. Mapped reads were removed if they did not have a minimum mapping quality of 30, were a (secondary) multimapper or aligned inside the ENCODE blacklist [108]. RNAseq sample counting and summarizing to gene level was performed on filtered bam using [109]. Sample sequencing strandness was inferred using [110] in order to improve quantification accuracy.

ATAC samples were tn5 bias shifted by seq2science. ChIP, CUT&RUN, and ATAC sample peaks were called with macs2 v2.2.7 [111] with options ’—shift -100—extsize 200—nomodel—keep-dup 1—buffer-size 10000’ in BAM mode. The effective genome size was estimated by taking the number of unique kmers in the assembly of the same length as the average read length for each sample. Narrowpeak files of ChiP-seq and CUT&RUN biological replicates belonging to the same condition were merged with the irreproducible discovery rate v2.0.4.2 [112]. ATAC-seq samples were correlated seq2science its DESeq2 reads per peak spearman correlation clustering (S5F Fig).

Single-cell RNA-seq data preprocessing

Single-cell libraries were preprocessed using the cellseq2 pipeline. Briefly, reads were aligned using star to the GRCh38.p13 genome. After which cells were quality controlled using Seurat, filtering cells on ERCC reads, genes measured and transcripts per cell. After visualization of the lack of heterogeneity by Umap, pseudobulk count data was generated by summing all the cells their UMI counts. Cellular heterogeneity was assessed using the analysis file Generate_scRNAseq_pseudobulk.Rmd. Finally, single-cell and bulk gene count tables were merged for a combined bulk and pseudobulk analysis.

RNA-seq data analysis and normalization

The bulk and pseudobulk count tables were merged on gene names, keeping all gene names detected. Due to potential sex differences between donors, genes located on chromosome X and Y were removed. Finally, genes with less than 10 counts per row were removed. Variance visualization was performed using sample distance and PCA. For quality control, sample variance and distance were visualized before and after removing technical variance due to different sequencing methods. Limma [113] was used to remove these batch effects.

Rld normalization was used for normalizing gene intensities. Between all conditions differentially expressed genes were detected using Deseq2 [114]. Non-batch corrected count tables were used for identifying the differentially regulated genes. Ashr log2 fold change shrinkage [115] was used to shrink the Log2 fold change values. Differentially regulated gene detection cutoffs were set as an adjusted p-value of 0.01 or lower, and an absolute log2 FC of 0.58 and larger.

The packages complex Heatmaps [116] and circlize [117] were used to visualize the differentially expressed genes. Subsequently, Progeny enrichment [47] was performed to quantify signaling pathway target gene enrichment. Clusterprofiler [44] was run for GO term enrichment on differentially expressed genes of each comparison. Finally, foldchange of all genes were used to generate a gene list for GSEA enrichment of the MSigDB collections [46]. Gene names were mapped to ENTREZID using AnnotationDbi [118] and these were used to run KEGG pathway enrichment. The enriched pathways were visualized using pathview [119].

Identification of CREs

In order to identify CREs, ATAC-seq was used. Bulk and scATAC data were merged from in vitro expanded KCs and LSCs. Next to the generated datasets, publicly available data were incorporated. To prevent a sequencing depth bias, the top 100.000 ATAC peaks from each cell type were combined, the overlapping peak summits were merged, and histone modifications in varying window sizes around these ATAC peaks were quantified using histone ChIP-seq datasets (Fig 2A). For ATAC signal quantification, the ATAC intensity was quantified in 200 bp around the peak summits, and for the promoter mark H3K4me3 and the enhancer mark H3K27ac, a 2kb window was used and finally for the repressive H3K27me3 mark, a 5 kb window was used for quantification (Fig 2A). This resulted in an extensive dataset containing CREs and their respective histone modification signal intensity.

Differential CREs were identified for the ATAC-seq and H3K27ac reads by running DESEQ2 on the read counts within the defined windows and identified regions (adjusted p-value <0.05). For H3K4me3 and H3K27me3 signals, differential CREs were identified with 2 steps. First, the histone mark distribution was plotted, and CREs with a low to no histone signal were disregarded. Next, high activity regions with variable signal were selected (S5 Fig).

Variable CREs were linked to genes with 2 approaches:

1. CREs were linked to all the TSS regions within a 100 kb window using bedtool window [120] after which the CREs per TF were distance weighted and summed based on the ANANSE distance weighing approach including promoter peaks [53].
2. CREs were linked to the closest TSS region within 20 kb using bedtool closest [120].

After linking the regions to genes in both approaches, intensity scores were printed to a CSV file. Heatmaps and GO term enrichments were generated in R using clusterprofiler and complex heatmap.

Single-cell ATAC-seq

A single-cell suspension was made using trypsin. After which cells were filtered using a 40 μm filter to remove cell clumps. The protocol on from Chen and colleagues [121] was used to sequence single-cell ATAC. Briefly, 50.000 cells were tagmented in bulk in 45 μl of tagmentation mix (20 mM Tris (pH 7.6), 10 mM magnesium chloride 20% Dimethylformamide), 5 μl of tagmentation protein and 0.25 μl of Digitonin. Cells were tagmented for 30 min at 37°C and 800 rpm. Tagmentation was stopped by adding 50 μl of tagmentation stop buffer (10 mM Tris-HCL (pH 7.8) and 20 mM EDTA). Cells were stained with DAPI and DAPI positive cells were FACs sorted in 384-well plates containing Nextera primers with unique molecular identifiers, NACL ProteinaseK and SDS page. Plates were spun down (1200 × g, 1 min, 4°C) and were incubated for 15 min at 65°C; 4 μl of Tween20 was added, and 2 μl of H20 was added, and finally 10 μl of NEBNext High-Fidelity 2X PCR Master Mix was added to each well. Thermal cycling conditions were set at 72°C for 5 min, 98°C for 5 min, and then, 20 repeats of 98°C for 10 s, 63°C for 30 s, and 72°C for 20 s.

Plate libraries were pooled and purified using a Qiagen PCR purification kit with adjusted buffer volumes according to Chen and colleagues [121]. After the column cleanup, a final beads cleanup was performed using AmpureXP (New England BioLabs) beads and the sample cDNA libraries were eluted with DNAse free water. Libraries were quantified using the KAPPA quantification kit following manufacturer’s protocol after which the plates were sequenced on the NextSeq 500 (Illumina) for 30 million reads per plate.

Motif analysis

The Gimme motifs database was pre-filtered [52] to include only motifs linked to TFs which were expressed in either KC and/or LSCs (using a cutoff of at least 10 counts in total). When multiple motifs mapped to a TF, the most variable motif was used. In case multiple TFs mapped to a motif, the most differential TF on the transcriptome was annotated to the motif. All highly variable CREs their log10 quantile normalized values were used as an input for Gimme maelstrom motif enrichment analysis.

ANANSE analysis

For the gene regulatory network analysis, all the called ATAC peaks were used, merging summits and excluding peaks on the chromosomes GL, Un, KI, MT, X, and Y due to potential donor sex differences. Next, ANANSE binding was ran using all the peaks as potential enhancer regions and using both ATAC and H3K27ac signals to predict potential TF binding. To select the TF binding model, a Jaccard similarity score of min 0.2 was used, to minimalize the false-positive models used to predict TF binding. For ANANSE network, the ANANSE binding files were combined with the RNAseq TPM files. This included all bulk RNAseq samples of KCs, LSCs, and ESCs (S1 Table). In the case of the in vivo pseudobulk data, FPKM values were used based on the UMI tables.

Finally, ANANSE influence was ran using the top 500.000 differential edges between networks. Deseq2 was ran on the countfiles of each comparison to identify differential genes needed for ANANSE influence. To prevent missing values, for the final ESC-KC, ESC-LSC, KC-LSC, and LSC-KC comparison, all differential edges were taken from each comparison and used to reran each comparison with all these edges included. This prevented missing values in the differential networks while comparing different differential networks.

TF hierarchy was estimated using the TF-target TF binding score generated by the influence command running the–full-output flag. This represents the motif, ATAC&H3K27ac signal intensity in the target TF locus and is excluding the difference in expression. The Delta binding score was calculated by subtracting the score of a TF-gene interaction within 1 GRN with the score of the interaction within the other GRN.

The delta binding score of the ESC-KC and ESC-LSC comparisons were averaged. If this average was higher than the delta binding score of KC-LSC and LSC-KC, the interaction was classified as “shared epithelial,” if the delta binding score was highest in LSC-KC, it was classified as “KC specific,” if the delta binding score was highest in KC-LSC, it was classified as “LSC specific.”

Single-cell RNA-seq analysis of the epidermis and the cornea

The raw sequencing data was downloaded from GEO and split it into fastq files using seq2science. Cellranger count was run with Cellranger 6.0.1 to retrieve the matrix, barcodes, and features files necessary for Seurat [122] analysis in R. scRNA-seq cells were selected with a minimum count of 2,000, a feature number above 1,000, and a mitochondrial percentage below 30 percent. Cell cycle scoring was performed using Seurat CellCycleScoring() feature with the cell cycle genes from Tirosh and colleagues [123]. All cells not in the G1 phase were removed. Leiden Clustering was performed and cell clusters were annotated based on described marker genes.

For the data of the epidermis, cell clusters were selected with high KRT14, KRT5, and low KRT1 and KRT10 expression as basal KCs. From the cornea dataset, cell clusters with high S100A2 with PAX6 and TP63 expression and without CPVL expression were selected as LSCs. The in vivo versus in vitro fold change difference plot was generated by loading deseq2 result tables to identify the TF fold changes.

ChIP-seq and CUT&RUN data analysis

ChIP-seq and CUT&RUN peaks were called with MACs2 and validated by IDR (preprocessing). Next for each peak summit reads were counted in 200 bp windows across each summit. Values were log-transformed and quantile-based normalized. Peaks were linked to TSS regions in 100 kb, using bedtools window. Afterwards, they were distance weighted using the ANANSE distance weighing approach. When genes did not have ChIP-seq peaks within a 100 kb window, they got an intensity score of 0.

Disease gene lists were collected from the EyeDiseases database [57], including all disease gene lists of more than 20 genes. A one-sided Mann–Whitney U test was performed to test the hypothesis that the disease genes have more TF binding than the other genes in the genome. Of each significant hit, the top 5 of most bound gene loci were outputted to a list. And the final list was used to generate a dotplot in R (hipseq_intensity_npeak_dotplot.Rmd). Alternatively, ChIP-seq/CUT&RUN peaks were mapped to the gene TSS start site using bedtool closest. Next, disease genes mapped versus non-mapped disease genes were compared to all genes mapped versus non-mapped. Using a Fisher exact test.

Cornea disease gene list

Curated cornea disease list was firstly compiled by retrieving all known genetic disorders affecting the cornea and respective affected genes from “Ophthalmic Genetic Diseases” [124] and then confirmed using available literature in Pubmed (https://pubmed.ncbi.nlm.nih.gov/) and online eye disease (https://gene.vision/) databases. Diseases were grouped as (1) corneal diseases; (2) systemic (or other ocular) disorders with corneal phenotypes; and (3) diseases with secondary cornea involvement (due to exposure or unclear involvement). Genes associated with multifactorial disease keratoconus were added based on literature search mainly of published genome-wide association (GWAS) and linkage (GWLS) studies [124–126]. Curated gene list is available as S7 Table.

Variant discovery

Participants of the 100,000 Genomes Project were identified for our analyses who had at least one of the following HPO terms or daughter terms present: corneal opacity (HP:0007957), corneal scarring (HP:0000559), Opacification of the corneal stroma (HP:0007759), central opacification of the cornea (HP:0011493), band keratopathy (HP:0000585), central posterior corneal opacity (HP:0008511), corneal crystals (HP:0000531), generalized opacification of the cornea (HP:0011494), peripheral opacification of the cornea (HP:0008011), punctate opacification of the cornea (HP:0007856), and sclerocornea (HP:0000647). A total of 33 probands were identified who remain genetically unsolved. The whole genome sequence data was interrogated for single-nucleotide variants (SNVs) and indels (insertions or deletions), copy number variants (CNVs), and structural variants as previously described [127]. Filtered variants were annotated using Ensembl Variant Effect Predictor (VEP v99) and prioritized identified variants using scores available from CADD, MutationTaster, Provean, Sift, polyphen2, MetaRNN, DANN, fathmm-MKL. Variant nomenclature was assessed using Variant Validator.