Allele-specific open chromatin in human iPSC neurons elucidates functional disease variants

Abstract

Most neuropsychiatric disease risk variants are in noncoding sequences and lack functional interpretation. Because regulatory sequences often reside in open chromatin, we reasoned that neuropsychiatric disease risk variants may affect chromatin accessibility during neurodevelopment. Using human induced pluripotent stem cell (iPSC)-derived neurons that model developing brains, we identified thousands of genetic variants exhibiting allele-specific open-chromatin (ASoC). These neuronal ASoCs were partially driven by altered transcription-factor-binding, overrepresented in brain gene enhancers and expression quantitative-trait-loci, and frequently associated with distal genes through chromatin contacts. ASoCs were enriched for genetic variants associated with brain disorders, enabling identification of functional schizophrenia risk variants and their cis-target genes. This study highlights ASoC as a functional mechanism of noncoding neuropsychiatric risk variants, providing a powerful framework for identifying disease causal variants/genes.

One Sentence Summary

Allele-specific open chromatin informs functional disease variants

The difficulty of assigning molecular function to noncoding disease risk variants has hindered the identification of causal variants/genes of complex disorders. Noncoding regulatory sequences are marked by open chromatin, and open chromatin regions (OCRs) are enriched for single nucleotide polymorphisms (SNPs) associated with gene expression and complex disorders/traits (1, 2). For neuropsychiatric disorders, disease-associated variants are also enriched in OCRs found not only in adult brains, but also in developing brains and iPSC-derived neuronal (iN) cells (3–6), suggesting the importance of neurodevelopmental chromatin regulation. However, as not all OCR variants are functional, it remains largely unknown whether neuropsychiatric disease risk variants affect chromatin accessibility during neurodevelopment.

To tie risk alleles to function, we identified allele-specific open chromatin (ASoC) variants displaying allelic imbalance of chromatin accessibility in Assay for Transposase-Accessible Chromatin sequencing (ATAC-seq) at heterozygous SNP sites (Fig. S1) (7, 8). We focused on 20 individuals carrying heterozygous (i.e., informative for ASoC assay) GWAS index SNPs at ~70/108 schizophrenia (SZ) loci (Fig. S2, Table S1) (7, 9, 10).

We differentiated iPSCs of these 20 individuals into neural progenitor cells (NPC), glutamatergic (iN-Glut), GABAergic (iN-GA), and dopaminergic (iN-DN) neurons with high purity (75–90%) (Figs. 1A, S3) (7). We performed ATAC-seq and RNA-seq in each cell type for 8 lines (“core-8”), and ATAC-seq for 12 additional lines in NPCs and iN-Glut (Figs. 1A, S4–5; Table S2). Pooling the reads of “core-8” lines, we identified 256–337K OCRs (FDR<0.05) in each cell subtype (Fig. S6A–C; Table S3).

Fig. 1. Mapping ASoC variants.

(A) Schematic of iPSC differentiation strategies. (B) PCA of OCRs from iNs, fetal brains/day-30 organoids and postmortem (PTM) adult brain. (C) Cell-type-specific ASoC rs2027349 at the SZ-associated VPS45 locus. The dashed box shows the ATAC-seq read pileups flanking rs2027349 and cell-type-specific allelic ratios. (D) Pairwise π1 estimation of ASoC sharing across cell types (7). ASoC SNPs were ascertained in the leading cell type, and π1 was estimated in the matched one. (E) Chromatin accessibility in iN-Glut vs iPSC for ASoC SNP-flanking peaks. The shaded area shows comparable (< 2-fold difference) accessibility between cell types.

The OCRs of cell-type-specific marker genes were confirmed (Fig. S6D–E) and principal component analysis (PCA) of ATAC-seq and RNA-seq samples showed cell-type-specific clustering (Figs. 1B, S7A–D). OCRs of iPSC-derived neural cells were more similar to fetal brains (11) and cortical organoids (12) than to adult brains (3) (Fig. 1B, S7C–D). Neuronal OCRs overlapped with ~53% of the 117,935 adult brain OCR peaks from PsychENCODE (3); however, the 62,813 overlapping peaks only accounted for ~20% of our neuronal OCRs (Table S3), consistent with the observation that the fetal cerebral cortex possesses more active enhancers than the adult brain (12).

We then tested ASoC variants in each iPSC-derived cell type and examined cell-type-specificity (Fig. S8) (7). We used a sample pooling approach (8, 13) after confirming the inter-individual concordance of the allelic imbalance directions of chromatin accessibility (Fig. S9A). We identified 920–2,392 ASoC SNPs (FDR<0.05) in each “core-8” cell type, and 5,611 and 3,547 ASoC SNPs in iN-Glut-20 and NPC-20 cells (n=20 lines each), respectively (Figs. S9B–E, Tables S4–11). These ASoC SNPs showed cell-type-specific patterns (Figs. 1C, S9B). Using Storey’s π1 analysis (7), we found a low percentage of estimated pairwise ASoC SNP sharing between neurons and iPSCs (10~20%; Fig. 1D). Even within neuronal cell types, ASoCs differed substantially (30–70% sharing) (Fig. 1D), suggesting cell-type-specificity of ASoC.

Cell-type-specific ASoC may be driven by either OCRs or different SNP effects (allelic ratios) across cell types. About half of the ASoC SNPs in neurons exhibited low or undetectable OCRs (read-depth<20 at SNP sites) in iPSCs (Fig. S10). On the other hand, about 50% of neuron-specific ASoC SNPs showed comparable or more accessible OCR in iPSCs (Figs.1E, S11), suggesting that variations in OCR read-depth and SNP effect size both contribute to cell-type-specific ASoC. Thus, cell-type-specific regulation likely occurs even in the absence of chromatin state changes.

To examine the cell-type-specific regulatory potential of ASoC SNPs, we compared the genomic/epigenomic features of cell-type-specific vs shared ASoC variants (7). ASoC SNPs were found enriched in both brain promoters and enhancers (Figs. 2A, S12A) (14). However, compared to shared ASoC SNPs, cell-type-specific SNPs showed stronger enrichment (2232-fold) in enhancers (Figs. 2A, Fig. S12B–E). This indicates that ASoC SNPs are associated with brain regulatory sequences and likely exhibit cell-type-specific enhancer activity.

Fig. 2. Characteristics of ASoC SNPs.

(A) Enrichment of ASoC SNPs (vs. genomic background) in the chromatin-state-based annotations of promoter and enhancer. (B) Enrichment of ASoC SNPs (vs genomic background) for brain QTL. (C) Venn diagram showing iN-Glut-20 ASoC SNPs mapped to cis-target genes through brain eQTL (17), brain/neural Hi-C (18), and promoter annotations. (D) An ASoC SNP mapped to RERE through eQTL and Hi-C. ATAC-seq peak: iN-Glut. (E) Selected TFs with significant correlations (FDR<0.05) of motif-disrupting scores and ASoC allelic imbalance. (F) SOX9 as an example of a putative chromatin activator showing positive correlation in (E). (G) TFs with binding-footprints enriched for neural ASoC SNPs (red dots; vs non-ASoC SNPs). Blue: neurodevelopmental TFs.

The effect of ASoC SNPs on enhancer/promoter activity was confirmed by reporter gene assay (7). In a massive parallel reporter assay (MPRA) SNP dataset from non-neural cells (15), 26~29% of neural ASoC SNPs altered reporter gene expression (~2-fold enrichment vs non-ASoC, p=0.001, Fisher’s exact test) (Fig. S13A–D, Tables S12–13). Using a reporter gene assay in NPCs, we found that 5 of 8 SZ-associated ASoC SNPs affected reporter gene expression (Fig. S13E–F, Table S14). These results suggest that a proportion of ASoC SNPs affect gene expression.

We then jointly analyzed our data with recent studies of genetic variants associated with brain gene expression (eQTL), histone modification (haQTL), and DNA methylation (meQTL) (7, 16). Putative causal variants underlying these QTL are enriched (> 30 fold) in iN-Glut-20 ASoC SNPs (Fig. 2B, Table S11). Intersecting with brain eQTL (17) and brain/neuronal Hi-C chromatin contact (18) annotations (7), we found that out of 5,611 iN-Glut-20 ASoC SNPs, 26% are also eQTL and 35% demonstrate chromatin contacts with distal promoters (Figs. 2C, S14A–B; Table S15). Together with promoter annotation, ~59% of ASoC SNPs were resolved to affect at least one cis-target gene (Figs. 2C, S14C). Integrating ASoC SNPs with brain eQTL and Hi-C data, we delineated putative causal eQTL SNPs and their affected cis-target genes (e.g.: Fig. 2D). These results support that many ASoCs affect distal genes through long-range chromatin contacts.

Mapping the transcription factor (TF)-binding footprints from ATAC-seq (7), we found that 32% of the 5,611 ASoC SNPs of iN-Glut-20 neurons were inside TF-binding sites (TFBSs), representing a 1.6-fold enrichment (vs non-ASoC SNPs, Fisher’s exact test, p=2.6×10⁻⁵⁸) (Fig. S15A). Furthermore, most ASoC SNPs were within 200 bp of TF-binding footprints, and footprint analyses in other cell types gave similar results (Figs. S15B–G).

To test whether TF-motif disruptions by ASoC SNPs near footprints caused matched changes in chromatin accessibility, we performed a “motif-break analysis” (7). In iN-Glu-20 and NPC-20 cells, we identified 48 TFs with SNP motif-disruption scores correlated with allelic imbalances of ASoC (Figs. 2E–F, S16). Comparing our ASoC SNPs with an allele-specific TF-binding SNP dataset (19), we found that 70–80% of overlapping SNPs showed matched direction of allelic effects on chromatin accessibility and TF-binding (Fig. S17, Table S16). These results suggest that genetic variants that alter TF-binding result in ASoC.

We then performed TF-motif enrichment analysis for ASoC SNPs (±50 bp) and OCR (7), and identified cell-type-specific patterns of enriched TF motifs (Fig. S18; Tables S17–18). Restricting the ASoC analysis specifically to TF footprints gave similar patterns (Figs. 2G, S19) and included TCF4 and SP4, which are both SZ risk genes (10). Thus, ASoC SNPs may affect cell-type-specific binding of TFs important for cell fate commitment and neurodevelopment.

Given the brain-relevant regulatory effects of ASoC SNPs, we reasoned that ASoCs could help infer noncoding risk variants for neuropsychiatric disorders. Of the 5,611 iN-Glut-20 ASoC SNPs, 21 were SZ-associated SNPs (10) at 17 independent SZ loci (Fig. 3A, Tables S19–20). Accounting for the uncertainty of GWAS risk variants due to linkage disequilibrium (LD) with TORUS (7), neuronal ASoC SNPs showed ~70-fold enrichment for SZ risk variants, contrasting to ~3–6-fold enrichments in neuronal OCRs (Fig. 3B).

Fig. 3. Enrichment of ASoC variants for genetic risk of neuropsychiatric disorders/traits.

(A) iN-Glut-20 ASoC SNPs were enriched for SZ GWAS SNPs (Fisher’s exact test; vs. non-ASoC SNPs). (B) Enrichment of SZ risk variants in ASoC SNPs, OCRs and other functional annotations (vs genomic non-ASoC SNPs). Bar: 95% confidence interval. (C) Enrichment of risk variants of 22 disorders/traits in ASoC SNPs and OCRs (vs genomic non-ASoC SNPs). ADHD: Attention-Deficit/Hyperactivity Disorder, BMI: Body mass index, LDL: low-density lipoprotein, HDL: high-density lipoprotein, T2D: type 2 diabetes, IBD: Inflammatory bowel disease.

Expanding this analysis to 9 other brain disorders/traits (7), we observed enrichment of GWAS variants for bipolar disorder (BP), major depressive disorder (MDD), intelligence, educational attainment, and neuroticism in ASoC variants (iN-Glut and NPC) (Figs. 3C, S20; Table S21). Enrichment of ASoC SNPs (only iN-Glut) was also seen for Alzheimer’s disease (AD), but not Parkinson’s disease (PD) (Figs. 3C). In contrast, 12 control disorders/traits showed little or no enrichment in neuronal ASoC SNPs or OCRs (Fig. 3C), although this may reflect limited statistical power. These results suggest that neuronal ASoCs are linked to neuropsychiatric disorders and highlight the value of utilizing ASoCs in prioritizing studies of noncoding risk variants.

We selected the top 20 ASoC SNPs associated with SZ (Tables S19–20) to assess the regulatory potential of their tagged ASoC sequences and cis-target genes. We used a modified CRISPR droplet sequencing (CROP-seq) approach (20) (Fig. 4A) in NPCs stably expressing dCas9/Krüppel associated box (KRAB) which represses possible cis-genes in specific ASoC sites targeted by gRNAs (Fig. S21A–C, Table S22) (7). We analyzed 4,099 cells, of which 2,522 were assigned to unique gRNAs (Fig. S21D–F). We identified putative cis-targets as differentially expressed genes (permutation p<0.05) for a given gRNA/ASoC sequence (Figs. 4B, S22; Table S23), and then validated by independent CRISPRi/qPCR (Fig. S23; Table S24) (7). In total, CROP-seq identified cis-targets of 10 ASoC sequences, of which four share targets with brain eQTL and brain/neural Hi-C contacts (18, 21, 22) (Fig S24; Table S25).

Fig. 4. Functional characterization of SZ-associated ASoC SNPs.

(A) Schematic of the modified CROP-seq (7) to screen cis-targets of the 20 ASoC sites. (B) Reduced VPS45 expression in single NPCs expressing gRNAs targeting rs2027349. (C) Fine-mapping the VPS45 (rs2027349) locus. Left y-axis: −log₁₀ p-values from SZ GWAS (10) (points above the x-axis); right y-axis: PIPs (points below the x-axis, red). Vertical bar: rs2027349. (D) CRISPR-editing rs2027349 altered local OCR accessibility in isogenic NPCs (AA vs GG; 2 lines/genotype). The G allele disrupts the EGR3 motif; note that AA lines showed higher (more accessible) peak but with a lower ATAC-seq signal at the EGR3-binding site likely due to A allele’s stronger EGR3-binding that prevents transposase from accessing. (E) Editing rs2027349 altered VPS45 expression in NPCs (major transcript ENST00000369130; by qPCR). VPS45 expression was normalized to GAPDH. Two iPSC lines (AG at rs2027349) were edited to isogenic lines with AA or GG genotype, each with 2–3 clones (n=2–3 independent cultures). Student’s t-test (Welch’s correction) was used.

CROP-seq screening of cis-targets of an ASoC SNP-flanking sequence did not inform SNP allelic effect. To link SNP risk allelic function to a cis-target gene, we carried out precise SNP editing by CRISPR/Cas9 (7). We prioritized an ASoC SNP for editing based on the support from ASoC mapping, CROP-seq screening, and the Posterior Inclusion Probability (PIP) (7), the statistical evidence of an SNP being a disease-causal variant based on fine-mapping (23). Among the 20 targeted ASoC SNPs, 9 had PIP>0.10 (Table S26) of which 6 had at least one CROP-seq cis-gene (Table S23). Of these, rs2027349 at the VPS45 locus showed the strongest ASoC in NPC-20 (Table S20), exhibited a repressive effect on VPS45 in CROP-seq (p=0.008, permutation; Fig. 4B), and was the only local SNP with a high PIP (0.45, Fig. 4C). rs2027349 is within a TF-binding footprint, with the G allele predicted to disrupt the motif of EGR3/4 (Fig. 4D), a TF involved in neurodevelopment and SZ (24). rs2027349 was thus prioritized for CRISPR/Cas9 editing.

We edited two iPSC lines heterozygous for rs2027349 from A/G to A/A and G/G (Fig. S25A–B, Table S27) and characterized the SNP allelic effect on VPS45 expression in NPCs. Allele A was associated with increased expression of VPS45, especially its major transcript isoform (Figs. 4E, S25C–D). This is consistent with the allelic effect of rs2027349 on VPS45 in PsychENCODE brain RNA-seq data (Fig. S25E). In addition, rs2027349-flanking OCR in CRISPR-edited A/A lines showed higher chromatin accessibility (vs G/G) (Figs. 4D). We did not observe similar effects on two other cis-genes (ANP32E and SF3B4) nominated by CROP-seq (Fig. S22A–B). Taken together, we identified rs2027349 as a likely SZ risk variant that affects the expression of VPS45, a gene potentially relevant to SZ given its involvement in vesicle-mediated protein trafficking and neurotransmitter release (25).

We also performed CRISPR/Cas9 editing of rs12895055 (C→T) (Fig. S26), an SZ-associated SNP that did not have a CROP-seq cis-target, but showed the strongest ASoC in iN-Glut (Table S19) and a high probability of being a causal SZ variant (PIP=0.33) (Table S26). Because rs12895055 is in an intron of BCL11B (a deep-layer neuronal marker) with undetectable expression in NPCs (Table S23), we examined its allelic effect on BCL11B in CRISPR-edited iN-Glut cells (Fig. S26, Table S28). The risk allele T was associated with reduced expression of BCL11B and is predicted to alter binding-motifs of multiple TFs (Fig. S26). Notably, rs12895055 is in perfect LD with the BP-associated rs11624408 (p=5×10⁻⁸) (26). BCL11B is thus a candidate risk gene for both SZ and BP, warranting follow-up.

In summary, we have found that genetic variation commonly affects chromatin accessibility and exhibits cell-type-specific ASoC in iPSC-based neurodevelopmental model. The enrichments of neuronal ASoC SNPs for brain enhancers, TFBSs, and brain QTL suggest mechanistic links between chromatin accessibility and gene expression. Our study identified ASoC variants that have plausible causal associations with SZ, and demonstrates that some neuropsychiatric disease risk variants alter chromatin accessibility.