We aimed to evaluate whether variants that alter chromatin accessibility in LCLs may explain genetic associations to IMD. First, we verified whether accessible regions in LCLs are enriched for IMD heritability. We reanalyzed 100 LCL ATAC-seq samples³⁰ to define accessible regions in this cell type. With stratified LD score regression (S-LDSC)³⁸, we estimated their heritability enrichment across 13 IMDs, including 11 autoimmune diseases — autoimmune thyroid disease (ATD)³⁹, celiac disease (CEL)⁴⁰, Crohn’s disease (CD)⁴¹, inflammatory bowel disease (IBD)⁴¹, juvenile idiopathic arthritis (JIA)⁴², multiple sclerosis (MS)⁴³, primary biliary cholangitis (PBC)³⁹, rheumatoid arthritis (RA)⁴⁴, systemic lupus erythematosus (SLE)⁴⁵, ulcerative colitis (UC)⁴¹, and vitiligo (VIT)⁴⁶ — and 2 allergic diseases — allergy (ALL)⁴⁷ and asthma (AST)⁴⁷. We also analyzed GWAS data for 3 non-immune diseases — type 2 diabetes (T2D)⁴⁸, coronary artery disease (CAD)⁴⁹ and schizophrenia (SCZ)⁵⁰ — for comparison. Single nucleotide polymorphisms (SNPs) in accessible regions in LCLs were significantly enriched for IMD heritability (p < 0.003125 (Bonferroni-corrected), S-LDSC; Figure 1A) and there was no significant enrichment for non-immune diseases (p > 0.05, S-LDSC). These results indicate that accessible regions in LCLs harbor many variants specifically associated with IMDs, and therefore that LCLs share IMD-associated accessible regions with those of the causal cell type(s).

Next, we applied mediated expression score regression (MESC)¹⁶ to investigate the causal relationship between caQTLs in LCLs and IMD associations. Compared to S-LDSC analysis that tests for heritability enrichment of SNPs with some functional annotation (e.g., accessible regions), MESC analysis specifically estimates the heritability that is mediated (i.e., h²_med) by the SNPs’ cis-effects on a molecular phenotype (e.g., caQTLs). We estimated that caQTLs in LCLs mediate 16.3–42.7 % of autoimmune disease heritability and 8.5–9.4 % of allergic disease heritability (Figure 1B). For non-immune diseases, the estimates were lower and not significant (p > 0.003125 (Bonferroni-corrected), MESC). Interestingly, schizophrenia showed a nominally significant proportion of caQTL-mediated heritability in LCLs (p < 0.05, MESC), consistent with the hypothesis that B cells may play some role in schizophrenia pathogenesis^50,51. Our results indicate that LCLs are a valid cell type in which to search for caQTLs that mediate genetic risk for 7 IMDs — CD, IBD, MS, PBC, RA, SLE, and UC — but not for allergic diseases. In subsequent analyses, we focused on 7 IMDs – CD, IBD, MS, PBC, RA, SLE, and UC – that showed significant caQTL-mediated heritability (p < 0.003125 (Bonferroni-corrected), MESC).

To understand which features characterize accessible regions that mediate IMD heritability, we estimated h²_med enrichment (proportion of h²_med / proportion of peaks)¹⁶ in specific sets of accessible regions. We found that peaks with a larger number of nonredundant sequencing reads (i.e., “stronger” peaks) in LCLs showed stronger h²_med enrichment (Figure 1C) and thus likely affect IMD-relevant gene expression more than “weaker” peaks do. This observation is consistent with the “Activity-by-Contact” model⁹, in which peaks with greater chromatin accessibility and H3K27ac ChIP-seq signal are predicted to have proportional effects on target gene expression.

Next, we considered peaks with the active histone marks H3K27ac, H3K4me1, or H3K4me3^28,52. Consistent with prior observations that putative cell-type-specific regulatory elements marked with H3K27ac and H3K4me1 are enriched for relevant disease associations^34,35, we found that caQTLs with H3K27ac and H3K4me1 marks in LCLs were significantly enriched for mediated IMD heritability (q-value < 0.05, MESC; Figure 1D). Strikingly, both peak sets explained almost all of caQTL-mediated IMD heritability (Supplementary Figure 1). Peaks with H3K4me3 marks, representative of promoters⁵³, also showed significant h²_med enrichment for most IMDs (q-value < 0.05, MESC). Peaks with promoter-like signatures (i.e., H3K27ac and H3K4me3)⁵⁴ and those with enhancer-like signatures (i.e., H3K27ac, but no H3K4me3)⁵⁴ were also enriched for all IMD heritability (q-value < 0.05, MESC). Conversely, peaks without any of the three active histone marks were completely depleted of caQTL-mediated IMD heritability (Supplementary Figure 1). These ‘ATAC only’ peaks were shorter, weaker, and further away from the TSS compared to peaks with active histone marks (Supplementary Figure 2). Altogether, these results indicate that peaks characterized as putatively active regulatory elements explain nearly all of caQTL-mediated IMD heritability.

The model that gene regulatory activity explains a significant fraction of noncoding genetic associations to IMDs is supported by our findings that caQTLs mediate a significant proportion of IMD heritability and that those with active histone marks show strong h²_med enrichment. This is in contrast with relatively low average h²_med / h²_SNP estimates (∼11%) previously having been observed for eQTLs across 48 human tissues in GTEx and various human traits⁵⁵. To directly compare the proportion of IMD heritability mediated by caQTLs and eQTLs in the same cell type (i.e., LCLs), we additionally applied MESC to gene expression data from LCLs (i.e., Geuvadis data)⁵⁶.

Across the 7 autoimmune diseases, the estimated proportion of heritability mediated by eQTLs (h²_{med; eQTL} / h²_SNP) ranged from 9% to 22% (Figure 2A). For all 7 diseases, we estimated that eQTLs mediated less heritability than did caQTLs, even though the caQTLs’ smaller sample size would potentially bias the estimates toward zero¹⁶. A possible explanation is that some IMD-associated regulatory variants may show detectable effects on chromatin accessibility, but not on gene expression, in LCLs at the current sample size (n = 373); such loci may account for the missing regulation¹⁵.

We anticipated that IMD-associated variants that affect gene expression in cis do so by modulating regulatory element’s activity. Therefore, we investigated whether eQTL-mediated IMD heritability is shared by caQTL-mediated signals (Figure 2B). We performed MESC on both caQTLs and eQTLs together to estimate the amount of IMD heritability mediated by both collectively (Methods). For the 7 IMDs, the combined h²_{med; caQTL ∪ eQTL} / h²_SNP was only slightly higher (2.2–9.0 %) than the estimates for just caQTLs (h²_{med; just caQTL} / h²_SNP; Figure 2A), suggesting that approximately 56–82% of eQTL-mediated heritability is shared with caQTL-mediated heritability (i.e., h²_{med; caQTL ⋂ eQTL} / h²_{med; eQTL}; Supplementary Figure 3). These estimates are consistent with substantial sharing of caQTL- and eQTL-mediated IMD heritability. Nevertheless, 9–27% of IMD heritability is explained just by caQTLs, while only 2–9% of IMD heritability is explained just by eQTLs.

We applied colocalization analysis³⁷ to identify IMD-associated loci that share genetic signals with caQTLs or eQTLs in LCLs. We selected candidate loci of 200-kb windows for each IMD with the following conditions: (1) lead IMD association at p < 10^-6, (2) lead caQTL or eQTL association at p < 10^-4, and (3) at least one variant simultaneously showed caQTL or eQTL χ² statistics greater than 0.8 × lead χ² statistics for the caQTL or eQTL, respectively, and IMD association χ² statistics greater than 0.8 × χ² statistics for the IMD lead variant in the locus. We applied gwas-pw³⁷ and considered loci with posterior probability of colocalization (PPA3) > 0.98 to be colocalized⁵⁷. Some loci colocalized with only either a caQTL or an eQTL, while others colocalized with both (Figure 2C and Supplementary Tables 1-2).

To confirm that the colocalized genes are relevant to IMD, we tested for their enrichment of Gene Ontology (GO) annotation terms for specific biological processes⁵⁸. Considering all genes within 500 kb of the IMD GWAS lead variants at colocalized loci as background, the genes that showed eQTL colocalization for any IMD were enriched for various immune responses and signaling processes, such as “positive regulation of immune system process” and “regulation of lymphocyte activation” (Supplementary Figure 4), indicating that the colocalized genes in LCLs are involved in immune function. For example, IL6R and IL12A encode direct or indirect targets of approved drugs – Tocilizumab and Ustekinumab – for autoimmune diseases like RA⁵⁹ and CD⁶⁰. These two genes showed colocalization with both caQTLs and eQTLs in CD and PBC GWAS, respectively (Supplementary Figure 5A-B). Increased IL6R expression was associated with higher risk for CD, and increased IL12A expression was associated with lower risk for PBC and SLE. The former observation is in line with Tocilizumab, a monoclonal antibody to IL-6 receptor, showing efficacy in CD patients⁵⁹, although it is not pursued for approval because of potential side effects⁶¹. Interestingly, IL6R and IL12A eQTLs did not colocalize with the association signals of RA and CD, respectively, which are the diseases for which these drugs are approved (Supplementary Figure 5C-D). Moreover, ELMO1, which previously had not been associated with autoimmune diseases, showed eQTL colocalization with RA, PBC, and MS association signals (Figure 2D). In all three, decreased ELMO1 expression was associated with increased disease risk. In mice, Elmo1 was required for polarization and migration of B and T lymphocytes⁶².

Across the IMDs, there were many loci that colocalized with a caQTL but not with an eQTL (Figure 2C). These “caQTL only” loci showed enrichment for immune response genes in cis compared to all accessible regions in LCLs^63,64, even though the colocalized eQTLs were enriched for immune response genes as well (Figure 2E and Supplementary Figure 4), indicating that IMD-relevant genes without eQTL colocalization in Geuvadis LCL data⁵⁶ are likely found in these loci.

Why might there be loci with caQTL colocalization only, despite the caQTL data having fewer samples than the eQTL data (100 vs. 373)? Limited statistical power can prevent some eQTLs from being detected and showing significant colocalization¹⁹. As hypothesized by Mostafavi and colleagues, disease-relevant eQTLs may be weaker and more distal¹⁸. To understand the extent to which this effect may result in many loci showing colocalization only with caQTLs, we compared the cis-heritability (h²_cis) of caQTLs and eQTLs depending on the distance from the ATAC peak to the transcription start site (TSS) of the gene (i.e., peak-to-TSS distance). We considered all caQTLs and eQTLs regardless of disease association.

We identified pairs of caQTLs and eQTLs that colocalized with each other³⁷, which implies that the regulatory variant modulating chromatin accessibility also affects gene expression. Then, the distance between the ATAC peak and the TSS of the eQTL gene (i.e., eGene) is the distance between a regulatory element and its target gene’s TSS. We stratified the pairs into peak-to-TSS distance quintiles and compared the eQTL h²_cis distribution of the first quintile (i.e., closest pairs) with that of the later quintiles. We observed that eQTL h²_cis distribution decreased with increasing distance of the paired ATAC peaks from the TSS (p = 1.0×10^-4, 2.1×10^-10, 1.6×10^-15, and 1.7×10^-20, respectively, one-sided Wilcoxon rank-sum test; Figure 3A), consistent with the negative relationship between promoter-enhancer genomic distance and impact on gene expression^65,66. This result also explains why discovered eQTLs are concentrated near the promoter, where the variants are more likely to show stronger effects⁶⁷. In contrast, caQTL h²_cis distribution was similar across peak-to-TSS distances (p > 0.05, one-sided Wilcoxon rank-sum test; Figure 3A). For all distance quintiles, caQTL h²_cis was significantly higher than that of the paired eQTLs (p = 4.0×10^-23, 1.3×10^-24, 1.5×10^-19, 1.1×10^-28, and 2.4×10^-49, respectively, one-sided paired Wilcoxon rank-sum test; Figure 3A), and the contrast between them was greater at more distant quintiles, suggesting that the statistical power to detect and colocalize eQTLs is increasingly lower than that for caQTLs for regulatory effects far from the TSS.

Overall, caQTL h²_cis had a significant positive effect (p = 6.6×10^-11, linear regression) and peak-to-TSS distance had a negative effect (p = 2.4×10^-23, linear regression) on eQTL h²_cis (Figure 3B). Thus, for regulatory variants that showed both caQTL and eQTL signals, those with larger effects on chromatin accessibility tended to exhibit larger effects on gene expression, but their eQTL effects diminished with increasing distance from TSSs.

Next, we investigated how caQTL-mediated IMD heritability is distributed with respect to TSS. If caQTLs beyond the typical cis-eQTL window of 1 megabase (Mb) around the genes’ TSS explain some proportion of IMD heritability, then cis-eQTL analyses might require a wider window to detect disease-associated eQTLs. Across the 7 diseases, caQTLs within 500 kb of the TSS of expressed genes explained almost all of the caQTL-mediated IMD heritability (92–100%; Figure 3C), indicating that regulatory variants are most likely within 500 kb of the target gene’s TSS and supporting the use of a 1 Mb window for cis-eQTL analyses. Depending on the disease, 41–66% of the caQTL-mediated IMD heritability was detected in distal peaks further than 10 kb from the TSS of expressed genes, further supporting the analysis of regulatory variants beyond promoter regions.

In sum, the power to detect eQTLs diminishes with increasing distance of the variant from the TSS, but the power to detect caQTLs is largely invariant regardless of peak-to-TSS distance (Figure 3D). Since h²_{med; caQTL} are distributed mostly within 500 kb of genes’ TSS, the IMD loci colocalizing only with caQTLs could still be weak, undetected eQTLs. Under this model, we predicted that increasing the power to detect eQTLs in LCLs, such as increasing sample size, may lead to further eQTL colocalizations in loci in which we observed only caQTL colocalization.

For genetic association studies, increasing the sample size is a way to increase statistical power. Therefore, we meta-analyzed four LCL eQTL summary statistics^55,56,68,69, leading to a total sample size of 1,128 individuals. We performed colocalization analysis using the meta-analyzed summary statistics to evaluate whether effectively increasing the sample sizes would uncover more disease-associated eQTLs in LCLs, especially in loci where a caQTL already showed IMD colocalization. Up to 6 additional loci showing eQTL colocalization were thus detected for each IMD (Figure 4A and Supplementary Table 3). For example, CIITA is the class II major histocompatibility complex transactivator, which causes severe immunodeficiency if dysfunctional⁷⁰. The CIITA locus is associated with IBD right below the genome-wide significance level (rs10445003, p = 7.5×10^-8), and it colocalized with a caQTL signal, but initially not with any eQTL (Supplementary Figure 6A). However, the meta-analyzed statistics showed a stronger association to CIITA expression (p = 6.7×10^-8) than without meta-analysis (p = 5.4×10^-4) and exhibited a significant colocalization. Interestingly, two of the causal CD genes that previously lacked colocalized eQTLs¹⁵, CARD9 and ATG16L1, showed significant colocalization in the meta-analyzed LCL eQTL data (Supplementary Figure 6B-C).

We hypothesized that increased sample size would improve the power to detect weaker and distal eQTL colocalization. Comparison of the accessibility peak’s distance to the paired eQTL gene’s (eGene) TSS showed that the newly detected eQTLs tended to be more distal (p = 0.06, one-sided Wilcoxon rank-sum test; Figure 4B). However, compared to the distribution of the peak-to-TSS distance for all caQTL-eQTL pairs showing colocalization, IMD-associated loci that showed caQTL and eQTL colocalization had greater peak-to-TSS distance on average (p = 0.002 for Geuvadis and p = 5.9×10^-6 for the meta-analyzed data; Figure 4B). For example, a RA-associated locus near POU3F1, a neuronal transcription factor that is also induced by interferon⁷¹, colocalized with a distal eQTL located about 126 kb upstream of its promoter, after meta-analysis strengthened the eQTL association (p < 10^-10; Figure 4C). These results suggest that additional IMD-associated loci with distal, weaker eQTLs in LCLs might be found if eQTL data were generated for a larger number of individuals.

Despite uncovering more eQTL colocalizations through meta-analysis, more than 40% of the caQTL-colocalized loci nevertheless showed no colocalization with an eQTL in LCLs (Figure 4B). We investigated whether these “caQTL only” peaks might be inactive regulatory elements. We quantified the active histone mark levels for H3K27ac, H3K4me1, and H3K4me3 at colocalized caQTL peaks in LCLs and then compared their levels between the “caQTL and eQTL” and “caQTL only” loci. On average, H3K27ac marks were stronger at “caQTL and eQTL” peaks, supporting that the corresponding regulatory elements might be more active (Supplementary Figure 7). H3K4me3 marks were detected more often at “caQTL and eQTL” peaks, leading to stronger average signal. In contrast, H3K4me1 levels were highly similar between the two sets of peaks. Although “caQTL only” peaks generally showed lower levels of active histone marks, several individual “caQTL only” peaks showed comparable levels. These peaks could be inactive cis-regulatory elements in LCLs that affect gene expression in a different cellular context. Therefore, we next examined whether those caQTLs appear as eQTLs in other immune cell types.

We downloaded eQTL summary statistics generated from 26 naïve and stimulated immune cell types^{22,23,72–74} to search for eQTLs that may correspond to the remaining, IMD-colocalized caQTLs (Supplementary Table 4). The profiled cell types range from B cells and monocytes to subtypes of T cells, as well as stimulated T cells and macrophages. We tested for colocalization of these eQTLs to IMD associations.

25–42% of the caQTL-colocalized IMD loci that were missing eQTL colocalizations in LCLs showed eQTL colocalizations in at least one immune cell type (Figure 5A and Supplementary Table 5). The overlap of LCL caQTLs with non-LCL immune cell eQTLs was greater than expected by chance for 5 of the 7 IMDs (p < 0.00714 (Bonferroni-corrected), Fisher’s exact test, for CD, IBD, PBC, RA, and UC; Supplementary Table 6), suggesting that the caQTLs found in LCLs may also show regulatory function in those immune cell types. Comparing across the datasets, we found that the number of loci with eQTL colocalization varied depending on the cell type, but that the effect of the sample size was more profound (r² = 0.60–0.79; Figure 5B and Supplementary Table 7). We meta-analyzed eQTL data of 3 immune cell types with multiple sources – naïve CD4⁺ T cell, monocyte, and memory regulatory T cell (Treg) – and this also increased the number of loci with significant eQTL colocalization (Supplementary Table 8). Altogether, these results suggest that although generating eQTL data in more cell types and cell states uncovers context-specific eQTLs, increasing the sample size should also be a priority to ensure sufficient statistical power.

We investigated the potential reasons why IMD loci that colocalized with caQTLs in LCLs showed eQTLs not in LCLs but in other immune cells. First, LCL caQTLs may correspond to gene regulatory elements that exert their effects on gene expression in a different cellular context. For instance, monocyte H3K27ac levels in the “caQTL only” loci where monocyte eQTLs colocalized were higher than those with no monocyte eQTLs (Supplementary Figure 8). Second, some examples, such as for TNFSF15, were due to cell-type specific gene expression (Figure 6): despite a significant colocalization of IBD with a caQTL in LCLs in the TNFSF15 locus, disease-associated eQTL signal was detected only in monocytes (Figure 6A). Tumor necrosis factor-like cytokine 1A (TL1A), the protein encoded by TNFSF15, is secreted by monocytes and many other cells to activate helper T cells, regulatory T cells, and B cells⁷⁵. TNFSF15 expression was low in LCLs (mean transcript per million [TPM] = 0.30), but higher in monocytes (mean TPM = 2.23). Of the genes that showed exclusively monocyte eQTL colocalization, those with low expression (mean TPM < 1) in LCLs generally showed higher expression in monocytes (Figure 6B-C).

However, low expression in LCLs was likely not the explanation for most cases of “monocyte only” eQTLs (blue points in Figure 6B) because most “monocyte only” eQTL genes were expressed at level higher than 1 TPM in LCLs.

Overall, expanding the eQTL search to various immune cell types increased the number of eQTL-colocalized loci among those that previously colocalized only with caQTLs (Figure 5A). On average, approximately 75% of the caQTL-colocalized loci ultimately showed eQTL colocalization. These results highlight the utility of eQTL data across a range of immune cell types for discovery of IMD-associated eQTLs.

The h²_med / h²_SNP estimates can be biased because of insufficient sample size of the QTL data. Thus, the proportion of mediated IMD heritability could change based on the specific caQTL and eQTL data. Colocalization analysis tests whether the QTL and GWAS data likely share genetic signal, and such shared signal could arise from either causal mediation or pleiotropy. Therefore, further experiments are needed to establish causality of colocalized eQTL genes. Lastly, we hypothesized that caQTL effects are often shared across cell types, but we did not have access to relevant data to test this hypothesis.

We downloaded LCL ATAC-seq data of British (GBR) samples (n = 100) from European Nucleotide Archive (ENA) under accession ERP110508³⁰. The available files were cram alignment files mapped to the b37 reference genome, so we extracted unique read pairs using SAMtools⁷⁶ and bamtofastq command from bedtools⁷⁷. The reads were paired-end and each 75 base pairs (bp) long. The data contained reads with Nextera transposase adapters, so we removed the adapter sequences and bases of poor quality at the 3’ end using cutadapt. Trimmed reads with both pairs shorter than 20 bp were discarded. The command was “cutadapt -a file:${forward} -A file:${reverse} -e 0.25 -j 2 -q 15 --pair-filter=both -m 20”. (${forward} and ${reverse} files contain the forward and reverse Nextera transposase adapters). We mapped the reads to the GRCh38 reference genome using Bowtie 2⁷⁸ with the “GRCh38_noalt_decoy_as” index provided on the tool’s website. The command was “bowtie2 --very-sensitive --no-mixed --no-discordant -I 20 -X 2000”. We kept only read alignments with mapping quality greater than 1. We also removed reads aligning to the mitochondrial genome, those overlapping ENCODE exclusion regions (file ID: ENCFF356LFX)⁷⁹ and potential PCR duplicates using scripts from WASP⁸⁰.

To represent peaks across the samples, we subsampled 3 million read pairs from each and pooled them. Then, we used MACS2⁸¹ with the BAMPE option for peak calling. The command was “macs2 callpeak -f BAMPE -g hs -q 0.05”. We further used the ‘bdgcmp’ and ‘bdgpeakcall’ subcommand to find peaks that are at least 100 bp long (-l) and merge those that are less than 100 bp apart (-g). We also merged peaks similarly derived from individual samples using the ‘merge’ command in bedtools.

Furthermore, for the sake of comprehensiveness, we repeated these steps with LCL ATAC-seq data of Yoruban (YRI) samples⁸² and merged the peaks with the earlier peak set derived from GBR samples. In total, there were 443,403 peaks genome-wide.

We downloaded RNA expression level data of the LCL samples⁵⁶ from the Expression Atlas⁸³. The source data is available at European Nucleotide Archive under accession ERP001942. We retained TPM values of protein-coding and long noncoding RNAs in European samples for downstream analyses, including QTL analysis, mediated heritability analysis, and colocalization.

To utilize the genotype calls of the highest quality, we downloaded the high-coverage 1000 Genomes (1kG) Project data⁸⁴. There was a total of 493 samples that had either ATAC-seq or RNA-seq data in this study. Of those, 479 samples had genotypes derived from the high-coverage whole-genome sequencing data. 14 samples had genotypes derived from microarrays⁸⁵. The samples that needed imputation are listed in Supplementary Table 9. We lifted over the microarray data based on the hg19 reference genome to GRCh38 and filtered for variants present in the high-coverage 1kG data. We first imputed microarray genotype data using the TOPMed imputation server⁸⁶ and extracted SNPs with imputation R² ≥ 0.5 and imputed the rest of the variants, including short indels, using high-coverage data with Beagle5.2^87,88. After keeping only variants with DR2 ≥ 0.7, we merged the imputed genotypes with the high-coverage 1kG genotypes.

We downloaded GWAS summary statistics for 13 IMDs, including 11 autoimmune diseases (autoimmune thyroid disease (ATD)³⁹, celiac disease (CEL)⁴⁰, Crohn’s disease (CD)⁴¹, inflammatory bowel disease (IBD)⁴¹, juvenile idiopathic arthritis (JIA)⁴², multiple sclerosis (MS)⁴³, primary biliary cholangitis (PBC)³⁹, rheumatoid arthritis (RA)⁴⁴, systemic lupus erythematosus (SLE)⁴⁵, ulcerative colitis (UC)⁴¹, and vitiligo (VIT)⁴⁶) and 2 allergic diseases (allergy (ALL)⁴⁷ and asthma (AST)⁴⁷). For each disease, we searched for more recent studies with larger sample sizes and prioritized those with genome-wide statistics, rather than those with only Immunochip variants⁸⁹. To compare with non-immune diseases, we also downloaded summary statistics for 3 non-immune diseases (type 2 diabetes (T2D)⁴⁸, coronary artery disease (CAD)⁴⁹ and schizophrenia (SCZ)⁵⁰). In this study, we analyzed only those GWAS summary statistics derived from cohorts of individuals with European ancestries.

Since the LCL samples’ genotypes are based on the GRCh38 reference genome, we lifted over any GWAS data based on b37 reference genome to the GRCh38 genomic coordinates. Briefly, we formatted the summary statistics as bed files and used the liftOver tool⁹⁰ to convert them to GRCh38 genomic coordinates. Then, to ensure that the reference alleles match the sequences of the GRCh38 reference genome, we used gwas2vcf tool⁹¹.

First, we quantified the chromatin accessibility levels at ATAC-seq peaks identified earlier. We counted the number of read fragments overlapping each peak using featureCounts⁹². For each sample, the read counts were normalized for library size using trimmed mean of M-values⁹³ so that the values are comparable across the samples. Then, the phenotype values were further normalized to follow a standard normal distribution across the samples, using quantile normalization. Peaks with counts per million (CPM) < 0.8 or counts < 10 for more than 20% of the samples were discarded.

Next, we performed a principal component analysis (PCA) on the phenotype matrix to derive potential latent covariates. We selected the number of principal components (PC) to incorporate in the regression model based on the Buja and Eyuboglu algorithm⁹⁴ that is implemented in PCAForQTL⁹⁵. Ultimately, we accounted for sex, library size, 3 genotype PCs and 13 phenotype PCs in the QTL analysis. We performed genetic association tests on variants within 200 kilobases (kb) of the peak using tensorQTL⁹⁶. We discarded variants with minor allele frequency less than 5%.

We evaluated the relevance of accessible regions in LCLs to IMD heritability using stratified LD score regression (S-LDSC)³⁸. We used the baselineLD v2.2 annotation in hg38 and the European LD reference from the 1000 Genomes Project (downloaded from the S-LDSC website, https://alkesgroup.broadinstitute.org/LDSCORE/GRCh38/). We used the set of filtered ATAC-seq peaks that we tested for QTL associations. We accounted for 16 diseases (13 IMDs and 3 non-IMDs) for the Bonferroni-corrected p-value threshold.

We estimated the heritability mediated by QTLs (h²_med) using mediated expression score regression (MESC)¹⁶. We denote heritability of tested SNPs as h²_SNP.

To test whether colocalized genes are likely relevant to autoimmune diseases, we surveyed which Biological Process Gene Ontology (GO) terms were overrepresented compared to all the genes within 500 kb of each IMD association signal. Enrichment of biological processes was evaluated using Protein Analysis Through Evolutionary Relationships (PANTHER)⁹⁷. The foreground list comprised all of the genes whose eQTL signal colocalized with one of the 7 autoimmune diseases (CD, IBD, MS, PBC, RA, SLE, and UC). The background list of genes was all of the genes within 500 kb of each IMD lead variant for which we observed colocalization. Moreover, we tested whether colocalized caQTLs without eQTLs are closer to genes related to immune processes than expected by chance. For this, we used Genomic Regions Enrichment of Annotations Tool (GREAT)⁶³. The foreground list comprised ATAC-seq peaks at IMD loci showing colocalization only with caQTLs and not with eQTLs. The background list is all ATAC-seq peaks identified in LCLs that we tested for caQTL association.

The distance between the colocalized caQTL peak and the eGene’s TSS (i.e., peak-to-TSS distance) was determined to be the shortest distance from one end of the caQTL peak to the TSS. The pairs of caQTLs and eQTLs were split into quintiles based on their peak-to-TSS distance from closest to farthest.

MESC analysis uses REML implemented in GCTA⁹⁸ to estimate QTL cis-heritability. We referred to its output and compared the cis-heritability of caQTLs and eQTLs based on peak-to-TSS distance. To visualize the distribution of cis-heritability estimates, we grouped pairs of colocalized caQTLs and eQTLs based on peak-to-TSS distance quintiles.

We downloaded LCL eQTL data from 3 studies through eQTL Catalogue release 6⁹⁹. The sample sizes were 190, 147, and 418 for GENCORD⁶⁸, GTEx⁵⁵, and TwinsUK⁶⁹, respectively. We meta-analyzed the summary statistics using the inverse variance weighted fixed effects model. If a variant was missing in a subset of the studies, then only the available statistics were meta-analyzed. We used these meta-analyzed statistics to perform colocalization the same way as earlier colocalization analyses.

We downloaded eQTL data for various immune cell types from the eQTL Catalogue release 6⁹⁹. The source studies from the eQTL Catalogue include BLUEPRINT⁷², DICE²², Alasoo et al.²³ and Bossini-Castillo et al.⁷⁴. The represented immune cell types include T cell subtypes, B cells, neutrophils, and monocytes. We also separately downloaded data for CD4⁺ T cells with and without stimulation⁷³. The selection of candidate loci and colocalization analysis on them followed the same procedure as that for other QTLs.

We also evaluated whether meta-analysis of eQTL data can increase the number of loci with eQTL colocalization. We meta-analyzed eQTL data for 3 immune cell types with multiple sources – naïve CD4⁺ T cell, monocyte, and memory regulatory T cell (Treg) – using the inverse variance weighted fixed effects model. Specifically, we meta-analyzed naïve CD4⁺ T cell eQTL summary statistics from Soskic et al.⁷³, BLUEPRINT⁷², and DICE²². We meta-analyzed monocyte eQTL summary statistics from BLUEPRINT⁷² and DICE²². We meta-analyzed memory Treg eQTL summary statistics from Bossini-Castillo et al.⁷⁴ and DICE²². We used these meta-analyzed statistics to perform colocalization the same way as earlier colocalization analyses.