Distinguishing causal from tagging enhancers using single-cell multiome data

Abstract

Methods that analyze single-cell RNA-seq+ATAC-seq multiome data have shown promise in linking enhancers to target genes by correlating chromatin accessibility with gene expression across cells. However, correlations among ATAC-seq peaks may induce non-causal tagging peak-gene links (analogous to tagging associations in GWAS); indeed, we confirm that tagging effects induced by peak co-accessibility are pervasive in peak-gene linking. We defined two scores for each ATAC-seq peak: co-accessibility score, the sum of squared correlations with each nearby peak; and co-activity score, the sum of squared correlations with each nearby gene. We compared these scores in 4 multiome data sets (spanning 86k cells and 6 immune/blood cell types) and determined that co-accessibility score and co-activity score were strongly correlated across peaks (r = 0.57-0.73); these correlations were not explained by read depth, cell subtypes, or measurement noise, but are consistent with tagging. Indeed, non-causal peak-gene correlations were strongly correlated to a peaks tagging correlation with a causal peak in CRISPRi data (r = 0.92). We further determined that causal peak-gene associations are concentrated in specific functional categories of peaks, by regressing co-activity scores on stratified co-accessibility scores (S-CASC): e.g. 2.91x (s.e. 0.67) enrichment for peaks closest to a genes TSS and 1.41x (s.e. 0.11) enrichment for peaks overlapping H3K27ac marks. Co-accessibility scores were substantially driven by the number of transcription factor binding sites (TFBS) within a peak, and peak-peak correlations were substantially driven by the number of TFBS pairs within the two peaks with a shared TF. These effects were concentrated in a small number of pioneer TFs, which activate repressed chromatin regions. Consistent with widespread tagging, peak-gene links that we fine-mapped using SuSiE significantly outperformed marginal peak-gene links in evaluation sets derived from CRISPRi and eQTL data. We provide examples demonstrating the impact of tagging effects at specific peaks and genes implicated in GWAS of blood cell traits. Our findings underscore the importance of accounting for tagging effects when linking enhancers to target genes.