Extended Data Fig. 6: Examples of interacting pairs of predictive motifs and motif co-occurrence statistics.

(a) Example regions demonstrating interacting motifs. Top row: putative enhancer affecting LAMC2 gene expression with an interacting NFKB2 motif and RUNX1 motif (chr1:183147408-183170430). Assay ranges are ATAC-seq: 0–600; H3K27ac: 0-300; H3K4me1: 0–50. The highlighted region in the signal tracks (left) demonstrates correctly predicted ATAC signal by the neural net (top middle heatmap). Base-resolution contribution score tracks are shown for the wild-type (genomic) sequence and sequences with marginal and joint perturbation of both motifs (middle tracks). The model predicts a super-multiplicative effects of the motif pair on chromatin accessibility (right plot). Bottom row: Analogous plots for a putative enhancer affecting MUC15 gene expression with an interacting GRHL2 motif and ATF4 motif (chr11:26590539-26610606). Assay ranges are ATAC-seq: 0–800; H3K27ac: 0–150; H3K4me1: 0-70. (b) Co-occurrence statistics (size of circle represents number of instances) for motif pairs based on all motif instances identified solely using sequence match scores (left) and motif pairs based on predictive, active motif instances based on contribution-weighted sequence match scores (right). Predictive motif instances highlight less promiscuous, more specific co-occurrence statistics. (c) Analogous co-occurrence statistics for motif pairs using all motif instances (left) and predictive motif instances (right) after filtering for pairs that show significant GO term enrichments for associated target genes. Once again, more specific co-occurrence patterns are observed for the predictive motif instances.