Fig. 5: A 1-5 kb shift in co-predictivity and DNA contact highlights preferential DNA conformations connecting two co-predictive regions.
a Predictive test-set accuracy of additive and non-additive ChromatinHD-pred models across all genes (n = 5000). Percentage of genes with out-of-sample-R2 ratio higher or lower than 1.25 are indicated. b Examples of co-predictive regions for BCL2. Immune-related GWAS SNPs are shown and colored according to haplotype (LD r2 > 0.9). c Odds-ratio for finding high co-predictivity (higher than median) and high Hi-C signal (higher than median) within a slice of genomic distances (1kb-10kb, 10kb-20kb, …) performed for the original Hi-C data (Hi-C max-pool distance = 0), and for max-pooled Hi-C data where we took the maximal Hi-C signal at various genomic distances around the original position. B-cell genes were defined as being differentially expressed in naive, memory or plasma cells compared to all other cell types in the dataset. d Hi-C pileups of potential DNA contact points (C1 and C2) close to two co-predictive regions (E1 and E2, distance 20-25 kb, corΔz-score > 0). Shown is the relative Hi-C signal centered on the co-predictive pair divided by a random pair around the same gene with the same genomic distance. The numbers 1-8 refer to various putative conformations of enhancers and contact points, further described in panels (e–l). e–l Illustrations of how different distances between predictive regions and DNA contact points from d may inform on DNA conformation. m Difference in log contact frequency between up and down-regulated genes in B-cells. n–o Same as d but with E1-E2 distances of 5–10 kb and 45–50 kb respectively. Source data are provided as a Source Data file.
