We downloaded 32 publicly available deep learning annotations for European SNPs from phase 3 of 1000 Genomes and used S-LDSC to compute LD scores. The 32 annotations were comprised of Basenji56 and DeepSEA54 deep learning predictions corresponding to DHSes, H3K27ac, H3K4me1 and H3K4me3 meta-analyzed separately for blood and brain cell types and computed for both allelic effect and variant level models57. Additionally, we analyzed 78 new tissue-specific variant level and allelic effect annotations from DeepSEA and Basenji models (Supplementary Note).These 78 annotations corresponded to cell types that we identified as drivers of any of the five representative traits (asthma, height, MCV, RA and PrCa). These 78 annotations extend beyond histone marks and DHS features used previously57, accounting also for TF binding (DeepSEA) and CAGE features (Basenji). All 78 annotations are reported in Supplementary Table 11.
