For our five representative traits asthma, RA, MCV, PrCa and height, we further compared functionally informed PRS-EUR using IMPACT to models using 123 DeepSEA and Basenji deep learning annotations54–57, 220 cell-type-specifically expressed genes (SEG)52, and 513 cell-type-specific histone modification tracks (CTS)3 (Fig. 5c, Supplementary Table 20 and Methods). To our knowledge, deep learning annotations have not been applied previously to improving PRS model performance. IMPACT explained greater phenotypic variance on average (mean R2 = 4.2%, s.e.m. = 1.0%) than the top deep learning annotations (3.2%, s.e.m. = 0.8%, one-tailed paired Wilcoxon P = 0.03). This observation was individually consistent for four of five traits (four one-tailed difference of means P < 0.006), while only trending higher for asthma (P=0.13). IMPACT also explained greater phenotypic variance on average than SEG (0.9%, s.e.m. = 0.2%, one-tailed paired Wilcoxon P = 0.03) and this difference was individually detected for each of five traits (all one-tailed difference of means P < 3.4 × 10−6). This trend was not as strong when comparing IMPACT to CTS (R2 = 2.6%, s.e.m. = 0.5%, one-tailed paired Wilcoxon
