Fig. 2.

Penalized regression model identifies genes relevant to platelet and hematopoietic biology. a SNP scores for platelet model training SNPs, or a set of validation platelet trait SNPs, were significantly higher than genome-wide SNP scores. Bars represent mean ± SEM, ****p < 0.0001 by ANOVA. b Performance comparison of our platelet trait model to DeepSEA [28], GWAVA [17], and CADD [29] for training platelet trait SNP identification. AUC values are shown in the legend. c Performance comparison for validation platelet trait SNP identification. There was substantial variation in the number of validation SNPs recognized and scored by each model. AUC values shown in the legend represent prediction accuracy in identifying validation SNPs for LASSO (n = 94 SNPs), GWAVA (n = 16), CADD (n = 18), and DeepSEA (n = 94) vs ~ 10,000 random control SNPs. d Platelet and hematopoiesis pathways [30] identified by the highest-scoring (top 1%) SNPs genome-wide for the indicated models, excluding established platelet trait loci [6] (FDR, false discovery rate)