Fig. 5: Variants prioritized by deep learning models.
a Dosage box plots showing covariate-adjusted quantile-transformed phenotypes against minor allele counts for variants in SLC9A5 and ANGPTL3/DOCK7. A predicted splice variant 16:67270978:G:A is negatively associated with HDL cholesterol (p = 7.83 × 10−12, score test), whereas intronic 1:62598067:T:C is negatively associated with Triglycerides (p = 1.37 × 10−25, score test). The numbers in brackets denote the number of carriers of at least one alternative allele. Center lines denote the medians. The lower and upper hinges indicate 25th and 75th percentiles. Whiskers extend to the largest/lowest values no further than 1.5 × IQR away from the upper/lower hinges and black points denote outliers. Maxima and minima, from left to right: −5.24 to 5.18, −4.62 to 4.5. b DeepRiPe binding probabilities for 1:62598067:T:C for three RBPs in HepG2 cells. While predicted probabilities for the reference sequence are ambiguous, the alternative allele shifts binding probabilities in favor of QKI and HNRNPL. All RBPs with absolute predicted variant effects above 0.2 and binding probabilities greater than 0.5 for either reference or alternative alleles are shown.
