To investigate the effect of purifying selection on nonsynonymous variants, we examined the relationship between MAFs of non-synonymous SNVs and functional prediction scores from SIFT, Polyphen2, a likelihood ratio test statistic, and MutationTaster (18). Each prediction score showed a significant (P < 10−16) negative correlation with MAF in the combined sample (Fig. 3A) as well as in each sample separately (18). Moreover, the proportion of predicted deleterious changes was negatively correlated with MAF (Fig. 3B). We next mapped 31,115 nonsynonymous SNVs to known protein structures and classified them into different structural categories (Fig. 3C) (18). Nonsynonymous SNVs in all categories, except sites that contact other protein chains, showed a significant excess of rare variants compared with synonymous sites (Fig. 3C), as expected if subjected to purifying selection. The relative effect sizes show that categories of variants with direct functional importance (e.g., active sites, enrichment of 2.8%; ligand-binding residues, enrichment of 1.7%) are under stronger constraint than categories important for structural stability. The exception is residues that form side-chain hydrogen bonds, which show a 2.8% enrichment of rare variants, suggesting that hydrogen bonds make a large contribution to protein folding and stability.
