Common sequence variants affect molecular function more than rare variants?

Abstract

Any two unrelated individuals differ by about 10,000 single amino acid variants (SAVs). Do these impact molecular function? Experimental answers cannot answer comprehensively, while state-of-the-art prediction methods can. We predicted the functional impacts of SAVs within human and for variants between human and other species. Several surprising results stood out. Firstly, four methods (CADD, PolyPhen-2, SIFT, and SNAP2) agreed within 10 percentage points on the percentage of rare SAVs predicted with effect. However, they differed substantially for the common SAVs: SNAP2 predicted, on average, more effect for common than for rare SAVs. Given the large ExAC data sets sampling 60,706 individuals, the differences were extremely significant (p-value < 2.2e-16). We provided evidence that SNAP2 might be closer to reality for common SAVs than the other methods, due to its different focus in development. Secondly, we predicted significantly higher fractions of SAVs with effect between healthy individuals than between species; the difference increased for more distantly related species. The same trends were maintained for subsets of only housekeeping proteins and when moving from exomes of 1,000 to 60,000 individuals. SAVs frozen at speciation might maintain protein function, while many variants within a species might bring about crucial changes, for better or worse.

Figure 1

60KE SAVs predicted to have more effect than cross-species variants. SNAP2 predicts the effect of single amino acid sequence variants (SAVs) upon protein function: the higher the score, the more reliable the prediction (horizontal x-axis, toward right); the more negative, the stronger the prediction that the variant is neutral (horizontal x-axis, toward left). The top panel (A) gives cumulative percentages, i.e. the percentage of SAVs in a data set predicted above a certain value, e.g. for SNAP2-score ≥+75, about 6% of all 60KE SAVs are predicted to have an effect; at the same threshold about half of all disease-causing SAVs are predicted to affect function. For 60KE, denisova and chimp, 99.7% confidence intervals (SNAP2-score ±3 standard error of mean) are indicated by dotted lines (indistinguishable for 60KE, barely distinguishable for chimp, clearly visible for denisova). Lower panel (B) gives cumulative accuracy (red: effect-SAVs correctly predicted to have effects, green: neutral-SAVs correctly predicted); here the values accumulate from the extremes to 0, i.e. left-to-right for neutral (green −100 to 0) and right-to-left for effect (red +100 to 0); estimates from cross-validation using only molecular function. For instance, at SNAP2-scores ≥+75 about 88% of all effect-SAVs are correctly predicted. On the other hand, variations between homologs in human and other species (human-denisova, human-chimp, human-mouse, and human-fly) were predicted to be much more neutral (all curves shifted toward lower left corner of neutral variants).

Figure 2

Higher SNAP2-scores imply stronger effect upon molecular function. We classified SAVs from the Protein Mutant Database (PMD) according to their impact upon molecular protein function into three classes (mild, moderate, and severe). Here, we repeat this analysis applying SNAP2 to the subset of human SAVs in PMD. We show density distributions, instead of cumulative. Although the three curves overlap, the shift is significant and consistent (black curve with most effect highest shift to right, orange curve with weakest shift most to the left). Thus, the SNAP2-score correlated with the strength of the effect upon molecular function.

Figure 3

Subsets of “house-keeping” proteins confirmed findings for entire proteomes. We reduced the analysis to SAVs from subsets of orthologs between three organisms (human, chimp, mouse), and with SAVs observed in the 1KG data. For brevity, we referred to those as to “house-keeping” proteins. With respect to the observation for the entire data set (Fig. 1), the curves shifted less strongly, but the main trend remained: a higher fraction of the SAVs in cross-species comparison (human-chimp and human-mouse) was predicted as neutral than for the SAVs between healthy individuals (1KG). Furthermore, the shift between cross-species and 1KG was higher for larger evolutionary distances (more neutral for larger distance).

Figure 4

Common SAVs predicted with more effect than rare SAVs. We grouped SAVs by their observed frequency in 1KG and 60KE exome data: rare (LDAF < 1%: dark blue triangles), uncommon (1% ≤ LDAF < 5%: not displayed), and common (LDAF ≥ 5%: black squares). The potential mutational background for human was estimated by randomly selecting a set of SNV-possible SAVs (gray circles). The curves for rare SAVs were similar to the results for all SAVs (Fig. 1, purple triangles for 60KE) since counting only unique SAVs the results were dominated by rare SAVs. Rare SAVs were predicted below randomly chosen SNV-possible SAVs, although the recent 60KE set came close to random. In contrast, the set of common SAVs remained substantially above the random curve for both common-1KG and common-60KE (Kolmogorv-Smirnov, estimated p-value < 2.2e-16 in both cases).

Figure 5

Methods correlated more for rare than for common 1KG SAVs. Each plot shows the correlation of functional effect scores between one pair of prediction methods for two samples of 1,000 rare and 1,000 common SAVs from 1KG. Results for common SAVs are shown above diagonal, those for rare SAVs are given below the diagonal. With the order of the plots being 1 = SNAP2, 2 = CADD, 3 = PolyPhen-2, and 4 = SIFT, this implied that the plot corresponding to matrix element Pmn compared common SAVs between methods m and n (above diagonal), and the element Pnm rare SAVs between those two (below diagonal). For instance, row = 1/column = 2 gave the correlation between SNAP2 and CADD for common SAVs, while the transposed element row = 2/column = 1 correlated rare SAVs for SNAP2 and CADD. Each point represents a pair of scores for a single SAV, e.g. from SNAP2 and CADD. The predicted score for SIFT has been inverted (1-SIFT) to ease the comparisons. The shape and color reflect the overall method agreement. We use the following code: black squares mark SAVs for which all four methods agree on the binary classification. Blue circles mark SAVs for which all methods but SNAP2 agree; orange triangles mark all other points. The Pearson Correlation Coefficient for all 1,000 SAVs was added above each plot, along with the corresponding value for the full set of all SAVs (in brackets, as in Table 2).