Mutation rate heterogeneity biases variant effect prediction and reveals genuine mutational robustness.

Abstract

Variant effect predictors (VEPs) are widely used to interpret the functional consequences of human genetic variation. Because most methods rely on sequence conservation, they implicitly treat conservation as evidence of functional constraint. However, substitution patterns across a phylogeny reflect not only selection but also differences in underlying mutation rates. Here, we show that this creates a systematic confounding: most VEPs capture mutation rate variation and misinterpret it as variation in functional importance. Widely used conservation metrics exhibit a related bias; in particular, phyloP scores correlate strongly with mutation rate even at putatively neutral sites. Consequently, variants at low-mutation-rate sites tend to be predicted as more damaging, and variants at highly mutable sites as more tolerated, than warranted by their true functional impact. We also identify a distinct biological signal in experimental measurements of mutational effects on protein stability: amino acid substitutions that are more likely to arise are, on average, less destabilizing than rarer substitutions. This provides empirical support for mutational robustness in the context of protein stability. However, this relationship is insufficient to explain the mutation-rate dependence observed in current VEP outputs. Together, our findings show that mutation rate heterogeneity systematically biases current variant effect prediction frameworks, highlight the need to model mutation probabilities explicitly in future VEPs, and reveal a genuine biological signal of mutational robustness.