Abstract
<jats:p>Evolutionary conservation is one of the most widely used signals for predicting whether a missense mutation is deleterious: the more strongly a position resists substitution across evolution, the more damaging a mutation there is assumed to be. We tested this assumption using 28,229 missense variants with conservation scores across ten deep mutational scanning (DMS) datasets, including the hereditary-cancer gene BRCA1. Among variants at the most conservation-flagged positions (top quartile of the PolyPhen-style conservation-violation score, Δpsic), 70.7% of those with an unambiguous experimental outcome were well tolerated (scaled fitness ≥ 0.8) rather than damaging, a result replicated (70.1%) at an even stricter top-decile threshold. Whether a strongly conservation-flagged variant proved tolerated or damaging was predicted by two structural properties: relative solvent accessibility (common-language effect size 0.69, Mann-Whitney p = 2.6×10⁻⁶⁹) and crystallographic B-factor, a proxy for conformational flexibility (effect size 0.78, p = 4.3×10⁻¹⁸⁹). Both remained independently predictive in logistic regression (pseudo-R² = 0.13, both p < 10⁻³²), and the pattern replicated across six of seven proteins (all p < 0.001). A complementary analysis showed that conservation retains predictive power even at exposed and flexible positions (Spearman ρ = −0.26 to −0.44), indicating that false-positive signals are concentrated among variants with extreme experimental outcomes rather than reflecting a general failure of conservation. These results suggest that conservation-based severity predictions should be interpreted cautiously at solvent-exposed and conformationally flexible positions, providing a structurally interpretable criterion for a known but rarely quantified limitation of conservation-based prediction.</jats:p>