The Conservation-Tolerance Paradox: Solvent Exposure and Conformational Flexibility Predict False-Positive Deleteriousness Signals Across Ten Deep Mutational Scanning Datasets

<title>Abstract</title> <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 how often this assumption is contradicted by direct experimental measurement, using 28,229 missense variants with conservation scores across ten independently studied 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 in fact well tolerated (scaled fitness ≥ 0.8) rather than damaging — a figure that replicated almost exactly (70.1%) at an even stricter top-decile threshold. Whether a strongly conservation-flagged variant turned out to be tolerated or damaging was predicted with a large effect size by two independent, easily measured 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 local conformational flexibility (effect size 0.78, p = 4.3×10⁻¹⁸⁹). Both properties remained independently predictive in a joint logistic regression (pseudo-R² = 0.13, both p &lt; 10⁻³²), and the pattern replicated in direction across six of seven individually testable proteins (all p &lt; 0.001), including bacterial enzymes, a molecular chaperone, and eukaryotic protein-interaction domains. A complementary, threshold-free analysis showed that conservation retains real, if more modest, predictive power even at exposed and flexible positions (Spearman ρ from −0.26 to −0.44 across structural strata), indicating that the sharp false-positive pattern is concentrated specifically among variants with extreme, unambiguous experimental outcomes rather than reflecting a wholesale breakdown of conservation as a signal. These results suggest that conservation-based severity calls should be treated with specific, quantifiable caution at solvent-exposed and conformationally flexible positions — a structurally interpretable and testable criterion for a known but rarely quantified failure mode of one of the field's most foundational predictive methods.</p>