Protein stability predictors no longer contradict themselves: EvoPhi-DDG turns an alignment model into a potential, so forward and reverse scores cancel exactly
Synopsis
Protein stability-change predictors often disagree with themselves: swapping the two sequences does not flip the sign of the prediction, and paths that detour through intermediate states at the same site fail to close. Without retraining, EvoPhi-DDG rewrites the family-alignment predictor MSAesm_ddG as a difference of an effective potential conditioned on the mutation site and alignment context, giving exact sign reversal and exact same-site additivity under a fixed alignment context. Averaged with equal weights with the published JanusDDG, it reaches a Pearson correlation of 0.59 on S669, against 0.55 for JanusDDG and 0.53 for EvoPhi-DDG.
Interpretation
Stability-change prediction can now reverse sign exactly and satisfy path additivity exactly at a fixed mutation site and alignment context, while predictive accuracy stays essentially unchanged. On S669 the original MSAesm_ddG had a mean absolute reversal defect of 0.20 kcal/mol, a largest defect of 1.19 kcal/mol, and an identity prediction of −1.38×10⁻³ kcal/mol rather than zero; with nine intermediate residues its direct-versus-path Pearson correlation fell from 0.96 to 0.84 and closure RMSE rose from 0.26 to 0.61 kcal/mol. On the 669 direct S669 variants, the antisymmetrized MSAesm_ddG eliminated both the reversal defect and the identity defect exactly, giving a direct-versus-path Pearson of 1−1.38×10⁻⁶ and a closure RMSE of 1.50×10⁻³ kcal/mol at nine intermediates; S669 Pearson moved from 0.54 to 0.53 and S96 from 0.47 to 0.48.
The construction rewrites an existing alignment-conditioned prediction head as the difference of an effective potential for each sequence, with no additional training. The same parameters are evaluated in both mutation directions and half their difference is taken, after which the head's hidden biases are set to zero; with zero hidden bias the antisymmetrized ReLU head becomes a linear functional of its input, so intermediate-state contributions cancel in the sum. Zeroing the hidden biases changed predictions on S2450 by a mean of 5.84×10⁻⁴ kcal/mol and at most 2.88×10⁻³ kcal/mol; multiple substitutions are handled by scoring each substitution on the wild-type background and each reversion on the fully mutated background, then antisymmetrizing the whole variant, which retains Pearson 0.66, MAE 1.45 kcal/mol and RMSE 1.95 kcal/mol on the 82 PTmul-NR variants.
Perspective
The demonstrated scope is the three single-substitution benchmarks S669, S461 and S96 plus the multiple-substitution benchmark PTmul-NR, all using published checkpoints, fixed alignment contexts and one shared set of random paths, with no additional training, score calibration or ensemble fitting; exact additivity holds only at a fixed mutation site and alignment context.
Closure error remains on multisite paths, because the readout depends on the mutation site and a fixed-site potential cannot define a common potential across different sites; PTmul-NR contains only 13 wild-type sequences, so those comparisons are sample-sensitive; S461 has only 23 stabilizing variants; and on S96 the ensemble's Pearson gain over JanusDDG has an interval that includes zero.