Integrating evolutionary and protein language model representations for thermodynamically consistent stability prediction

EvoPhi-DDG and JanusDDG averaged into a thermodynamically constrained ensemble preserve exact antisymmetry and improve stability-change prediction, reaching Pearson 0.66 on PTmul-NR.

• Why it matters: Learned stability predictors may violate reversal and path-independence properties even when physics demands them, and evolutionary alignment signals could complement single-sequence protein language models.
• What they did: The authors reformulate MSAesm_ddG into EvoPhi-DDG by antisymmetrizing predictions and zeroing hidden biases to yield a potential-difference form, then extend it to multiple substitutions and ensemble it with JanusDDG using fixed equal-weight averaging.
• The result: The ensemble reduces multisite pathway-closure RMSE relative to JanusDDG and improves Pearson correlations on S669 (0.59) and S96 (0.55), and on PTmul-NR achieves 0.66, supporting training-free integration with physical consistency.

AI summary of the abstract · may contain errors · not medical advice · read the paper.