New Method Improves Ancestral Protein Sequence Reconstruction by Accounting for Coevolution

A preprint from arXiv introduces a constraint-aware optimization framework for predicting how protein mutations affect thermodynamic stability (ΔΔG), targeting three known weaknesses in current models: poor generalization to out-of-distribution proteins, systematic bias when predicting forward versus reverse mutations, and underrepresentation of rare stabilizing mutations. The framework adds Balanced Mean Squared Error, a Siamese anti-symmetric regularizer, and an OOD-margin consistency loss to the existing SPURS model backbone without changing its architecture, improving Spearman correlation on key benchmarks such as S669 (0.486 to 0.540) and S461 (0.653 to 0.711). Notably, the authors find that anti-symmetric training does not fully eliminate forward-reverse bias, suggesting improvements stem from implicit regularization rather than strict thermodynamic constraint enforcement. Separately, a bioRxiv preprint addresses ancestral sequence reconstruction (ASR), a method used to infer ancient protein sequences and study molecular evolution. Most ASR methods assume protein sites evolve independently, ignoring epistatic interactions between residues that are critical for protein stability and function; the new framework combines standard phylogenetic inference with Direct Coupling Analysis (DCA) to enforce residue-residue coevolutionary constraints. Tested on beta-lactamases and DNA-binding domains, the coevolution-aware approach produces ancestral sequence ensembles that are both phylogenetically consistent and statistically compatible with natural protein families, bridging the gap between overly deterministic and unconstrained probabilistic reconstruction methods.