Protein mutation effects better predicted using multiple energy and language models
MutAtlas: A PDB-Wide Energy-Guided Atlas of Protein Mutation Effects
Machine Learning
Summary
Predicting how changes in proteins affect their form and function is important but data from experiments are very limited. The authors created a large dataset covering many protein structures with all possible single changes and combined signals from physics-based models and AI language models. They found that different models often disagree, not because of random errors but due to different assumptions. To address this, they designed a new method that learns from multiple models while accounting for their disagreements, improving prediction accuracy without using experimental labels. They also provide their dataset and tools for others to use.
What this means in practice
- •For protein engineers: Use combined model predictions to better predict effects of mutations on protein stability and design without needing extensive experimental data.
- •For clinical variant interpreters: Improve interpretation of disease-related protein variants by using multi-model mutation effect predictions for variants lacking experimental evidence.
Authors
Ruihan Guo, Chaoran Cheng, Zhanghan Ni, Neil He, Bangji Yang, Ge Liu
Abstract
Protein mutation effect prediction is fundamental to protein engineering and disease variant interpretation, yet experimentally measured mutation data remain accurate but extremely sparse. To provide scalable supplementary mutation signals, we construct a PDB-wide mutation augmentation dataset that exhaustively enumerates single-site substitutions on experimentally resolved protein structures and aligns mutation signals from physics-based energy models, protein language models, and inverse folding models. Large-scale analysis under a unified mutation preference representation reveals substantial differences in the consistency, concentration, and substitution patterns of mutation distributions across models, indicating that disagreement is pervasive and reflects conflicting inductive biases rather than random noise. Motivated by these observations, we propose an unsupervised multi-source mutation preference distillation framework that learns from relative mutation preferences while explicitly modeling cross-source disagreement. Without using any experimental mutation labels during training, our approach achieves the best overall performance among the evaluated zero-shot baselines and naive multi-source fusion strategies on ProteinGym. We release the dataset and evaluation pipeline to support reproducible studies of protein mutation effects.