A new map of protein changes helps predict how mutations work
Researchers built a large dataset to align different ways of predicting protein mutations.
Scientists created MutAtlas, a new resource for understanding protein mutations. This tool uses a large dataset to align different types of mutation signals. It covers many protein structures found in the Protein Data Bank.
Predicting how amino-acid substitutions change a protein is vital for protein engineering and disease variant interpretation. However, accurate experimental data is very sparse. It only covers a few proteins and specific sites.
To help, researchers built a PDB-wide mutation augmentation dataset. This dataset lists every single-site substitution on resolved protein structures. It aligns signals from physics-based energy models, protein language models, and inverse folding models.
Physics-based energy models like FoldX estimate changes in stability. These models use force-field models to look at local structural stability. Protein language models and MSA-based models use evolutionary history to find statistical preferences. Inverse folding models use backbone geometry to estimate amino-acid compatibility.
The study found that different models often disagree. This disagreement is pervasive. The differences in mutation patterns reflect conflicting inductive biases rather than random noise.
The researchers proposed an unsupervised multi-source mutation preference distillation framework. This method learns from relative mutation preferences. It also explicitly models the disagreement between different sources. It does not use any experimental mutation labels during training.
This new approach achieved the best overall performance among zero-shot baselines on ProteinGym. The researchers released the dataset and evaluation pipeline to help other studies.
Key facts
- MutAtlas aligns mutation signals from physics, evolution, and structure models.
- The dataset covers single-site substitutions across the Protein Data Bank.
- The new framework performs well on ProteinGym without using experimental labels.
- Disagreement between models reflects different inductive biases.