BayeStab: Predicting effects of mutations on protein stability with uncertainty quantification.

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in protein science

Background:

• Accurate prediction of protein thermostability changes upon mutation is essential for understanding protein function, disease mechanisms, and therapeutic development.
• Existing computational methods often face challenges in generalizing to proteins with no sequence homology and providing reliable uncertainty estimates for their predictions.

Purpose of the Study:

• To develop a robust and generalizable computational method for predicting protein thermostability changes upon mutation.
• To leverage graph neural networks (GNNs) for enhanced protein feature extraction and Bayesian neural networks (BNNs) for uncertainty quantification.

Main Methods:

• Utilized graph neural networks for extracting structural features from proteins.
• Implemented concrete dropout enabled Bayesian neural networks to infer predictive models and quantify prediction uncertainty.
• Tested the developed method, BayeStab, on diverse datasets including S669, S611, S350, and Myoglobin.

Main Results:

• BayeStab demonstrated high generalization performance across multiple independent test datasets.
• The method exhibited strong performance in predicting protein thermostability changes, outperforming existing approaches.
• Bayesian uncertainty decomposition provided insights into data noise and model limitations, indicating the inherent task difficulty.

Conclusions:

• BayeStab offers a powerful and reliable tool for predicting protein thermostability changes, addressing limitations of previous methods.
• The integration of GNNs and BNNs provides accurate predictions and crucial uncertainty estimates, advancing structure-property prediction.
• The developed web server and open-source code facilitate broader application and reproducibility in the scientific community.