BayeStab

BayeStab predicts changes in protein thermostability resulting from mutations by estimating Gibbs free energy (ΔG) changes and quantifying predictive uncertainty.

Key Features:

• Graph Neural Networks for Feature Extraction: Employs graph neural networks to extract structural and sequence-derived protein features for structure–property prediction.
• Bayesian Neural Networks with Concrete Dropout: Integrates Bayesian neural networks with concrete dropout to produce probabilistic predictions and decompose uncertainty into data noise and model-induced uncertainty.
• High Generalization and Symmetry Performance: Demonstrates high generalization and symmetry performance across benchmark datasets S669, S611, S350, and Myoglobin.
• Uncertainty Quantification: Estimates upper bounds of prediction uncertainty to inform confidence in computed Gibbs free energy changes upon mutation.

Scientific Applications:

• Disease Research: Assesses how specific mutations alter protein thermostability and Gibbs free energy to inform studies of disease mechanisms.
• Therapeutic Design: Supports prediction of mutation effects on protein stability to guide design of therapeutics targeting unstable or misfolded proteins.
• Protein Engineering: Enables engineering of proteins with tailored thermostability for industrial or medical applications by predicting stability changes from mutations.

Methodology:

Uses graph neural networks for protein feature extraction and Bayesian neural networks with concrete dropout for calibrated probabilistic inference and uncertainty decomposition, with performance evaluated on S669, S611, S350, and Myoglobin.