PROST

PROST predicts protein stability changes resulting from amino acid substitutions by integrating sequence-derived descriptors and AlphaFold2 structural features into an ensemble machine-learning model.

Key Features:

• Multidimensional Descriptors: Incorporates sequence-based characteristics, physicochemical properties, evolutionary information, structural features from AlphaFold2, and evolutionary-based physicochemical attributes.
• Ensemble Modeling: Combines XGBoost decision trees with an extra-trees regressor to produce ensemble predictions.
• Diverse Training Set: Trained on direct and hypothetical reverse mutations, including the S5294 dataset (S2647 direct + inverse).
• Optimized Parameters: Uses grid search and feature importance analysis to prioritize features and tune model parameters.
• Structure-agnostic Prediction: Produces accurate stability-change predictions without requiring detailed experimental structural data.

Scientific Applications:

• Predicting Stability Changes: Demonstrates improved prediction of stability changes across benchmark datasets, including frataxin and S276, relative to existing predictors.
• Frataxin Case Study: Identifies stability changes at wild-type residues in the frataxin protein.

Methodology:

Extracts descriptors from multiple sequence-based predictors and AlphaFold2 structural features, integrates them into an ensemble model combining XGBoost and an extra-trees regressor, and trains on datasets containing direct and hypothetical reverse mutations (including S5294) with grid-search hyperparameter optimization and feature importance analysis.