MLSTA

MLSTA predicts protein stability changes caused by single-site amino acid substitutions by integrating sequence- and structure-based features with machine learning integration strategies to estimate the sign and magnitude of stability alterations.

Key Features:

• Feature Integration: Incorporates sequence- and structure-based features including amino acid substitution likelihoods, equilibrium fluctuations of alpha- and beta-carbon atoms, packing density, local compositional packing, and mobility extent of mutated residues.
• Machine Learning Approaches: Implements early integration of raw input features, intermediate integration using kernels over feature subsets, and late integration that merges decision outputs.
• Classification vs. Regression: Employs classification methods to predict the sign of stability change and regression methods to estimate the magnitude of stability change.
• Reject Option: Applies a reject option to withhold low-confidence predictions to reduce false positives.

Scientific Applications:

• Protein engineering: Predicts effects of amino acid substitutions to guide protein design and stability optimization.
• Functional genomics: Interprets mutation effects on protein stability for studies of phenotype and variant impact.

Methodology:

Uses early, intermediate and late machine learning integration strategies and both classification and regression models; evaluated on datasets S1615 and S2783 (S2783 extracted from ProTherm as of July 2, 2009) with cross-validation and testing: S1615 early integration (sequence+structure) achieved cross-validation accuracy 0.842 and testing accuracy 0.904, intermediate integration benefited from added features such as local compositional packing, S2783 (sequence only) achieved cross-validation 0.835 and testing 0.832, and rejecting 10% of low-confidence predictions with late integration reduced false positives to <0.005.