Techniques for Developing Reliable Machine Learning Classifiers Applied to Understanding and Predicting...

Area of Science:

• Computational biology
• Biophysics
• Bioinformatics

Background:

• Machine learning is transforming scientific prediction, but success hinges on data quality, feature selection, and robust validation.
• Predicting protein interaction hot spots is crucial for understanding biological function and designing therapeutics.
• Existing methods may lack robustness or require extensive data for accurate predictions.

Purpose of the Study:

• To develop and share protocols for building robust datasets for machine learning in biology.
• To rigorously compare predictive classifiers and evaluate model robustness.
• To create a reliable classifier, Hotspotter, for predicting protein interaction hot spots.

Main Methods:

• Utilized the scikit-learn Python library for developing and comparing predictive classifiers.
• Implemented rigorous validation techniques, including 500-fold repartitioning of training and test sets.
• Developed an intuitive method for quantifying feature importance in predictions.
• Trained and tested Hotspotter on a curated dataset of 1046 alanine scanning mutation sites from 97 protein complexes.

Main Results:

• Accessible surface area and evolutionary conservation were identified as key features for predicting hot spots.
• A structure-based Hotspotter classifier demonstrated robust prediction of energetically important residues.
• A sequence-based variant of Hotspotter, using secondary structure, showed comparable robustness with fewer features.
• Hotspotter successfully identified critical hot spots on the ACE2 receptor relevant to COVID-19.

Conclusions:

• Hotspotter provides a robust and validated method for predicting protein interaction hot spots.
• The tool can guide protein interface design, ligand development, and protein-protein docking predictions.
• Accessible surface area and evolutionary conservation are powerful predictors of protein binding sites.