Machine learning framework for cost effective deep mutational scanning through targeted substitution profiling.

Area of Science:

• Biochemistry
• Computational Biology
• Genomics

Background:

• Deep mutational scanning (DMS) generates extensive protein variant effect maps but is labor-intensive.
• Machine learning (ML) offers a promising avenue to predict functional impacts from limited experimental data, potentially reducing experimental costs.

Purpose of the Study:

• To develop and evaluate an ML classifier for predicting the functional impact of amino acid substitutions.
• To assess the feasibility of using ML with targeted DMS to reduce experimental burden while maintaining accuracy.

Main Methods:

• Trained an ML classifier on normalized DMS scores from SARS-CoV-2 main protease (Mpro).
• Employed brute-force feature selection to identify minimal feature subsets for accurate classification.
• Evaluated classifier performance using zero-shot prediction on independent datasets, including other enzymes and protein interaction/folding systems.

Main Results:

• Achieved classification accuracy exceeding 90% with minimal feature subsets (six substitution scores per residue).
• Demonstrated strong zero-shot prediction performance on other enzymes.
• Observed more modest performance on systems assessing protein folding and/or protein-protein interactions.

Conclusions:

• Targeted DMS combined with ML significantly reduces sequencing and reagent costs.
• This integrated approach offers a practical and efficient method for accelerating variant effect prediction.
• The findings highlight the potential of ML to streamline functional genomics studies.