Predicting mutant outcome by combining deep mutational scanning and machine learning.

Area of Science:

• Protein Engineering
• Computational Biology
• Biophysics

Background:

• Deep mutational scanning generates extensive data on protein variant function.
• Predicting the impact of mutations computationally is challenging but crucial for protein design.
• Understanding mutation effects aids in developing proteins for therapeutic and industrial applications.

Purpose of the Study:

• To develop a novel machine learning approach for predicting functional mutation outcomes from deep mutational scanning data.
• To leverage sequence and structural features for accurate prediction of protein variant properties.
• To demonstrate the generalizability and utility of the predictive models across diverse protein targets.

Main Methods:

• Utilized sequence (one-hot) and structural features of known protein variants.
• Trained machine learning models to predict functional properties of uncharacterized variants.
• Validated the approach on five diverse datasets: APPI protease specificity, IGBPG stability, and GFP fluorescence.

Main Results:

• The machine learning models successfully predicted functional mutation outcomes across various protein targets.
• Performance was measured by correlation between predicted and observed properties, and by enrichment of potent variants.
• The models demonstrated generalization capabilities, accurately predicting properties of variants with single and multiple mutations.

Conclusions:

• The developed machine learning approach effectively predicts functional mutation outcomes in deep mutational screens.
• This method facilitates exploration of the protein mutational landscape and guides experimental design.
• The approach accelerates the discovery of novel protein variants with desired properties.