Modeling mutational effects on biochemical phenotypes using convolutional neural networks: application to SARS-CoV-2

Area of Science:

• Virology
• Computational Biology
• Biochemistry

Background:

• SARS-CoV-2 spike protein and human ACE2 are critical for viral entry and evolution.
• Deep mutational scanning (DMS) quantifies mutation effects on protein function.
• Understanding these interactions is key for antibody evasion and therapeutic strategies.

Purpose of the Study:

• To model biochemical phenotypes of SARS-CoV-2 spike protein and ACE2 mutations using neural networks.
• To assess the predictive power of neural networks for binding affinity, protein expression, and antibody escape.
• To explore the utility of deep learning in dissecting molecular mechanisms of viral-host interactions.

Main Methods:

• Massively parallel assays were used for deep mutational scanning of SARS-CoV-2 spike RBD and human ACE2.
• Neural networks were trained on protein sequence mutations to predict biochemical phenotypes.
• Physicochemical properties of amino acids were integrated to enhance predictive accuracy.

Main Results:

• Neural networks accurately predicted binding affinity, protein expression, and antibody escape.
• Integration of physicochemical properties significantly improved prediction accuracy (empirical p < 0.01).
• Neural network predictions showed concordance with molecular dynamics simulations of the spike protein-ACE2 interface.

Conclusions:

• Deep learning models, particularly neural networks, are powerful tools for predicting mutation effects on protein function.
• These models can capture complex interactions missed by conventional methods.
• The findings have significant implications for using deep learning to understand viral evolution and design therapeutics.