Machine Learning-Driven Simulations of the SARS-CoV-2 Fitness Landscape from Deep Mutational Scanning Experiments.

Area of Science:

• Computational biology
• Virology
• Protein engineering

Background:

• Predicting protein variant effects is essential for understanding viral evolution, disease mechanisms, and protein design.
• Complex interactions like allosteric and epistatic effects make modeling protein sequence-structure-function relationships challenging.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting the effects of protein mutations, specifically focusing on the SARS-CoV-2 receptor binding domain (RBD).
• To assess the models' ability to predict combinatorial mutation effects and their performance in extrapolating to new viral variants.

Main Methods:

• Trained supervised ML models using deep mutational scanning (DMS) data of SARS-CoV-2 RBD sequences with associated ACE2 binding affinity.
• Combined ML with Markov Chain Monte Carlo (MCMC) simulations to characterize the RBD fitness landscape and predict evolutionary patterns.
• Introduced Mavenets, a software package for applying these modeling strategies.

Main Results:

• ML models outperformed traditional methods (additive/averaging effects) in predicting combinatorial mutation impacts.
• Models demonstrated strong extrapolative power, accurately ranking Omicron variants even when trained on near wild-type (WT) sequences.
• MCMC simulations generated sequence profiles comparable to high-fitness sequences in DMS data and predicted mutations in previously unseen variants.

Conclusions:

• The developed ML models offer valuable insights into RBD sequence-function relationships and the prediction of emerging viral strains.
• This approach, utilizing DMS and ML, provides a novel perspective for predicting viral evolution and can be applied to other evolutionary prediction tasks.
• The Mavenets tool facilitates the application and further development of these predictive modeling strategies.