Machine Learning on the Impacts of Mutations in the SARS-CoV-2 Spike RBD on Binding Affinity to Human ACE2 Based on...

Area of Science:

• Virology
• Biochemistry
• Computational Biology

Background:

• Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) accumulates mutations in the spike receptor-binding domain (RBD).
• These mutations can alter binding affinity to the human angiotensin converting enzyme 2 (hACE2) receptor, impacting viral transmissibility.
• Deep mutational scanning (DMS) is a key experimental method for assessing mutation effects on protein function.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting the impact of SARS-CoV-2 mutations on spike protein function.
• To improve physics-based models by incorporating local environment information.
• To assess the performance of a dual-encoding convolutional neural network (CNN) model against existing protein language models.

Main Methods:

• Built ML models using SARS-CoV-2 DMS data, with input features from Rosetta-computed energy terms and local residue environment information.
• Employed a CNN model utilizing amino acid sequence, physicochemical, and biochemical properties.
• Applied transfer learning to fine-tune CNN models for specific SARS-CoV-2 variants (Alpha, Delta, Omicron subvariants).

Main Results:

• ML models showed good agreement with experimental DMS data.
• The dual-encoding CNN model outperformed three popular protein language models on multiple DMS datasets.
• Fine-tuned CNN models successfully predicted variant-specific effects, including for Omicron subvariants.

Conclusions:

• ML models trained on DMS data can predict the effects of single and multiple point mutations.
• These models provide valuable insights for viral surveillance and understanding SARS-CoV-2 evolution.
• The dual-encoding CNN model offers a robust, alternative ML approach for DMS studies without requiring 3D structural information.