Can we predict whether a mutation in the SARS-CoV-2 spike protein is likely to:

• Increase ACE2 binding affinity?

• Escape neutralizing antibodies?

• Persist in the population (i.e., be evolutionarily fit)?

Data Sources

• GISAID

• ProteinGym

• EMBL EBI PDBe-KB, NExtstrain, or COV-GLUE

Approach

• Preprocess protien sequences of spike proteins into 3-mer or 6-mer tokens

• Fine-tune ESM-1b or ProtBERT using

  • classification (beneficial vs neutral vs deleterious mutation)
  • regression (fitness scores, binding energy)

• Optional: Include position-aware embeddings for mutations (attention to RBD)

• Evaluate against CNN or MLP baselines

• Visualize embeddings or attention heatmaps to identify important residues