Few-Shot Viral Variant Detection via Bayesian Active Learning and Biophysics.

Area of Science:

• Virology
• Computational Biology
• Machine Learning

Background:

• Early detection of high-fitness viral variants is crucial for pandemic response.
• Limited experimental resources often hinder timely identification of emerging variants.

Purpose of the Study:

• To develop an active learning framework for rapid, few-shot prediction of viral variant fitness.
• To accelerate the identification of concerning viral variants, including SARS-CoV-2, before widespread circulation.

Main Methods:

• Integration of a protein language model (ESM3), Gaussian process with uncertainty estimation, and a biophysical model.
• Few-shot learning approach applied to predict variant fitness using limited experimental data.
• Benchmarking on historical SARS-CoV-2 data and deep mutational scans.

Main Results:

• Accelerated identification of high-fitness variants by up to fivefold compared to random sampling.
• Reduced experimental characterization to less than 1% of possible variants.
• Identified key mutation sites associated with antibody escape and preserved ACE2 binding, with a two-year predictive advantage.

Conclusions:

• The active learning framework serves as an effective early warning system for identifying potentially dangerous viral variants.
• Incorporating uncertainty in variant selection allows for broader exploration of the sequence landscape.
• The framework can aid in the early detection of emerging viruses with pandemic potential.