A predictive language model for SARS-CoV-2 evolution

Area of Science:

• Virology
• Computational Biology
• Genomics

Background:

• Predicting viral mutations is crucial for pandemic preparedness, but current models struggle with data requirements and integrating mutation patterns.
• Existing models often fail to capture both the inherent regularity and randomness of viral evolution.

Purpose of the Study:

• To develop a data-efficient language model for predicting SARS-CoV-2 variants and mutations.
• To integrate both the regularity and randomness of viral mutations for improved predictive accuracy.
• To forecast viral evolution and identify emerging variants with public health implications.

Main Methods:

• Constructed "grammatical frameworks" of S1 sequences for dimension reduction and semantic representation to capture latent regularity.
• Incorporated the mutational profile (mutation frequency) to model randomness.
• Utilized sequence data from three time points to detect circulating and predict emerging strains.

Main Results:

• Successfully identified and validated SARS-CoV-2 variants with enhanced infectivity and immune evasion through wet-lab experiments.
• Detected key mutations and circulating strains (XBB.1.16, EG.5, JN.1, BA.2.86) prior to their widespread emergence.
• Predicted previously unknown variants with the potential to cause future epidemics.

Conclusions:

• The developed language model is a fast-responding, concise, and promising tool for forecasting viral evolution.
• The model's approach is potentially generalizable to other viral pathogens for early detection of concerning variants.
• This study offers a valuable method for identifying critical mutation hotspots and warning of emerging public health threats.