CovTransformer: A transformer model for SARS-CoV-2 lineage frequency forecasting.

Area of Science:

• Virology
• Computational Biology
• Epidemiology

Background:

• Hundreds of SARS-CoV-2 lineages circulate globally, necessitating accurate forecasting of lineage frequencies.
• Predicting lineage dominance is crucial for understanding pathogenicity and immune escape of future variants.

Purpose of the Study:

• To develop a reliable machine learning model for SARS-CoV-2 lineage frequency forecasting.
• To address limitations of traditional regression-based approaches due to noisy and biased lineage data.

Main Methods:

• Developed CovTransformer, a machine learning model based on the transformer architecture.
• Trained and tested the model on SARS-CoV-2 lineage data from the UK and USA, then evaluated generalization to other countries and US states.
• Compared CovTransformer's performance against the multinomial regression model used in Nextstrain.

Main Results:

• CovTransformer accurately predicts lineage frequencies up to two months into the future globally and at the US-state level.
• The model significantly outperformed the Nextstrain multinomial regression model.
• Retrospective analysis showed CovTransformer identifies dominant lineages an average of 7 weeks in advance.

Conclusions:

• Transformer models offer a promising approach for accurate SARS-CoV-2 forecasting and pandemic surveillance.
• CovTransformer provides a robust tool for identifying rapidly expanding SARS-CoV-2 lineages.
• This advancement aids in proactive research on variant pathogenicity and immune escape.