Explainable machine learning reveals evolutionary signals in influenza haemagglutinin

Area of Science:

• Virology
• Computational Biology
• Machine Learning

Background:

• Identifying amino acid changes linked to phenotypic shifts is vital for viral surveillance.
• Traditional metrics like evolutionary rates and entropy measure variability, not necessarily phenotypic importance.

Purpose of the Study:

• To demonstrate that explainable machine learning (ML) models offer a complementary approach for viral sequence analysis.
• To identify mutations associated with host adaptation and control for confounding factors.

Main Methods:

• Utilized 39,121 hemagglutinin (H3) protein sequences with passage annotations.
• Trained gradient-boosted decision trees with encoded amino acids.
• Employed SHapley Additive exPlanations (SHAP) values to quantify site importance.

Main Results:

• Achieved 81% accuracy in a passage classifier, distinguishing egg-grown from unpassaged isolates with 90% recall.
• Predicted sample collection dates with R2=0.98 and a mean absolute error of 74.5 days.
• Identified sites highly enriched for experimentally validated antigenic sites, outperforming traditional metrics.

Conclusions:

• Explainable ML can reveal critical substitutions and provide tree-free molecular dating.
• This methodology can transform passage metadata into a valuable experimental probe for viral evolution studies.