#18 · ML / AI methods

Richer input representations sharpen pathogen resistance and antigenic site prediction

The dominant thread is input representation rather than model architecture. The HIV study reports that ESM-2 protein language model embeddings with attention pooling outperform XGBoost across 18 drugs, with attention providing interpretability. The H3N2 study applies interpretable gradient-boosted models with SHAP attribution to two hemagglutination inhibition datasets, yielding a consolidated ranking of candidate antigenic sites. What is new is the use of learned protein embeddings and SHAP- or attention-based attribution to make predictions interpretable. The TB atlas differs in method: it is a computational scoring of first-line resistance mutations across 31,073 isolates for effects on viability and drug binding, not a learning approach, and is included only as adjacent evidence.