Selective Benefits of Sequence-Drug Multimodal Learning for Antimicrobial Resistance Prediction

Antimicrobial resistance (AMR) prediction from genomic data is typically treated as a unimodal problem based on sequence-derived features, yet the reliability of genomic signal alone remains unclear. We evaluate Escherichia coli AMR phenotype prediction using proteome-level protein language model embeddings with multiple instance learning, comparing single-drug (sequence-only) models to multimodal models that incorporate drug representations. Proteome-based models achieve strong performance, with antibiotic-specific models performing best overall, but exhibit a consistent gap between resistant and susceptible prediction. Incorporating drug representations improves susceptibility prediction in settings with weaker or imbalanced genomic signal, while providing little benefit when genomic features are already sufficient. Overall, multimodal learning provides conditional, rather than universal, gains, highlighting when additional biological context is required beyond sequence alone.