TACTIC: A transfer learning framework to predict drug interactions in emerging pathogens

Machine learning (ML) is necessary to efficiently identify potent drug combinations within a large candidate space to combat drug resistance. However, existing ML approaches cannot be applied to emerging and understudied pathogens with limited training data. Hence, we developed a computational framework (TACTIC [transfer-learning and cross-species training to infer combination therapies]) to train ML models on data from multiple model bacteria. TACTIC was built using 2,965 drug interactions from 12 bacterial strains and outperformed traditional ML models in predicting drug interaction outcomes for species with limited training data. Upon analyzing ∼600,000 predicted drug interactions across 18 pathogenic and commensal strains, we identified interactions that are selectively synergistic against Gram-negative and mycobacterial pathogens. We experimentally validated synergistic drug combinations containing clarithromycin, ampicillin, and mecillinam, against clinical strains of M. abscessus, an emerging drug-resistant pathogen. We also confirmed 11 synergistic combinations against Pseudomonas aeruginosa and Staphylococcus aureus strains that cause endophthalmitis.