Robust AI Framework for Comprehensive Tuberculosis Drug Resistance Profiling with Rapid Adaptability

Tuberculosis remains the leading cause of death from a single infectious agent, with drug-resistant tuberculosis, particularly multidrug-resistant and extensively drug-resistant strains, posing major challenges for timely treatment. Whole-genome sequencing can accelerate resistance detection, but current genomic and machine-learning approaches typically predict resistance to individual drugs, do not directly infer regimen-relevant resistance profiles, and generalise poorly across regions or newly introduced drugs. We developed MuseAMR, a multimodal, multi-label deep-learning framework that predicts both individual-drug resistance and clinically actionable composite phenotypes from Mycobacterium tuberculosis genomes, with robust cross-regional performance and few-shot adaptation to emerging drugs. Trained on 10,886 isolates and externally validated on 18,334 isolates from six global regions, MuseAMR improved sensitivity for second-line drug resistance (0.857 versus 0.655) and MDR/pre-XDR profiles compared with WHO catalogue-based prediction while maintaining high specificity. It also showed robust cross-regional performance and few-shot adaptation to bedaquiline, delamanid and linezolid using 5-20 resistant isolates, with attribution analyses recovering established resistance loci. These results support its potential for regimen-level tuberculosis resistance profiling and surveillance.