Multi-Drug Generalization of Genomic Tuberculosis Resistance Prediction: Does a Permutation-Invariant Encoder Transfer to Underrepresented Cohorts?

Whole-genome machine-learning predictors of Mycobacterium tuberculosis (MTB) drug resistance are trained and benchmarked almost entirely on genomes from Europe, North America, and East Asia, yet the highest burden of multi-drug-resistant TB falls on populations that are largely absent from these datasets. We ask a deployment-critical question : does a modern genomic resistance model actually transfer to the populations it is meant to serve? Building on the permutation- invariant variant-set representation introduced by VAMP-Net, we distill its Set-Transformer pathway (Path-1) into a single-path, four-drug multi-task predictor for rifampicin (RIF), rifabutin (RFB), isoniazid (INH), and ethambutol (EMB), and stress-test it under geographic distribution shift. Using the 10,310-isolate merged CRyPTIC compendium, we auto-detect a 710-isolate cohort of isolates from Muslim-majority, largely Global-South countries (6.9% of the data) and evaluate three regimes: in-distribution cross-validation, a strict Leave-Country-Out zero-shot transfer, and few-shot domain adaptation with bootstrap confidence intervals. The model is strong in-distribution (AUROC 0.91–0.97) and, encouragingly, transfers with only minor loss for RIF, RFB, and INH (AUROC drop ≤ 0.036). Ethambutol, however, collapses under geographic shift (AUROC −0.14, recall −0.30). Few-shot adaptation at 10–50% of local data did not significantly recover EMB performance under our compute-reduced settings. Our findings argue that per-drug transferability audits on underrepresented cohorts should be a standard reporting requirement for clinical genomic models.