Open-access genomic drug resistance prediction tools for Mycobacterium tuberculosis: a systematic review and meta-analysis

Whole-genome sequencing (WGS) accelerates drug-susceptibility testing (DST) in Mycobacterium tuberculosis (Mtb). Open-access software tools have become widely available, but the sources of real-world performance variability remain uncharacterized. We performed a systematic review and meta-analysis of the performance of open-access, independently validated WGS-based DST prediction tools. Bivariate random-effects meta-analysis was performed for six maintained tools (TBProfiler, Mykrobe, PhyResSE, MTBseq, GenTB, and SAM-TB). Bivariate meta-regression identified covariates associated with performance variation. Thirty-nine studies comprising 144,623 genomes were included. For the two most extensively validated tools, TBProfiler and Mykrobe, pooled rifampicin sensitivity was 95.4% (95% CI: 93.5-96.7) and 93.7% (92.0-95.1), with a specificity of 97.3% (95.7-98.3) and 97.0% (94.8-98.3), respectively. For isoniazid, the sensitivity was 92.0% (90.4-93.3) and 88.2% (85.5-90.4) and specificity 97.3% (96.0-98.2) and 97.5% (95.8-98.5). For ethambutol, the specificity was heterogeneous across tools (86.5%-95.4%); for pyrazinamide, the sensitivity varied widely (49.9%-80.6%). For fluoroquinolones, both sensitivity and specificity approached 90%, with heterogeneity. For newer agents, data scarcity precluded meaningful assessment. Meta-regression identified rifampicin resistance prevalence as the dominant predictor of decreased specificity across first-line drugs (beta -1.5 to -3.6 on logit scale, false discovery rate [FDR] q < 0.05), while lineage composition effects were small and confounded. Current open-access WGS prediction tools achieve clinically useful accuracy as rule-out tests for rifampicin, isoniazid, and fluoroquinolone resistance. Predictive performance for second-line drugs is limited by data scarcity. Methodological limitations, including lineage bias, data leakage, and selective sampling, may undermine the tools' generalizability across diverse global tuberculosis populations.