Whole-genome sequencing reveals lineage-associated drug resistance and enables machine learning-based prediction in Mycobacterium tuberculosis clinical isolates.

BACKGROUND: Drug-resistant tuberculosis (TB) remains a major obstacle to global TB control, particularly in high-burden settings where timely detection of resistance is limited. This study aimed to integrate whole-genome sequencing and machine learning approaches to characterize the genomic architecture of drug resistance and develop predictive models for phenotypic resistance in clinical Mycobacterium tuberculosis isolates from Kashi, Xinjiang, China. METHODS: A total of 160 clinical M. tuberculosis isolates were collected from culture-confirmed pulmonary tuberculosis patients at First People's Hospital of Kashi over a five-year period (January 2020-December 2024). Whole-genome sequencing was performed, and sequencing reads were mapped to the H37Rv reference genome for high-confidence single nucleotide polymorphism (SNP) and insertion/deletion (INDEL) calling. Resistance-associated mutations were identified using curated resistance mutation catalogues, and comparative genomic analyses were conducted to assess mutation burden across phenotypic resistance groups. Phylogenetic reconstruction based on genome-wide SNPs was performed to evaluate lineage distribution and clustering of resistant strains. In addition, random forest machine learning models were trained using WGS-derived coding and promoter mutations to predict phenotypic resistance to first-line anti-tuberculosis drugs. RESULTS: Phenotypic drug susceptibility testing classified 63 isolates (39.4%) as drug-sensitive, 67 (41.9%) as single-drug resistant (SDR), and 28 (17.5%) as multidrug-resistant (MDR), with a small number of extensively drug-resistant isolates. Resistance was dominated by first-line drugs, particularly isoniazid, streptomycin, and rifampicin. Phylogenetic analysis revealed non-random distribution of resistant isolates, with enrichment of MDR strains within Lineage 2 (Beijing lineage). Genomic profiling demonstrated a progressive increase in resistance-associated mutational burden from drug-sensitive to SDR and MDR isolates, characterized by accumulation of canonical mutations in rpoB, katG, embB, gyrA, pncA, and rrs. Machine learning-based prediction achieved strong performance for rifampicin, isoniazid, and streptomycin resistance, with key resistance-conferring mutations and promoter variants emerging as dominant predictive features. CONCLUSION: This study demonstrates the potential utility of integrating phenotypic testing, whole-genome sequencing, and interpretable machine learning approaches for characterizing and predicting drug-resistant M. tuberculosis. The observed burden of resistance-associated mutations and lineage-associated phylogenetic clustering of resistant isolates support the value of WGS-based surveillance and predictive modelling for tuberculosis control strategies.