iTB-DRS: a multi-stage feature selection framework to identify Mycobacterium tuberculosis drug-resistance using single-nucleotide polymorphisms

Rapid and accurate prediction of Mycobacterium tuberculosis drug resistance is crucial for effective tuberculosis control. However, data from genome-wide association studies are often characterized by high dimensionality and many redundant features, posing significant challenges to prediction models. To address these challenges, we propose a novel computational framework called iTB-DRS, which employs a hierarchical multi-stage feature selection strategy. The workflow begins with data encoding followed by a four-step selection mechanism, including (i) filtering noise using a Mean Decrease Accuracy based on Random Forest; (ii) ranking features by Maximum Information Coefficient to capture nonlinear association; (iii) performing pairwise predictive evaluation to identify SNP-SNP combinations with increased joint predictive value; and (iv) optimizing feature subsets through backward feature elimination. Finally, a set of Random Forest classifiers is trained to predict drug resistance to different drugs. When using an independent data set, iTB-DRS generally outperformed RF, LASSO + RF, and mRMR + RF baselines across multiple drug metric combinations. Simultaneously, the developed framework achieved good predictive accuracy on the training data set close to that of the computational pipeline TB-Profiler frequently used to predict Mycobacterium tuberculosis lineage and drug resistance. However, a performance gap between these two predictive tools emerged on the independent data set for several second-line drugs, which is mainly due to the catalog-based prior knowledge used by TB-Profiler. Taken together, the data obtained showed that iTB-DRS provides a low-complexity and interpretable solution that balances predictive accuracy and computational feasibility, offering a complementary research tool for genomic surveillance and candidate-marker discovery.IMPORTANCEDrug-resistant tuberculosis poses a significant threat to global health, narrowing the range of effective treatment options and increasing mortality rates. Rapid identification of drug resistance is crucial for selecting appropriate treatments during therapy. However, existing genomic tools often rely on predefined mutation catalogs, which may not fully capture complex or emerging resistance patterns. Therefore, studies on the development of novel effective methods for prediction of drug-resistant tuberculosis are of great importance. Here, we introduce a data-driven computational framework abbreviated as iTB-DRS that predicts Mycobacterium tuberculosis drug resistance using whole-genome sequencing data and a multi-stage feature selection strategy. Unlike catalog-based tools that use embedded prior knowledge, the proposed framework can be retrained on newly accumulated data to evaluate mutations and drugs not represented in existing catalogs. This flexibility positions iTB-DRS as a complementary, hypothesis-generating tool for genomic surveillance rather than a stand-alone replacement for validated catalog-based clinical software.