An explainable artificial intelligence framework reveals mutations associated with drug resistance in Mycobacterium tuberculosis

• Scientific questions. • The emergence of drug resistance in Mycobacterium tuberculosis (MTB) has significantly increased the complexity of tuberculosis treatment and transmission risks, underscoring the critical need to elucidate resistance mechanisms for rapid diagnostics, optimized therapeutic strategies, and global tuberculosis burden mitigation • Evidence before this study. • While genomics-based statistical association studies have identified many resistance-associated mutations, MTB continues to develop new resistance mutations under the persistent selective pressure of anti-tuberculosis drugs, emphasizing the necessity to further refine existing catalogues of resistance mutations. Although deep learning models have been applied to explain resistance-associated mutations at the population level, individual-level resistance mechanisms remain underexplored. • New findings. • We propose an explainable artificial intelligence framework called xAI-MTBDR. It combines multiple machine learning models with the SHAP method, aiming to identify new drug resistance-associated mutations and predict the drug resistance of MTB. To our knowledge, this is the largest study utilizing the MTB dataset (39,145 isolates) to assess the performance of similar models for drug resistance prediction. Compared to existing methods, xAI-MTBDR not only significantly improves the prediction accuracy of drug resistance but also provides detailed explanations of each mutation’s contribution to resistance. The predictions of xAI-MTBDR are highly consistent with the MTB drug resistance mutation catalogue published by the WHO, validating its broad applicability and reliability. At the individual level, xAI-MTBDR can deeply analyze the unique drug resistance mechanisms of each isolate. We revealed 27 potential resistance markers, some of which are closer to the corresponding drug targets in protein structure than known resistance mutations. • Significance of the study. • xAI-MTBDR not only predicts the drug resistance of MTB but also provides the rationale for each isolate’s prediction, helping researchers decide how to utilize the results. This framework can also be used to identify new drug resistance-associated mutations for the study of resistance mechanisms. Understanding the mechanisms of drug resistance in Mycobacterium tuberculosis (MTB) is essential for the rapid detection of resistance and for guiding effective treatment, ultimately contributing to reducing the global burden of tuberculosis (TB). Under anti-TB drugs pressure, MTB continues to accumulate resistance loci. The current repertoire of known resistance-associated mutations requires further refinement, necessitating efficient methods for the timely identification of potential resistance sites. Here, we introduce xAI-MTBDR, an explainable artificial intelligence framework designed to identify potential resistance-associated mutations and predict drug resistance in MTB. It outperforms state-of-the-art methods in predicting drug resistance for all first-line drugs, and scoring each mutation’s contribution to resistance. By leveraging public whole-genome sequencing data from nearly 40,000 MTB isolates, the framework identified 788 candidate resistance-related mutations and revealed 27 potential resistance markers, several of which are positioned closer to their respective drugs in protein structures than known resistance mutations, suggesting a potentially more direct role in mediating resistance. Furthermore, these scores enabled the framework to efficiently subgroup isolates with different resistance mechanisms and reflect varying levels of resistance. The framework serves as a valuable tool for accurate detection of drug-resistant MTB and offers new insights into its underlying mechanisms.