Explainable AI reveals drug-resistance mutations in tuberculosis

Drug resistance in Mycobacterium tuberculosis (MTB) is one of the biggest barriers to successfully treating tuberculosis and reducing its global burden. Under the pressure of anti-TB drugs, MTB accumulates genetic changes that can prevent medicines from working, and the current catalogue of resistance-linked mutations still needs refinement. Rapid, reliable detection of resistance and a clearer picture of how specific mutations alter drug response are essential for getting the right treatment to patients quickly. To tackle this problem, a team led by Guoping Zhao developed xAI-MTBDR, an explainable artificial intelligence framework built to discover resistance-associated mutations and to predict drug resistance in MTB. Rather than treating the prediction as a black box, xAI-MTBDR assigns scores that quantify how much each mutation contributes to resistance. The researchers applied this approach to a large public resource, drawing on whole-genome sequencing data from nearly 40,000 MTB isolates, with the explicit goal of finding new candidate resistance sites and improving the accuracy of resistance detection.

xAI-MTBDR was tested against existing approaches and reported outperforming state-of-the-art methods in predicting drug resistance for all first-line drugs, while also providing a per-mutation score that reflects its contribution to resistance. Using public whole-genome sequencing data from nearly 40,000 MTB isolates, the framework flagged 788 candidate resistance-related mutations and highlighted 30 potential resistance markers deserving closer attention. Structural analysis—comparing the location of these markers to known resistance mutations—showed that several of the new candidates sit closer to their respective drugs, suggesting a plausible mechanistic link. Beyond listing mutations, the mutation scores produced by xAI-MTBDR allowed the researchers to subgroup isolates by different resistance mechanisms and to capture varying levels of resistance among strains. Together, these results indicate that xAI-MTBDR can both improve predictive performance and offer interpretable insights into which mutations matter most.

The explainable nature of xAI-MTBDR makes it more than a prediction engine: it provides interpretable evidence that can help researchers and clinicians understand how specific genetic changes drive resistance. By refining the repertoire of resistance-associated mutations and flagging new candidate markers, the framework offers a resource that could speed up laboratory detection, improve surveillance of emerging resistance, and shepherd the design of more targeted diagnostic tests. Its ability to subgroup isolates by mechanism and to indicate different resistance levels adds granularity that may be useful for research into drug action and for public health decision-making. While further validation and clinical integration will be needed, xAI-MTBDR represents a promising tool for accurate detection of drug-resistant MTB and for uncovering the molecular underpinnings of resistance.