AI reads bacterial images to reveal how TB drugs work

Tuberculosis (TB) remains the world’s deadliest infectious disease caused by a single pathogen, and rising antimicrobial resistance (AMR) makes finding new drugs with novel mechanisms of action (MoAs) urgent. Traditional phenotypic screens of chemical libraries can find compounds that kill or slow bacteria, but they usually do not reveal how those compounds work. That gap hampers the targeted selection and development of drugs with new MoAs. To address this problem, Daniel Krentzel and colleagues developed a deep learning (DL)-based screening approach that reads high-throughput microscope images to connect chemical effects with biological pathways. They used Corynebacterium glutamicum ( Cglu ) as a lab-friendly surrogate model for Mycobacterium tuberculosis ( Mtb ) and trained a model to recognize imaging patterns caused by different perturbations. The idea is simple but powerful: if a drug produces the same image signature as a genetic mutation, the drug may act on the same pathway. By teaching a computer to see those signatures, the team aimed to make it possible to predict targets and discover compounds with genuinely novel MoAs directly from images.

The method is built around a convolutional neural network that takes high throughput images as input and is trained to distinguish between different MoAs. Using this DL model the researchers showed it could robustly differentiate between the MoAs of established antibiotics and could correctly recognize the MoA of antibiotics that had not been included in training. They also found that inhibitors sharing the same MoA — including MoAs not previously seen by the model — cluster together and separate from other reference drugs, a behavior that supports new MoA discovery. Crucially, the model linked images of chemical (drugs) and genetic (mutants) perturbations that target similar pathways, demonstrating a route toward mutant-based target prediction of compounds that act through novel MoAs, directly from high-content images. Finally, features extracted with the DL model recovered known biological relationships from high-throughput images alone, illustrated by analysis of the cell cycle of Cglu as a case study.

The approach described by Krentzel’s team suggests a practical way to add mechanistic insight to phenotypic screening without extra biochemical assays. By classifying image-derived signatures of MoA and by mapping drug effects onto mutant phenotypes, the technique could help researchers prioritize compounds that work by new mechanisms, reducing wasted effort on redundant targets. That capability is important for TB drug discovery, where novel MoAs are needed to outpace AMR. Beyond drug discovery, the ability to recover known biological relationships such as stages of the cell cycle from images alone points to broader applications in fundamental microbiology and in studies that seek to understand how genes and chemicals shape bacterial physiology. In short, the work opens a path from high-throughput imaging to actionable hypotheses about drug targets and bacterial biology.