Deep learning extracts MoA-specific signatures from high-throughput images of chemically and genetically perturbed Corynebacteria.

To address drug-resistant tuberculosis, the leading single-pathogen infectious killer, drugs with novel modes of action (MoAs) are urgently needed. Phenotypic screening of chemical libraries can identify antimicrobial compounds, but standard screens cannot reveal the MoA of hits, limiting targeted selection of compounds with novel MoAs. Here, we develop a deep learning (DL) model to screen drug-treated Corynebacterium glutamicum (Cglu), a surrogate for Mycobacterium tuberculosis. We train our DL model to distinguish between MoAs directly from high-throughput images. Our model robustly classifies MoAs of established antibiotics and recognizes the MoA of previously unseen antibiotics. Inhibitors with a previously unseen MoA cluster together and apart from reference drugs, enabling the detection of novel MoAs. Moreover, our model links images of chemical (drugs) and genetic (mutants) perturbations targeting similar pathways, supporting mutant-based target prediction of compounds with novel MoAs directly from images. Last, our DL model recovers known biological relationships from images alone using the Cglu cell cycle as a case study.