Deep learning recognises antibiotic modes of action from brightfield images.

The antimicrobial resistance crisis urgently calls for antibiotics with novel modes of action (MoAs). While growth inhibition assays can identify antibiotic molecules, they miss promising compounds below inhibitory concentrations and cannot reveal their MoA. Imaging-based profiling of drug-treated bacteria can inform on MoA, but current approaches generally require fluorescent labelling and/or inhibitory concentrations and it remains unclear whether compounds with novel MoAs can be robustly detected. Here, we demonstrate a deep learning approach to recognise antibiotic MoAs from unlabelled images. We train a convolutional neural network to predict treatment conditions from brightfield images of Escherichia coli exposed to antibiotics covering eight MoAs. Our approach can detect drug exposure at subinhibitory concentrations and allows near-perfect MoA recognition, even when trained on only eight images per treatment condition. Previously unseen compounds are assigned to their correct MoA with good accuracy if the MoA is represented in the training data, and our method can robustly detect MoA novelty in five out of six considered MoAs, enabling microscopy-based identification of new antibiotic classes. We also achieve near-perfect MoA recognition in Klebsiella pneumoniae, suggesting applicability to other species. Our approach complements growth inhibition assays and is poised to improve the search for innovative antibiotic compounds.