Pasteur AI: when artificial intelligence learns to ‘read’ bacteria to fight antibiotic resistance

Why is it so difficult to discover new antibiotics?

Finding a molecule that kills bacteria is complicated. Understanding how it kills them is even more difficult. Yet it is this question of ‘mechanism of action’ that makes all the difference: it enables us to determine whether a candidate antibiotic is truly innovative and therefore capable of overcoming existing resistance.

Current methods often rely on fluorescent markers, are expensive, and only detect an effect when the molecule is concentrated enough to inhibit growth or eliminate bacteria. As a result, some promising molecules, which are active at very low doses, go completely unnoticed.

AI that can ‘see’ the effect of antibiotics

Several teams at the Pasteur Institute (Christophe Zimmer, Ivo G. Boneca, Anne-Marie Wehenkel and Mark Brönstrup) have worked together to overcome these limitations using deep learning. Their artificial intelligence model – a convolutional neural network inspired by the way our brain processes images – was trained to detect the subtle changes in shape that an antibiotic causes in bacteria.

To train their AI model, the scientists used bright-field microscopy images – without fluorescent labelling – of Escherichia coli bacteria exposed to 22 antibiotics with eight different modes of action.

Very promising results, even at very low doses

The results are particularly encouraging. Indeed, artificial intelligence recognises an antibiotic’s mode of action with near-perfect accuracy. Better still, it detects the effect of a molecule at sub-inhibitory concentrations – that is, even before the bacteria stop multiplying. This capability is a major asset for identifying candidates that conventional tests would have ruled out.

Even more surprising is that the model is capable of indicating when a molecule acts in an unprecedented way. This ability could prove decisive in discovering new classes of antibiotics – the very ones most needed to overcome current resistance.

The approach has also been validated on another bacterium, Klebsiella pneumoniae, which causes serious infections and is often multi-drug resistant. This provides initial evidence that the method could be applied to other microorganisms.

A boost to the fight against antibiotic resistance

When used alongside conventional testing, this AI could enable the screening of thousands of potential antibiotic molecules more quickly and at a lower cost.

The stakes are high: antibiotic resistance is a public health issue that lies at the heart of the Pasteur Institute’s scientific priorities. Artificial intelligence models such as this one could therefore change the outcome of this race against time.

This work has been carried out as part of an ERC-funded research programme (ERC Synergy AI4AMR) involving the research groups led by C. Zimmer, I. Boneca and M. Brönstrup.

Reference: Krentzel, D., Zimmer, C., Kho, K., Petit, J. et al. Deep learning recognises antibiotic modes of action from brightfield images. Nature Communications (2026). https://doi.org/10.1038/s41467-026-76355-0.