This is also a milestone for our ERC Synergy consortium (AI4AMR) — hopefully the first of many. Many thanks to all who contributed to or funded this work.

Work led by @danielkrentzel.bsky.social, together with @kelvinkho.bsky.social and Julienne Petit. Great collaboration with the labs of @ivo-boneca.bsky.social, @awehenkel.bsky.social and Mark Brönstrup, with help from the microscopy and screening platforms of @pasteur.fr

We believe that this approach is a useful complement to traditional growth assays and will speed up the discovery of badly needed antibiotics with novel MoAs.

This is not restricted to E. coli. We obtained similar results on Klebsiella pneumoniae, classified by the WHO as a critical priority pathogen.

Our model assigns an unseen drug to the correct MoA with high accuracy if the MoA was seen before. Most importantly, it can generally flag a drug with a new MoA, a key advantage in the search for drugs against antimicrobial resistant bacteria.

Unlike traditional growth assays, our method can detect drug effects at subinhibitory conditions (well before bacteria stop growing), enabling more sensitive drug screens.

Can you tell how an antibiotic kills bacteria just by watching them under a regular microscope — without any dyes ? Yes, it turns out, at least if you are a deep neural net.