AI reads an antibiotic’s action from images of treated bacteria

An AI model can use brightfield (transmitted-light) microscope images to recognise how antibiotics affect E. coli. A study first published in September trained on 22 drugs spanning eight target classes.[1]

Cell shape was a useful clue long before this model. In a 2013 study, Poochit Nonejuie and colleagues stained bacterial membranes and DNA, then measured features including length, width, circularity and chromosome organisation. Compounds affecting protein production, DNA replication or cell-wall construction produced different profiles. Tetracycline, for example, left wide cells with ring-shaped chromosomes; rifampicin left chromosomes more spread out. The measurements grouped compounds by their biological effects even when their chemical structures differed.[2]

A profile narrowed the affected pathway; establishing an unfamiliar compound’s precise molecular target still required complementary experiments.[2]

The new network learnt from drug and concentration labels, receiving only the brightfield channel. Researchers trained on three plates and tested on the fourth, rotating that role. For familiar drugs at a concentration that halved growth, action-class accuracy averaged 95.7% per well across four plates, with a standard deviation of 3.6 percentage points.[1]

Discovery poses a tougher question: what happens when the drug itself was absent from training? After withholding a compound, the researchers assigned its action by comparing its image features with those of reference drugs on the same test plate. Accuracy at the same concentration averaged 75.0%, with a 4.1-point standard deviation across four test plates; other drugs still had to represent that action.[1]

A distinctive cellular response can steer the next biochemical experiment, helping researchers decide which mechanism to investigate.[1][2]

Report an error or suggest a correction