AI learns cellular signatures across laboratory boundaries

A study published in Nature Methods on 29 October tests whether AI can recognise related chemical effects across laboratories. Its models learn representations directly from Cell Painting images, without first outlining individual cells for feature extraction.[1]

The experiment begins with biology made visible. In the original Cell Painting protocol, six fluorescent dyes illuminate structures including nuclei, mitochondria and internal membranes. Five imaging channels collect their signals. Software then measures cell shape, brightness, texture and spatial relationships, assembling a profile of the treated sample.[2]

That profile makes a useful comparison possible: chemicals with different molecular structures may produce similar cellular changes. Matching their profiles can suggest a shared biological action. The assay surveys many features together, so investigators can look beyond a single change chosen in advance.[2]

The new comparison pretrained four learning approaches on about 1.77 million images from seven laboratories. DINO learns by matching representations of differently cropped and altered views of the same image, without biological labels. DINO and DINOv2 performed particularly well, but correction for experimental batches remained important: raw representations could group samples by their experimental history.[1]

The older protocol explains why that history matters. Uneven illumination can distort intensity measurements; differences in cell seeding and growth can alter the resulting profiles. Correcting these effects helps distinguish a treatment response from the conditions under which it was photographed. Some biological effects, however, produce no visible signature under the chosen stains, cell type and observation time.[2]

A separate test used a cultured cell type called HepG2 at four sites. After excluding toxic and phenotypically inactive compounds, DINO retrieved 33% more mechanisms of action than CellProfiler with agreement across at least three sites. Yet two inhibitor classes were recovered only by CellProfiler. The result broadens the set of recognisable responses; it does not make either representation exhaustive.[1]

Report an error or suggest a correction