Self-supervised models improve Cell Painting profiling in drug-discovery benchmarks
Published in Artificial Intelligence, Bioinformatics, Drugs development & discovery, Journals, Tools.
Published October 31, 2026.
Self-supervised AI models can extract useful cellular profiles directly from microscopy images, according to a large benchmark published October 29 in Nature Methods.
Researchers pretrained four model families on roughly 1.77 million Cell Painting images covering 119,876 compounds from seven laboratories. Their pipeline operates without the cell segmentation and handcrafted feature extraction used in classical morphological profiling.
DINO and DINOv2 generalized across several datasets and cell types. On the EU-OPENSCREEN benchmark, the authors reported 31–37% gains over CellProfiler in site-averaged mean average precision for retrieving compound mechanisms of action.
These are improvements on specific retrieval benchmarks, not evidence of a corresponding increase in drug approval rates or clinical effectiveness. The models were pretrained on chemically perturbed U2OS cells; batch correction, interpretability and transfer to more complex cell types remain challenges.
Pretrained models and code are available through DeepCellMorpho. All authors disclose employment at Johnson & Johnson and salary and/or stock interests.
Source: Abdelmoula, W.M., Albuquerque, T., Jaensch, S. et al., Nature Methods, October 29, 2026. Read the original study.