Researchers trying to identify how a potential drug works can expose cultured cells to thousands of compounds and examine how the cells change. This technique, known as morphological profiling or Cell Painting, has traditionally depended on software measuring predefined properties such as shape, texture and intensity. A Nature Methods study published on 29 October instead asks whether image models trained without labels can discover more informative patterns directly from microscopy pictures.[1]
Walid Abdelmoula and collaborators conducted a large comparison of self-supervised learning approaches on the Joint Undertaking for Morphological Profiling Cell Painting dataset. Their pretraining pool covered 119,876 compounds and approximately 1.77 million images collected at seven laboratories. These figures describe data scale; they are not the number of validated drugs or successful therapies. The study’s scientific question is how well the resulting representations generalize to compounds and biological settings that were not supplied as annotated targets during pretraining.[1]
The comparison separates learning methods
The researchers evaluated contrastive approaches including DINO, masked-image reconstruction methods such as MAE and iTPN, and a hybrid represented by DINOv2. They compared the learned image features with CellProfiler, a conventional system built around engineered measurements. In replicate retrieval, where different images from the same compound should be recognized as related, the strongest self-supervised models outperformed CellProfiler on the held-out JUMP–Target2 evaluation with 302 annotated compounds; the study reports statistical significance at P at or below 0.001.[1]
In a separate EU-OPENSCREEN Bioactives collection of 2,464 compounds, DINO and DINOv2 improved site-averaged mean average precision for mechanism-of-action retrieval by 31% to 37% relative to CellProfiler. The researchers also report that the contrastive approaches consistently surpassed the masked-image models across their benchmarks. This distinction is valuable because more complicated pretraining is not automatically better: what matters is whether the representation preserves the biological relationships the laboratory intends to query.[1]
Retrieving a mechanism is not proving it
A retrieval benchmark asks whether compounds known to share a mechanism tend to appear near each other in the model’s embedding space. It does not prove that the model has identified a novel molecular target, demonstrated therapeutic benefit or predicted an adverse reaction in a patient. An imaging system may group drugs because they produce similar visible cell stress, not because they bind the same protein. Mechanistic interpretation still needs orthogonal biological assays.[1]
The authors’ use of data from multiple laboratories addresses an important source of failure. Cell appearance can shift because of staining protocols, optics, plate layouts and processing conditions. A model that uses those technical signatures rather than chemistry may perform impressively in one facility and poorly elsewhere. Multi-site pretraining and evaluations on different perturbation sets help test this, but they do not eliminate all possible batch effects.[1]
Self-supervised pretraining also changes the economics of dataset preparation. Labeling every image with a validated biological mechanism is expensive and often impossible for exploratory screens. Learning from unannotated images allows researchers to use a much larger pool, then apply a smaller annotated test set to assess whether the features are useful. The study shows such transfer can work on specified tasks; its comparison should not be stretched into a claim that expert annotations are no longer needed at any stage.[1]
The immediate use is to compare cellular responses and prioritise follow-up experiments. The authors have released pretrained models and inference code, but the weights carry a CC BY-NC-SA 4.0 licence, including a non-commercial restriction. Laboratories can inspect and reproduce the reported representation-learning work under those terms; identifying a new drug mechanism still requires biological confirmation.[1]
Sources & further reading
- Advancing morphological profiling with self-supervised learning through comparative insights into representation learning of cellular perturbations Nature Methods · 29 Oct 2026 · Accessed 30 Oct 2026