DeepCellMorpho benchmarks self-supervised models for Cell Painting

DeepCellMorpho compares four self-supervised models trained on roughly 1.77 million Cell Painting images. In the Nature Methods study, DINO and DINOv2 performed strongly on compound and mechanism-of-action retrieval, using image representations without cell segmentation or hand-engineered features. The authors provide inference code, pretrained models and embeddings; model weights and embeddings are restricted to noncommercial use. Cross-laboratory batch effects and interpretability remain limitations.