Cell-DINO Learns Rich Cellular Morphology Embeddings
Researchers led by Moutakanni et al. publish Dec. 29, 2025 a study introducing Cell-DINO, adapting DINOv2 to fluorescent microscopy to learn image embeddings without labels. Across Human Protein Atlas and Cell Painting datasets, Cell-DINO outperforms supervised and self-supervised baselines — for example, 70% higher accuracy with 1% labels and 24% over another SSL method — improving low-label phenotyping and MoA prediction.
Key Points
- 1Shows DINOv2-based Cell-DINO learns rich cell-image representations without manual annotations.
- 2Delivers large accuracy gains—up to 70% with 1% labels—over supervised baselines on HPA.
- 3Enables low-label phenotyping and improved mechanism-of-action prediction from Cell Painting and HPA images.
Scoring Rationale
Strong empirical gains with public code validate applicability; limitation is domain-focused adaptation rather than a novel learning algorithm.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
