Method / tool

scDINO

Self-supervised Vision Transformer framework adapted for multi-channel fluorescence microscopy that learns cell-state embeddings from unlabeled single-cell image crops, enabling phenotype discovery without annotation.

What it is — scDINO (Pfaendler, Hanimann et al., 2023) adapts the DINO self-distillation framework — student/teacher Vision Transformers trained without labels — for non-RGB, multi-channel fluorescence microscopy. The key engineering change is handling arbitrary channel counts and greyscale inputs that have no natural three-channel encoding. Trained on high-content imaging data of human PBMCs, scDINO's frozen encoder separates immune cell identities and morphological states with accuracy that matches or exceeds supervised baselines on the same embedding, without a single manual annotation at training time.

Role in image-based profiling — scDINO occupies the features stage of the Cell Painting pipeline: it takes single-cell crops produced downstream of segmentation and returns compact embedding vectors. That makes it a direct alternative to CellProfiler's named morphology descriptors — and the trade-off is the same one the From Masks to Morphology: Features & Self-Supervised Embeddings note maps out. Learned embeddings capture signal that named features miss, but the dimensions have no biological label, and the encoder is just as happy to absorb plate-to-plate variation as it is to absorb genuine phenotype. Self-supervised does not mean batch-invariant.

Where it breaks — Because scDINO is pretrained on a specific imaging modality and cell type, transferability to new assays is not guaranteed. A model trained on five-channel PBMC images may embed Cell Painting crops into a poorly structured latent space unless fine-tuned or at minimum evaluated against a percent-replicating control. The Batch Correction Without Erasing Biology problem is, if anything, sharper with learned embeddings than with CellProfiler features, because the nuisance axes are unnamed and harder to audit.

Self-supervised pretraining removes the annotation bottleneck but not the validation requirement. Every scDINO embedding shipped to a downstream readout must clear a reproducibility gate — % Replicating on positive controls — before the signal can be trusted as biology rather than acquisition noise.

References

Appears in these notes

← Back to the constellation