One model maps cells across tissues and species
A Nature paper from Stanford researchers introduces a universal cell embedding, a foundation model trained without cell-type labels on a corpus of more than 36 million cells. New experiments can be placed in that space without retraining the model. On a later human cell atlas it separated cell types while mixing experimental batches, and it could embed cells from species absent from training, including green monkey, naked mole-rat, and chicken. In one green-monkey lymph node sample, a group labeled as B cells sat near T cells and preferentially expressed T-cell marker genes. The model is a representation for analysis and hypothesis generation, not a diagnosis.
The record ends here. Everything below this line is invented.
What if every cell, from any species, had one address?
Part one
The Nearest Cell
Tomas worked the night desk of the atlas, which was not a room of slides anymore but a map with one address for every cell that had ever been sequenced into it. A new vial from a field station came in unlabeled except for the animal and the tissue. He did not retrain anything. He dropped the counts into the map and waited for neighbors.
The lymph-node cells fell where lymph-node cells fall. One tight group, marked B on the packing slip, sat against the T-cell shore instead. Tomas pulled the genes the map did not need in order to place them. The group was bright for the markers he associated with T cells, and dim for the B-cell ones.
He could correct the slip and go home. The station had already booked the cells as a control for a vaccine study that started at dawn. A wrong shore would make the control look like a response.
He highlighted the breakaway group. The nearest named centroid, from a different species entirely, was a T cell.
Tomas deleted the B from the slip and did not type a replacement.
What’s real
- The model embeds new cells without retraining
- It was trained without cell-type labels
- A mislabeled green-monkey cluster expressed T-cell markers
- It is a research map, not a diagnosis
What’s invented
- A staffed night atlas that files live vials
- Tomas and the vaccine control at dawn
- A packing slip he can erase
- A rule that the nearest centroid becomes the name
Part two: The Name He Would Not Type
Part two isn’t on sale yet. Check back soon.
Filed under biology, single cell, foundation models, cell biology.