A Stanford model trained on 112 million cells sorted a sponge cell in with worm neurons

The design is a masked-prediction task. TranscriptFormer was trained the way a large language model is trained, by looking repeatedly at inputs with parts missing and adjusting parameters until the missing parts come out right, with gene expression values standing in for words ^[5]. That buys generality. A model that never sees a cell-type label can still learn which genes appear together and at what levels, so it can place an unfamiliar cell near familiar ones. It also sets the limit. Proximity in that space is a statement about expression similarity, and about nothing else.

Stephen Quake, the Stanford bioengineering professor who co-led the work, calls the result a universal space, mathematical rather than physical ^[6]. "All the cells of every organism on Earth can sit in that space, and you can look at their relationship to each other," he said ^[6].

Twelve species and 112 million cells work out to an average of about 9.3 million cells per species ^[14]. The Stanford Medicine account does not break down the split, and the split governs how much of the geometry the thinly sampled organisms get to shape. The twelve include mice, rabbits, chickens, zebrafish, fruit flies, the African clawed frog, the parasite that causes malaria, a sea urchin, a sponge and the roundworm C. elegans ^[3].

Two of the reported abilities are the ones that would change how much animal work a research program needs. The model identifies cell types when handed expression data from a species outside its training set ^[9], and the team reported that it can separate healthy cells from diseased ones ^[10]. The account describes both without reporting accuracy, naming the held-out species, or saying which reference labels the calls were scored against ^[15]. The figures that would settle the question are per-species accuracy on organisms left out of training and the provenance of the labels used to score them.

The sponge results show the same gap from the other side. A choanocyte sitting near roundworm and frog neurons is consistent with neurons and choanocytes sharing an ancestor, and equally consistent with two lineages converging on similar expression; the embedding does not separate those two stories ^[7]. The neuroid result is the more immediately useful one, because it disagrees with a name assigned on morphological grounds and proposes a specific function to test, digestion ^[8]. Quake also expects the model to help design cell therapies eventually, on the grounds that it can imagine functional cells that do not currently exist ^[16].

"How do we think about all these genes at once? The human brain can't," Quake said ^[13]. He described what the two papers, published in Nature and Science, are for in narrow terms: "This is a tool to help skilled biologists think about complex problems" ^[11]. On timing he said, "These papers mark the beginning of, we hope, a whole new field and a decade of work" ^[12].