A human immune cell and a sponge feeding cell can be described through the genes they are using. The lists are enormous, the species are distant and their gene names do not align neatly. TranscriptFormer turns those expression profiles into a common mathematical space in which cells can be compared across the tree of life.[1][2]
The model family was trained on up to 112 million cells from 12 species, spanning an estimated 1.53 billion years of evolution. Its training set ranges from humans, mice and frogs to a freshwater sponge, yeast and the malaria parasite. The Science paper was published in July; Stanford highlighted the wider cell-model programme in September.[1][2][3]
The model learns a cell by predicting its expressed genes
A language model learns structure by predicting missing or next words. TranscriptFormer jointly models gene identities and their expression counts. Each cell becomes a sequence of expressed genes plus an assay token that records how the sample was measured. Expression levels influence attention directly, allowing the model to distinguish which genes are merely present from which are highly active.[3][4]
The representation is useful because cell identity is an activity pattern. A pancreatic beta cell expresses insulin-related machinery; a B cell expresses antibody-related programmes. Disease, development and species change those patterns. Learning them across many atlases gives the model a coordinate system for comparison.[2]
Three released TranscriptFormer variants
| Variant | Training coverage | Published size and best fit |
|---|---|---|
| TF-Metazoa | 112 million cells across all 12 species | 444M trainable + 633M frozen parameters; broadest cross-species generalisationSource 3 |
| TF-Exemplar | 110 million cells from human and four model organisms | 542M trainable + 282M frozen parameters; strong multi-species work with smaller coverageSource 3 |
| TF-Sapiens | 57 million human cells | 368M trainable + 61M frozen parameters; human-only tasksSource 3 |
An evolutionary map can challenge a familiar label
Sponge biologists have debated which sponge cell is most closely related to the neurons of other animals. TranscriptFormer placed the sponge choanocyte near neurons from a roundworm and a frog. The sponge’s so-called neuroid cell, named for neuron-like features, landed closer to a frog gland cell, suggesting a possible digestive role. These are model-supported biological hypotheses to investigate, not a replacement for comparative experiments.[2]
This is exactly where a cross-species representation becomes scientifically interesting. It does not merely attach familiar names to cells. It can reveal that morphology, historical terminology and gene-expression programmes tell different stories, giving researchers a specific relationship to test.[2]
What the published evidence establishes
| Measure | Value | Context |
|---|---|---|
| Training scale | Up to 112 million cells | Twelve species spanning vertebrates, invertebrates, a fungus and a protist.Source 1Source 3 |
| Evolutionary span | 1.53 billion years | The model places highly distant organisms in one representation space.Source 1 |
| Cross-species test | More than 685 million years apart | The paper reports state-of-the-art cell-type classification across that separation.Source 1 |
| Human disease task | Zero-shot identification | The model distinguished disease states without task-specific retraining.Source 1Source 3 |
| Emergent structure | Development, phylogeny and cell hierarchy | These patterns appeared without explicit annotation training for those concepts.Source 1 |
| Published access | MIT-licensed repository and weights | The Virtual Cells Platform provides model variants and download routes.Source 3 |
A model that can be downloaded still needs serious hardware
The published card recommends an NVIDIA A100 with 40 GB of memory for efficient inference and embedding extraction. It says a 16 GB GPU can work with an inference batch size of two. The final checkpoint is about 3 GB, but weights are not the entire runtime budget: activations, input length, batch size and embedding extraction add memory.[3]
That access detail matters. “Open weights” creates a path for laboratories to examine the model, but it does not make a 112-million-cell atlas a phone application. Researchers need the cell data, biological metadata and compute to ask a meaningful question and interpret the answer.
The most promising output is a testable biological connection
TranscriptFormer can classify cells from species absent from training and transfer annotations across species. As a generative model, it can also propose transcription factors and gene-gene relationships that align with independent evidence. The productive workflow is to use those outputs to prioritise a comparison or experiment, then return to the measured biology.[1][4]
Cell atlases are becoming too large for any person to compare gene by gene. TranscriptFormer gives that abundance a navigable structure. Its value will be measured in discoveries like the sponge-cell result: a relationship hidden across species, expressed clearly enough that a biologist can decide what to test next.
Sources & further reading
- TranscriptFormer: A generative cell atlas across 1.5 billion years of evolution James D. Pearce and colleagues / Science via PubMed · 2 Jul 2026 · Accessed 9 Oct 2026
- AI model opens new horizons for cell biology Stanford Report · 14 Sep 2026 · Accessed 9 Oct 2026
- TranscriptFormer model card Chan Zuckerberg Initiative Virtual Cells Platform · Accessed 9 Oct 2026
- A Cross-Species Generative Cell Atlas Across 1.5 Billion Years of Evolution James D. Pearce and colleagues / bioRxiv · 25 Apr 2025 · Accessed 9 Oct 2026