AI Potluck
Back to Gap Map Model components / Scientific AI models

Geneformer

Theodoris Lab
open source / Overall score: 2.4

Geneformer is the Theodoris lab's transformer model pretrained on human single-cell transcriptomes to make context-aware predictions in network biology. It ships 10M, 104M, and 316M-parameter checkpoints plus a cancer-tuned variant, and the same repository carries the tokenizer, the DeepSpeed pretraining example, fine-tuning notebooks, embedding extraction, and in silico perturbation code.

The weights, the code package, and the Genecorpus training data are all distributed from Hugging Face; there is no separate GitHub repository. Verified 2026-09-10 via the Hugging Face API and model card.

Openness

5 high confidence
5.0
weights
open(Apache-2.0 ungated V1-10M, V2-104M, and V2-316M checkpoints)
data
open(Genecorpus-30M and Genecorpus-104M published as Hugging Face datasets)
code
open(tokenization, DeepSpeed pretraining, fine-tuning, and in silico perturbation code in the model repository)
license
Apache-2.0(OSI)

Geneformer publishes the weights, both named pretraining corpora, and a pretraining script in one Apache-2.0 repository, which is the full from-scratch release the top rung asks for.

Adoption

1 high confidence
1.0

1,737 downloads in the trailing 30 days for the declared Hugging Face repository, which carries every checkpoint in the family.

Capability

3 high confidence
3.0

Two attributes are established directly. The long list of downstream fine-tuning tasks is not converted into extra features, the three parameter sizes are not scales, and gene-level and cell-level outputs are treated as one modality rather than two.

Verified 2026-09-10