Single-cell language model that writes a cell's ontology type and tumour meta-program profile from scRNA-seq, and names unseen types in context.
No providers recorded yet. Browse all providers
Single-cell RNA sequencing of a tumour raises two questions about every cell: what lineage it belongs to (a CD8 T cell, a fibroblast, a malignant epithelial cell) and what it is currently doing (cycling, responding to hypoxia, undergoing epithelial-mesenchymal transition). Annotation tools usually answer the first with a classifier head fixed to a predefined label set, and the second with signature scores such as ssGSEA or AUCell, or with NMF programs that must be refit on each cohort. Neither produces one portable description of a new cell, and a classifier head cannot name a type it was never trained on.
CellScribe, from InstaDeep and the Laboratory of Mathematics and Computer Science (MICS) at CentraleSupélec, Université Paris-Saclay, treats both questions as text generation. A frozen scFoundation encoder turns a cell's expression profile into an embedding, and that embedding modulates every layer of a small pretrained language model through adaptive layer normalization (AdaLN). Asked "what is the cell type?", the model writes a Cell Ontology term; asked about state, it writes the active meta-programs from the Curated Cancer Cell Atlas (3CA) with activity scores and a dominant hallmark. Because the output is language, the same model can also read several labelled cells in one prompt and assign a query cell to a type outside its training vocabulary.
The work was posted to bioRxiv in October 2026 as a paper for the GenAI for Health workshop at NeurIPS 2026. It joins InstaDeep's earlier omics models BulkRNABert and MOJO in the same open-source repository.
The encoder is the 100M-parameter scFoundation model, which maps counts over 19,264 genes to a 3,072-dimensional cell representation; the decoder is Qwen2.5-0.5B-Instruct (24 layers, hidden size 896). Both stay frozen, and the 79.9M trainable parameters are the cell projection and 48 AdaLN projectors; the released checkpoint holds 673.6M parameters in total. Cell typing was trained on CELLxGENE Census (62.6M cells, over 700 ontology types) and cell state on 3CA (4.07M cells from 99 cancer studies, 41 meta-programs in 11 hallmarks), with study-holdout splits throughout. Training ran in three stages: supervised fine-tuning on 20M sampled cells, GRPO alignment on 600k cells with 12 completions per prompt, and in-context post-training on 6M cells.
On held-out CELLxGENE studies (4.44M cells, 309 types) CellScribe reached macro-F1 0.223 and weighted-F1 0.484, against 0.194 and 0.366 for a supervised probe on frozen scFoundation embeddings and macro-F1 below 0.09 for Cell2Sentence, CellWhisperer and InstructCell. On held-out 3CA malignant cells it reached 0.344 hallmark accuracy and 0.330 NDCG@5 against 0.227 and 0.233 for the probe. For 50 cell types excluded from all training, one labelled example per candidate raised exact-match accuracy from 0.020 to 0.364 under free generation and to 0.882 when decoding was constrained to the three demonstrated types.
CellScribe is aimed at research-grade profiling of tumour ecosystems: annotating lineages across malignant and microenvironment cells while scoring each malignant cell's meta-program activity, without refitting NMF per cohort. The in-context mode covers datasets containing rare or fine-grained types absent from training, where a handful of expert-labelled cells stand in for retraining. The Hugging Face release ships a raw-counts-in predictor and example cells for each mode.
CellScribe shows that a generative interface supports what fixed classifiers cannot: ontology-aware reward alignment, multi-label state output and demonstration-based typing of new classes. The paper also reports that a simple probe on frozen scFoundation embeddings is already a strong typing baseline, which places much of the discriminative power in the upstream encoder. State prediction is limited to the 41 NMF-derived 3CA meta-programs, the evaluation is entirely in silico, and no journal version has been published. The code is Apache-2.0, but the bundled scFoundation encoder weights remain under BioMap's non-commercial model license, which restricts commercial use of the full checkpoint.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.