Single-cell foundation models: what each record says it is, and what evaluation papers report

Research question

• Which single-cell foundation models are in the library, and how does each record describe its training scale and design?
• What do the evaluation papers in the library report about zero-shot use, scale and interpretability of these models?

Current understanding

Our reading of the library records: construction claims and evaluation claims about single-cell foundation models are different kinds of evidence. The construction records describe models by training scale and design; the evaluation records (zero-shot, perturbation, scale and interpretability studies) report limitations. We have not reconciled them against the full texts. Certainty: high that the records say this; not assessed beyond that. Several models appear only as preprints.

Scope and how to read this page. Scope: library records only (title, venue, one-line summary and verification source); we have not read the full texts for this page, and every statement below is limited to what those records state. Preprints are marked and are not peer reviewed. This page makes no claim about glioma or other tumours. VC0008 VC0009

Construction claims (peer-reviewed records). scGPT is recorded as a generative pretrained transformer trained on 33M cells (Nature Methods 2024). Geneformer is recorded as a context-aware transformer enabling in silico deletion and network predictions (Nature 2023). scFoundation is recorded as a 100M-parameter model trained on about 50M human cells (Nature Methods 2024). GeneCompass is recorded as a knowledge-informed cross-species foundation model (Cell Research 2024). Certainty: high that these are the records' descriptions; the training-data composition (tissues, cell types) is not recorded in the library. VC0008 VC0009 VC0011 VC0029

Construction claims (preprints, not peer reviewed). UCE is recorded as a cross-species zero-shot cell embedding model (preprint). TranscriptFormer is recorded as a generative cross-species cell atlas of 112M cells across 12 species; the record is the bioRxiv preprint, and its verification note says a peer-reviewed version was published in Science, which we have not separately logged. AIDO.Cell is recorded as dense models with transcriptome-scale context (preprint). Certainty: moderate; preprint records only. VC0012 · preprint VC0021 · preprint VC0022 · preprint

Other designs in the library. Cell2Sentence (cells expressed as text for large language models; preprint) and C2S-Scale (scaling such models to 27B parameters; preprint); CellPLM, a cell language model that goes beyond single cells to include spatial context (preprint); scBERT, an early pretrained language model for cell-type annotation (preprint; its record notes a peer-reviewed version in Nature Machine Intelligence). Two application records: a Nature Protocols paper on discovering candidate therapeutic targets with Geneformer in silico perturbation, and the ISP3 platform for Geneformer-based sequential multi-gene in silico perturbation (preprint). Certainty: high that the records say this; no performance claims are made here. VC0030 · preprint VC0031 · preprint VC0032 · preprint VC0023 · preprint VC0010 VC0038 · preprint

What the evaluation records report. A Genome Biology paper reports zero-shot limitations of scGPT and Geneformer (VC0050). A BMC Genomics benchmark compares foundation cell models with simple baselines for post-perturbation RNA-seq prediction (VC0049). A PLoS Computational Biology paper assesses scale and predictive diversity of Geneformer-based models (VC0057). A BMC Genomics study reports that attention-derived edge scores in single-cell foundation models add no gene-regulatory-network information beyond baselines (VC0056). Taken together, the records describe limitations of current evaluations for these models, but each study uses its own tasks and metrics, so they are not interchangeable. Certainty: high that the papers report this as summarised in the library; we have not compared their protocols. See also the companion pages on linear baselines and evaluation pitfalls. VC0050 VC0049 VC0057 VC0056

Evidence map

Primary studies & datasets

• VC0008 scGPT: toward building a foundation model for single-cell multi-omics using generative AI. — Cui H, Wang C, Maan H, Pang K, Luo F, Duan N, Wang B., Nature methods, 2024
• VC0009 Transfer learning enables predictions in network biology. — Theodoris CV, Xiao L, Chopra A, Chaffin MD, Al Sayed ZR, Hill MC, Mantineo H, Brydon EM, Zeng Z, Liu XS, Ellinor PT., Nature, 2023
• VC0011 Large-scale foundation model on single-cell transcriptomics. — Hao M, Gong J, Zeng X, Liu C, Guo Y, Cheng X, Wang T, Ma J, Zhang X, Song L., Nature methods, 2024
• VC0029 GeneCompass: deciphering universal gene regulatory mechanisms with a knowledge-informed cross-species foundation model. — Yang X et al., Cell research, 2024
• VC0012 · preprint Universal Cell Embeddings: A Foundation Model for Cell Biology — Rosen Y, Roohani Y, Agrawal A, Samotorcan L, Tabula Sapiens Consortium, Quake SR, Quake SR, Leskovec J., bioRxiv, 2023
• VC0021 · preprint A Cross-Species Generative Cell Atlas Across 1.5 Billion Years of Evolution: The TranscriptFormer Single-cell Model — Pearce JD, Simmonds SE, Mahmoudabadi G, Krishnan L, Palla G, Istrate A, Tarashansky A, Nelson B, Valenzuela O, Li D, Quake SR, Karaletsos T., bioRxiv, 2025
• VC0022 · preprint Scaling Dense Representations for Single Cell with Transcriptome-Scale Context — Ho N, Ellington CN, Hou J, Addagudi S, Mo S, Tao T, Li D, Zhuang Y, Wang H, Cheng X, Song L, Xing EP., bioRxiv, 2024

Related models & tools

• VC0010 Discovery of candidate therapeutic targets with Geneformer. — Zhang Y, Venkatesh MS, Theodoris CV., Nature protocols, 2026
• VC0023 · preprint CellPLM: Pre-training of Cell Language Model Beyond Single Cells — Wen H, Tang W, Dai X, Ding J, Jin W, Xie Y, Tang J., bioRxiv, 2023
• VC0030 · preprint scBERT as a Large-scale Pretrained Deep Language Model for Cell Type Annotation of Single-cell RNA-seq Data — Yang F, Wang W, Wang F, Fang Y, Tang D, Huang J, Lu H, Yao J., bioRxiv, 2021
• VC0031 · preprint Cell2Sentence: Teaching Large Language Models the Language of Biology — Levine D et al., bioRxiv, 2023
• VC0032 · preprint Scaling Large Language Models for Next-Generation Single-Cell Analysis — Rizvi SA et al., bioRxiv, 2025
• VC0038 · preprint ISP 3 Platform powered by Geneformer: Framework for Cross-Species, Sequential, and Multi-Gene In Silico Perturbation Screens with Application to iPS Cell State Transitions — Sanaki Y, Shigenobu S, Niwa R., bioRxiv, 2026

Perturbation & validation evidence

• VC0050 Zero-shot evaluation reveals limitations of single-cell foundation models. — Kedzierska KZ, Crawford L, Amini AP, Lu AX., Genome biology, 2025
• VC0049 Benchmarking foundation cell models for post-perturbation RNA-seq prediction. — Csendes G, Sanz G, Szalay KZ, Szalai B., BMC genomics, 2025
• VC0057 Assessing scale and predictive diversity in models for single-cell transcriptomics based on Geneformer. — Chen J, Schmidt F, Henao R., PLoS computational biology, 2026
• VC0056 Systematic evaluation of single-cell foundation model interpretability: attention-derived edge scores add no incremental value over gene-level features for perturbation-target prediction. — Kendiukhov I., BMC genomics, 2026

17 library items cited on this page; 8 are preprints (not peer reviewed).

Related reviews

• Glioblastoma cell states and AI virtual cells
• GBM cell-state atlases and the virtual-cell foundation (draft, pending review)

Entities on this page

• AIDO.Cell Models
• Cell2Sentence / C2S-Scale Models
• Single-cell foundation model Methods & concepts
• GeneCompass Models
• Geneformer Models
• In silico perturbation Methods & concepts
• Simple linear baseline Methods & concepts
• scBERT Models
• scFoundation Models
• scGPT Models
• TranscriptFormer Models
• Universal Cell Embeddings Models
• Zero-shot evaluation Methods & concepts

What remains uncertain

The library records do not state training-data composition (tissues, species mix, disease states) for most models, so how far any model's behaviour carries beyond its training distribution cannot be read from the library alone. VC0008 VC0011

Several models are represented only by preprints; the evaluation papers cover only some of the models listed, so the absence of an evaluation here is not evidence of good or poor performance. VC0012 · preprint VC0022 · preprint VC0050

Related questions

• What is an AI virtual cell model?
• Which virtual cell models have been applied to glioblastoma?
• When papers report that a mean or linear baseline matches deep models, are they comparing the same object, with the same metric, on the same kind of perturbations?
• Which mechanisms could make a perturbation-agnostic predictor look good (shared response, noise, target gene, metric), and which papers actually test which?
• How are virtual cell models validated?
• Do deep learning models beat linear baselines for perturbation prediction?
• Which general single-cell and perturbation data platforms are in the library, and under what data licence is each available?
• What do the records say about scale, curation method and access?
• Which evaluation pitfalls do the benchmark and critique papers in the library identify?
• For each pitfall, which record is the source and how mature is that source (peer reviewed or preprint)?