scGPT¶
Family: sc-fm Modality: Single-cell RNA-seq (gene-expression value tokens) Released: 2024 (Nature Methods) License: MIT (code); checkpoints distributed via Google Drive under the same terms Code/checkpoint: github.com/bowang-lab/scGPT (whole-human, organ-specific, and pan-cancer checkpoints) Also surfaced in: agentic-ai, single-cell-spatial-omics, virtual-cells
scGPT is a generative single-cell foundation model from the Bo Wang Lab at the University of Toronto that treats each cell as a sequence of gene–value tokens and pretrains a transformer to reconstruct masked expression values. It is trained on more than 33 million human cells from CELLxGENE and supports cell-type annotation, integration, perturbation prediction, and gene-regulatory inference after lightweight fine-tuning. Organ-specific (brain, blood, lung, heart, kidney) and pan-cancer (5.7M cells) checkpoints are released alongside the whole-human default.
Architecture & training¶
scGPT uses a transformer encoder with a custom generative attention mask: gene tokens are paired with binned expression-value tokens and condition tokens (batch, modality), and the model is trained to predict masked expression values given the unmasked context. The flagship "whole-human" checkpoint is pretrained on ~33M normal human cells aggregated from CELLxGENE; organ-specific variants range from ~0.8M to ~13M cells, and a pan-cancer variant is trained on ~5.7M tumor cells. Fine-tuning uses a small task head for cell-type classification, batch integration, or perturbation response, and the paper reports competitive or improved performance over scBERT, Geneformer, and scVI baselines across these tasks.
Use in the AACR 2026 corpus¶
Posters mentioning scGPT (n=2):
- 4212 — Baseline peripheral blood scRNA-seq AI estimator framework predicts solid-tumor response and adverse events via molecular foundation models and cell-to-patient learning …e PBMC scRNA-seq counts (with optional cell-type annotations), pre-trained molecular foundation models (FMs) (scGPT v1.0 [1], scFoundation [2]), cell-to-patient Multi-Instance Learning (MIL) [3] with RECIST labels (CR/PR Resp…
- 5478 — Benchmarking gene expression foundation models on bulk RNA-Seq data …ble models with released weights, including six single-cell models (CellFM, GeneFormer, scBERT, scFoundation, scGPT, and scLong) and BulkFormer, a model trained on bulk RNA-seq data. Embeddings were extracted following each m…
Sessions mentioning scGPT (n=1):
- 2026-04-20-am-ai-revolution-in-cancer-research …k and send us any feedbacks you may have. All right, so what's next? I talk about virtual cell models such as scGPT and ExCell. I talk about bioreason models, particularly bioreasons through DNA sequences, bioreasons through…
What's missing / where evidence is weak¶
- The pan-cancer checkpoint is absent. Despite a 5.7M tumor-cell pan-cancer variant being released alongside the whole-human default, no corpus mention uses or evaluates it. Both poster mentions reach for the v1.0 whole-human checkpoint or a generic "scGPT" baseline.
- No fine-tuning protocol detail in the corpus. Poster 4212 (PBMC peripheral-blood AI estimator) calls out scGPT v1.0 as a feature extractor, and poster 5478 benchmarks it, but neither describes fine-tuning hyperparameters, organ-specific checkpoint choice, or perturbation evaluation.
- The session mention is a name-check. The April-20 AI Revolution session lists scGPT alongside ExCell as "virtual cell models" without methodological detail; it adds no evidence beyond the posters.
- No tumor-microenvironment or perturbation work. scGPT's headline capabilities (perturbation prediction, GRN inference) do not appear in any corpus mention.
Takeaway¶
scGPT is present in AACR 2026 as a representative single-cell foundation model — invoked in two posters and one session — but always as a context entry in a foundation-model lineup (typically alongside scFoundation or Geneformer) rather than as the primary engine of an analysis. The corpus pattern mirrors Geneformer's: scGPT is widely known and consistently named, but no AACR 2026 contribution leans on its generative or perturbation capabilities to drive a finding.
FM comparison & 2026 status¶
Where it sits in the FM landscape. scGPT is a generative, decoder-style transformer in the single-cell FM family — distinguished from Geneformer (encoder-only, rank-based) and Cell2Location (Bayesian, not an FM at all) by a gene-value tokenization that lets it run masked expression regression rather than just rank-token masking. The flagship whole-human checkpoint is ~51M parameters, with a 53M-parameter pan-cancer variant and several organ-specific releases; this puts scGPT at the smaller end of the modern single-cell FM cohort (UCE: 650M; scFoundation: 100M; CellPLM: 80M; Geneformer-V2-Deep: 316M).
Direct peers.
| Model | Architecture | Pretraining corpus | Public weights | Headline benchmark |
|---|---|---|---|---|
| scGPT v1.0 | Decoder transformer, gene-value tokens, ~51M params | 33M human cells (CELLxGENE) + 5.7M pan-cancer variant | MIT, Google Drive | Cell-type annotation, integration, perturbation (Nature Methods 2024) |
| UCE (Universal Cell Embeddings) | Encoder transformer, protein-sequence-conditioned gene embeddings, 650M params | 36M cells, 8 species | MIT, GitHub + HF | Strongest cross-dataset zero-shot embedding in 2025 scFM benchmarks |
| scFoundation | xTrimoGene encoder, continuous-value embedding, 100M params | 50M human cells | Code MIT, weights gated | Tumor-tissue tasks (best in pooled-data eval) |
| GET | Sequence + chromatin-accessibility transformer | 213 fetal/adult cell types (multi-omics) | MIT, GitHub | Cross-cell-type expression prediction (Nature 2025) |
| CellPLM | Cell-as-token encoder, ~80M params | 11M cells with spatial context | MIT, GitHub | Spatial-context cell-state inference |
What changed in 2025-2026. Three things matter. First, the Bo Wang lab released scGPT-spatial (bioRxiv Feb 2025), a continually pretrained variant on SpatialHuman30M (~30M spatial profiles) with a Mixture-of-Experts decoder for protocol-aware deconvolution and imputation — the spatial successor that AACR 2026 has not yet caught up to. Second, a series of 2025 benchmark papers (Kedzierska et al., Genome Biology 2025; Wenkel et al., Nature Methods 2025) showed that scGPT, Geneformer, and scFoundation do not outperform simple linear baselines on perturbation prediction and have weak zero-shot embedding quality — a credibility hit that has not surfaced in any AACR 2026 poster framing. Third, the scGPT-cancer (pan-cancer) checkpoint has been reported to underperform the generic whole-human checkpoint on several non-cancer tasks, raising the question of whether domain-specific continual pretraining was worth the 5.7M-cell investment. No formal retraction or concern flag has been issued on the published paper.
AACR-relevant use cases. Three live use modes appear in the 2026 corpus or are visible immediately adjacent to it: (1) PBMC response prediction — poster 4212 uses scGPT v1.0 + scFoundation embeddings as MIL features to predict RECIST response from peripheral-blood scRNA-seq, the only "primary engine" usage in the corpus; (2) Bulk-RNA-seq benchmarking — poster 5478 evaluates scGPT alongside five other scFMs on bulk-RNA-seq embeddings, an off-label use that the recent zero-shot critique literature would predict to fail; (3) Virtual-cell scaffolding — the April 20 AI Revolution session name-checks scGPT alongside ExCell as a virtual-cell substrate, framing it as a building block for cell-simulation agents rather than an end-user analysis tool. The pan-cancer 5.7M-cell checkpoint and scGPT-spatial are absent from both posters and sessions — a substantive gap given that scGPT-spatial would directly compete with Cell2Location on spot-resolution deconvolution.