The DOI matches its Crossref record; 110 of the 110 references it cites were found in a registry or an index; none of the 12 sources its claims cite is checked for support yet.
scGPT: Towards Building a Foundation Model for Single-Cell Multi-omics Using Generative AI
Open accessAbstract
Generative pre-trained models have achieved remarkable success in various domains such as natural language processing and computer vision. Specifically, the combination of large-scale diverse datasets and pre-trained transformers has emerged as a promising approach for developing foundation models. Drawing parallels between linguistic constructs and cellular biology — where texts comprise words, similarly, cells are defined by genes — our study probes the applicability of foundation models to advance cellular biology and genetics research. Utilizing the burgeoning single-cell sequencing data, we have pioneered the construction of a foundation model for single-cell biology, scGPT, which is based on generative pre-trained transformer across a repository of over 33 million cells. Our findings illustrate that scGPT, a generative pre-trained transformer, effectively distills critical biological insights concerning genes and cells. Through the further adaptation of transfer learning, scGPT can be optimized to achieve superior performance across diverse downstream applications. This includes tasks such as cell-type annotation, multi-batch integration, multi-omic integration, genetic perturbation prediction, and gene network inference. The scGPT codebase is publicly available at https://github.com/bowang-lab/scGPT .
Overview
CoChat Summary
TL;DR
scGPT is a generative pretrained transformer foundation model for single-cell biology, pretrained on over 33 million human scRNA-seq cells and fine-tuned for cell-type annotation, integration, perturbation prediction, and gene network inference. It reports state-of-the-art or competitive performance across these tasks and identifies a scaling law where larger pretraining data improves downstream fine-tuning.
Key findings
- Pretraining used 33 million normal human scRNA-seq cells from CELLxGENE, spanning 51 organs/tissues and 441 studies; UMAP of a 10% subset showed cell types separated into distinct clusters despite over 400 studies.
- Cell-type annotation: scGPT achieved high precision (>0.8) for most hPancreas cell types except rare types, around 0.85 accuracy on multiple sclerosis, and ≥0.7 precision for most tumor-infiltrating myeloid cell types; it outperformed TOSICA and scBert on Accuracy, Precision, Recall, and MacroF1.
- Perturbation prediction: scGPT achieved the highest correlation on both Adamson and Norman datasets versus GEARS and CPA; in reverse perturbation prediction it identified 71.4% correct perturbations (5/7 test cases) within top 5 predictions, a 95.2% speedup over random tryouts (105.5 average attempts out of 210).
- Integration and GRN: PBMC 10K AvgBIO was 0.821, 5–10% higher than compared methods; ASAP PBMC AvgBATCH was 0.951 versus scMoMaT’s 0.916; scGPT uniquely identified 22 Reactome pathways, 14 immune-related, and attention analysis found all top 20 DDIT3-influenced genes and 19/20 BHLHE40-influenced genes as ChIP-Atlas targets.
Methods
- Design: two-stage generative pretraining and fine-tuning of a transformer that jointly learns cell and gene embeddings; a specialized attention mask enables generative prediction on non-sequential gene expression, with fine-tuning objectives including GEP, GEPC, ECS, DAR, and CLS.
- Data: pretraining on 33 million normal human scRNA-seq cells from CELLxGENE; downstream datasets included hPancreas, multiple sclerosis, myeloid, Adamson, Norman, COVID-19, PBMC 10K, Perirhinal Cortex, 10X Multiome PBMC, BMMC, ASAP PBMC, and Immune Human.
- Analysis: benchmarks against TOSICA, scBert, GEARS, CPA, scVI, Seurat, Harmony, scGLUE, and scMoMaT; metrics included Accuracy, Precision, Recall, MacroF1, Pearson correlation and corr(Δ), AvgBIO, AvgBATCH, Overall, and Reactome pathway enrichment.
Limitations
- Stated: the scaling law appeared to saturate after 3M pretraining cells; the authors hypothesize larger model sizes could improve performance at 33M and beyond but leave larger model sizes to future work.
- Stated: future work plans larger-scale, more diverse pretraining including multi-omic, spatial omics, diseased conditions, perturbation and temporal data, broader biological validation, and in-context instruction learning.
- Evident: this is a bioRxiv preprint not certified by peer review; results are computational benchmarks, and the text does not report prospective wet-lab validation of predictions.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
0 Valid · 2 Not valid · 11 Not checked · 67 Own results and interpretations
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“the fine-tuned scGPT consistently showcases performance gain for tasks such as integration376 and cell type annotation.”
§ 1 3 > 2.6 Analyzing the transfer learning of pretrained knowledge: the scaling370
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“scGPT achieves a higher overall batch correction287 score AvgBATCH of 0.951 compared to scMoMat (AvgBATCH = 0.916).”
§ 1 3 > 2.4 scGPT enables accurate multi-batch and multi-omics integration231
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“scGPT334 consistently demonstrates a significantly higher number of enriched pathways across all clustering335 resolutions.”
§ 1 3 > 2.5 scGPT uncovers gene regulatory networks for specific cell states and292
boostedstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“scGPT stands264 out as the only method that successfully generates a distinct cluster for CD8 Naive cells, while the265 other two methods fail to do so.”
§ 1 3 > 2.4 scGPT enables accurate multi-batch and multi-omics integration231
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The top performers in data integration tasks were models pretrained415 on whole-human, blood, and lung datasets”
§ 1 3 > 2.6 Analyzing the transfer learning of pretrained knowledge: the scaling370
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
32 more claims
Create a free account to see all 80 claims and to verify the 8 references still open.
- Probably not a substantial claim. 26
- It may state two separate claims as one. 17
- The quote only partly states this claim. 13
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Empirical benchmarking studies have documented substantial performance degradation for biological foundation models under cross-platform transfer, preprocessing variability, and rare-population evaluation [20–24].”
From The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery (2026) · cites this paper as [20]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Compared to tasks such as cell annotation and batch correction, perturbation prediction offers experimentally generated model input/output with well-defined objectives, making it a more robust benchmark for assessing foundation models (15, 20, 21).”
From A systematic comparison of single-cell perturbation response prediction models (2026) · cites this paper as [15]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“In oncology, transformer-based methods support clinical text processing, whole-slide and radiologic image interpretation (vision transformers), and genomic sequence modeling [24,25,26,27].”
From Artificial Intelligence in Clinical Oncology: From Productivity Enhancement to Creative Discovery (2025) · cites this paper as [27]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“We chose these models because they offer pretrained weights and have been trained using unsupervised objectives on extensive datasets (ca. 30M single-cell transcriptomes).”
From Zero-shot evaluation reveals limitations of single-cell foundation models (2025) · cites this paper as [7]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“However, scGPT directly learns a cell embedding, which is integrated into its pretraining loss of predicting masked genes: scGPT first predicts a masked gene expression bin and a cell embedding from unmasked genes; then, in a second step, it further iteratively refines masked gene expression using the cell embedding predicted in the first step.”
From Zero-shot evaluation reveals limitations of single-cell foundation models (2025) · cites this paper as [7]
Experimental — accuracy not yet measured
21 more claims about this paper
Create a free account to see every claim other papers make about this one.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Supplementary Notes1264
“AvgBATCH = 0.6 ∗AvgBIO + 0.4 ∗AvgBATCH”
§ Supplementary Notes1264
“AvgBATCH = (ASWbatch +GraphConn)/2”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 4 Methods462 > 4.7 Benchmarking Experiment Setup860
“batch correction and cell type clustering on four integration datasets: COVID-19 [13], PBMC 10K930 [44], and Perirhinal Cortex [45].”
§ 1 3 > 2.4 scGPT enables accurate multi-batch and multi-omics integration231
“The evaluation was conducted on three integration236 datasets, namely the COVID-19 (18 batches) [13], PBMC 10K (2 batches) [44], and Perirhinal237 Cortex (2 batches) [45].”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 1 3 > 2.2 scGPT improves the precision of cell type annotation120
“fewer than 50 cells belong to the mast and MHC class II cell types out of the 106,000134 cells in the reference set.”
§ 4 Methods462 > 4.6 Datasets775
“The reference set contains 10,600 cells of 13 cell groups (alpha, beta, ductal,807 acinar, delta, PSC, PP, endothelial, macrophage, mast, epsilon, schwann, and t cell). The query808 set contains 4,218 cells of 11 cell groups (alpha, beta, ductal, PP, acinar, delta, PSC, endothelial,809 epsilon, mast, MHC class II).”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 1 3 > 2.2 scGPT improves the precision of cell type annotation120
“out of the 106,000134 cells in the reference set.”
§ 4 Methods462 > 4.6 Datasets775
“The reference set contains 10,600 cells of 13 cell groups”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 1 3 > 2.4 scGPT enables accurate multi-batch and multi-omics integration231
“subgroup annotations (48 cell types).”
§ 4 Methods462 > 4.6 Datasets775
“annotations consist of 45 detailed immune cell subtypes.”
1 more finding
Create a free account to see every finding of this check.
Citations
Cites
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. (2020) · arXiv · doi:10.48550/arxiv.2010.11929 · Cited by 47,472
Highly accurate protein structure prediction with AlphaFold
Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A., et al. (2021) · Nature · doi:10.1038/s41586-021-03819-2 · Cited by 43,764
Language Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372 · cited 2×
“The generative pre-training used in OpenAI GPT3 [71] and GPT4 [19] employs a unified 585”
MethodRoBERTa: A Robustly Optimized BERT Pretraining Approach
Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., & Stoyanov, V. (2019) · arXiv · doi:10.48550/arxiv.1907.11692 · Cited by 16,515
Integrated analysis of multimodal single-cell data
Hao, Y., Hao, S., Andersen-Nissen, E., Mauck, W. M., Zheng, S., Butler, A., Lee, M. J., Wilk, A. J., Darby, C., Zagar, M., et al. (2020) · doi:10.1101/2020.10.12.335331 · Cited by 15,508 · cited 5×
“In the paired data integration experiment, we benchmarked scGPT with scGLUE 951 [14] and Seurat v4 [46] on the 10X Multiome PBMC [47] dataset with RNA and ATAC-seq data as 952 the first example.”
MethodInfluentialComprehensive Integration of Single-Cell Data
Stuart, T., Butler, A., Hoffman, P., Hafemeister, C., Papalexi, E., Mauck, W. M., Hao, Y., Stoeckius, M., Smibert, P., & Satija, R. (2019) · Cell · doi:10.1016/j.cell.2019.05.031 · Cited by 15,453
Comprehensive integration of single-cell dataPreprint
Stuart, T., Butler, A. W., Hoffman, P. J., Hafemeister, C., Papalexi, E., Mauck, W., Hao, Y.-H., Stoeckius, M., Smibert, P., & Satija, R. (2018) · bioRxiv · doi:10.1101/460147 · Cited by 13,858
Fast, sensitive and accurate integration of single-cell data with Harmony
Korsunsky, I., Millard, N., Fan, J., Slowikowski, K., Zhang, F., Wei, K., Baglaenko, Y., Brenner, M., Loh, P.-R., & Raychaudhuri, S. (2019) · Nature Methods · doi:10.1038/s41592-019-0619-0 · Cited by 10,488 · cited 5×
“In our benchmarking experiments, we compared scGPT with three popular integration methods: scVI [41], Seurat [42], and Harmony [43].”
MethodInfluential
Research works printed but not linked: 6
- 2019, In: Chinese Computational Linguistics: 18th China National Conference, CCL 2019, Kunming, China, October 18–20, 2019, Proceedings, 18, 194
- National Library of Medicine. Bhlhe40 basic helix-loop-helix family, member e40 [Mus mus-culus (house mouse)]. https://www.ncbi.nlm.nih.gov/gene/20893. Online; accessed 29 May 2023. 2023.
- 2019, Language models are unsupervised multitask learners, In: OpenAI blog, 1, 9
- 10.1186/S13059-020-02000-8/FIGURES/6 · doi:10.1186/s13059-020-02000-8/figures/6
- 2013, Auto-encoding variational bayes, In: arXiv
- 2021, Joint probabilistic modeling of single-cell multi-omic data with to-talVI, In: Nature methods, 18, 272
Cited by
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. (2024) · Nature Methods · doi:10.1038/s41592-024-02305-7 · Cited by 628
Program, C. C. S., Abdulla, S., Aevermann, B., Assis, P., Badajoz, S., Bell, S. M., Bezzi, E., Cakir, B., Chaffer, J., Chambers, S., et al. (2025) · Nucleic Acids Research · doi:10.1093/nar/gkae1142 · Cited by 399
How to build the virtual cell with artificial intelligence: Priorities and opportunities
Bunne, C., Roohani, Y., Rosen, Y., Gupta, A., Zhang, X., Roed, M., Alexandrov, T., AlQuraishi, M., Brennan, P., Burkhardt, D. B., et al. (2024) · Cell · doi:10.1016/j.cell.2024.11.015 · Cited by 393
Self-citationProfiling cell identity and tissue architecture with single-cell and spatial transcriptomics
Gulati, G. S., D’Silva, J. P., Liu, Y., Wang, L., & Newman, A. M. (2025) · Nature Reviews Molecular Cell Biology · doi:10.1038/s41580-024-00768-2 · Cited by 293
Deep learning-based approaches for multi-omics data integration and analysis
Ballard, J. L., Wang, Z., Li, W., Shen, L., & Long, Q. (2024) · BioData Mining · doi:10.1186/s13040-024-00391-z · Cited by 258
“Recently, Cui et al. [52] proposed scGPT, a foundation model for single-cell omics data.”
Background
Metadata
- Authors
- Cui, H., Wang, C., Maan, H., Pang, K., Luo, F., & Wang, B.
- Affiliations
- University Health Network (CA)
- University of Toronto (CA)
- Corresponding author
- Wang, Bo
- Venue
- Nature Methods
- Publisher
- openRxiv
- Published
- 2023-05-01
- Volume
- 21
- Issue
- 8
- Pages
- 1470-1480
- ISSN
- 1548-7091 (ISSN-L)
- OpenAlex
- W4392168151
- Cited by
- 1,694
- References
- 98
- Citations by year
- 2026: 918 · 2025: 521 · 2024: 234 · 2023: 19
- FWCI
- 186.38
- Type
- journal article
- Language
- en
- Open access
- green
- doi.org
- Licence
- cc-by-nc-nd
- Locations
- Nature Methods (publishedVersion)
- bioRxiv (Cold Spring Harbor Laboratory) (submittedVersion, cc-by-nc-nd)
- PubMed (publishedVersion)
- Funders
- University Health Network
- Canadian Network for Research and Innovation in Machining Technology, Natural Sciences and Engineering Research Council of Canada (DGECR-2020-00294, RGPIN-2020-06189)
- Topics
- Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Single-cell and spatial transcriptomics
- Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Bioinformatics and Genomic Networks
- Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Gene Regulatory Network Analysis
- Keywords
- scGPT, single-cell multi-omics, foundation models, generative pre-trained transformer, single-cell RNA sequencing, cell type annotation, multi-omics integration, perturbation response prediction, gene network inference, transfer learning, single-cell biology
- MeSH terms
- Multiomics, Algorithms, Artificial Intelligence, Humans, Computational Biology, Gene Regulatory Networks, Single-Cell Analysis
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-11-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-11-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: scGPT: Towards Building a Foundation Model for Single-Cell Multi-omics Using Generative AI. https://app.cochat.ai/papers/10.1101/2023.04.30.538439
- The paper
- Cui, H., Wang, C., Maan, H., Pang, K., Luo, F., & Wang, B. (2023). scGPT: Towards Building a Foundation Model for Single-Cell Multi-omics Using Generative AI. Nature Methods. https://doi.org/10.1101/2023.04.30.538439
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.