Single-Cell Foundation Models

Published

July 7, 2026

Single-cell biology produced its foundation-model wave between 2022 and 2024. scGPT, Geneformer, scFoundation, scBERT, UCE, and tGPT each trained on tens of millions of single cells with the explicit goal of producing general-purpose cell representations that transfer to many downstream tasks. Nicheformer extends the frame by pretraining on both dissociated single-cell and targeted spatial transcriptomics data (Tejada-Lapuerta et al., 2025). The capability is genuine for representation learning, batch correction, zero-shot annotation, and selected spatial-context tasks. Independent evaluations published in 2024 and 2025 (Ahlmann-Eltze et al., Boiarsky et al.) also show that for perturbation prediction and several other tasks, the deep approaches do not yet outperform simple PCA plus linear regression. The evidence supports scFMs as real tools for parts of single-cell analysis and leaves other claims unverified.

Learning Objectives

Use this chapter to:

  • Represent cell states, annotations, perturbation responses, and atlas relationships from high-dimensional single-cell data.
  • A cell embedding is only as good as the donor, tissue, protocol, batch, species, and label structure behind it.

Prerequisites: Foundation Models for Biology for the general pretrain-then-transfer framework; Evaluation Principles for Life Sciences AI for the biology-aware split discipline.

Summary: Represent cell states, annotations, perturbation responses, and atlas relationships from high-dimensional single-cell data. Single-cell foundation models support useful representations and selected transfer tasks, but independent evaluations show that simple baselines still matter.

Key point: A cell embedding is only as good as the donor, tissue, protocol, batch, species, and label structure behind it. Open question: how much transfer remains after stronger baselines, donor splits, tissue splits, and perturbation tests.

Bottom line: Single-cell models connect molecular regulation to tissues, disease states, systems biology, perturbation biology, aging, neuroscience, and therapeutics.

Field Guide

What is this field trying to solve? Represent cell states, annotations, perturbation responses, and atlas relationships from high-dimensional single-cell data.

What is the core idea? A cell embedding is only as good as the donor, tissue, protocol, batch, species, and label structure behind it.

What is the current state of the field? Single-cell foundation models support useful representations and selected transfer tasks, but independent evaluations show that simple baselines still matter.

What do we know, and what remains open? Known reference points include scGPT, Geneformer, scFoundation, UCE, scBERT, SCimilarity, CELLxGENE, Human Cell Atlas, OpenProblems, scIB, and perturbation datasets. What remains open is how much transfer remains after stronger baselines, donor splits, tissue splits, and perturbation tests.

Why does this matter? Single-cell models connect molecular regulation to tissues, disease states, systems biology, perturbation biology, aging, neuroscience, and therapeutics.

Introduction

Three claims drove the single-cell foundation-model wave between 2022 and 2024: that pretraining on tens of millions of cells would produce useful general-purpose representations, that those representations would transfer to many downstream tasks (annotation, integration, perturbation), and that the resulting models would become the default analytical layer for single-cell biology the way structure-prediction models had changed structural biology.

The first claim is well supported. scGPT (Cui et al., 2024), Geneformer (Theodoris et al., 2023), and scFoundation (Hao et al., 2024) demonstrably produce cell-state embeddings that capture cell type, perturbation context, and developmental state on a wide range of pretraining-similar settings. Nicheformer extends that claim to spatial-context-aware representations by training on a corpus that includes both dissociated and spatially resolved transcriptomics (Tejada-Lapuerta et al., 2025).

The second and third claims are more contested. Independent evaluations published in 2024 and 2025 (Boiarsky et al., 2024; Ahlmann-Eltze et al., 2025) find that on perturbation prediction and several other core tasks, the deep approaches do not consistently outperform PCA plus linear regression. The infrastructure layer (CZ CELLxGENE, Tabula Sapiens, the November 2024 Human Cell Atlas Nature collection) continues to mature; single-cell foundation models have not yet earned the role of default analytical layer in the same way.

This chapter applies the evidence framework to scFMs specifically. The discipline matters because the data scale is real, the architectural innovation is real, and the marketing tendency to overclaim is also real.

What is demonstrated?

Representation learning at scale

The clearest demonstrated capability is general-purpose cell-state representation. Pretraining a transformer on tens of millions of single-cell transcriptomes produces embeddings that:

  • Cluster cells by canonical cell type with limited fine-tuning
  • Capture developmental trajectories without explicit supervision
  • Reflect perturbation state (control vs. treated) in many bulk-scale settings
  • Support cell-atlas search (the explicit SCimilarity contribution, Heimberg et al., 2025)

scGPT trained on more than 33 million cells with a generative pretraining objective and demonstrated transfer to cell-type annotation, perturbation prediction, multi-batch integration, and gene-network inference (Cui et al., 2024). Geneformer trained on approximately 30 million cells using a rank-order gene encoding and demonstrated transfer to network biology tasks including dosage-sensitivity prediction (Theodoris et al., 2023). scFoundation extended the scale and added a different pretraining recipe (Hao et al., 2024). scBERT preceded the wave with a smaller pretraining corpus and remains a useful baseline (Yang et al., 2022).

Nicheformer trained on SpatialCorpus-110M, combining dissociated single-cell data with spatially resolved transcriptomics, and evaluated tasks such as spatial composition and spatial label prediction (Tejada-Lapuerta et al., 2025). It should be read as evidence that spatial context changes the representation problem. It should not be promoted to general tissue simulation or perturbation prediction without matched dynamic evidence.

Integration and batch correction

Pretrained representations transfer well to integration tasks (combining datasets across batches, donors, technologies, and labs). Performance is typically benchmarked against scIB (Luecken et al., 2022), which combines batch-correction quality and biological-signal preservation. scFM methods are competitive with classical methods (scVI, Harmony, Seurat integration) but not uniformly better; the choice depends on dataset size and integration scenario.

The baseline layer is not old infrastructure to skip. Seurat v3 introduced anchor-based integration and label transfer across single-cell modalities (Stuart et al., 2019). Harmony remains a strong shared-embedding method for integrating large single-cell datasets across experimental and biological factors (Korsunsky et al., 2019). scVI and scANVI provide probabilistic latent-variable baselines for representation, harmonization, and cell-state annotation (Lopez et al., 2018; Xu et al., 2021). Any scFM adoption decision should report whether it improves the actual downstream endpoint over these methods, not only whether it improves an embedding visualization.

Baseline What It Tests Why It Matters
Seurat anchors Cross-dataset alignment and label transfer Strong practical baseline for reference mapping
Harmony Shared latent embedding across batch and biological covariates Fast, widely used, and hard to beat on routine integration
scVI / scANVI Probabilistic integration and annotation Captures uncertainty and supports downstream differential analysis

Data infrastructure underneath the models

The model wave depended on the data wave that preceded it:

  • CZ CELLxGENE Census (CZ CELLxGENE team, 2023, preprint) provides approximately 100 million curated single-cell observations in a standardized, queryable platform. This is the single largest training source for the scFM wave.
  • Tabula Sapiens (Tabula Sapiens Consortium, 2022) produced 1.1 million cells from 28 organs of 24 donors as a benchmark first-draft human cell atlas.
  • Human Cell Atlas (Regev et al., 2017) continues to release organ-specific and developmental atlases, with the November 2024 Nature collection including a roadmap to a unified foundation model (Rood et al., 2025) and several major organ atlases (neural organoids, prenatal skin, thymus, embryonic skeleton).

The data layer’s quality, curation, and metadata determine the scFM ceiling. Pretraining choices about which cells to include propagate to every downstream claim.

Evidence anchor summary

Evidence Anchor What It Supports Practical Constraint
scGPT, Geneformer, scFoundation General-purpose cell-state representation at pretraining-similar settings Out-of-distribution transfer needs separate evaluation
Nicheformer Spatially aware cell representations Spatial context is represented, not simulated dynamically
SCimilarity Atlas-scale cell-similarity search Reference quality bounds search quality
Integration on scIB Competitive batch correction Not always better than classical methods
CZ CELLxGENE Census Standardized aggregated single-cell platform Curation choices shape downstream transfer
Tabula Sapiens Benchmark first-draft human atlas One donor population; not yet a population reference
HCA Nov 2024 collection Continuing organ-specific atlases Coverage uneven across rare cell types

What is theoretical?

Several capabilities are plausible but not yet established.

Perturbation prediction at production quality. GEARS (Roohani et al., 2024) demonstrated transcriptional perturbation prediction for single and combinatorial CRISPR perturbations in bounded settings. Other approaches: scGen (Lotfollahi et al., 2019), CPA (Lotfollahi et al., 2023), scPRAM (Jiang et al., 2024). The Ahlmann-Eltze critique below applies here: on the canonical out-of-distribution settings, simple linear baselines do nearly as well as the deep approaches. Treat current performance as a research result, not a production capability. See Perturbation Prediction and Virtual Cells for the focused chapter.

Novel-tissue and novel-species transfer. Pretraining is dominated by human (and to a smaller extent mouse) blood, brain, and tumor data. Transfer to underrepresented tissues, developmental stages, and species (especially non-mammalian) is partial. The HCA roadmap toward a unified foundation model (Rood et al., 2025) targets this gap.

Mechanistic interpretation. Attention maps and probing studies suggest scFMs capture some gene-regulatory structure. Whether they capture mechanism in a way that generates testable biological hypotheses is an open research question.

Multi-omics integration. Most current scFMs are transcriptome-only. Joint training with proteomics, chromatin accessibility (ATAC), and spatial information is an active area; see Microbiome and Multi-Omics AI and Spatial Omics and Tissue Models.

Single-cell foundation model as standard atlas-search infrastructure. SCimilarity points in this direction for atlas-scale search. Production-grade infrastructure that handles versioning, reproducibility, and provenance for atlas-scale queries is still emerging.

What is beyond current capability?

A few framing claims are not supported by current evidence.

Foundation models replace experimental single-cell biology. They do not. Cell-state representation is a research input, not a substitute for measurement. The cells the model has never seen are the cells whose biology you do not yet know.

Single-cell models predict in vivo phenotypes from in vitro data. Current models are trained on the cells available, which are predominantly in vitro or freshly dissociated. Cellular context, microenvironment, and systems-level interactions are not in the training distribution.

Foundation models are the right default for any single-cell task. They are sometimes the right tool and sometimes worse than PCA. The choice should be evidence-driven for each task, not by class. Ahlmann-Eltze et al. (2025) and Boiarsky et al. (2024) are the clearest current evidence on this point.

Mechanistic gene-regulatory inference from attention maps alone. Attention patterns are a useful interpretability signal; they are not a substitute for direct perturbation evidence.

What would make this more promising?

The claim would strengthen if scFMs repeatedly beat PCA, scVI, scANVI, Seurat, and Harmony on donor-held-out, tissue-held-out, perturbation-held-out, and laboratory-held-out tasks, with independent groups reproducing the gains. It would strengthen further if model outputs predicted perturbational or spatial outcomes that were confirmed prospectively.

The claim would weaken if reported gains disappear under simple baselines, stricter split design, pretraining-corpus audit, or benchmark reuse through OpenProblems and scIB. Failure on rare tissues, disease states, or unseen perturbation classes should narrow the claim rather than be treated as a general model limitation.

What should researchers, biotech teams, funders, and program leaders do with this?

For researchers evaluating or adopting single-cell foundation approaches:

  • Run the linear baseline first. PCA followed by linear regression (or scVI for batch-aware embedding) is a strong baseline that should appear in any honest comparison. If the published paper omits it, run it locally before drawing conclusions.
  • Match the task to the evidence. For representation, integration, and atlas search, the published results are reasonable evidence. For perturbation prediction and novel-tissue transfer, the published results overstate what the field has shown.
  • Read the pretraining distribution. A pretraining corpus weighted toward blood and brain transfers differently to liver, kidney, and immune-cell subsets in disease tissue. Match pretraining coverage to the deployment biology.
  • Use atlas-scale benchmarks honestly. scIB (Luecken et al., 2022) and OpenProblems (Luecken et al., 2025) are the community references. Self-reported numbers in scFM papers should be triangulated against these.
  • Treat zero-shot claims as task-curated. Most reported zero-shot performance is on tasks curated from cells very similar to pretraining. True out-of-distribution transfer (new species, new tissues, new perturbation classes) is a different evaluation and a different result.
  • Cite both the model and the data foundation. A claim about scGPT is also a claim about the cells it was pretrained on. Report the pretraining corpus alongside the method name.
  • Watch for cell-line leakage. Train and test cells that come from the same biological samples, the same donor cohort, or closely related lineages do not test transfer. Apply donor-level or tissue-level splits where the deployment will require them.
  • Keep classical methods in the toolkit. scVI, scANVI, Seurat integration, Harmony, and PCA remain the right tool for many real analyses. The foundation-model wave changes what is possible at the upper end, not what is required for the median single-cell project.