Virtual Cell Modeling

Introduction

A "virtual cell" is a computational model that predicts a cell's state and its response to interventions — genetic perturbations, drugs, or environmental changes — without running the wet-lab experiment. The field spans two historically separate lineages that are now converging: mechanistic simulation (ODE/SBML kinetics, flux-balance metabolism, agent-based tissue models, and the whole-cell simulations pioneered by Karr et al.), and data-driven machine learning (single-cell foundation models pretrained on tens to hundreds of millions of cells, and perturbation-response predictors). A 2024 Cell roadmap by Bunne, Roohani, Leskovec, Quake, Regev and 30+ others crystallized this into the "AI Virtual Cell" (AIVC) vision. Crucially, the most important recent development is not a new model but a credibility correction: multiple 2025 benchmarks (Kedzierska et al.; Ahlmann-Eltze et al.) show that today's large foundation models, used zero-shot or for perturbation prediction, do not yet reliably beat simple linear or mean baselines — and the Arc Institute's Virtual Cell Challenge was created in part to force honest, held-out evaluation. This survey organizes the field by approach family, giving mechanism, key tools, evidence, strengths, limits, and verified citations for each.

Family 1 — Single-Cell Foundation Models (SCFMs)

What it is. Transformer models pretrained self-supervised on massive unlabeled single-cell transcriptomes, producing a general-purpose "cell embedding" that can be fine-tuned or used zero-shot for downstream tasks (cell-type annotation, batch integration, perturbation, gene-network inference).

How it models/predicts (mechanism). A cell (~20,000 gene counts) is tokenized and fed to a transformer with a masked-prediction pretraining objective, but the tokenization differs by model:

• scGPT (Cui et al. 2024) — value-binning tokens; a generative masked-attention objective; ~33M cells. The reference "GPT-style" SCFM. Nat Methods 2024;21(8):1470–1480. DOI 10.1038/s41592-024-02201-0.
• Geneformer (Theodoris et al. 2023) — rank-value encoding (genes ordered by expression rank, no absolute values); pretrained on ~30M cells ("Genecorpus-30M"). Pitch: transfer learning for network biology. Nature 2023;618:616–624. DOI 10.1038/s41586-023-06139-9.
• scFoundation / xTrimoGene (Hao et al. 2024) — continuous value embedding (no lossy binning), an asymmetric encoder–decoder processing only non-zero genes, and Read-Depth-Aware (RDA) pretraining; ~100M params, >50M human cells, 19,264 genes. Nat Methods 2024;21(8):1481–1491. DOI 10.1038/s41592-024-02305-7. (The backbone BioMap adapted to win VCC 2025.)
• scBERT (Yang et al. 2022) — early BERT-style SCFM for cell-type annotation. Nat Mach Intell 2022;4:852–866. DOI 10.1038/s42256-022-00534-z.
• UCE (Universal Cell Embeddings) (Rosen, Roohani et al. 2023) — 33-layer, ~650M-param model on >36M cells across 300+ CELLxGENE datasets, using ESM2 protein embeddings so genes from any species map into one universal space. bioRxiv 10.1101/2023.11.28.568918.
• TranscriptFormer (CZI, 2025) — cross-species generative SCFM (~112M cells, 12 species); Apache-2.0 licensed. bioRxiv 10.1101/2025.04.25.650731 (UNVERIFIED DOI — confirm on click).

Evidence / strengths. Strong when fine-tuned on a labeled downstream task; competitive cell-type annotation, batch integration, useful embeddings; enable transfer to data-scarce settings. UCE/TranscriptFormer add cross-species generalization.

Limitations (load-bearing). In zero-shot use, Kedzierska et al. 2025 (Genome Biology 26:101; DOI 10.1186/s13059-025-03574-x; PMID 40251685) showed scGPT and Geneformer perform inconsistently and often worse than plain highly-variable-gene selection, scVI, or Harmony for cell-type clustering across five datasets. Recurring lesson: single-cell data is dominated by technical noise (dropout, batch), the biological signal is near-linear, and huge models overfit that noise — so curation + statistical structure often beats raw scale.

Family 2 — Perturbation-Response Prediction

What it is. Models that predict the transcriptome-wide response to a specific perturbation (silence gene X, apply drug Y at dose Z), including generalization to unseen single or combinatorial perturbations.

Mechanism.

• scGen (Lotfollahi et al. 2019) — VAE doing latent-space vector arithmetic (learn a "perturbation direction," add to unseen contexts). Nat Methods 2019;16:715–721. DOI 10.1038/s41592-019-0494-8.
• CPA (Compositional Perturbation Autoencoder) (Lotfollahi et al. 2023) — adversarially disentangles basal state, perturbation, dose, covariates into composable latent factors → drug-combination and dose-response prediction. Mol Syst Biol 2023;19:e11517. DOI 10.15252/msb.202211517.
• chemCPA (Hetzel et al. 2022) — extends CPA to unseen drugs via molecular-structure encoding. NeurIPS 2022; arXiv:2204.13545.
• GEARS (Roohani et al.) — GNN over a GO / co-expression knowledge graph, predicting effects of multigene (combinatorial) perturbations including genes never perturbed alone. Nat Biotechnol 2024;42:927–935. DOI 10.1038/s41587-023-01905-6.
• STATE (Arc Institute; Adduri et al. 2025) — set-attention state-transition model on >100M perturbed cells across 70 contexts + a cell-embedding model on 167M cells; reports >50% improvement in perturbation-effect discrimination. bioRxiv 10.1101/2025.06.26.661135 (non-commercial license).
• BioMap xTrimoSCPerturb — the VCC 2025 grand-prize winner; fused the scFoundation backbone with explicit statistical baseline features and a metric-aligned (PDS+DES+MAE) loss. Documented via the Arc VCC 2025 wrap-up + BioMap's post; no peer-reviewed methods paper verified yet — internal recipe details are vendor-claimed / UNVERIFIED.

Key datasets & benchmark. Genome-scale Perturb-seq — Replogle et al. 2022, Cell 185:2559–2575.e28, DOI 10.1016/j.cell.2022.05.013 — is the foundational training/transfer resource (K562/RPE1, ~8K genes). Norman et al. 2019 (combinatorial CRISPRa, GSE133344) is the standard combinatorial testbed. The Arc Virtual Cell Challenge (2025; H1-hESC, 300 CRISPRi perturbations split 150/50/100) is the flagship held-out evaluation, scored on PDS (perturbation-discrimination), DES (differential-expression), and MAE.

Strengths. GEARS and STATE genuinely push combinatorial and cross-context generalization.

Limitations. See Family 5 — for unseen and combinatorial perturbations, deep predictors have not yet beaten additive / mean baselines in controlled benchmarks. The winning VCC recipes were hybrids (statistics + curation + selective DL), not pure end-to-end nets.

Family 3 — The "AI Virtual Cell" (AIVC) Vision

What it is. A field-defining roadmap: Bunne, Roohani, Rosen, … Regev, Lundberg, Leskovec, Quake et al. 2024, "How to build the virtual cell with artificial intelligence: Priorities and opportunities," Cell 2024;187:7045–7063, DOI 10.1016/j.cell.2024.11.015.

How it frames the problem. Proposes the AIVC as a multi-scale, multi-modal system of "universal representations" of molecules, cells, and tissues, learned from large atlases, that can be queried and simulated — predict unmeasured states and responses (virtual instruments / virtual perturbations) and be validated against real experiments.

Operationalization. Arc Institute turned the vision into a measurable target with the Virtual Cell Challenge — "Virtual Cell Challenge: Toward a Turing test for the virtual cell," Cell, June 26, 2025, PII S0092-8674(25)00675-0 (numeric DOI UNVERIFIED — cited by verified publisher URL). Supporting infrastructure: the Arc Virtual Cell Atlas and Tahoe-100M (100M cells, ~1,100 drugs × 50 cancer lines), CZI's CELLxGENE Census and cz-benchmarks/cell-eval harness.

Strengths. Aligns a fragmented field on shared objectives, metrics, and open data; catalyzed the honest-evaluation movement.

Limitations. Aspirational — no current system is a true virtual cell; the gap between vision and demonstrated accuracy (Family 5) is large.

Family 4 — Whole-Cell & Mechanistic Models

What it is. Bottom-up simulations grounded in molecular mechanism: ODE/SBML kinetics, flux-balance metabolism (FBA), agent-based/spatial tissue models, integrative whole-cell models.

Mechanism.

• Whole-cell model — Karr et al. 2012, "A Whole-Cell Computational Model Predicts Phenotype from Genotype," Cell 150:389–401, DOI 10.1016/j.cell.2012.05.044. Simulates every characterized gene function of Mycoplasma genitalium via ~28 integrated submodels, predicting phenotype from genotype. The landmark proof-of-concept.
• Flux-balance analysis (FBA) — genome-scale metabolic models (Recon3D, Human-GEM) solved by linear programming under stoichiometric/steady-state constraints; tool COBRApy. (Orth et al. 2010, Nat Biotechnol 28:245–248.)
• ODE/SBML kinetic simulation — mass-action / Michaelis-Menten networks (Tellurium, AMICI, COPASI); used for signaling and cardiac safety (CiPA O'Hara-Rudy action-potential model).
• Agent-based / spatial — PhysiCell, Chaste, CompuCell3D, Morpheus, FLAME-GPU2 for tumor microenvironment, morphogenesis, drug penetration.
• Hybrid mechanistic + ML — SBML-to-JAX, adjoint-sensitivity fitting (AMICI + pyPESTO), composable multi-scale frameworks (Vivarium).

Strengths. Interpretable, causal, extrapolate to unseen conditions in principle, and are the regulatorily-accepted end of the spectrum (PBPK, CiPA/QSP appear in real FDA/EMA submissions).

Limitations. Require extensive prior knowledge and parameterization; do not scale to whole human-cell transcriptomic response; the Karr-style approach remains confined to minimal organisms and is computationally heavy.

Family 5 — Honest Negative Results, Benchmarks & the Evaluation Debate

What it is. Rigorous benchmark studies whose central finding is that the hype has outrun the evidence — the field's most important recent contribution.

• Ahlmann-Eltze, Huber & Anders 2025 — "Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines," Nature Methods 22:1657–1661, DOI 10.1038/s41592-025-02772-6, PMID 40759747. For combinatorial perturbations DL did not beat a simple additive model; for unseen genes DL did not beat predicting the training mean.
• Kedzierska, Crawford, Amini & Lu 2025 — "Zero-shot evaluation reveals limitations of single-cell foundation models," Genome Biology 26:101, DOI 10.1186/s13059-025-03574-x, PMID 40251685.
• PerturBench (Wu et al. 2024) — arXiv:2408.10609 — ~27 perturbation methods across ~29 datasets with rank-based specificity metrics; strong non-DL baselines.
• The VCC as a benchmark. The Arc wrap-up reports a trivial cell-mean baseline beat most deep models on MAE, and that "pure end-to-end neural networks have yet to outperform hybrid models" — which is why the challenge weights PDS (perturbation-specificity) as the deciding metric (a constant prediction scores chance = 0.5).

The debate. Cross-paper score comparisons are usually invalid (different tasks/data/metrics). The defensible recipe the community is converging on: your own task + your own metric + a permuted null + a simple baseline + an oracle/ceiling, all on the same held-out split.

What's Winning / The Honest State

• Mechanistic models win where they're accepted: PBPK, CiPA/QSP, and FBA are the only virtual-cell approaches with real regulatory standing today.
• For perturbation-response, hybrids win — not pure deep learning. All three VCC 2025 winners (BioMap xTrimoSCPerturb; Altos flow-matching; Team Outlier's TransPert) beat brute-force scale by combining curated data + explicit statistical priors + metric-aligned loss. TransPert (3rd overall, #1 on PDS) used no neural net at all.
• Foundation models are useful but oversold. Fine-tuned they help; zero-shot they frequently lose to HVG/scVI (Kedzierska 2025). For unseen/combinatorial perturbation they don't yet beat linear baselines (Ahlmann-Eltze 2025).
• The honest gap: no system today is a true "virtual cell." STATE and the AIVC roadmap point the direction; the near-term frontier is cross-context / cross-cell-type transfer with rigorous held-out evaluation, where statistics-first hybrids currently lead.

BioMate Positioning

(Accurate, from BioMate's internal docs; several items are plans/experiments, not shipped capabilities.)

• Mechanistic virtual-cell infrastructure (built). Four Phase-1 workflows run on BioMate's AWS-Batch engine: CiPA cardiac-safety ODE (hERG mechanistic confirmation), ODE signaling simulation, metabolic-flux FBA, and LINCS-L1000 transcriptomic MoA — integrated with existing ADMET/PBPK/RNA-seq workflows.
• Perturbation-prediction infrastructure (staged). GEARS, CPA, scGen, scFoundation/TranscriptFormer are catalogued with licenses/GPU sizing/Docker recipes as planned Phase-2 GPU workflows.
• A rigorous evaluation posture. Consistent with this survey's evidence, BioMate measures any model against simple baselines and an oracle ceiling, and favors mechanistic and statistically-grounded methods over headline claims that don't replicate.
• Honest scope. BioMate does not claim a trained perturbation model beating SOTA, nor a whole-cell simulator; its differentiators are the execution engine (mechanistic + ML workflows on managed AWS Batch) and a rigorous, baseline-anchored evaluation approach.