Perturbation Prediction and Virtual Cells
Perturbation prediction is the most direct route from representation learning to biological action. The model must answer what changes after an intervention, not only what a cell resembles. GEARS (Roohani et al., Nature Biotechnology 2024) is the canonical demonstrated example, with subsequent work on scGen, CPA, scPRAM, and neural optimal transport extending the toolbox. Independent evaluation (Ahlmann-Eltze et al., Nature Methods 2025) shows that on many key tasks the deep approaches do not consistently outperform simple PCA plus linear regression. The Virtual Cell Challenge now has a peer-reviewed Cell benchmark framing, and CZI Virtual Cells Platform is infrastructure for the broader research agenda (Roohani et al., 2025). The virtual cell ambition is valuable as framing; reliable executable cell models that forecast all molecular, phenotypic, and temporal consequences of arbitrary interventions remain beyond current capabilities.
Use this chapter to:
- Predict how cells respond to genetic, chemical, environmental, or combinatorial perturbations in measured contexts.
- The counterfactual question is the hard part: held-out genes, combinations, cell types, donors, laboratories, and readouts test different claims.
Prerequisites: Single-Cell Foundation Models for the representation context; Evaluation Principles for Life Sciences AI for the linear-baseline rule.
Summary: Predict how cells respond to genetic, chemical, environmental, or combinatorial perturbations in measured contexts. Perturbation prediction is useful for bounded prioritization, but independent evaluations show that broad virtual-cell claims need stronger benchmarks.
Key point: The counterfactual question is the hard part: held-out genes, combinations, cell types, donors, laboratories, and readouts test different claims. Open question: whether perturbation models generalize across donors, cell types, laboratories, combinations, and functional readouts.
Bottom line: Virtual-cell work links single-cell models to systems biology, target discovery, toxicology, cell therapy, and automated experiment selection.
What is this field trying to solve? Predict how cells respond to genetic, chemical, environmental, or combinatorial perturbations in measured contexts.
What is the core idea? The counterfactual question is the hard part: held-out genes, combinations, cell types, donors, laboratories, and readouts test different claims.
What is the current state of the field? Perturbation prediction is useful for bounded prioritization, but independent evaluations show that broad virtual-cell claims need stronger benchmarks.
What do we know, and what remains open? Known reference points include GEARS, scGen, CPA, scPRAM, Perturb-seq, CRISPR screens, Virtual Cell Challenge, Ahlmann-Eltze benchmarks, and OpenProblems resources. What remains open is whether perturbation models generalize across donors, cell types, laboratories, combinations, and functional readouts.
Why does this matter? Virtual-cell work links single-cell models to systems biology, target discovery, toxicology, cell therapy, and automated experiment selection.
Introduction
GEARS uses graph-enhanced modelling to predict transcriptional outcomes of novel multigene perturbations in selected settings (Roohani et al., 2024). scGen (Lotfollahi et al., 2019) and CPA (Lotfollahi et al., 2023) preceded and extended this work with VAE and dose-aware architectures. scPRAM (Jiang et al., 2024) added attention-mechanism approaches. Neural optimal transport framed perturbation response as counterfactual transport between measured single-cell states (Bunne et al., 2023). The Arc Institute Virtual Cell Challenge gives the field a public benchmark for perturbation response prediction.
The published headline performance is bounded by the Ahlmann-Eltze et al. (2025) critique: deep perturbation predictors do not consistently beat PCA plus linear regression on the canonical out-of-distribution settings.
The data foundation is perturbational, not merely observational. Perturb-seq linked pooled CRISPR perturbations to single-cell RNA profiles (Dixit et al., 2016), and parallel CRISPRi single-cell screens showed how perturbation barcodes can support pathway dissection (Adamson et al., 2016). Later work used rich single-cell phenotypes to map genetic interaction structure (Norman et al., 2019) and expanded Perturb-seq toward genome-scale genotype-phenotype maps (Replogle et al., 2022). Perturb-CITE-seq added paired RNA and protein readouts in patient-derived models, showing why virtual-cell claims need modality-specific endpoints rather than one generic “response” label (Frangieh et al., 2021). Virtual-cell claims should be read against this measured perturbation layer.
What a perturbation assay actually measures
The word perturbation hides several different experiments. CRISPR knockout changes DNA sequence and can create complete loss of function. CRISPRi represses transcription without cutting DNA. CRISPRa increases expression. RNA interference, base editing, and prime editing change different molecular layers. Compounds add dose, exposure time, target engagement, off-target effects, metabolism, and cell-state dependence. Environmental perturbations such as hypoxia, cytokines, nutrient changes, infection, or co-culture add still more context.
The readout matters as much as the intervention. Single-cell RNA-seq measures transcript abundance. CITE-seq and Perturb-CITE-seq add protein markers. Cell Painting measures morphology. Pooled viability screens measure growth or death. Secreted-protein assays measure a different phenotype again. A model trained to predict transcription after CRISPRi has not demonstrated prediction of morphology after drug exposure, or viability after a combination treatment.
The counterfactual problem
Perturbation prediction is a counterfactual task. The observed cell state is one outcome; the model is asked to estimate the state that would have occurred if the same biological context had received a different intervention. This is harder than annotation because the target outcome is often unmeasured for the exact donor, cell type, genotype, and environmental state of interest. The model must borrow information from related perturbations, related genes, related cell states, or prior biological structure.
This is why perturbation models need explicit guardrails. The causal claim is not “the embedding changed.” The causal claim is this intervention would produce this measured response in this context. The difference matters for experimental design. A gene-expression prediction that is accurate for highly expressed marker genes may still be weak for the pathway, cytokine, surface protein, or viability endpoint that drives the biological decision.
Split design is the evidence test
The evaluation split defines the scientific claim. A random cell split mostly tests whether the model can interpolate within known perturbations and known contexts. A held-out gene split asks whether the model generalizes to an unseen target. A held-out gene-combination split asks whether it can compose interactions. A held-out cell-type or donor split asks whether the prediction transfers across biological context. A held-out dataset or laboratory split asks whether it survives platform and protocol drift.
Weak perturbation papers often report one metric across one split and then discuss the result as if it established a general virtual cell. Stronger papers name the held-out unit, run simple baselines, stratify by effect size, and show where performance collapses. The Ahlmann-Eltze critique is important because it makes this discipline explicit: a deep method has not earned its claim until it has beaten a strong simple comparator on the split that matters.
What is demonstrated?
Demonstrated capability includes transcriptional perturbation prediction for selected genes, cell contexts, and benchmark designs. GEARS demonstrated prediction of some novel multigene perturbation outcomes (Roohani et al., 2024). scGen, CPA, scPRAM, and neural optimal transport demonstrate adjacent approaches to transfer, dose-aware response modelling, attention-based response modelling, and counterfactual transport. Public challenge efforts demonstrate growing demand for independent evaluation of virtual cell models. The critical-evaluation literature (Ahlmann-Eltze et al., 2025) now sets the baseline discipline for new claims.
The 2025 Virtual Cell Challenge is useful because it tests models under a shared benchmark rather than a paper-specific split. The peer-reviewed Cell paper frames the challenge as a shared benchmark for virtual-cell claims, while the challenge wrap-up reported that perturbation prediction models were not yet consistently outperforming naive baselines across all metrics and did show improvements on perturbation discrimination and differentially expressed gene identification (Roohani et al., 2025; Arc Institute, 2025). This aligns with the linear-baseline critique: the field is progressing, but the honest benchmark is still hard.
The strongest demonstrations share a common pattern. They are anchored in measured perturbational data, not only pretrained embeddings. They define the readout, because RNA, protein, morphology, and viability are not interchangeable. They specify the held-out unit, because gene holdout and context holdout test different biology. They compare against a simple model, because perturbational response often has strong low-rank structure. These details are not statistical housekeeping. They are the difference between a model that organizes known screens and a model that can guide the next experiment.
| Evidence Anchor | What It Supports | Practical Constraint |
|---|---|---|
| GEARS | Multigene perturbation outcome prediction | Generalisation depends on graph priors and cell context |
| scGen / CPA / scPRAM | Adjacent perturbation modelling approaches | Performance varies; linear baselines often competitive |
| Neural optimal transport | Counterfactual single-cell response transfer | Needs measured source and target domains |
| Perturb-seq lineage | Measured input data for perturbation prediction | Assay design and guide efficacy shape the label |
| Perturb-CITE-seq | Multi-modal perturbation readouts | Protein and RNA effects can diverge |
| Ahlmann-Eltze 2025 critique | The discipline that bounds claims | Linear-baseline comparison is mandatory |
| Virtual Cell Challenge | Public benchmark framing | A challenge metric is not a full cell model |
What is theoretical?
Theoretical capability includes models that rank interventions before CRISPR, drug, or combination screens. This is plausible for defined cell systems and measured outputs, especially when active learning adds new experiments. Cross-cell-type, cross-tissue, and cross-species transfer is partial. Multi-modal perturbation prediction, combining transcriptomics, proteomics, imaging, and metabolomics, is an active research direction.
The high-value near-term use case is experimental prioritization. A model can reduce the number of perturbations that need to be screened if it reliably identifies unpromising regions of the design space and nominates high-information experiments. That use case does not require full cellular simulation. It requires calibrated uncertainty, active-learning discipline, and a laboratory workflow that records failed predictions as carefully as successful ones.
The AI virtual-cell roadmap frames the long-term target as multi-scale and multi-modal, spanning molecules, cells, tissues, perturbations, and time (Bunne et al., 2024). That roadmap is a useful research agenda, not a claim that a general executable cell currently exists.
What is beyond current capability?
Beyond current capabilities includes a complete executable cell model that forecasts all molecular, phenotypic, and temporal consequences of arbitrary interventions. Current data cover slices of cellular behaviour. Replacement of experimental perturbation screens by virtual cell predictions remains beyond current capabilities.
Also beyond current capabilities is unqualified cross-context response prediction. A predictor that works in one cancer cell line, one immune-cell activation state, or one organoid system has not shown that it works across donors, disease stages, species, or tissue microenvironments. The wider the claim, the more the benchmark must resemble the intended biological use.
What would make this more promising?
The claim would strengthen if perturbation models beat PCA plus linear regression on preregistered held-out genes, combinations, cell types, donors, laboratories, and readouts, with uncertainty that remains calibrated under shift. Stronger virtual-cell evidence would include prospective predictions that change experimental selection and are then confirmed by RNA, protein, morphology, viability, or functional assays.
The claim would weaken if gains vanish on simple baselines, if results depend on one cell line or screen, or if claimed response prediction fails when the perturbation type, dose, time point, or biological context changes.
What should researchers, biotech teams, funders, and program leaders do with this?
- Name the perturbation type: knockout, knockdown, activation, compound, dose, timing, or combination.
- Name the output: expression, morphology, viability, secretion, or functional assay.
- State the held-out unit: gene, combination, cell type, donor, species, laboratory, or time point.
- Use held-out perturbations and held-out cell contexts separately.
- Always include a PCA plus linear regression baseline in any perturbation prediction comparison.
- Stratify results by effect size; small average errors can hide failures on biologically important large responses.
- Preserve negative evidence. Failed predictions are training data for the next screen.
- Avoid causal claims from observational pretraining alone.
- Watch the Virtual Cell Challenge results as the strongest evidence source for the field.
- Treat virtual cell predictions as research prioritisation, not as substitutes for experiments.