If AI could tell you, before running an experiment, which genes a drug would silence in a given cancer cell line, which pathways it would activate, and whether the cell would die — how much of the roughly 90% clinical failure rate in drug development could it move? On August 31, 2026, the virtual cell model State, led by Arc Institute, was formally published in *Cell* after peer review. Its training data covers 267 million human cells across 70 cell lines, and the accompanying Cell-Eval evaluation framework is becoming an industry benchmark for virtual cell research.
A Model That "Predicts" Cell Fate
Drug development is the world's most expensive trial-and-error game: about 90% of drug candidates fail in clinical trials, usually due to insufficient efficacy or unexpected side effects in real human bodies. Arc Institute's bet is an AI model that predicts how cells respond to perturbations it has never seen — a "virtual cell."
State's architecture has three layers:
- Gene embeddings: one embedding vector per gene, capturing functional relationships across experimental conditions.
- State Embedding (SE): a bidirectional Transformer pretrained on 167 million observational human cells. It compresses each cell's transcriptomic state into a compact representation, filtering out technical noise from different labs, reagents, and sequencing depths.
- State Transition (ST): a self-attention Transformer trained on 100 million perturbed cells. It observes differences and commonalities among cells under identical conditions, precisely separating true perturbation effects from background noise.
- Sponsors: NVIDIA and 10x Genomics; prize: $100,000;
- Task: predict CRISPRi knockdown responses on 6 independent, previously unseen cell lines;
- Evaluation: Cell-Eval, with zero-shot as the key word. The first edition drew 1200+ teams.
- It currently predicts expression changes only for a set of highly variable genes, not the whole genome;
- Zero-shot performance on entirely unseen datasets remains limited by inter-experiment technical variation;
- Theoretical guarantees of optimal transport mapping rest on regularization conditions.
- Roohani et al., *Predicting cellular responses to perturbation across diverse contexts with State*, Cell, 2026-08-31, Open Access, https://www.cell.com/cell
- Arc Institute official announcement and Virtual Cell Challenge 2026 launch (sponsored by NVIDIA + 10x Genomics, $100,000 prize)
- Tahoe-100M, Parse-PBMC, Replogle-Nadig public datasets (from the Roohani et al. paper)
- Public references: DrugComb (drug combination synergy scores), DepMap (cancer dependency map)
Cell-Eval: Splitting "How Well Does It Predict" into Three Metrics
The open-source Cell-Eval framework measures virtual cell models along three dimensions:
1. Gene expression counts: accuracy of predicted values (Pearson correlation); 2. Differential expression genes (DEG): correctly identifying up/downregulated genes; 3. Perturbation effect size: accuracy in estimating overall perturbation strength (Spearman correlation).
Performance improvements over baselines on three public datasets:
| Dataset | Perturbation Type | Discrimination Gain | Correlation Gain | |---|---|---|---| | Tahoe-100M | Chemical drugs | +66% | +91% | | Parse-PBMC | Cytokines | +29% | +48% | | Replogle-Nadig | Genetic perturbations | +10% | +10% |
On perturbation effect size, State reached 0.9 Spearman correlation on Tahoe-100M — nearly perfectly ranking the potency of different drugs.
Four "Clinically Relevant" Capabilities Already Validated
1. Zero-shot prediction on unseen cell types
State predicted drug effects on A549 lung cancer cells with no prior dose data, closely matching real experiments and correctly identifying core pathway genes (PTPN11, RAF1, SHOC2) as the most similar genetic perturbations. It also predicted FDA-approved trametinib's gene expression changes in C32 melanoma cells it had never seen at any dose — where naive averaging baselines completely failed.
2. Indirect prediction of cell viability
State predicts gene expression profiles, which a regression model then maps to viability estimates, achieving a Pearson correlation of 0.52 — far above simple baselines (0.2–0.31). This means the clinical endpoint that matters most (does the cell die?) can be predicted through gene expression as a mediator.
3. Simulating combination therapies
With abiraterone as the first drug, State's predictions of second-drug effects matched known results in the DrugComb synergy database. Since wet-lab testing of combinations scales quadratically with the number of drugs, in-silico screening is structurally valuable.
4. Cross-modal transfer learning
Even without genetic perturbation training data, State transferred gene regulatory relationships learned from drug perturbations to predict knockout responses, correlating significantly with DepMap (the Cancer Dependency Map) survival data.
From bioRxiv to Cell: 14 Months of Peer Review
State was released as a preprint in June 2025 and formally published in *Cell* (Open Access) on August 31, 2026, under the title *Predicting cellular responses to perturbation across diverse contexts with State*. Notably, Arc publicly released the Cell-Eval framework so other virtual cell teams can benchmark against it.
In August 2026, Arc launched the second Virtual Cell Challenge 2026:
The Next Hurdle for Virtual Cells
The paper honestly notes its limitations:
The more important contribution is the explicit architectural choice: separating cell heterogeneity from technical noise across layers, and decoupling SE and ST training objectives. The authors note that as larger perturbation atlases are built, State's performance should improve further, and its attention mechanisms show sensitivity to population heterogeneity — opening new possibilities for understanding gene regulation.
The real bottleneck for virtual cells may not be model architecture, but establishing bidirectional feedback between real experiments and in-silico simulation. By open-sourcing Cell-Eval and running zero-shot challenges, Arc is pushing the "evaluation standard" out of any single company's hands into public space — and each public benchmark raises the credibility of virtual cells by a notch.