Overview
Arc Institute's virtual cell model State passed peer review and was formally published in *Cell* on August 31, 2026, under the title *Predicting cellular responses to perturbation across diverse contexts with State*. Its training data covers 267 million human cells across 70 cell lines. The accompanying Cell-Eval evaluation framework is becoming an industry benchmark for virtual cell research.
Drug development is the world's most expensive trial-and-error game: roughly 90% of candidate drugs fail in clinical trials, usually due to insufficient efficacy or unexpected side effects in real human bodies. Arc's bet is an AI model that predicts how cells respond to perturbations they have never seen — a "virtual cell."
Model Architecture
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 cells, compressing each cell's transcriptomic state into a compact representation that filters out technical noise from different labs, reagents, and sequencing depths.
- State Transition (ST): a self-attention Transformer trained on 100 million perturbed cells, jointly observing differences and commonalities among cells under the same conditions to precisely separate true perturbation effects from background noise.
- Sponsors: NVIDIA and 10x Genomics
- Prize: $100,000
- Task: predict CRISPRi knockdown responses across 6 independent, fully unseen cell lines
- Evaluation: Cell-Eval
- Predictions currently cover a set of highly variable genes, not the whole genome;
- Zero-shot performance remains limited on fully unseen datasets due to technical variation between experiments;
- Optimal transport mapping guarantees depend on certain 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 announcements and Virtual Cell Challenge 2026 launch (sponsored by NVIDIA and 10x Genomics, $100,000 prize, 1,200+ prior entrants)
- Public datasets: Tahoe-100M, Parse-PBMC, Replogle-Nadig (from Roohani et al.)
- Reference databases: DrugComb (combination synergy scores), DepMap (cancer dependency map)
Cell-Eval: A Unified Yardstick
Cell-Eval measures virtual cell models along three dimensions:
1. Gene expression counts — accuracy of predicted values; 2. Differentially expressed genes (DEGs) — correctly identifying upregulated/downregulated genes; 3. Perturbation effect size — estimating overall perturbation strength.
State's performance on three public datasets:
| Dataset | Perturbation type | Discrimination gain | Correlation gain | |---|---|---|---| | Tahoe-100M | Chemical drugs | +66% | +91% | | Parse-PBMC | Cytokines | +29% | +48% | | Replogle-Nadig | Gene editing | +10% | +10% |
On perturbation effect size, State achieved a Spearman correlation of 0.9 on Tahoe-100M — nearly perfect ranking of drug potencies.
Four Clinically Relevant Capabilities
1. Zero-shot prediction on unseen cell types
State predicted drug effects on A549 lung cancer cells with no dose data or priors, 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 effects on C32 melanoma cells — with no dose data — where simple averaging baselines failed completely.
2. Indirect prediction of cell viability
State predicts expression profiles rather than viability directly. A regression model mapping predicted profiles to viability estimates achieved a Pearson correlation of 0.52, far better than simple baselines (0.2–0.31) — meaning clinically relevant endpoints can be predicted via expression as a mediator.
3. Simulating drug combinations
When abiraterone was the first drug, State's predicted second-drug effects matched known results in the DrugComb synergy database. Since wet-lab testing of combinations scales quadratically with drug count, in-silico combination screening is structurally valuable.
4. Cross-modal transfer learning
Gene regulatory relationships learned from drug perturbation data transferred to gene perturbation scenarios: State predicted knockout responses consistent with DepMap survival data, despite never being trained on gene perturbations.
From bioRxiv to Cell: 14 Months
State appeared as a preprint on bioRxiv in June 2025 and reached *Cell* on August 31, 2026 — a typical 14-month peer-review window. Notably, Arc open-sourced Cell-Eval so other virtual cell teams can benchmark against it, not just Arc's internal tool.
In August 2026, Arc launched the Virtual Cell Challenge 2026:
Zero-shot prediction is the key theme; the previous challenge drew 1,200+ teams.
Acknowledged Limitations
More importantly, the paper makes an explicit architectural choice: separating cell heterogeneity from technical noise across layers, and decoupling SE and ST training objectives. This points State's ceiling not at single-model scaling but at finer-grained modeling of cell biology.
Positioning
The core promise of virtual cell models like State is to move part of the wet-lab screening pipeline into the dry-lab stage — simulating in silico before deciding which experiments are worth running. Clinical trials won't be eliminated (the human body's complexity far exceeds cell cultures), but the leverage of early screening, early elimination, and early focus is real. Even a few percentage points of improvement in preclinical hit rates would translate to billions of dollars saved and years of faster drug launches.
The real bottleneck for virtual cells is not model architecture but establishing bidirectional feedback between real experiments and in-silico simulation. By open-sourcing Cell-Eval and emphasizing zero-shot evaluation in the Virtual Cell Challenge, Arc is pushing "evaluation authority" from a single company into the public domain.