Loading…
Virtual cell transformer that predicts how cells respond to genetic, chemical, or signaling perturbations, generalizing to unseen cellular contexts.
A CRISPR screen run in one cell line says surprisingly little about what the same perturbation does in another: effect sizes shift, some genes respond only in particular contexts, and the technical variation between two experiments can rival the biological signal between them. Perturbation prediction models have mostly been fit to a single dataset, and degrade as soon as the context changes — precisely when a prediction would be worth having.
State, developed at Arc Institute by Abhinav K. Adduri, Yusuf H. Roohani and colleagues, splits that problem in two. The State Embedding (SE) model learns what a cell is from observational single-cell RNA-seq — 167 million human cells, no perturbations involved. The State Transition (ST) model then learns how a perturbation moves a cell within that representation, training on over 100 million perturbed cells across 70 cellular contexts. Because the two are trained on different data and composed only at inference, an embedding derived purely from observational data can carry a transition model into a cell type where no perturbation was ever measured. Arc styles the name State; the preprint's small capitals are often transcribed STATE.
The second departure is that ST predicts over sets of cells rather than individual transcriptomes: self-attention runs across the cells in a perturbation arm, so the model sees the spread of responses instead of being forced into a distributional assumption about a population mean. State appeared as a bioRxiv preprint in June 2025 and was published in Cell in August 2026.
SE is a 600M-parameter encoder-decoder. Its encoder stacks 16 pre-normalized transformer layers with 16 attention heads and hidden dimension 2,048; the decoder is an MLP reconstructing log-normalized gene expression from a cell embedding plus target gene embeddings. Training ran four epochs over 14,420 AnnData files on 32 NVIDIA H100 GPUs, split at the dataset level to avoid leakage. All inputs are harmonized to 19,790 human protein-coding Ensembl genes and normalized to a total UMI depth of 10,000. ST is far smaller — roughly 244M parameters in the Tahoe-100M and Parse-PBMC configurations — a four-layer bidirectional transformer on a LLaMA-style backbone with SwiGLU blocks and RMSNorm, operating on cell sets of 256 to 512 with positional embeddings disabled because cell order carries no meaning.
Evaluation covered three regimes: chemical perturbations from Tahoe-100M (380 drugs, 50 cancer cell lines), cytokine signaling from Parse-PBMC (90 perturbations, 12 donors, 18 cell types) and genetic perturbations from Replogle-Nadig (2,024 perturbations). On a perturbation discrimination score, State improved on baselines by more than 30% on the large datasets — 54% on Tahoe-100M and 29% on Parse-PBMC — while matching the perturbation-mean baseline on the genetic data, where effects are subtler. For differential expression it led all baselines on p-value AUPRC, most sharply on the genetic dataset, and ranked perturbations by effect size with Spearman correlations 70% higher than baselines on Tahoe-100M, 53% higher on Parse-PBMC and 22% higher on Replogle-Nadig.
State is aimed at researchers who need transcriptomic response predictions at a scale their bench cannot reach — screening compounds or genetic targets in silico before committing to a large CRISPR or chemical screen, or asking how a disease-relevant cell type would respond to an intervention profiled only elsewhere. Generalization to unobserved contexts matters most where the relevant cells are scarce or hard to culture. Cell-Eval and the cell-load preprocessing library give methods developers the same evaluation and data pipeline, and the SE checkpoints are usable for annotation and dataset integration on their own.
State is Arc Institute's first public perturbation model, following its Evo 2 genome language model, and it makes a concrete case that generalization improves when cell identity and perturbation effect are modeled separately rather than end to end. Its release anchored the Virtual Cell Challenge, which used Cell-Eval as the scoring framework. The limitations are real: coverage is human transcriptomics only, with no protein-level, spatial or cross-species data; the genetic perturbation results sit much closer to simple mean baselines than the chemical and signaling ones; and the weights carry a non-commercial license that rules out industrial use without a separate arrangement.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.