What is a virtual cell, and what can it predict?

A virtual cell is a computational model of a specified cellular behaviour. In the research discussed here, it predicts how gene expression changes after a genetic, chemical or signalling intervention. Researchers can compare those predictions with measurements from physical cells.

A model trained to predict RNA measurements gives an answer about RNA. Protein activity, metabolism and tissue function each need their own data and experimental test.

Imagine a researcher deciding which genes to suppress in an ageing cell culture. A model could rank proposed interventions by their predicted effect on an expression pattern. That would help choose experiments. Whether the cells repair damage, divide appropriately or perform their specialised job remains a question for those experiments. This is an illustrative use, not a reported rejuvenation result.

The experiment behind the training data

Perturb-seq combines genetic interventions with RNA sequencing of individual cells. In the 2016 work by Dixit and colleagues, expressed barcodes connected a cell's RNA measurements with its assigned CRISPR perturbation. Pooling many interventions allowed researchers to examine responses across large numbers of cells.

Each training example links an intervention and cellular context to measured responses. A prediction model needs to learn how those parts relate. The 2022 genome-scale study by Replogle and colleagues extended this experimental approach across expressed genes and supplied data used in subsequent model evaluations.

An expression profile records which RNA molecules the assay detected and their measured abundance. Different cells receiving the same intervention can produce different profiles. The distribution of responses therefore deserves attention alongside the average. A small population responding strongly can disappear in an average dominated by cells that hardly respond.

Suppressing expression with CRISPR interference, activating expression and removing a gene create different interventions. Measurement time affects what the assay captures. A model's answer should name the intervention and the time represented by its training data.

What State attempts

Arc's State paper appeared online in Cell on 31 August 2026, following its earlier preprint. The authors report improved prediction against their chosen baselines across genetic, chemical and signalling datasets.

Arc's tool description separates an embedding component, which represents individual cells, from a transition component operating on sets of cells. The design addresses variation within a population and between contexts. These are defined evaluations of a particular model and dataset collection. They do not establish reliable prediction of every response in an arbitrary human tissue.

A developer evaluation is useful evidence, especially when the methods and comparison code are available. An external test asks a further question: will the result persist when another group chooses the experiments, keeps their results hidden and applies a declared scoring method?

Four ways to make a prediction test harder

A test should explain what the model has never seen. Randomly withholding some cells from an experiment leaves the model much closer to the answer than withholding the entire intervention. Changing the cellular context adds another demand.

A public test in unfamiliar cells

Arc's 2026 Virtual Cell Challenge asks entrants to predict CRISPR-interference responses across six cell lines. Participants receive unperturbed profiles and target-gene identifiers, while the matching experimental responses stay hidden. Three lines support validation; three are reserved for final testing. The final scoring combines six metrics.

As checked on 4 October, the final test release is scheduled for 22 October and submissions for 5 November. Final results are not yet available. The test asks whether models can predict responses to familiar interventions in cellular contexts lacking matching response examples.

For longevity research, the useful next step is an experiment selected because of a prediction, followed by a functional result in the stated model. A paper can show that the model narrowed a search even when most candidates fail. Report the failures and the number actually tested; those figures tell readers what was gained.