Can a Virtual Cell Tell Us Which Experiment to Run Next?
A biologist considering an experiment faces a familiar problem: there are more plausible questions than there is time to test them.
What happens if a gene becomes less active? Does a chemical change a cell’s behaviour? Will the same intervention have a different effect in another type of cell?
Each question can lead to more questions. Testing combinations multiplies the possibilities further.
The ambition behind a virtual cell is to make some of these questions computationally tractable. A researcher would describe a starting condition and an intervention, and a model would predict a useful aspect of the response.
The idea is sometimes presented as a digital replica of life. That image can obscure both the scientific challenge and the nearer-term opportunity. A model would not need to reproduce every molecular event to become valuable. It would need to make predictions reliable enough to improve the next experimental decision.
Whether current systems can cross that threshold is a much harder question than whether they can produce convincing-looking biological data.
A cell is observed through measurements
The phrase “virtual cell” suggests a complete object. In research, the model usually has access to particular measurements of a cell.
Gene expression data, for example, describes the RNA molecules detected under specified conditions. It provides an informative view of cellular activity, but it is not an inventory of everything the cell is doing.
Protein activity, location, interactions with neighbouring cells, and changes over time can all matter. Different experimental methods expose different parts of the system.
A 2024 perspective on building AI virtual cells described a broad programme integrating biological data across scales and types. It was an account of priorities and opportunities, rather than evidence that a comprehensive predictive cell had already been achieved.
This distinction helps clarify the field. A model that predicts a particular molecular readout can be scientifically useful while representing only a limited aspect of cellular behaviour.
The important questions concern what it predicts, under which conditions, and how its predictions are tested.
Predicting an intervention is harder than recognising a pattern
A model can learn that certain measurements tend to occur together. Researchers often want something more demanding: a prediction of what will happen when they deliberately change the system.
Those are different tasks.
Imagine a model trained on observations of healthy and diseased cells. It might distinguish the two groups accurately. That does not establish which intervention would move a diseased cell towards a healthier state.
A characteristic associated with disease might contribute to it, result from it, or reflect some third process. Recognising the association does not resolve these possibilities.
Intervention data helps by recording responses to controlled changes. In experiments combining genetic perturbation with single-cell measurements, researchers can reduce the activity of selected genes and observe the resulting expression patterns.
The data still has boundaries. Results from one cellular setting may not carry over to another. The size of an effect may depend on the starting state, the experimental method, and when the measurement is taken.
A useful predictor must learn something that survives at least some of those changes.
The competition that exposed the measurement problem
Arc Institute’s first Virtual Cell Challenge made this ambition concrete. Participants were asked to predict gene expression responses to genetic perturbations.
The benchmark used approximately 300,000 single-cell profiles from human embryonic stem cells, with 300 targeted perturbations. Its construction required decisions about cell type, intervention method, measurement depth, and which responses would be represented. Those choices were part of the scientific work, not merely preparation for model training.
The results were instructive.
In its December 2025 assessment, Arc reported that perturbation prediction models were not consistently outperforming simple baselines across all metrics. There were improvements in particular capabilities, but the competition also revealed how strongly rankings depended on what the evaluation rewarded.
Almost all models performed worse than the baseline on mean absolute error, while competitive optimisation concentrated on other scores. Arc introduced an additional prize using a broader evaluation.
This does not mean the models learned nothing. It means “better prediction” required a more precise definition than a leaderboard position could provide on its own.
Why a simple average can be a serious competitor
A baseline might predict a typical response or average expression pattern. At first glance, that seems an embarrassingly low bar for an advanced model.
But biological measurements contain substantial structure that is shared across conditions. Much of a measured profile can remain similar even when a smaller, important set of features changes.
A model rewarded mainly for overall numerical similarity may do well by reproducing what usually stays the same. A scientist looking for the consequences of a specific intervention needs to know whether the model captures the meaningful differences.
An independent study published in Nature Biotechnology, introducing the Systema evaluation framework, examined this issue using multiple datasets and simple baselines. It showed why average effects and systematic variation need to be considered when assessing perturbation prediction methods.
The lesson extends beyond choosing a tougher scoring system. Researchers must decide what kind of mistake would make a prediction scientifically misleading.
Missing the direction of an important response may matter more for one use than a small error distributed across many measurements. A single universal score is unlikely to capture every intended application.
The harder test: a cell the model has not seen perturbed
Arc’s 2026 challenge moved towards another difficult question: transferring predictions across cellular contexts.
The announced task asks models to predict responses in six cell lines without seeing measured perturbation responses from those contexts. Participants receive baseline profiles and the genes to be targeted; the experimental outcomes are withheld for evaluation.
As of October 2026, the challenge remains an ongoing test, rather than a completed demonstration of success.
The design matters because scientific value often lies precisely where measurements are missing.
Interpolating between familiar experiments can be useful. Predicting in a context where experiments are difficult, costly, or sparse would extend what a researcher can reasonably investigate.
But a model that performs well in a few cell lines has not thereby demonstrated reliability in a patient’s tissue. Each extension requires evidence.
Useful predictions do not have to replace the laboratory
A practical virtual cell might begin as a tool for prioritisation.
Suppose a research group has hundreds of candidate interventions and resources to investigate only a small subset. A model could help select candidates, identify uncertain cases, or suggest which experiment would be most informative.
Its value would then be assessed through the decisions it improves: whether recommended experiments reveal reproducible effects, whether failed predictions expose useful gaps, and whether the process makes better use of laboratory resources.
This also creates a productive feedback loop. Experiments challenge the model; the model helps choose further experiments.
The danger is allowing predictions to become evidence merely because generating them is inexpensive. A model can produce thousands of hypotheses without making any one of them biologically sound.
The more consequential achievement would be narrower and easier to miss in a headline: a model reliably identifying an experiment that researchers would otherwise have overlooked, followed by a laboratory result that confirms something new.