The 2026 Virtual Cell Challenge: predicting perturbation responses in cell contexts a model has never seen
Registration for the 2026 Virtual Cell Challenge is open at virtualcellchallenge.org. This year the task is zero-shot: we are not releasing a training set, and the evaluation dataset is far more expansive than last year’s challenge. Models must predict CRISPRi knockdown responses in six cell lines they have never seen perturbed, using only the unperturbed state of those cells and a list of genes to knock down. The grand prize is $100,000, and the Challenge is again sponsored by NVIDIA, 10x Genomics, and Ultima Genomics.
What we learned from 2025
The inaugural Virtual Cell Challenge drew more than 5,000 people registered from 114 countries, more than 1,200 teams submitted results, and more than 300 teams made a final submission. Participants ranged from students entering the field to established industrial research groups.
Year one gave the field a shared task focused on predicting how cells respond to targeted interventions with a purpose-built benchmark dataset, and a community willing to argue in public about how these models should be evaluated (which was one of the more enjoyable parts to watch). It also showed where the field still has room to grow: model accuracy, metric design and biological generalization. Two resources from last year remain relevant. The full 2025 dataset, including training, validation, and the held-out test perturbations, is now downloadable through the Arc Virtual Cell Atlas. The Generalist Prize-winning team from Altos Labs has also posted a preprint describing their flow-matching approach, which we recommend reading before building this year.
The 2026 task: zero-shot across cellular contexts
The 2025 Challenge asked models to predict within a single measured context based on a small set of training data from the same context. We provided H1 human embryonic stem cell perturbation data and asked participants to generalize to held-out perturbations in the same cells.
That was hard enough! But generalizing to a fully unseen cell type is harder, and closer to what biologists actually need. Many of the cellular contexts we care about in human biology are difficult or impossible to perturb directly: rare cell types, primary cells that do not expand outside the body, diseased cells that cannot be cultured faithfully, and tissue states that depend on their native microenvironment. A virtual cell becomes useful when it can take data from the experiments we can run and make better predictions in the contexts where new experiments are sparse, expensive, or unavailable.
We are now moving closer to this practically useful task. There is no Challenge-specific training set this year and the prediction task is zero-shot.
At Arc we ran Perturb-seq across six cell lines from different tissues of origin, using CRISPR-interference, 10x Flex single-cell profiling, and high-throughput sequencing on Ultima’s UG100, with perturbations chosen to provide a strong set of perturbations and responses for a robust Challenge. Three of these cell lines will be used for the validation phase and the live leaderboard. The remaining three are held back for final testing and determining our winners.
For each cell line, participants receive:
- expression profiles from cells expressing non-targeting guides, which define the basal state of the context, and
- the gene identifiers for the CRISPRi knockdown targets to be predicted. Each model must predict the post-perturbation expression profiles that would be measured after CRISPRi knockdown followed by 10x Flex profiling. Arc’s experimental measurements themselves are withheld and serve as ground truth.
The Challenge is designed so that participants see what the new cellular context looks like unperturbed, but never see a perturbation response measured in it. Doing well means inferring how perturbation effects transfer and change between cell types, instead of interpolating among examples in the same one.
Participants may use any modeling strategy and train their models on any data. Public perturbation data is abundant and growing, and teams can also use their own datasets.
Evaluation
The live leaderboard runs on the validation dataset and uses a suite of metrics. This year, we are running a new version of our scoring tool cell-eval, in collaboration with NVIDIA.
For this year, final scoring changes in one important respect. Last year showed that no single metric captures model quality, and that a scoring function with a narrow surface invites optimization against the metric rather than the biology. This year, final rankings use an aggregate across a broader panel of six metrics (read more about the scoring methodology on the Challenge website).
Timeline
- Thursday, August 20: Validation data live, leaderboard active, submissions open.
- Thursday, October 22: Final test set released.
- Thursday, November 5: Final submissions due (11:59 pm UTC, check your timezone)
- Mid-late November: Winners announced
Prizes and sponsors
The grand prize is $100,000, with $50,000 and $25,000 for second and third place. The prizes are a mix of cash and NVIDIA Brev credits.
The Challenge is sponsored by NVIDIA, 10x Genomics, and Ultima Genomics. NVIDIA's involvement this year extended well beyond sponsorship: their team worked with us directly on accelerating and comparing evaluation metrics for cell-eval.
Who should enter
One of our goals for the Challenge is to build a new community out of three groups:
Machine learning researchers who are interested in building models that generalize across high-dimensional biological data without task-specific training examples. For those who have not previously worked in biology, this is a good time to start!
Computational biologists who have experience considering what meaningful generalization looks like, and who know when a model is wrong for biological rather than numerical reasons.
Experimental and translational researchers who think about the biology behind gene expression networks and will have insights on refining models to absorb this logic.
Individuals, academic labs, companies, and independent research organizations are all eligible.
What we are after
Gene perturbation followed by expression profiling has been a workhorse of cell biology for over two decades. If the field can learn to move that single assay in silico, reliably and across contexts, it fundamentally changes the scale at which we can reason in biology, and the complexity of the questions we can answer. That is the AlphaFold and ImageNet moment we are aiming at for cellular modeling. While we would be pleasantly surprised if anyone reaches this level in 2026, we believe the only way to establish how far away it is, is to measure it in public.
Register at virtualcellchallenge.org. We hope to see many of you back for round two.
More AIxBio from Arc Institute
The Virtual Cell Challenge is one piece of a broader effort at Arc to build the full stack of interconnected AI and biology. On the AI side, these resources range from the Evo series of models to learn the language of DNA, to virtual cell models that predict how cells respond to perturbations, to agentic mining of public data, to tools for training models and applying their predictions and designs.
Learn about:
Arc's Virtual Cell Initiative — Our Institute-wide effort taking a full-stack approach to generate training data and build virtual cell models.
Evo 2 — A biological foundation and generative AI model that reads and writes DNA, trained on genetic data from across the tree of life.
State — Arc's first virtual cell model, trained on large perturbational datasets to predict how genetic, chemical, and environmental changes shift gene expression across cell types.
Stack — A single-cell foundation model that uses in-context learning to predict cellular responses to perturbations never directly measured.
scBaseCount — AI agents that find, clean, and uniformly process single-cell data for model training, part of Arc's Virtual Cell Atlas.
Proto — A framework that integrates disparate AI tools for protein, RNA, and gene-regulation design into a single pipeline for multi-modal biological design.
CodonFM — A family of open-source AI models developed with NVIDIA that reveal the grammar underlying codon choice.
Arc’s AIxBio Fellows Program — A remote fellowship for undergraduates working at the interface of AI and biology, pairing student teams with Arc mentors on 6–12 month projects.
Subscribe to Arc’s newsletter to stay up to date on our AIxBio work.