The $1.8 Billion Bet That More Data Will Enable Progress in Virtual Cells

In October, Biohub, the Chan Zuckerberg science organization, expanded its Virtual Biology Initiative to $1.8 billion. The money pays for open, standardized data to train virtual cells: AI models that predict how a cell will respond to a drug or a gene edit before anyone runs the experiment.

Today's models don't yet work reliably on cells and perturbations they haven't seen. Biohub's head of science, Alex Rives, expects accurate predictive models within five years, not now. More and better data is supposed to close the gap.

Whether it will is genuinely contested. Benchmarks published this year disagree on how good the models are, largely because the field hasn't agreed on how to measure them. This week brings the closest thing to a common yardstick: final entries for Arc Institute's 2026 Virtual Cell Challenge close Thursday, November 5.

Who is betting, and on what

Biohub builds its own models, such as rBio, and with Arc and Tahoe Therapeutics it is producing a perturbation dataset of more than 120 million cells. In November 2025 the EvolutionaryScale team joined, and its co-founder Rives now leads Biohub's science.

Not all of the $1.8 billion is new cash. Biohub committed $500 million in April. The US Department of Energy is adding more than $500 million over five years, NIH is contributing datasets built with earlier federal funding, and Google DeepMind, Isomorphic Labs and Meta are putting in $300 million together.

There is a catch for everyone else. Rives told Reuters that commercial funders get an embargo period before the data goes public, and The Next Web reports that he told Axios it lasts a year. The first dataset is due in about a year.

Biohub isn't alone. Illumina's billion-cell CRISPR atlas is co-funded by AstraZeneca, Merck and Lilly. A growing share of the money in this field is going into data, not just models.

The evidence cuts both ways

The developers' own numbers look strong. Arc's STATE paper calls it, to the authors' knowledge, the first model to consistently outperform simple linear baselines at generalizing across cell contexts, scored with Cell-Eval, an evaluation framework the same paper introduced. In a talk abstract, Xaira executives say X-Cell beats state-of-the-art models by up to five-fold on metrics such as Pearson Δ. Both figures come from the developers' own evaluations.

Independent benchmarks from 2026 disagree with each other:

Skeptical. An April benchmark of models "in the wild" tested STATE and seven other deep models on unseen cells and perturbations. It found performance in common setups is often overestimated and drops markedly under strict conditions. "Signal, Bounds, and Baselines" found across seven datasets that deep models often fail to meaningfully beat simple linear baselines.
More positive. A July preprint comparing 11 methods found STATE improved on a cell-mean baseline by 26% for unseen single-gene perturbations, against 19% for a linear model: ahead, but not dramatically. It found clear gains for deep models on gene combinations and on transfer to unseen cell types. A Nature Biotechnology paper by Shift Bioscience and collaborators, reported by GEN, found that on calibrated metrics, deep models often beat uninformative baselines, and that PRESAGE and scLambda did better still.

Much of the disagreement comes down to measurement. The April benchmark found that different metrics produce substantially different model rankings. The Nature Biotechnology team found that common metrics, such as mean squared error, are often poorly calibrated, especially on datasets where perturbations have weak effects. The July preprint reported that a widely used correlation metric scores best for a model that predicts no change at all. Two Stanford computer scientists argued that current benchmarks fail to reliably measure performance, and proposed judging models by discovery outcomes, such as finding hits, instead.

Arc, which runs the Challenge, gave a mixed verdict on the first round. In its wrap-up, it said models were "not yet consistently outperforming naive baselines across all metrics," while showing significant improvements on core capabilities.

Why this week's Challenge matters

The 2026 round is stricter than last year's. In 2025, teams predicted held-out knockdowns in a cell line they had training data from. This year, models must predict knockdown effects in six cancer cell lines with no Challenge-specific training data.

The scoring speaks to the measurement problem. Each of six metrics is scaled between two anchors: a baseline that guesses the average response for that cell context, and a real replicate of the experiment. A score shows how far a model gets from "average guess" toward "as good as rerunning the experiment." Final entries are due November 5, and scores stay hidden until the winners are announced.

The results will show how far today's models get on cells they weren't trained on, measured the same way for everyone. They won't show whether Biohub's data will help, because that data doesn't exist yet. And Arc also builds STATE and Stack, so for Arc's own models the Challenge is a common test, not an independent one.

Where it's already useful

While the general question stays open, pharma is already paying for narrow uses:

GSK and Noetik. GSK licensed Noetik's OCTO-VC for $50 million upfront, one of the first big-pharma licenses of a virtual cell model. In August, Noetik announced the first milestone. GSK uses the model for patient stratification and to enrich its pathology data.
X-Cell. Xaira says the regulators X-Cell predicted for T cells were confirmed in a separate primary T-cell screen.
C2S-Scale. Google and Yale's model generated a cancer-therapy hypothesis that was confirmed in living cells.

All three results come from the developers or their partners, and each is tied to a specific use rather than a claim to simulate any cell. For now, virtual cells are tools for choosing which experiments to run, not replacements for them.

What to watch

The Challenge results. Strong zero-shot scores would weaken this year's skeptical findings. Weak ones would strengthen the case for narrow, validated uses over general-purpose models.
Biohub's first dataset, due in about a year, and the published terms of the one-year embargo. Only then can anyone test the bet that data is what's missing.
X-Cell outside Xaira. Xaira says the weights are coming soon, though only a smaller "Mini" version is slated for release. An outside test of its claims would be a strong signal either way.
A second pharma license. Watch whether Noetik signs another, or another big pharma strikes a comparable model deal.

We will update the tracker when the Challenge results are out, and publish a follow-up on how the models did.

The full picture: The Virtual Cell Tracker compares 14 models, 11 datasets and 16 deals, with license status and an evidence label for every performance claim.

Topic: AI in Bio