The $1.8 Billion Bet on a “Virtual Cell”
A new U.S.-anchored coalition wants to build the biological data layer that could let AI predict how living cells respond before researchers run the experiment. The ambition is enormous. So are the caveats.
For years, artificial intelligence in biology has been sold with a seductive analogy: if large language models can learn patterns from vast libraries of text, perhaps new models can learn the grammar of life from vast libraries of cells. The catch is that the biological library is nowhere near as complete, standardized or easy to copy.
That is the problem a newly expanded coalition is trying to attack. On October 7, the U.S. Department of Energy, the National Institutes of Health and Biohub announced a coordinated effort to build what they describe as an open, AI-ready foundation of biological data. Biohub says the broader Virtual Biology Initiative now aligns about $1.8 billion in new funding, prior federal investments, data resources, computing power and measurement technology. Google DeepMind, Isomorphic Labs and Meta are joining with a combined $300 million commitment, while NVIDIA is supporting the effort with accelerated-computing infrastructure, software and technical expertise.
The destination is often described as a “virtual cell”: a model that can take a biological state, apply a hypothetical intervention such as a drug or genetic perturbation, and predict what the cell will do next. If that works reliably, researchers could use computation to narrow the search space before committing time, money and scarce samples to laboratory experiments. But the most important word is if. Today’s best systems are still far from a universal, fully faithful digital cell, and independent reviews warn that bigger models alone will not solve the harder problems of biological context, validation and generalization.
What “virtual cell” means — and what it does not
The phrase covers several kinds of models, from mechanistic simulations to neural networks trained on single-cell measurements.
A literal digital replica of a human cell would have to represent an almost absurd range of interacting systems: genes switching on and off, proteins folding and binding, membranes moving, metabolites flowing, organelles communicating, signaling pathways firing, DNA being repaired, cells dividing and neighboring cells changing the outcome. A useful scientific model does not need to reproduce every molecule at every instant. It needs to make reliable predictions about a defined question.
That distinction matters because much of the current “virtual cell” field is narrower than the name suggests. Some models predict changes in gene expression after a drug treatment. Others infer cell state from single-cell RNA sequencing, connect multiple “omics” layers, model trajectories over time, or predict how a genetic edit changes downstream pathways. Nature Biotechnology noted earlier this year that many AI systems branded as virtual cells are not yet representations of entire cells. The editorial argued that the field will likely need to combine AI’s pattern-finding power with mechanistic models that encode known biology.
Rank better experiments
A strong model could help scientists choose which perturbations, targets or combinations deserve scarce laboratory time.
Generalize to new contexts
Predictions must hold across cell types, donors, laboratories, doses, time points and biological states that were not in the training set.
Replace wet-lab proof
Current reviews do not support treating AI predictions as a substitute for rigorous experimental and clinical validation.
A virtual cell can be valuable before it is complete.
The practical threshold is not perfection. It is whether the model improves scientific decisions: finding more promising targets, eliminating weak hypotheses earlier, or revealing which experiment would be most informative next. That is a much more testable claim than “AI will simulate biology.”
Why the bottleneck is data, not just compute
Text is abundant and standardized enough to scrape at internet scale. Biology is physical, expensive, heterogeneous and deeply context-dependent.
Language models became dramatically more capable in part because they could ingest enormous quantities of digitized text. Biology has no equivalent corpus. A gene-expression profile from one cell line, measured on one instrument at one dose and one time point, cannot automatically stand in for another tissue, donor or disease state. The same intervention can produce different outcomes depending on age, environment, cell history, neighboring cells and the molecular state at the moment the intervention begins.
That is why the coalition is emphasizing new measurements rather than only bigger neural networks. DOE says its contribution will use national-laboratory capabilities including exascale supercomputing, genome and environmental-molecular facilities, X-ray and neutron scattering, cryo-electron microscopy and tomography, and automated laboratories. NIH says it will coordinate biomedical datasets, repositories and knowledge bases, while working with Biohub to standardize suitable data for training. Biohub’s own five-year commitment includes new imaging systems, cellular engineering, data infrastructure and outside grants intended to expand the kinds of measurements available.
This is the strongest part of the new strategy. Rather than assuming the next model architecture will magically overcome missing evidence, the program is trying to manufacture a richer evidence base: more interventions, more cell types, better spatial data, faster imaging, standardized identifiers and shared data access. A February position paper coauthored by Biohub, University of Chicago and Northwestern researchers made essentially the same argument in sharper terms: virtual cells need coverage of biological contexts, not merely more parameter count or more of the same data.
The ambition also creates a coordination problem. Different laboratories name things differently, use different instruments, process samples differently and store metadata at different levels of detail. If those differences are hidden or poorly documented, a model may learn laboratory artifacts rather than biology. The initiative therefore treats common identifiers, standards and a unified access layer as core scientific infrastructure rather than administrative cleanup.
The $1.8 billion headline is not one new check
The number combines new commitments with previously funded federal data resources and other forms of support.
Biohub describes the expansion as a $1.8 billion commitment in funding, data, computation and new measurement technology. That number is meaningful as a picture of scale, but it should not be read as $1.8 billion in newly appropriated cash flowing to one program. The components are different in timing, form and accounting.
A five-year commitment announced in April: $400 million for internal technologies, infrastructure and data generation, plus $100 million to support external research.
New five-year investment spanning data collection, AI analytics, laboratory measurement, imaging, modeling and computation.
Relevant datasets, repositories and knowledge bases built through prior federal investment and now being aligned to the initiative, not a newly announced $500 million grant.
Combined commitment from Google DeepMind, Isomorphic Labs and Meta. NVIDIA is separately contributing infrastructure and technical support without a dollar figure in the expansion release.
The accounting nuance does not make the effort small. Even after separating prior investment from new commitments, the program connects major federal scientific infrastructure with philanthropic money and some of the largest AI research organizations in the world. The question is whether the coalition can turn that scale into interoperable evidence rather than a patchwork of impressive but disconnected assets.
The new scientific loop: predict, perturb, measure, repeat
The most consequential idea is not a single model. It is a closed loop between AI predictions and automated physical experiments.
A virtual cell becomes scientifically useful when it can be challenged. The program’s implied workflow starts with a model prediction, turns that prediction into a physical perturbation, measures the response, feeds the result back into the data commons, and retrains or revises the model. In machine-learning terms, this is an active-learning loop. In laboratory terms, it is a way to make the next experiment depend on what the model is most uncertain about.
Integrate molecular, imaging, spatial and contextual measurements into a computational description of the cell.
Ask what happens after a drug, gene edit, environmental change or combination of interventions.
Use robotic and human-guided laboratory systems to apply the perturbation under controlled conditions.
Capture expression, structure, location, phenotype and time-dependent responses rather than a single endpoint.
Compare prediction with reality, expose failure modes and choose the next experiment based on what was learned.
Measurement technology is part of the AI stack
The initiative’s emphasis on cryo-electron tomography, high-throughput microscopy and cellular engineering reflects a deeper shift: in biology, better models depend on better instruments.
- Microscopy can add spatial context that sequencing alone loses.
- Perturbation platforms reveal cause-and-effect relationships more directly than static atlases.
- Automation makes it possible to generate repeatable experimental series at a scale human-only workflows cannot match.
Where this could change drug discovery
The biggest near-term prize is not an AI-designed cure on demand. It is reducing wasted experiments and making biological hypotheses more testable.
Drug discovery is full of expensive filters. A target that looks compelling in a database can fail in cells. A result in cells can fail in animals. A mechanism that seems clean in a narrow experiment can become messy in the human body. By the time a therapy reaches clinical trials, the financial cost of being wrong is enormous.
A decision-grade virtual cell could push some failures earlier. Suppose a research team has 2,000 plausible genetic targets for a disease pathway and enough budget to test 100 deeply. A useful model does not need to identify the final drug. It may create value simply by improving the ranking so those 100 experiments contain more of the truly informative targets. The same logic applies to drug combinations, dosing windows, biomarkers and mechanisms of resistance.
This is why the field is paying more attention to perturbation data — measurements taken after deliberately changing a cell. Static atlases are excellent for describing what exists. They are less powerful for learning what causes change. A trustworthy predictive system needs examples that include dose, timing, combinations, cell identity and environmental context, followed by validation in real experiments.
There is already evidence that AI can help specific parts of the therapeutic pipeline. Stanford researchers recently described a “Virtual Biotech” system of cooperating AI agents that can integrate genetics, single-cell data, disease biology and clinical-trial evidence to support drug-development decisions. That work is not the same as a virtual cell, and it does not eliminate human review, but it illustrates the direction of travel: computational systems are moving from passive analysis toward choosing and prioritizing scientific actions.
The hardest test is generalization
A model that performs well on held-out data from a familiar experiment may still fail on the unfamiliar biology that matters most.
AI benchmarks often reward a model for reconstructing patterns inside a dataset that resembles its training data. Biology asks a more punishing question: what happens when the cell type changes, the donor changes, the disease state changes, the lab protocol changes, the drug combination changes, or the response unfolds for days rather than hours?
A September review in Frontiers in Cell and Developmental Biology argued that virtual-cell systems should be judged by their ability to handle unseen contexts, perturbations, doses, time points and combinations, with uncertainty estimates and independent experimental validation. The authors warned that many advanced models show smaller-than-expected gains over simpler baselines when evaluation moves outside familiar settings. A separate Biohub-affiliated position paper reached a similar conclusion: context coverage is a fundamental bottleneck.
That uncertainty matters. A model that is right 85 percent of the time but knows which 15 percent it is unsure about can be more useful than a model with slightly higher average accuracy that is confidently wrong in the most unusual cases. For science, calibrated uncertainty can tell researchers where to spend experimental effort. For medicine, it can become a safety requirement.
Open science will be tested by access, provenance and privacy
The initiative promises an open resource, but openness is not a single switch. The details of who can use what, when and under which safeguards will matter.
Biohub says the generated datasets are intended to become an open resource for the global scientific community, continuing an approach it has used with projects such as CELLxGENE and the CryoET Data Portal. That promise could be highly consequential. A public data commons would allow smaller academic groups to build and test models they could never afford to train on proprietary datasets collected behind corporate walls.
Yet “open” biological data is more complicated than open text or images. Some experiments use human samples. Genomic and health information can be sensitive even when obvious identifiers are removed. Provenance must record consent, collection conditions and allowed uses. Datasets also need demographic and biological diversity; otherwise models may perform best for the populations most heavily represented in training data and worst for groups that are already underserved.
There is also a commercial-timing question. Reuters reported that industry backers are expected to receive an early access period before datasets are fully released to the public. That does not necessarily negate the open-science goal, but it means the timeline and conditions for release deserve scrutiny. Public resources create the most scientific value when access rules are transparent, predictable and designed to support reproducibility rather than simply advertise openness.
If donors, tissues or disease states are unevenly represented, a model can encode those gaps and still look impressive on an average benchmark.
Differences in instruments, sample preparation or processing can create patterns that a model learns unless metadata and standardization are unusually rigorous.
Scientific users need to know when a prediction is extrapolating beyond its evidence, not just receive a single confident answer.
Publishing files is not enough. Other groups need documentation, stable identifiers, versioning and benchmark protocols that make results comparable.
How to tell whether the initiative is working
The best milestones will be scientific, not promotional: better predictions, reproducible datasets and fewer failed hypotheses carried too far.
The Virtual Biology Initiative is built around five-year commitments, which is long enough to generate major infrastructure and short enough that the field should be able to show measurable progress. The first signs of success will probably not be a perfect universal cell model. They will be narrower and more concrete.
A practical scorecard for the next five years
Are new datasets large, diverse, intervention-rich, well documented and genuinely usable outside the laboratories that produced them?
Can researchers combine measurements across institutions without spending months reverse-engineering metadata and resolving incompatible identifiers?
Do models beat strong simple baselines on experiments they did not see, including new cell types, doses, combinations and time horizons?
Are successful predictions confirmed by independent wet-lab experiments and, where appropriate, animal or clinical evidence?
Can the system show that it led researchers to a target, mechanism or experiment they would have been less likely to prioritize otherwise?
Do public researchers receive timely, documented access to the core datasets and evaluation tools promised by the initiative?
There should also be negative results. A credible scientific program will publish where the models fail, which biological contexts remain underrepresented, and which seemingly sophisticated architectures do not outperform simpler methods. That kind of transparency can look less glamorous than a headline claiming a virtual cell has been “built,” but it is exactly how a field turns an ambitious metaphor into a trustworthy instrument.
The most important part of this AI project may be the new biology it forces scientists to measure.
The $1.8 billion figure will attract attention because it is large and because it joins federal laboratories with some of the biggest names in AI. But the deeper bet is methodological. The initiative assumes that a major limitation of today’s biological AI is not simply that models are too small. It is that scientists have not yet measured living systems with enough breadth, context and consistency to train truly predictive models.
If that diagnosis is right, the payoff could extend far beyond one virtual-cell architecture. New imaging systems, perturbation datasets, automated labs, standards and open repositories would become scientific infrastructure that future models can reuse. Even if the first generation of virtual cells falls short of the rhetoric, the measurement ecosystem could still accelerate research.
If the diagnosis is wrong, the program will discover something equally important: that the scaling laws that transformed language and image generation do not transfer cleanly to living systems. Either outcome would teach the field something fundamental. The test now is whether the coalition measures its success by reproducible scientific prediction rather than by model size, partnership count or press-release ambition.
What comes next
The next phase is less about announcing the destination than proving the feedback loop works.
Over the next several years, researchers will need to turn the coalition’s assets into a usable sequence: standardized datasets, interoperable access, validated benchmarks, models that generalize and autonomous-lab loops that produce new evidence faster than traditional workflows. The hardest part will be resisting the temptation to call intermediate milestones a solved virtual cell.
The history of biomedical science is full of tools that were transformative precisely because they made some questions cheaper and faster without making biology simple. DNA sequencing did not eliminate experiments. High-resolution imaging did not eliminate uncertainty. Protein-structure prediction did not remove the need to test function in living systems. A successful virtual cell would likely fit the same pattern: not a replacement for the lab, but a new instrument that changes what scientists decide to do in the lab.
That is why this week’s announcement matters beyond the participating companies and agencies. It is a test of whether the United States can build a shared biological data commons at a scale normally associated with frontier AI, while preserving the standards, validation and openness that make scientific claims trustworthy. The result may determine whether “virtual cell” becomes a useful research instrument — or remains one of the most attractive metaphors in modern biotech.
Quick answers
Is the $1.8 billion all new government spending?
No. Biohub’s total combines several categories. DOE says it will invest more than $500 million over five years. Biohub has a $500 million five-year commitment. Google DeepMind, Isomorphic Labs and Meta together are committing $300 million. NIH is aligning datasets and knowledge resources created through more than $500 million in prior federal investment.
Will researchers be able to run a whole human cell on a computer soon?
That is not what current evidence supports. Today’s systems generally model specific layers or responses. The realistic near-term goal is to predict useful aspects of cell behavior well enough to prioritize experiments, while validating those predictions in real biological systems.
Why is the Department of Energy involved in biology?
DOE’s national laboratories operate some of the country’s most powerful scientific computing systems and major imaging, scattering, genomics and automation facilities. Those capabilities can generate and process the large, multi-modal datasets needed for predictive biological models.
What is the biggest scientific risk?
Generalization. A model may look strong on familiar datasets but fail when cell type, disease state, dose, time, lab protocol or intervention changes. Independent experimental validation and calibrated uncertainty will be essential.
What would count as a real breakthrough?
A model that repeatedly predicts previously unseen cellular responses better than strong baselines, identifies high-value experiments, survives replication in independent laboratories and measurably accelerates a discovery program without hiding uncertainty.
Sources and reporting notes
- Biohub, October 7, 2026 — expansion of the Virtual Biology Initiative and the $1.8 billion combined commitment.
- U.S. Department of Energy, October 7, 2026 — five-year DOE investment and national-laboratory capabilities.
- National Institutes of Health, October 7, 2026 — NIH role in data coordination and standardization.
- Biohub, April 29, 2026 — original $500 million five-year Virtual Biology Initiative commitment.
- Reuters, October 7, 2026 — independent reporting on the biological-data bottleneck, funding structure and access model.
- Nature Biotechnology, April 7, 2026 — editorial context on AI, mechanistic modeling and the limits of current “virtual cell” terminology.
- Frontiers in Cell and Developmental Biology, September 8, 2026 — review of validation, context and generalization requirements for trustworthy virtual-cell models.
Comments
Post a Comment