AI biology is entering a phase where better models are no longer the only scarce resource. In its dated October 7, 2026 announcement, Biohub described an expanded $1.8 billion effort involving federal agencies, Google DeepMind, Isomorphic Labs, Meta and other partners. That total combines funding, data, computation and measurement technology; it is not all newly committed cash. Biohub originally committed $500 million in April to build the measurement technology, datasets and infrastructure needed for predictive models of cells. The practical takeaway is bigger than the funding number: in biology, the next capability jump may depend as much on what can be measured in real experiments as on how large a model can become.
That changes how teams should evaluate “virtual cell” systems. A model can sound sophisticated and still be trained on narrow, observational data that says little about what happens when a drug, gene edit or environmental change is introduced. The useful question is not whether a model has seen millions or billions of cells. It is whether the underlying data captures the right interventions, biological contexts and validation needed for the decision you want to make.
What a virtual-cell model is trying to predict
Biohub describes virtual biology as the attempt to build high-fidelity digital representations of molecules, genomes, cells and living systems so scientists can ask some questions computationally before moving into the lab. Its Virtual Biology Initiative is explicitly built around creating multi-modal datasets for predictive models of the human cell, not around replacing laboratory science.
That distinction matters. A virtual-cell model is not automatically a digital twin of a patient, and a prediction is not a clinical result. At its best, the model can help rank hypotheses: which perturbation may change a cell state, which target may deserve an experiment, or which molecular interaction is worth measuring. The laboratory is still where the prediction has to survive contact with biology.
NextWith.ai’s recent coverage of AlphaFold’s viral protein-complex expansion shows the same boundary in a narrower domain. Predicted structures can help researchers decide what to investigate, but they do not by themselves establish viral behavior, treatment efficacy or clinical relevance. Virtual-cell models widen the prediction problem from structure toward behavior, which increases both the potential value and the number of ways a model can be wrong.
Why biology needs different data from a general AI model
Text models can learn from enormous collections of naturally occurring language. Biology does not offer an equivalent ready-made corpus for many of the questions researchers care about. A cell’s behavior depends on cell type, genetic background, tissue context, developmental state, disease state, time, dose and interactions with other cells and molecules. A dataset can be huge while still missing the variable that matters.
Biohub describes a need for biological data at unprecedented scale and is investing in imaging, molecular measurements, cellular engineering and tissue-level observation. Its AI models and datasets program describes protein-model research and a protein atlas. That is the core data problem: a predictive model needs examples that cover not only what cells look like, but how they respond when something changes.
Perturbation data is especially important. In January 2026, Biohub, Arc Institute and Tahoe Therapeutics announced a project to generate data from more than 120 million cells across 225,000 drug-patient perturbation interactions. The partners described the effort as a way to address the data bottleneck for virtual-cell models. Observational data can reveal patterns; perturbation data gives models a better chance to learn how systems respond to interventions.
Five checks before you trust an AI biology dataset
1) Match the measurement to the question. Transcriptomics measures RNA activity, proteomics measures proteins, imaging captures structure and spatial context, and other assays can record chromatin state, molecular interactions or functional responses. No single modality is “the biology.” If your task is to predict a treatment response, a dataset containing only static gene-expression snapshots may be insufficient even if it is very large.
2) Look for interventions, not only observations. Ask how many examples involve controlled perturbations such as drugs, gene knockouts, dosage changes or environmental conditions. A model trained mostly on naturally occurring states may be good at describing what is present while remaining weak at predicting what will happen after an intervention.
3) Check diversity across biological contexts. Cell type, tissue, disease state, ancestry, age, experimental platform and laboratory conditions can all shift the distribution a model sees. The relevant metric is not simply “number of cells.” It is whether the dataset covers the populations and conditions where you expect to use the model.
4) Separate data scale from data quality. Larger datasets amplify systematic errors as efficiently as they amplify signal. Look for replicate experiments, quality-control rules, batch-effect handling, metadata completeness and evidence that measurements from different labs or instruments can be compared. If those details are missing, a large number can create false confidence.
5) Verify the access terms and release path. Biohub describes a planned open research resource. An announcement of future openness does not establish that every dataset, model or access term is already available. For research planning, check when the actual files, metadata, benchmarks and usage terms become available rather than relying on the word open in an announcement.
How to evaluate the model built on top
A strong dataset does not guarantee a useful model. The next layer is task-specific validation. Ask what the model is predicting, how far its test data is separated from its training data, and whether the benchmark resembles the biological decision you care about. Predicting a held-out gene-expression profile is different from selecting a drug target that works in an animal, and both are different from a therapy that succeeds in people.
The model card discipline used for general-purpose AI still helps here. NextWith.ai’s model-card checklist recommends identifying the exact model version, intended use, limitations, training/evaluation evidence and licensing before adoption. For AI biology, add three questions: what experimental data generated the labels, what perturbations were held out for testing, and which predictions have been confirmed in the lab.
Independent replication is the strongest signal. A vendor can report a benchmark improvement; a research group can report that its model predicts an intervention; neither tells you whether another lab can reproduce the result under different conditions. The more consequential the decision, the more weight you should place on prospective experiments rather than retrospective benchmark scores.
Where government infrastructure fits
The data push is no longer only a private-sector project. Biohub’s October announcement says DOE will contribute more than $500 million over five years and NIH will coordinate datasets stemming from more than $500 million in prior federal investment. These are different forms of support, and the NIH figure should not be counted as a new cash grant. That matters because large biological datasets require more than GPUs: they need instruments, laboratories, standards, storage, metadata systems and long-lived coordination across institutions.
The easy-to-miss consequence is that AI biology may develop more like a scientific infrastructure market than a normal software market. Model developers will compete, but so will measurement platforms, data consortia, compute providers and organizations that can run high-throughput experiments. The valuable asset may be a feedback loop between model predictions and new measurements, not a single model checkpoint.
What happens next
Biohub’s original initiative is structured as a five-year effort, so the most important results will arrive as datasets, measurement technologies and validated models rather than as one launch event. The key milestones to watch are concrete: how much data is actually released, which modalities and perturbations it covers, whether metadata is standardized, what access restrictions apply, and whether models trained on it make predictions that hold up in prospective experiments.
For companies evaluating AI biology today, the safest approach is to treat dataset quality as part of model due diligence. Ask for the training-data scope, perturbation coverage, biological diversity, validation design and release terms before comparing benchmark scores. A virtual-cell model is useful when it helps you choose better experiments; it becomes risky when a polished prediction is mistaken for experimental evidence.
Before adopting or funding a virtual-cell system, ask for one concrete example where a prediction was tested prospectively in the lab, then trace that result back to the dataset, perturbation coverage and validation protocol that made the prediction possible.