Meta, Google and US government turn $1.8bn Biohub push toward AI models of human cells

Biohub’s Virtual Biology Initiative is expanding into a roughly $1.8 billion public-private effort to generate large biological datasets for artificial intelligence, bringing together the nonprofit founded by Mark Zuckerberg and Priscilla Chan with Meta Platforms, Inc., Google DeepMind, Isomorphic Labs and United States government agencies. Reuters reported that Meta Platforms, Google DeepMind and Isomorphic Labs are contributing approximately $300 million collectively, while the United States Department of Energy plans more than $500 million of support over five years and the National Institutes of Health will coordinate federally funded biological data resources. Biohub had already committed $500 million when the Virtual Biology Initiative was launched in April, meaning the $1.8 billion figure represents a combination of new commitments and existing programmes rather than a single newly announced financing round.

The initiative is attempting to solve a problem fundamentally different from the one that helped create today’s language models. Internet-scale text, images and software code already existed before modern artificial intelligence laboratories began training models on them, but comparable biological datasets often have to be generated experimentally inside laboratories at enormous cost. Biohub therefore wants to industrialise measurements of how cells behave across genetic, chemical and environmental conditions and make sufficiently standardised datasets available to researchers building predictive biological models.

The ambition is to create systems capable of forecasting how cells respond before researchers conduct every experiment physically. Reuters reported that the initiative expects its first major dataset within approximately a year and is targeting useful predictive models within five years, although those timelines are objectives rather than guaranteed scientific outcomes. If successful, the programme could improve target discovery, disease research and drug-development decisions, but it would not eliminate the need for laboratory validation or clinical trials.

Why does AI biology need more data rather than simply better models?

Large language models were able to improve dramatically because developers could train on huge collections of text and other digital information. Biology is much harder because the behaviour of a cell depends on genes, proteins, signalling pathways, environment, disease state and numerous interactions that cannot simply be scraped from websites. Researchers frequently lack comparable measurements across enough cells and experimental conditions to train models capable of predicting unseen biological behaviour reliably.

Biohub’s April plan devoted $400 million of its initial $500 million commitment specifically to large-scale data generation and measurement technology, while another $100 million was intended to coordinate a wider international effort. The organisation described the objective as creating multimodal datasets large enough to support predictive models of cells rather than merely cataloguing biological observations.

This creates a notable inversion of the standard AI economics. In language modelling, computing infrastructure has often been the most visible bottleneck because much of the raw training material already exists. In biology, additional GPUs are less useful if the experimental observations needed to teach the system have never been generated.

Why are Meta Platforms and Google interested in biological datasets?

Meta Platforms has spent heavily on open artificial intelligence research, while Alphabet Inc. already possesses one of the most important AI-biology franchises through Google DeepMind and Isomorphic Labs. DeepMind’s AlphaFold demonstrated that machine learning could solve a biological prediction problem considered extraordinarily difficult by using carefully structured scientific information at scale. Isomorphic Labs is attempting to extend related techniques toward drug discovery.

Participation in Biohub gives those organisations access to a much broader data-generation ecosystem than any technology company could efficiently build alone. Laboratories, government research facilities and scientific institutes possess measurement technologies and biological expertise that software companies typically lack. The technology participants bring modelling expertise and computing infrastructure, creating a division of labour between generating biological reality and learning from it.

Investors should nevertheless avoid assuming the $300 million contribution immediately produces a commercial drug pipeline for Meta or Alphabet. Biohub is structured around open scientific infrastructure, and the strategic value may initially appear through better research tools and shared datasets rather than directly attributable corporate revenue. The project is consequently better understood as infrastructure for the AI-biology ecosystem than as a conventional pharmaceutical collaboration.

Why is the United States government committing more than $500m through the Department of Energy?

The Department of Energy operates national laboratories containing some of the world’s most powerful scientific-computing resources and specialised research facilities. Modern biology increasingly overlaps with large-scale computing because imaging, genomics and molecular measurements generate enormous datasets whose analysis can require high-performance infrastructure. Connecting those capabilities with Biohub’s experimental programmes provides an opportunity to combine government science assets with private artificial intelligence expertise.

Reuters reported that the Department of Energy expects to contribute more than $500 million across five years for measurement, modelling and computing, while the National Institutes of Health will coordinate repositories and biological datasets supported through existing federal investment. This makes the initiative unusually broad because it spans philanthropy, technology companies and several government research systems rather than being controlled by one commercial sponsor.

Government participation also increases the likelihood that standardisation becomes an explicit objective. Biological datasets generated at different laboratories can use different protocols and metadata, making them difficult to combine even when all are scientifically valid. AI models benefit disproportionately from consistent data, so coordination may prove as important as the absolute amount of money being spent.

Could a predictive virtual cell genuinely accelerate drug discovery?

The potential is substantial because traditional drug development involves enormous numbers of experiments that eventually prove unproductive. A sufficiently accurate cell model could help researchers rank targets, predict responses or identify promising experimental conditions before committing time and laboratory resources. Even modest improvements in early-stage decision quality could be economically significant when late-stage clinical failures cost hundreds of millions of dollars.

The phrase “virtual cell” can nevertheless create unrealistic expectations if interpreted as a perfect digital replica of human biology. Cells behave differently depending on tissue, disease, genetics and environment, while many drug effects emerge through interactions across organs and entire biological systems. A useful model does not need to reproduce every aspect of human biology, but its predictions must generalise beyond the experiments used for training.

The initiative should therefore be measured through progressively harder benchmarks rather than promises to replace experiments. Researchers need to demonstrate that models predict previously unseen cellular responses accurately enough to influence real scientific decisions. The first major datasets will be important, but predictive performance against new experimental observations will ultimately determine whether the investment changes biotechnology research.

Why could open biological data alter competition in pharmaceuticals?

Large pharmaceutical companies possess enormous proprietary datasets accumulated through decades of experiments and clinical programmes. Smaller biotechnology companies and academic laboratories typically have much less information, creating a structural advantage for organisations with scale. Large publicly accessible datasets could narrow part of that gap by giving researchers a common foundation on which to train models.

That does not eliminate proprietary advantage because companies will still combine public information with internal experiments, clinical data and chemistry. The likely effect is to raise the baseline capability available to everyone while making unique private datasets more valuable at the margin. The dynamic resembles open-source software, where shared infrastructure can accelerate an entire industry while individual companies still build differentiated products above it.

The initiative may also create new specialist businesses around biological data generation, laboratory automation, imaging and model validation. If AI laboratories need orders of magnitude more experimental information, laboratory equipment and scientific-instrument companies could become indirect beneficiaries alongside cloud and semiconductor providers.

What are the largest risks surrounding the $1.8bn effort?

Data quality is the first. Generating enormous datasets provides little value if measurement errors, inconsistent protocols or batch effects cause models to learn artefacts rather than genuine biology. Standardisation and quality control therefore need to scale at the same pace as experimental volume.

The second challenge is scientific complexity. Biological systems contain causal relationships that may not be obvious even from enormous datasets, meaning scaling data alone may not reproduce the dramatic improvement curve seen in language models. Researchers could discover that predictive biology requires new model architectures, mechanistic knowledge or experimental techniques rather than simply larger training sets.

Governance is another issue because biological models can have dual-use implications. Systems that understand how cells respond to genetic or chemical manipulation may assist medicine while also raising biosecurity questions, requiring careful controls around particularly sensitive capabilities. The involvement of major government and technology institutions suggests those issues will become central as model performance improves.

The $1.8 billion Virtual Biology Initiative is therefore best understood as an infrastructure experiment. Artificial intelligence transformed digital fields partly because the internet had already created oceans of machine-readable information, while biology lacks an equivalent resource. Biohub and its partners are effectively spending billions to manufacture that missing training corpus, and whether they succeed could determine how quickly AI moves from describing biology to genuinely predicting it.