Biohub is putting serious money and computing power behind a difficult question in artificial intelligence and science: can AI learn enough about living systems to predict what happens inside a cell?
The organisation has expanded its Virtual Biology Initiative into a collaboration involving the U.S. Department of Energy (DOE), National Institutes of Health (NIH), Google DeepMind, Isomorphic Labs and Meta. Together, the effort brings funding, biological data, computing infrastructure and experimental technologies into one large programme.
The combined commitment reaches $1.8 billion, with the partners aiming to build the data and scientific infrastructure needed for predictive models of biology.
The idea is ambitious. Instead of using AI only to analyse biological information that already exists, researchers want models that can help predict how living systems respond when scientists change something.
Biohub Is Building the Data Foundation for AI-Powered Biology
The expansion of the Virtual Biology Initiative addresses one of the biggest limitations facing AI research in biology: models need much more high-quality experimental data if they are going to make reliable predictions.
Biohub says the initiative will combine biological measurements from multiple organisations and research programmes into an open resource that can be used to develop predictive models. The effort will bring together data covering different cell types, biological conditions and interventions, creating a much broader foundation than any individual research organisation could realistically build alone.
That scale is important because biology does not behave like a simple dataset. A change that produces one result in one cell type can produce something completely different elsewhere.
A Virtual Cell Is the Long-Term Goal Behind the Initiative
At the centre of the project is the idea of creating a virtual cell capable of predicting biological behaviour.
Such a system would not simply identify patterns in existing research. It would ideally allow scientists to explore what might happen when a cell, molecule or biological pathway is altered. Researchers could then use those predictions to decide which experiments deserve further investigation in the laboratory.
Biohub Head of Science Alex Rives has described the creation of a virtual cell as one of the major scientific challenges ahead. The organisation sees predictive biology as a possible way to study disease and biological processes with far greater speed and scale.
That capability does not exist yet. The current initiative is focused on building the experimental data and infrastructure required to make it possible.
Google DeepMind, Isomorphic Labs and Meta Are Bringing AI Expertise Into the Project
The participation of Google DeepMind, Isomorphic Labs and Meta gives the initiative access to organisations already working at the frontier of artificial intelligence.
The three companies are collectively contributing $300 million to develop technologies and multimodal datasets for predictive models of biology. Their involvement adds AI model development and large-scale computing expertise to Biohub’s biological research capabilities.
Isomorphic Labs, which focuses on AI-powered drug discovery, has argued that predictive systems biology requires more data than any single organisation can generate. That makes collaboration particularly important for a project attempting to model complex biological systems rather than one narrow research problem.
The US Department of Energy Will Add Supercomputing and Experimental Infrastructure
The U.S. Department of Energy is contributing more than $500 million over five years through its Genesis Mission, bringing national laboratory capabilities into the Virtual Biology Initiative.
Those resources include exascale computing, advanced imaging and measurement technologies, X-ray and neutron scattering, cryo-electron microscopy and tomography, as well as autonomous laboratory systems. Together, they can produce different types of information about biological structures and processes.
This matters because predictive AI models require more than static images or isolated measurements. They need information about biological systems under different conditions and, crucially, how those systems change over time.
NIH Will Turn Existing Biomedical Resources Into AI-Ready Data
The National Institutes of Health will contribute another major piece of the project by bringing together biomedical datasets and research infrastructure developed through years of federal investment.
NIH resources include biological repositories and databases operated through the National Library of Medicine and the National Center for Biotechnology Information, along with programmes focused on mapping cells and building shared biological data resources.
Biohub will work with NIH to standardise relevant datasets so researchers can use them more effectively for AI model training. That work may not attract the same attention as a new AI model, but it is critical. Data collected by different laboratories often follows different formats, standards and measurement methods.
Without common standards, even huge collections of biological information can remain difficult for AI systems to use effectively.
The Initiative Will Generate New Data From More Biological Conditions
Existing biological datasets leave significant gaps, particularly when researchers want to understand how cells respond to different interventions.
The expanded initiative aims to generate new measurements across more cell types and conditions. Biohub’s original commitment includes $400 million for measurement technologies, with research covering areas such as cryo-electron tomography, high-scale microscopy and tools for engineering and manipulating biological systems.
The purpose is to create richer datasets showing not only what biological systems look like, but how they respond to changes.
That distinction could ultimately determine whether a virtual biology model can make useful predictions or simply become another sophisticated system for analysing historical data.
An Open Biological Data Resource Will Connect Research From Multiple Organisations
Biohub is positioning the initiative as an open resource rather than a closed project controlled by a single company.
The partners are working toward shared standards, common identifiers and a unified access point for datasets generated through the programme. The initiative also brings together research organisations including the Allen Institute, Broad Institute, Gladstone Institutes, Human Cell Atlas, Human Protein Atlas and Wellcome Sanger Institute.
NVIDIA is contributing accelerated computing infrastructure, software and technical expertise, while Renaissance Philanthropy is supporting additional funding for data generation.
The open approach could prove important because predictive biology will require contributions from many fields. No single laboratory is likely to produce all the measurements needed to model an entire living system.
Biohub Is Extending a Research Infrastructure It Has Already Built
The Virtual Biology Initiative builds on years of Biohub work in biological data and measurement.
Projects including Tabula Sapiens, OpenCell and Zebrahub, together with resources such as CELLxGENE and the CryoET Data Portal, have already helped create large-scale datasets and tools for biological research.
The new programme takes that approach into a much bigger arena. Instead of focusing on one dataset or biological question, Biohub is trying to establish a foundation that AI researchers can use to build models across different levels of biology.
The difference is significant. A predictive model of a single protein is one thing. A system that attempts to model how cells respond to interventions is a much larger scientific problem.
AI in Biology Is Moving From Pattern Recognition Toward Prediction
AI has already made major advances in areas such as protein structure prediction, biomedical imaging and drug discovery. The Virtual Biology Initiative is aimed at the next step: predicting how biological systems behave when conditions change.
That requires models to understand relationships between biological structures, cellular states and interventions rather than simply recognising patterns in existing datasets.
Google DeepMind’s Pushmeet Kohli has emphasised the need for experimental data showing how living cells behave and respond to changes. A large shared biological data resource could provide the foundation for models capable of making those predictions.
But there is a major distinction between predicting something that resembles the training data and predicting a biological response that has never been observed. The latter is where these systems will ultimately be tested.
A Working Virtual Cell Could Change How Scientists Run Experiments
If predictive biological models become reliable enough, researchers could eventually use them to explore thousands of possible experiments digitally before selecting a smaller number for laboratory testing.
That could reduce the amount of time spent testing unlikely candidates and help researchers identify promising interventions earlier. Drug discovery, disease research and biotechnology could all benefit if models can accurately predict how biological systems respond.
The Virtual Biology Initiative is not claiming that this future has already arrived. Its current focus is much more basic: generate better data, improve biological measurements, connect research resources and provide the computing infrastructure needed to train increasingly capable models.
That may sound less dramatic than announcing a fully functioning virtual cell. It is also the part of the problem that has to come first.
Sources
- Chan Zuckerberg Biohub — Virtual Biology Initiative Expansion
https://biohub.org/news/virtual-biology-initiative-expansion/ - U.S. Department of Energy — Office of Science
https://science.osti.gov/ - National Institutes of Health
https://www.nih.gov/ - Isomorphic Labs
https://www.isomorphiclabs.com/ - Google DeepMind
https://deepmind.google/
