Biohub, the U.S. Department of Energy and the National Institutes of Health are joining technology and life-science organizations in a $1.8 billion effort to build the data infrastructure needed for predictive AI models of biology. The expanded Virtual Biology Initiative will combine biological datasets, advanced measurement technologies and high-performance computing to help researchers develop AI models capable of predicting how cells respond to drugs, genes and other interventions.
The next major AI breakthrough in biology may depend less on building a larger model and more on collecting the right data.
That is the premise behind an expanded international effort led by Biohub, the U.S. Department of Energy (DOE) and the National Institutes of Health (NIH), which has committed $1.8 billion in funding, data, computing resources and measurement technology toward AI-ready biological datasets.
The initiative is designed to give researchers the raw material needed to build predictive models of biology, including ambitious systems sometimes described as “virtual cells.” Instead of simply analyzing existing biological literature or predicting the structure of individual molecules, these models aim to simulate how living cells change in response to drugs, genetic modifications and other interventions.
The scale of the announcement is significant. DOE plans to invest more than $500 million over five years in cell research, measurement, modeling and computation. NIH will coordinate relevant datasets and knowledge bases supported by more than $500 million in previous federal investment. Meanwhile, Google DeepMind, Isomorphic Labs and Meta are collectively committing another $300 million to the Virtual Biology Initiative.
Biohub had already committed $500 million to launch the Virtual Biology Initiative in April 2026. That original commitment included $400 million for technologies capable of generating new biological measurements and $100 million for research outside Biohub.
The new funding expands the initiative from a major research program into something closer to biological AI infrastructure.
The problem it is attempting to solve is straightforward to describe but difficult to execute: AI models need enormous quantities of high-quality, standardized data, and biology has relatively little of it in the format required for training predictive models.
Protein structure prediction provides a useful comparison. Google DeepMind’s AlphaFold benefited from decades of structural biology research and databases containing experimentally determined protein structures. A predictive model of the cell faces a much larger information problem because cells are dynamic systems.
Researchers need to know not only what components exist inside a cell, but how those components interact, how cells change over time and how different interventions alter those processes.
The Virtual Biology Initiative therefore plans to generate multimodal and time-resolved datasets across more cell types and experimental conditions. The program will use technologies including cryo-electron tomography, advanced microscopy, molecular and cellular engineering, autonomous laboratories and large-scale computation.
DOE brings another critical component: computing infrastructure.
Through the Genesis Mission, the department says its contribution will draw on exascale computing, X-ray and neutron scattering, cryo-electron microscopy and tomography, and autonomous laboratories across the U.S. National Laboratory system.
That combination of laboratory automation and accelerated computing points toward a different kind of AI development cycle. Rather than training a model exclusively on historical datasets, researchers could generate experiments specifically to fill gaps in the model’s understanding, feed the results back into training and then use the improved model to select subsequent experiments.
In other words, the ambition is a closed loop between AI and physical biology.
Google DeepMind is already pursuing this broader AI-for-science strategy. Its research organization describes AI as a way to generate and test scientific hypotheses at substantially greater scale, while tools such as AlphaFold have demonstrated how machine learning can reshape biological research.
The commercial stakes are also growing. AI drug discovery has attracted billions of dollars as pharmaceutical companies and startups attempt to use machine learning for target identification, molecular design and biological prediction. A successful virtual-cell model could extend that trend upstream, allowing researchers to investigate disease mechanisms and treatment responses computationally before committing resources to physical experiments.
But a virtual cell would face challenges that differ from conventional generative AI. Biology is noisy, context-dependent and highly variable. A model that performs well on one cell type or experimental condition may not generalize to another. Training data can also contain measurement artifacts, missing information and biases toward heavily studied biological systems.
That makes the initiative’s emphasis on open standards and common identifiers particularly important.
Biohub says it is working with NIH to standardize datasets and build a unified layer through which different biological resources can work together. Existing projects and infrastructure, including Tabula Sapiens, OpenCell, Zebrahub, CELLxGENE and the CryoET Data Portal, provide pieces of that ecosystem.
The participating organizations also illustrate how the competitive landscape around AI for science is changing. Google DeepMind and Isomorphic Labs bring advanced AI and computational drug-discovery expertise, while Meta contributes funding and technology resources. NVIDIA is supporting the initiative with accelerated computing infrastructure, domain-specific software and technical expertise.
For pharmaceutical researchers, the eventual value could be substantial. A reliable predictive biological model could allow scientists to test hypotheses digitally, prioritize experiments and potentially identify ineffective approaches earlier in the development process.
Still, the initiative is not evidence that scientists can replace laboratory research with AI. The more realistic near-term outcome is an increasingly tight relationship between computational prediction and experimental validation.
That distinction is important because the value of a virtual cell will ultimately depend on whether its predictions hold up against real biological systems.
The broader AI industry is increasingly moving toward this model of scientific computing. Recent research describes AI for science as evolving from specialized analytical tools toward systems that can participate across literature research, data analysis, hypothesis generation and experimental execution.
Biohub’s $1.8 billion initiative is effectively an attempt to build the biological data layer for that future.
If the project succeeds, researchers could gain access to an open foundation for training models that do more than recognize patterns in biological data. They could begin asking AI systems to predict what will happen when biology is changed—and then test those predictions in the laboratory.
That would make the virtual cell less of a chatbot-like application and more of a new computational instrument for biomedical research.
Market Landscape
The AI-for-science market is shifting toward foundation models, autonomous laboratories and multimodal scientific datasets. Rather than treating AI as an analysis tool applied after an experiment, research organizations are increasingly integrating models into the full discovery cycle.
Biohub’s initiative is particularly notable because it addresses a bottleneck that can limit almost every biological AI application: training data.
The $1.8 billion commitment brings together philanthropic capital, federal research infrastructure, technology companies, pharmaceutical-AI expertise and scientific institutions. Google DeepMind, Isomorphic Labs and NVIDIA add commercial AI and computing capabilities, while DOE and NIH provide large-scale research infrastructure.
The competitive implications extend beyond drug discovery. Predictive biological models could eventually support genomics, cell biology, biotechnology, therapeutic development and personalized medicine.
However, the field remains experimental. Better models do not automatically produce better medicines. The difficult step will be validating whether predictions generalize across biological contexts and translate into experimentally and clinically useful outcomes.
Top Insights
- $1.8B commitment: Biohub, government agencies and technology companies are pooling funding, datasets, computing and measurement technology for AI-ready biology.
- Virtual cell ambition: The initiative aims to enable models that predict how cells respond to drugs, genetic changes and other biological interventions.
- Data is infrastructure: Standardized, multimodal and time-resolved biological measurements are becoming as important to AI biology as model architecture.
- AI meets laboratories: Autonomous experiments, advanced imaging and exascale computing could create a feedback loop between biological experimentation and AI prediction.
- Open science strategy: Shared standards, identifiers and public datasets could allow researchers worldwide to build models on a common biological data foundation.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
