How Open Science Can Help Researchers Prepare for the Next Pandemic

When COVID-19 broke out, scientists had a card up their sleeve: decades of coronavirus research had given them enough understanding of key proteins to design vaccines in record time. But the next pandemic may not grant the same head start. NVIDIA, together with Google DeepMind, EMBL-EBI and other institutions, has openly released predicted three-dimensional structures of over 2,800 viral protein complexes, accessible to any scientist through the AlphaFold database.

Joe Grove of the MRC-University of Glasgow Centre for Virus Research, a project collaborator, put it plainly: the next pandemic may involve a completely unfamiliar pathogen, and there will be no COVID-like stock of knowledge to draw on, so what needs to be done now is to stockpile knowledge in advance.

These structures are not experimentally determined; they were inferred by Google DeepMind's AlphaFold2 model and optimized using the NVIDIA BioNeMo Inference Runtime, thereby extending inference to thousands of viral proteomes and predicting the complexes encoded by each virus—that is, a set of interacting proteins.

NVIDIA has also open-sourced the BioNeMo Structure Prediction Pipeline, the GPU-accelerated workflow that generated this dataset, enabling researchers to predict the three-dimensional structures of their own targets starting from protein sequences. The significance of this step is that the dataset itself is a finished product, while the pipeline is a tool, and the latter determines whether this approach can be reused by outside teams on new targets.

Most proteins do not work alone; they assemble into multi-molecular complexes to carry out complex functions, and these structures are often precisely what vaccines or drugs need to target in order to disrupt viral function. Understanding the three-dimensional structure of the COVID-19 virus spike protein proved to be the foundation of vaccine design; for thousands of other viruses, such structural knowledge does not exist today.

The dataset annotates predictions by confidence, showing what viral complexes might look like and how individual proteins might interact within a viral proteome. Chris Dallago, who leads NVIDIA's digital biology applied research science team, calls it a "hypothesis generation engine," enabling biologists and the AI community to study protein interactions as complexes rather than as single molecules.

The publisher states that about 30% of the protein interactions added to the database are completely unknown to the scientific community, with interaction forms never recorded in the Protein Data Bank. The Protein Data Bank is the primary repository for experimentally determined protein structures, which means this portion is not a reorganization of existing knowledge but new leads for the biology community to explore.

This proportion is given by the publisher and has not been independently verified. It measures "newness," not "correctness"—whether the new interactions can actually translate into progress in diagnostics, drugs or vaccines is not supported by validation results in the source.

Traditional protein structure determination methods—crystallizing a protein and bombarding it with X-rays—can take years and cost thousands of dollars per structure; AlphaFold2 predicts a structure in minutes and can run at scale. This order-of-magnitude difference is precisely the precondition for extending structure prediction to thousands of viral proteomes.

The project systematically processed protein structures of viral families known to infect humans, covering everything from common cold viruses to emerging threats such as Mpox. The dataset contributes to the AlphaFold database, which currently contains over 260 million protein and protein complex predictions, covering nearly all known catalogued proteins.

Collaborators include the Coalition for Epidemic Preparedness Innovations (CEPI), EMBL-EBI, Google DeepMind, NVIDIA, Seoul National University, Sungkyunkwan University, the Swiss Institute of Bioinformatics and the University of Glasgow. The dataset release coincided with a World Economic Forum meeting on pandemic prevention, preparedness and response held in New York during the United Nations General Assembly.

Jo McEntyre, interim director of EMBL-EBI, said the dataset also covers less-studied viruses and lowers the barrier for scientists in low-resource regions on the front lines of outbreaks to access data. This accessibility benefit is currently a qualitative statement; the source provides no quantification.

First, the dataset provides predicted rather than experimentally determined structures. Scientists can validate high-confidence predicted structures through experimental methods, but validation itself requires additional work, and the confidence annotations also remind users that the structures show what things "might look like."

Second, all the above information comes from a press release on NVIDIA's official blog, which is the publisher's own account, and no independent third party has verified the scale or quality of the dataset. The figure of about 30% entirely new interactions and the estimate of about a 50% pandemic probability—the former given by the publisher, the latter from an analysis by the Center for Global Development—are both external or one-sided estimates, not measured conclusions of this project.

Third, the actual adoption and effectiveness of the open-sourced BioNeMo Structure Prediction Pipeline are unknown; the source also provides no validation results on the effectiveness of AlphaFold2-predicted structures in actual vaccine or drug development. The release coincided in time with the related United Nations General Assembly meeting, and the source does not explain its specific policy impact.