Viral Protein Structure Dataset Maps 2,800 Viruses With BioNeMo

A structural map drawn before the outbreak

When SARS-CoV-2 appeared, vaccine designers did not start from scratch. They had decades of coronavirus research behind them, including detailed knowledge of the spike protein. A coalition led by NVIDIA, Google DeepMind and EMBL's European Bioinformatics Institute is now trying to give future outbreak responders a similar head start for viruses that have never had that kind of attention.

On September 24, the group published predicted 3D structures for the protein complexes of more than 2,800 viruses. The data is freely available through the AlphaFold Database.1 The predictions were made with AlphaFold2, sped up by NVIDIA's BioNeMo Inference Runtime, so the team could work through thousands of viral proteomes and predict which proteins in each virus bind to one another.1 NVIDIA also released the GPU-accelerated workflow behind the dataset as open source. It is called the BioNeMo Structure Prediction Pipeline, and it lets researchers go from a protein sequence to a predicted structure for their own targets.1

The timing was deliberate. The release coincided with a United Nations General Assembly meeting on pandemic prevention, preparedness and response, convened by the World Economic Forum in New York.1 NVIDIA's announcement cites a Center for Global Development estimate that the world has roughly a 50% chance of facing a pandemic as severe as COVID-19 by 2050.12

What is actually in the dataset

The headline figure of "2,800 viruses" needs some unpacking. One detailed account, which cites the AlphaFold Database, says the release covers 2,812 viral proteomes across 23 virus families relevant to human health.3 By that count, the project produced 5,279 high-confidence heterodimers, meaning pairs of two different interacting proteins, and 2,749 high-confidence homodimers, meaning two copies of the same protein bound together.3 The families range from common-cold viruses to emerging threats such as Mpox.1

The new material sits inside a Pandemic Preparedness Portal on the AlphaFold Database. The portal combines the new viral complexes with viral predictions the database already held.3 Including this release, the AlphaFold Database now contains more than 260 million protein and protein-complex predictions.1

The claim that has drawn the most attention is novelty. According to NVIDIA, about 30% of the protein interactions being added have no counterpart in the Protein Data Bank, the main archive of structures determined by experiment.1 Every outlet covering the story repeated this figure. But they framed it differently, and the difference matters. NVIDIA's own post calls these interactions "completely new to science".1 Some secondary coverage went further and described the dataset as delivering "high-resolution" blueprints.9 More careful reports stressed that a missing PDB entry means only that no one has solved the structure experimentally. It does not mean the AI predictions have been confirmed.37

The careful framing is the right one. These are hypotheses, labeled with confidence scores.1 NVIDIA's Chris Dallago described the database as "an engine for hypothesis generation".1 That phrase is more accurate than some of the coverage built on top of it.

Why complexes, not single proteins

The most important scientific choice in this release is its focus on complexes. The original 2021 AlphaFold Database was mostly about single protein chains. This dataset predicts how viral proteins pair up and interact. NVIDIA's argument is that most proteins work in multi-molecule groups, and that these assemblies are often what a vaccine or drug has to disrupt.1 One report adds that many drugs, antibodies and vaccines work by interfering with specific molecular interactions. That makes interaction-level structural data especially useful in infectious disease research.3

For AI drug discovery, the implication is simple. Target selection usually comes before designing small molecules or biologics, and a target is much easier to reason about once its likely binding interface is known. A high-confidence predicted heterodimer interface is not a drug target in itself. But it narrows the search, so labs can prioritize which interactions to validate rather than surveying an entire proteome blind.3 Several outlets made the same practical point: the value lies in shrinking the number of structures scientists must investigate from scratch, not in replacing crystallography.237

Joe Grove, a molecular virologist at the MRC-University of Glasgow Centre for Virus Research who worked on the project, described the goal as stockpiling knowledge ahead of time for a pathogen that "comes out of the blue".2 He recalled that during his PhD there were no structures at all for the proteins his group studied.1

The pipeline matters more than the press release suggests

The dataset got the headlines. For practitioners, the open pipeline may be the more lasting contribution. The project's GitHub repository describes three containerized stages that run on a Slurm cluster with NVIDIA GPUs. First, the pipeline builds multiple sequence alignments using GPU-accelerated MMseqs2 search via ColabFold. Second, it predicts structures with OpenFold2 running AlphaFold2-Multimer weights, accelerated by the BioNeMo Inference Runtime. Third, it scores complex interfaces with metrics such as ipSAE and pDockQ2, checks for steric clashes, and exports files in standard formats.

This is where the reporting is least consistent. NVIDIA's blog and most coverage simply say AlphaFold2 was used.12 One outlet names both AlphaFold2 and AlphaFold-Multimer.3 The repository is more specific: OpenFold2 with AlphaFold2-Multimer weights is the main folding engine, and standard OpenFold2 or ColabFold are optional backends. In other words, the model weights trace back to DeepMind, while the inference code shipped in the pipeline is OpenFold-based. One article printed a neat one-line bionemo-inference command as an example.4 That does not match the repository's documented setup, which involves cluster jobs, large local alignment databases and object storage. Readers should take the repository as the authoritative description.

The repository is also frank about what it costs to run. Users must supply their own sequence databases and model weights. The UniRef30 search index alone takes up about 240 GB, and fast NVMe storage is recommended. It ships with a 1,000-target benchmark drawn from PDB structures released between 2022 and 2025, made up of 750 monomers, 150 dimers, 50 trimers and 50 tetramers. Because it uses recent structures, the benchmark gives users a way to check prediction quality on proteins the models are less likely to have seen during training.

The repository also shows that this release is part of a larger effort. The same pipeline was used to produce a separate proteome-scale dataset: about 1.8 million high-confidence complexes, selected from roughly 31 million predictions across 4,777 proteomes. The viral release is best understood as a focused, policy-friendly slice of a much bigger structural run.

The "open science" claim, and NVIDIA's position

The partner list goes well beyond two tech giants. It includes the Coalition for Epidemic Preparedness Innovations, Seoul National University, Sungkyunkwan University, the Swiss Institute of Bioinformatics and the University of Glasgow.3 EMBL-EBI's Jo McEntyre argued that open access especially helps scientists in low-resource settings who deal with outbreaks first-hand, and that the dataset deliberately covers lesser-studied viruses.1 Google DeepMind's Risha Patel framed it as part of the AlphaFold Database's long-standing goal of making foundational biology widely accessible.8

NVIDIA's motives are worth stating plainly. This kind of release is open science, and it is also a showcase for GPU-accelerated biology. NVIDIA's Anthony Costa said the company's researchers benchmarked and ran large-scale predictions to turn broad viral coverage into a usable public resource. The company has been building BioNeMo's reputation for years. A 2024 Scientific Data paper co-written by NVIDIA staff used BioNeMo-hosted AlphaFold2, OpenFold and ESMFold to model almost all of a curated human reference proteome. Its authors pointed to structure-based drug design and protein-design tools such as RFdiffusion as downstream uses. TechRepublic's coverage also links BioNeMo to NVIDIA's work with Lilly and Thermo Fisher in drug discovery.2

This does not reduce the value of the data. But it helps explain the design choices. An open pipeline that runs best on NVIDIA hardware spreads the method and the platform together.

The verdict

The coverage agrees on the essentials: the scale, the 30% novelty figure, the coalition behind it, and the open pipeline.1237 It differs mainly in tone. The better reports treat the dataset as a well-organized set of educated guesses. Others drift toward calling it a finished map of viral weak points.79

The careful reading is the correct one, and it is still a significant development. Predicted interfaces for thousands of under-studied viruses, labeled by confidence and freely downloadable, change where a research team can start when a new pathogen emerges. Because the pipeline is reproducible and comes with a benchmark, other groups can test and extend the predictions instead of just accepting them. The real test is still to come. The question is how many of these predicted complexes survive experimental validation, and whether any lead to diagnostics, antibodies or antivirals before the next outbreak arrives.