On September 24, 2026, NVIDIA announced on its official blog the free release of predicted 3D structures for the protein complexes of more than 2,800 viruses, produced together with Google DeepMind, the European Molecular Biology Laboratory's European Bioinformatics Institute (EMBL-EBI), and other global research organizations. Scientists anywhere in the world can access the data for free through the AlphaFold Database. The timing was deliberate: it coincides with pandemic-preparedness meetings held in New York during the United Nations General Assembly week, convened with the World Economic Forum.
The structures were predicted with AlphaFold2 — Google DeepMind's AI model for predicting how proteins fold into 3D shapes — further optimized with the NVIDIA BioNeMo Inference Runtime so thousands of viral proteomes could be inferred in bulk on GPUs. The team systematically worked through the viral families known to infect humans, from common-cold viruses to emerging threats like Mpox. What was predicted is not single proteins, but protein complexes: the "team structures" of multiple proteins working together — precisely the targets vaccines and drugs most need to aim at.
Beyond the volume, the "newness" is what deserves attention. Roughly 30% of the protein interactions are being documented by science for the first time, with no corresponding entries in the PDB, the main repository of experimentally determined structures. Chris Dallago, applied research science team lead in digital biology at NVIDIA, described the database as a "hypothesis generation engine": it lets biologists and AI researchers study not single molecules but entire interaction networks, moving the whole field forward.
It's not just the data that's open. NVIDIA also open-sourced on GitHub the BioNeMo Structure Prediction Pipeline used to generate the dataset, so researchers can run the full workflow — from protein sequence to predicted structure — on their own targets. Jo McEntyre, interim director of EMBL-EBI, noted that the dataset covers many little-studied viruses and lowers the barrier for frontline scientists in low-resource settings — the people who confront outbreaks first often have the least data.
Why does this deserve a serious look? COVID-19's lesson was that decades of accumulated knowledge about coronavirus proteins let vaccine development move at record speed. But the next pandemic may well not "follow the script." Joe Grove, professor of molecular virology at the University of Glasgow and a collaborator on the project, put it bluntly: what we're trying to do is stockpile that knowledge ahead of time. An analysis by the Center for Global Development estimates roughly a 50% chance that the world faces a pandemic as severe as COVID-19 by 2050.
Running opposite to the closed-model race in the AI world, NVIDIA is playing an "open science as productive force" hand here: no API sales, no locked ecosystem — data and toolchains laid open so labs worldwide start from the same line. Nor is the move pure philanthropy: the more the BioNeMo ecosystem thrives, the firmer GPUs' position in the life sciences becomes. It won't directly become a new drug, but it may determine how much "advance study material" humanity has when the next outbreak arrives. In the AI-for-Science narrative, this is rare tangible progress — downloadable today, usable tomorrow.
