AlphaFold Database Adds Protein Complexes From 2,800 Viruses

A new viral layer for the world's protein atlas

The AlphaFold Protein Structure Database, which covers almost every catalogued protein, now includes a section built for the next outbreak. On September 24, 2026, a group led by EMBL's European Bioinformatics Institute (EMBL-EBI), Google DeepMind and NVIDIA released AI-predicted protein complex structures for more than 2,800 viruses. The data is free to use and focuses on viral families known to infect humans.1 The partners also include Seoul National University, the University of Glasgow, the Swiss Institute of Bioinformatics, the Coalition for Epidemic Preparedness Innovations (CEPI) and Sungkyunkwan University. Lund University in Sweden contributed complexes it had computed separately.15

The timing was deliberate. The release came the day before the UN General Assembly's High-level Meeting on Pandemic Prevention, Preparedness and Response on September 25, which was set to review the preparedness commitments made in 2023.1 It was first announced at a roundtable hosted by CEPI and the World Economic Forum on how AI could speed up diagnostics, therapeutics and vaccines.5

The story is usually told as "AI meets pandemic preparedness." The more useful story is about open infrastructure. A shared, freely available library of predicted structures, combined with open tools and open benchmarks, is quietly becoming a major part of how drug and vaccine discovery will respond to a new pathogen.

What was actually released

The reports agree on the main facts but differ in detail, and the detail matters. NVIDIA describes the structures as predictions from AlphaFold2, sped up with its BioNeMo Inference Runtime. That optimisation let the team run predictions across thousands of viral proteomes and model complexes of interacting proteins, not just single chains.11 NVIDIA is also open-sourcing the GPU workflow it used, the BioNeMo Structure Prediction Pipeline, so researchers can take their own sequences and produce predicted 3D structures.11

The preprint behind the release gives exact numbers. The team used AlphaFold2 and AlphaFold-Multimer to model single proteins and pairs of proteins (dimers) across 2,812 viral proteomes from 23 virus families relevant to human health. They reported 5,279 high-confidence pairs of different proteins and 2,749 high-confidence pairs of identical ones.8 Grouping these by the shape of the contact surface between the two proteins reduced them to 1,598 clusters. Of those, 471, about 29.5 percent, looked nothing like any experimentally solved interface in the Protein Data Bank.8 The authors point to two early results: new information about glycoproteins that matter for vaccine design, and accurate predictions of how viral proteases recognise and cut their targets. Those proteases are classic targets for antiviral drugs.8

Nature's coverage shows how strict the quality filter was. The team predicted 40,746 identical pairs and nearly 1.7 million mixed pairs, but only about 8,000 in total met the bar for the curated database. The full set of predictions is still available to the public.6 At least one outlet described the update as adding roughly 1.7 to 1.8 million "high-confidence" complex predictions and dated it to March 2026.7 That conflicts with the primary accounts. The more defensible reading is that about 1.7 million pairs were modelled, a small fraction passed the confidence threshold, and the release took place in September.68 In practice, the database is a carefully filtered core sitting on top of a much larger, rougher set of predictions available for bulk download.

Why structure-first preparedness matters

The case for the project rests on one lesson from COVID-19. Knowing the 3D structure of the SARS-CoV-2 spike protein was central to vaccine design, and most other viruses have no comparable structural picture.11 The organisers link the release to the 100 Days Mission, which aims to have safe and effective vaccines and treatments ready within 100 days of identifying a pandemic threat. Meeting that goal depends on having structural knowledge in hand before a virus becomes a global problem.1 The project also used the UK Health Security Agency's priority pathogen tool to decide which families to cover, from common-cold picornaviruses to mpox.1

The partners frame the risk with a Center for Global Development estimate: roughly a 50 percent chance of a pandemic as severe as COVID-19 by 2050.11 They make an equity argument too. EMBL-EBI's interim director Jo McEntyre said the dataset covers neglected viruses and lowers barriers for scientists in low-resource settings who often face outbreaks first.1 The Swiss Institute of Bioinformatics noted that many viruses can only be studied computationally.2

That equity point lines up with a 2026 review of AI-driven protein design for vaccines and antibodies. It found that during COVID-19, regions with little sequencing or lab capacity gained much less from AI advances, because success depended on good data and experimental throughput as well as algorithms.12 Open structure data addresses part of that gap. It does not supply the wet labs.

The limits are stated plainly, and they are real

What stands out across the coverage is how much the partners themselves stress the limits. The predictions show what viral complexes may look like. They do not show how mutations change a virus or how a virus interacts with its host.1 Glasgow virologist Joe Grove put it plainly: a complex structure does not reveal what happens when a virus mutates, and it does not make dangerous pathogens easier to engineer.4

The technical gaps are significant. The models leave out the sugar molecules that coat many viral proteins and help them evade the immune system.6 Many key viral machines also involve more than two proteins. The SARS-CoV-2 spike and HIV's envelope protein are each built from three identical units, and in many such cases the two-protein predictions were not accurate enough to include.6 Confidence labels are attached to each prediction, but experimental validation is still essential.119

This caution also works as biosecurity messaging. By saying clearly that the data reveals structure rather than how dangerous a virus is, the partners are getting ahead of the obvious dual-use concern. That argument holds up, given what static, sugar-free, two-protein models can and cannot show.

The wider open ecosystem around the database

A structure database becomes useful when someone can build on it, and the open tools for doing that have grown quickly. OpenFold3, launched in October 2025, is a Columbia-led open reconstruction of AlphaFold3 that predicts how proteins bind nucleic acids and drug-like molecules, and it can be used commercially.20 MIT's BoltzGen, released the same month, builds on the open Boltz-2 model to design new protein binders from scratch. Its developers argue that fully open releases allow community-wide progress in drug design.14 In May 2026, Biohub, the Chan Zuckerberg-backed research organisation, released open protein models based on ESM and reported designed binders that reactivated immune cells in lab tests.16 Trade coverage noted that Biohub gave no quantitative performance data and no comparisons with existing platforms.13

The design side is moving too. POSTECH researchers and David Baker's lab used RFdiffusion to design a single protein that assembles into virus-like shells 70 to 220 nanometres across, confirmed by cryo-electron microscopy. They see these shells as possible platforms for displaying vaccine antigens.18 The COVID-era record suggests such work can reach the clinic. The same review highlights the computationally designed GBP510 nanoparticle vaccine and the miniprotein binder LCB1. In LCB1's case, cryo-EM showed it bound its target exactly as the model predicted.12

Open benchmarks show what AI can and cannot do yet

The best evidence on whether open AI actually speeds up antiviral drug discovery comes from blind competitions. The NIH-funded ASAP Discovery consortium, which takes an open-science approach to antivirals, shared unpublished data from its dual SARS-CoV-2 and MERS-CoV main protease (Mpro) programme. That programme had reached preclinical candidate nomination.21 The challenge, run with Polaris and the ARPA-H-funded OpenADMET project, asked teams to predict how compounds bind, how potent they are, and their ADMET properties (absorption, distribution, metabolism, excretion and toxicity), and it drew 66 teams.23

The results were humbling. In the ADMET track, Inductive Bio found that the biggest gains came from adding outside data specific to the task. Huge models pre-trained on general chemistry data helped only a little: the winning model beat a simple fingerprint baseline by about 1.7 percent.25 One physics-based docking approach correctly placed more than half of the compounds, within 2 Å of the crystal structure, for both protease targets. It still trailed methods that fine-tuned co-folding models.26

The verdict

Taken together, the coverage supports a clear conclusion. The new viral section of the AlphaFold Database does not cure anything, and its builders do not claim it does. What it offers is a head start: a filtered, openly licensed map of viral protein contacts, with nearly a third of the interface types never seen before, released alongside the open pipeline that produced it.811 Its value will depend on what happens next. That means experimental checks of the predicted complexes, models that handle sugars and three-protein assemblies, and more blind benchmarks like ASAP's that measure how well AI predictions match lab results. Open science is how the field gets ready. The experiments still have to show that the predictions hold up.