AlphaFold Database adds more than 8,000 predicted viral protein pairs

The AlphaFold Protein Structure Database, run by EMBL's European Bioinformatics Institute (EMBL-EBI) in Hinxton, UK, has added AI-predicted structures of viral protein complexes and a new "Pandemic Preparedness Portal", EMBL announced on Thursday 24 September 2026.

What was added

According to the database's own "What's new" note, the collaboration "predicted structures across 2,812 viral proteomes from 23 human-health-relevant families, yielding 5,279 high-confidence heterodimers and 2,749 high-confidence homodimers". That is 8,028 high-confidence pairs of interacting proteins, which Nature's news report describes as "more than 8,000 virus protein dimers".

• What is counted: dimers, meaning pairs of interacting protein chains. A homodimer is two identical chains; a heterodimer is two different ones.
• What was predicted in total: the database FAQ says the viral complexes dataset "contains 1,703,992 predicted protein–protein interactions across ~2,800 viral proteomes". Nature reports 40,746 homodimers and "nearly 1.7 million heterodimers" were predicted, and that "all of the predictions were made publicly available".
• A second viral set: the same note says "A further 4,681 high-confidence viral homodimers have been added from the Atkinson Lab's Viral AlphaFold Database (VAD)". EMBL says this set was "computed independently of this collaboration by colleagues at Lund University in Sweden".
• Which viruses: EMBL says the structures run "from 'common cold' viruses such as those in the Picornaviridae family, through to emergent viral threats such as Mpox". Nature reports the sequences include viruses that cause mpox, measles and hepatitis B. EMBL says the selection was "Guided by the UK Health Security Agency's priority pathogen tool".

The sources describe the virus families slightly differently. EMBL says the dataset "prioritises proteomes from viral families known to infect humans"; the database FAQ says "23 viral families known to infect humans and/or animals"; Nature says the 23 families all have "human-infecting members".

How the predictions were made

The database note says the work used "AlphaFold2 and AlphaFold-Multimer". The FAQ says the dataset "was generated using an accelerated implementation of AlphaFold2-Multimer". NVIDIA's blog says the structures "were inferred using AlphaFold2 — Google DeepMind's AI model for predicting how proteins fold into 3D shapes — with optimization from NVIDIA BioNeMo Inference Runtime".

Nature reports that researchers at the Swiss Institute of Bioinformatics "identified accurate sequences for thousands of viral proteins that are cut from polyproteins", which it says had left gaps for viruses such as Zika and dengue.

EMBL lists the partners as EMBL-EBI, Google DeepMind, NVIDIA, Seoul National University, the University of Glasgow, the Swiss Institute of Bioinformatics, the Coalition for Epidemic Preparedness Innovations (CEPI) and Sungkyunkwan University. None of the sources we read say who funded the work.

Confidence and limits

The database FAQ says complex entries shown on the website "are filtered using thresholds of pDockQ2max ≥ 0.23, and ipSAEmax ≥ 0.6", and that "A complete list of predictions and their associated scores is available in the FTP area."

EMBL's release has a "Responsible use and limitations" section. It says the predictions "do not predict the impact of genetic variation on a virus or host-pathogen interactions", and: "Predicted protein structures alone cannot tell scientists how a virus actually behaves; this requires experimental investigation in a laboratory." The same section says: "Scientists can't use these protein structure predictions to see how changes make a virus more deadly or transmissible, or to engineer viruses that infect humans."

Joe Grove of the MRC-University of Glasgow Centre for Virus Research, who worked on the dataset, said in the release: "The new dataset provides valuable foundational knowledge that can enable fundamental research and the development of countermeasures, but it doesn't shed light on why some viruses thrive and others don't, and it doesn't make it easier to engineer more dangerous pathogens."

Nature's subheading says the predictions "will need experimental confirmation". It reports that they "lack the sugar molecules that adorn many viral proteins", and that dimer predictions for larger complexes such as the SARS-CoV-2 spike "won't be included in the AlphaFold database in many cases, because their dimer predictions weren't accurate enough, says Grove."

Access

The database says its data "is available for academic and commercial use, under a CC-BY-4.0 licence." EMBL says the viral data is reached through the Pandemic Preparedness Portal on the database homepage. NVIDIA says it is also releasing the "BioNeMo Structure Prediction Pipeline" used to generate the dataset.

What the partners say

• Jo McEntyre, Interim Director of EMBL-EBI: "Making these data open is critical for understanding viral diagnostics and developing treatments and vaccines".
• Anna Koivuniemi, Head of Google DeepMind Impact Accelerator: "Adding these 3D models to the AlphaFold Database makes them immediately useful for real-world science."
• NVIDIA's blog states: "About 30% of the protein interactions being added to the database are completely new to science". This is NVIDIA's figure; we did not find it in EMBL's release or the database pages.

EMBL says the release coincides with the UN General Assembly high-level meeting on pandemic prevention, preparedness and response in New York on 25 September.

Why it matters

EMBL says the release "could help scientists understand what these proteins look like and which regions may interact with human cells", which "could, in turn, support the development of vaccines and other countermeasures if an outbreak were to occur". On our reading, the numbers show how selective the database was: Nature reports around 1.7 million predicted pairs, of which about 8,000 were "deemed accurate enough to include". EMBL says laboratory work is needed to show how a virus behaves; Nature says the predictions will need experimental confirmation.