Researchers investigating an unfamiliar virus often have protein sequences before they have experimentally determined structures. A new release in the AlphaFold Database gives them another place to start: predicted 3D structures of protein complexes drawn from more than 2,800 viral proteomes, including families relevant to human health.
Announced on September 24, 2026, the dataset is accessible through a Pandemic Preparedness Portal. Google DeepMind, EMBL-EBI, NVIDIA and academic research groups collaborated on the release. NVIDIA has also published the GPU workflow used to produce the predictions, allowing labs with suitable infrastructure to apply it to their own sequences.
These models are predictions. They can help scientists decide which viral proteins and possible interactions warrant closer investigation. They do not establish that a complex forms in an infected cell, explain a virus’s behavior or replace laboratory validation.
The Portal Puts Viral Complexes in One Place
The Pandemic Preparedness Portal brings viral structure predictions together in a browsable view. According to the database’s September update, the new collaborative effort used AlphaFold2 and AlphaFold-Multimer to analyze 2,812 viral proteomes across 23 families relevant to human health. It yielded 5,279 high-confidence predictions for complexes of different proteins, called heterodimers, and 2,749 for pairs of the same protein, called homodimers. The portal also includes another 4,681 high-confidence viral homodimer predictions from a separate dataset.
The figure describes viral proteomes, the sets of proteins encoded by viruses, rather than 2,812 experimentally characterized infections. The resource combines new predictions with existing viral data and labels models by confidence; its entries should not be counted as newly validated protein interactions.
EMBL-EBI’s account of the release says the collaborators prioritized families known to infect humans, guided by a UK Health Security Agency priority-pathogen tool. The collection ranges from viruses associated with the common cold to emerging threats such as mpox. It offers a starting point where researchers have sequence data but little structural information for the proteins they want to study.
A complex prediction proposes how protein chains might fit together, adding to what a model of one protein folded on its own can show. The proposed arrangement can suggest a surface or interface worth examining experimentally, even when it cannot establish whether the interaction occurs in living tissue.
Researchers Can Browse, Download and Check Candidates
A researcher can use the portal to locate a virus or protein of interest, inspect available models and their confidence information, then turn to the AlphaFold Database’s data access and download options for local analysis. The database states that its data is available for academic and commercial use under CC BY 4.0, which requires attribution. Researchers should retain the entry metadata and check the relevant dataset collection for any additional citation guidance.
A useful first question might be whether a protein with little prior structural coverage has a plausible predicted partner, or whether a proposed interface merits a laboratory test. A model could help a team choose what to examine next. It cannot answer that experimental question on its own.
For a concrete starting point, EMBL-EBI highlights a predicted mpox entry-fusion protein complex in its announcement and identifies the database entry as AF-0000000213001003. Readers can find and inspect that example without beginning with an open-ended search across thousands of proteomes.
NVIDIA says roughly 30% of the added protein interactions have shapes not documented in the experimentally determined Protein Data Bank, according to its release announcement. That indicates how much potentially useful structural territory the predictions cover; it does not mean 30% are confirmed new biological interactions. An absent experimental record may reflect how little a protein has been studied, and a predicted shape still needs testing.
NVIDIA Has Released the Prediction Workflow Too
The public database lets researchers examine existing models. The BioNeMo Structure Prediction Pipeline lets equipped groups generate predictions for additional protein sequences.
Its documented workflow builds multiple sequence alignments, which use related sequences to inform a prediction; predicts single-chain and multi-chain structures; and scores the results. For complexes, postprocessing includes interface scores and checks for steric clashes. NVIDIA says the large-scale viral predictions used AlphaFold2 with optimization through BioNeMo Inference Runtime. The repository documents folding backends built around OpenFold2 and AlphaFold2-Multimer weights, alongside other supported options. This workflow runs and assesses predictions; it does not mean the underlying structures were experimentally solved.
Running it requires substantial infrastructure. The repository lists a Slurm cluster configured with Pyxis and Enroot, NVIDIA GPUs with at least 80 GB of memory, sequence-search databases, model weights and storage for intermediate results. It puts the UniRef30 search index alone at about 240 GB. For many biologists, the portal will be the more accessible starting point.
The repository is published under the Apache 2.0 license, while its documentation separately identifies required inputs and third-party components. A lab considering a run should read those requirements before treating the release as a one-command service. Publishing the workflow also gives equipped researchers a way to apply the production process to other targets beyond the predictions already deposited.
Confidence Scores Are Not Biological Proof
The models carry confidence information, but different scores address different questions. The pipeline reports per-residue confidence and predicted aligned error, as well as measures intended to assess complex interfaces. These can help researchers prioritize predictions and spot uncertainty. A strong-looking interface score alone cannot show that two proteins meet inside a virus-infected cell.
Frequently Asked Questions
4 questions
1Where Can I Find the New Viral Protein Predictions?
The predictions are available through the Pandemic Preparedness Portal on the AlphaFold Database website. Researchers can inspect viral entries, including newly added complex models and their confidence information, then consult the database’s download options and collection metadata if they need files for local analysis.
2
Sources
- AlphaFold Databasealphafold.com
- EMBL-EBI’s account of the releaseebi.ac.uk
- Protein Data Bankrcsb.org
- release announcementblogs.nvidia.com
- BioNeMo Structure Prediction Pipelinegithub.com
