AlphaFold Database adds viral protein pairs from 2,800 human-infecting viruses
An international collaboration added more than 8,000 predicted viral protein pairs to the AlphaFold Database, spanning about 2,800 human-infecting viruses. The models stop at pairs and leave out the sugar coats, so trimeric targets like the coronavirus spike are largely absent.
The Scientist · Science desk
What happened
- An international collaboration published more than 8,000 predicted viral protein pairs in the AlphaFold Database, launching a pandemic-preparedness portal covering 23 human-infecting virus families.
- The vetted models are pairs only, so trimeric entry proteins such as the SARS-CoV-2 spike and HIV's envelope protein are largely left out.
- Collaborators planned to unveil the release on Sept. 24 at a CEPI and World Economic Forum roundtable, a day before the UN's high-level meeting on pandemic prevention.
Why it matters
- What's now possible capability Open, ready-made structures let any lab, including under-resourced ones facing outbreaks, look for the surface regions that make diagnostics, drugs and vaccines.
- What it limits constraint A drug or vaccine team cannot design against the assembled spike from this set yet; the highest-value targets still need experimental structures or a later trimer release.
- What it sets up precedent Trimer predictions are flagged as the next step, so the portal is likely to grow toward the multi-protein assemblies that dominate viral entry.
Next story · Invest · Oct 7
Meta, DeepMind and Isomorphic pay $300 million for first use of Biohub's open cell dataConfirmed11 publishers
Viruses were the blind spot in a database that already holds most known proteins, Grove said [9]. Their genomes often encode one long polyprotein that the cell then cuts into working pieces, so the edges of any single protein are hard to read from sequence alone and the resulting structure predictions come out incomplete [10]. Researchers at the Swiss Institute of Bioinformatics in Geneva worked through thousands of these cut proteins to fix accurate sequences first [11]. "Many viral proteins do not act individually, they act in concert with partners," said Joe Grove, a molecular virologist at the University of Glasgow who worked on the release [8].
The run was large and the filter was strict. The team took 41,774 proteins from about 2,800 viruses, among them the causes of mpox, measles and hepatitis B, and used AlphaFold2 to model 40,746 pairs of identical proteins and nearly 1.7 million pairs of different ones [12][13]. Only 2,749 of the identical pairs cleared the accuracy bar for the curated set, about one in fifteen, along with 5,279 of the mixed pairs [14][19]. That leaves 8,028 vetted pairs in the new portal, though every prediction was posted publicly [18][3].
Two limits matter for the pandemic use in view. The models stop at pairs, and they do not include the sugar molecules that coat many viral proteins and help them evade the immune system [4]. Many of the proteins a vaccine team would aim at sit in larger assemblies. The SARS-CoV-2 spike is three identical proteins, and so is HIV's entry protein; in many cases the two-protein predictions for these were not accurate enough to include, Grove said [5]. Trimer predictions are the obvious next step, according to EMBL-EBI's Sameer Velankar, though for bigger complexes it is not always clear how many copies of each part they contain [6].
"Making these data open is critical for understanding viral diagnostics and developing treatments and vaccines," said Jo McEntyre, interim director of EMBL-EBI. "They can support faster and more effective responses to future pandemics. They also cover lesser-studied viruses and lower the barriers for scientists in low-resource settings, who are confronting outbreaks firsthand." [20]
The push to have such structures early comes from pandemic-planning numbers. The Center for Global Development puts the chance of a COVID-scale pandemic by 2050 at nearly 50 percent [15]. The 100 Days Mission wants safe treatments and vaccines ready within 100 days of a new threat being identified, and the collaborators note that hitting that needs structural information in hand well before a virus turns global [16]. They planned to announce the release on Sept. 24 at a CEPI and World Economic Forum roundtable, a day ahead of the UN's high-level meeting on pandemic prevention in New York [17].
Grove wants to test it against his own work on viral entry proteins. "The proof of the pudding is going to be in putting that data out there and having teams such as my own start to drill down," he said [21].
What to watch
- Whether trimer predictions for spike proteins land in a future portal update, as EMBL-EBI's Velankar flagged.
- How the first analyses of the public data by Grove's and other teams hold up against experimental structures.
- What comes out of the Sept. 25 UN meeting reviewing the 2023 pandemic-preparedness commitments.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence70
- Adoption30
- Hype gap+25
- Incentives55
- Confidence70
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
AI-predicted protein complex structures for more than 2,800 viruses are now openly available in the AlphaFold Database, with the dataset prioritizing proteomes from viral families known to infect humans.
- [2]
Researchers added more than 8,000 virus protein dimers, pairs of interacting molecules, to the AlphaFold Protein Structure Database, from 23 virus families all with human-infecting members, as part of the launch of a pandemic preparedness portal within the database.
- [3]
The predictions are generated using the AlphaFold2 artificial-intelligence tool and made publicly available.
- [4]
The predictions lack the sugar molecules that adorn many viral proteins and help them evade immune detection.
- [5]
Many viral proteins work in complexes larger than dimers: the SARS-CoV-2 spike is made of three identical proteins, as is HIV's envelope-entry protein, and in many cases the dimer predictions for such complexes were not accurate enough to be included.
- [6]
Including trimer predictions is an obvious next step, but for bigger complexes it is not always clear how many copies of each component they are made of.
- [7]
The release came from an international collaboration among EMBL's European Bioinformatics Institute (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 in South Korea, plus complexes computed independently by Lund University in Sweden.
- [8]
Many viral proteins do not act individually, they act in concert with partners.
- [9]
Although the database holds predicted structures for most known proteins, viruses have been a blind spot, with high-quality entries lacking for many individual proteins such as those in flaviviruses, which include Zika and dengue.
- [10]
When some viruses replicate, their RNA is first translated into a polyprotein that is then cut into individual functional molecules, so it is not always clear from a virus's genetic sequence where a protein begins and ends, which can lead to incomplete structure predictions.
- [11]
Researchers at the Swiss Institute of Bioinformatics in Geneva identified accurate sequences for thousands of viral proteins that are cut from polyproteins.
- [12]
Grove and researchers at various organizations analysed the sequences of 41,774 proteins from around 2,800 viruses, including ones that cause mpox, measles and hepatitis B.
- [13]
The researchers used AlphaFold2 to predict structures for 40,746 homodimers, pairs of identical protein strands, and nearly 1.7 million heterodimers, in which the pairs are different molecules.
- [14]
Of the predicted pairs, only 2,749 of the homodimers and 5,279 of the heterodimers were deemed accurate enough to include in the AlphaFold database, although all of the predictions were made publicly available.
- [15]
An analysis by the Center for Global Development estimates a nearly 50% chance of the world facing a pandemic as severe as COVID-19 by 2050.
- [16]
The 100 Days Mission aims to deploy safe and effective treatments and vaccines within the first 100 days of identifying a new viral pandemic threat, which would require access to structural information about viruses long before they pose a global threat.
- [17]
Collaborators planned to announce the dataset for the first time Sept. 24 at a roundtable hosted by CEPI and the World Economic Forum, ahead of the United Nations General Assembly High-level Meeting on Pandemic Prevention, Preparedness and Response on Sept. 25 at UN headquarters in New York.
- [18]
A total of 8,028 predicted pairs were deemed accurate enough for inclusion in the curated set.
- [19]
About 6.7 percent of the homodimer predictions, roughly one in fifteen, cleared the accuracy bar for inclusion.
- [20]
Making these data open is critical for understanding viral diagnostics and developing treatments and vaccines. They can support faster and more effective responses to future pandemics. They also cover lesser-studied viruses and lower the barriers for scientists in low-resource settings, who are confronting outbreaks firsthand.
- [21]
The proof of the pudding is going to be in putting that data out there and having teams such as my own start to drill down.
Sources
2 independent publishers whose own reporting we read for this story.
- nature.comAlphaFold 'goes viral': database adds protein complexes of common viruses
1 article · September 23, 2026
- phys.orgNew AlphaFold Database release adds viral protein complexes to support pandemic preparedness
1 article · September 24, 2026
Topics and entities
Topics
- Protein structure prediction
- Pandemic preparedness
- Open scientific data and preprints
Entities
- 100 Days Mission
- Nvidia
- Google DeepMind
- Coalition for Epidemic Preparedness Innovations
- University of Glasgow
- Swiss Institute of Bioinformatics
- AlphaFold Database