AlphaFold Database adds predicted protein pairs for 2,800 viruses to aid pandemic research

Scientists preparing for the next pandemic have gained a large, free reference library. The AlphaFold Database, run by the European Molecular Biology Laboratory's European Bioinformatics Institute (EMBL-EBI), now holds AI-predicted 3D structures showing how proteins from more than 2,800 viruses may fit together.[2]

The release was announced on 24 September, a day before a United Nations General Assembly high-level meeting on pandemic prevention, preparedness and response in New York. The partners say the aim is to have structural knowledge ready before an outbreak, rather than building it after one starts.[2][1]

The database's new Pandemic Preparedness Portal says the team predicted structures across 2,812 viral proteomes from 23 virus families relevant to human health. It kept 5,279 high-confidence pairs of different proteins and 2,749 pairs of identical ones. Another 4,681 identical pairs came from a separate Viral AlphaFold Database project.[3]

Nature reports that the researchers worked from 41,774 viral protein sequences and predicted nearly 1.7 million pairs in total. Only the most reliable made it into the main database, although all predictions were made public. The viruses range from common-cold families to emerging threats such as mpox, measles and hepatitis B.[4][2]

The structures were inferred with AlphaFold2, Google DeepMind's protein-folding model, and sped up with NVIDIA's BioNeMo Inference Runtime so the team could run thousands of viral proteomes. NVIDIA has also released the GPU workflow it used as open source, so labs can predict structures for their own targets.[1]

The partners are EMBL-EBI, Google DeepMind, NVIDIA, Seoul National University, Sungkyunkwan University, the University of Glasgow, the Swiss Institute of Bioinformatics and the Coalition for Epidemic Preparedness Innovations. NVIDIA says about 30% of the protein interactions added are new to science, with no match in the main experimental structure archive.[2][1]

EMBL-EBI is explicit about the limits. The predictions do not show how genetic changes affect a virus or how it interacts with its host. They cannot show what makes a virus more deadly or transmissible, and they cannot be used to engineer viruses that infect humans. Real behaviour still has to be tested in a laboratory.[2]

Nature adds two gaps. The models leave out the sugar molecules that help many viral proteins hide from the immune system, and many complexes larger than pairs are missing, including the three-part coronavirus spike. Flaviviruses such as dengue and Zika have lacked good entries because their proteins are cut from one long chain, a problem this effort tried to address.[4]

Malaysia lives with viral disease every year. The Ministry of Health's iDengue portal recorded 72,049 dengue cases and 66 deaths between 4 January and 27 September 2026, with Selangor alone accounting for 30,523 cases. The new portal also has a separate view for neglected tropical diseases.[5][3]

The World Health Organization lists Malaysia among the countries that have reported Nipah virus, whose case fatality rate it estimates at 40% to 75%. Fruit bats are its natural host, and there is still no treatment or vaccine. EMBL-EBI's interim director says open data lowers barriers for scientists in low-resource settings who face outbreaks first-hand.[6][2]