AlphaFold Database Adds More Than 8,000 Viral Protein Structures to Support Pandemic Preparedness

The AlphaFold protein structure database has been upgraded to better represent some of the least understood and most dangerous viruses. Researchers have added more than 8,000 predicted viral protein dimers—pairs of interacting molecules—to the resource.

The new data cover 23 virus families, all of which include members that infect humans. The update is part of a new “pandemic preparedness portal” within the widely used, freely available AlphaFold database, which contains publicly accessible three-dimensional structure predictions generated with the AlphaFold2 artificial intelligence tool.

AlphaFold expands its protein complex predictions

Earlier this year, researchers added predicted structures for 1.7 million pairs of interacting proteins from 20 widely studied organisms, including humans, mice and the bacterium that causes tuberculosis. It was the first time that protein complexes such as enzymes made up of two identical protein chains had been included in the AlphaFold database.

The database is maintained by the European Molecular Biology Laboratory’s European Bioinformatics Institute (EMBL-EBI) in Hinxton, UK.

“Many viral proteins do not act individually, but in conjunction with partners,” says Joe Grove, a molecular virologist at the University of Glasgow, UK, who was involved in adding the viral complexes.

Viruses were a blind spot in the database

The database has more than 3 million users and stores predicted structures for most known proteins. But viruses were a blind spot, Grove says. Many individual viral proteins lacked high-quality entries, including proteins from flaviviruses such as Zika and dengue viruses.

One challenge is that some viruses replicate by first translating their RNA into a “polyprotein”, which is then cleaved into individual functional molecules. As a result, it is not always clear from a viral genetic sequence where each protein begins and ends. This can lead to incomplete structural predictions.

To address the problem, researchers at the Swiss Bioinformatics Institute in Geneva determined the precise sequences of thousands of viral proteins excised from polyproteins. Grove and researchers from organizations around the world analyzed 41,774 protein sequences from about 2,800 viruses, including viruses that cause mpox, measles and hepatitis B.

Thousands of viral dimers meet accuracy standards

Using these sequences, the researchers used AlphaFold2 to predict the structures of 40,746 homodimers, which are pairs of identical interacting protein chains, and about 1.7 million heterodimers, which are pairs of different molecules.

Of those predictions, 2,749 homodimers and 5,279 heterodimers were considered accurate enough for inclusion in the AlphaFold database. All of the predictions, including those that did not meet the database’s accuracy threshold, were made publicly available.

What the viral protein predictions do not show

The predictions do not include sugar molecules that decorate many viral proteins and can help viruses evade immune detection. Many viral proteins also operate in complexes larger than dimers.

For example, the spike protein that enables SARS-CoV-2 and other coronaviruses to infect host cells consists of three identical proteins. HIV’s envelope entry protein is also made up of three identical proteins. Predictions of these dimers are often excluded from the AlphaFold database because they are not accurate enough, Grove says.

Including such “trimeric” predictions is an obvious next step, says Sameer Velankar, a bioinformatician at EMBL-EBI involved in the project. For larger complexes, however, it is not always clear how many copies of each component are present.

Grove is eager to use the new data to study viral entry proteins. “The proof of the problem is when you make that data public and teams like myself start drilling down,” he says.