← Back to Home

Preparing a Molecular Map of Viruses in Advance: AlphaFold Releases Protein Complex Predictions for More Than 2,800 Viruses

How viral proteins assemble may offer entry points for vaccine and drug development. Newly released data provide extensive structural clues, but experiments must still determine whether the interactions shown in the models actually exist and can serve as effective targets.

By SURL BioNews

To prepare for the next epidemic, researchers need to know in advance not only which genes a virus has, but also how its proteins assemble and work together. On September 24, EMBL-EBI and its collaborators released protein complex structure predictions covering more than 2,800 viruses, prioritizing viral families known to infect humans. The aim is to prepare molecular clues for diagnostic, drug, and vaccine research before new threats emerge.

The data extend the focus from the shape of individual proteins to how proteins bind to one another. Binding interfaces and surface regions may help researchers select drug target sites or vaccine antigens worth testing. According to Nature, the team analyzed 41,774 proteins from approximately 2,800 viruses, including viruses associated with diseases such as mpox, measles, and hepatitis B. However, these models provide a starting point for research and have not demonstrated the therapeutic efficacy of any candidate target.

There is also a substantial gap between the scale of the predictions and the number that passed screening. Nature reported that the research generated 40,746 homodimer models composed of identical proteins and nearly 1.7 million heterodimer models composed of different proteins. Of these, 2,749 and 5,279, respectively, met the threshold for inclusion in the database, while the full set of predictions was released separately. This threshold screens for model quality and should not be interpreted as experimental confirmation of the interactions.

One key step in making these predictions possible was first determining which proteins viruses actually use. Many viruses produce a long “polyprotein” that is then cleaved into different functional proteins. The Swiss Institute of Bioinformatics (SIB) compiled cleavage information for 699 viruses, providing reliable sequences for 7,783 mature functional proteins, and assessed the plausibility of predicted interactions using knowledge of viral biology. SIB also established a new polyprotein annotation standard, making cleavage data available through ViralZone for subsequent research.

On the computational side, NVIDIA said the team used AlphaFold2 with BioNeMo Inference Runtime to accelerate large-scale predictions and released a GPU-accelerated structure prediction workflow for researchers to process their own protein targets. The company estimated that approximately 30% of the structural forms among the newly added interactions had not previously appeared in the Protein Data Bank (PDB), which contains experimental structures. This provides more hypotheses to explore, but whether they represent new biological mechanisms remains to be verified.

The models also leave important gaps that affect their application. Nature noted that the predictions did not include the sugar molecules found on the surfaces of many viral proteins. They also modeled only dimers consisting of two proteins and therefore cannot fully represent complexes composed of three or more proteins. When evaluating vaccine antigens or drug binding sites, researchers must therefore still confirm whether the models reflect the actual molecular environment; the confidence indicators accompanying the data cannot replace this step.

The value of this open resource lies in providing structural clues for comparison and testing even for less-studied viruses, helping scientists prioritize experiments. Moving from a predicted molecular map to usable medical tools still requires confirming whether the proteins actually interact, what roles they play during infection, and whether intervening produces an effect. EMBL-EBI also emphasized that structure predictions alone cannot determine a virus’s actual behavior or the effects of mutations on transmission and pathogenicity.

References

  1. EMBL-EBI
  2. NVIDIA Blog
  3. SIB Swiss Institute of Bioinformatics
  4. Nature