AlphaFold just mapped thousands of viral protein partnerships. The models still need a lab check

A new map of viral machinery

The AlphaFold Protein Structure Database on Sept. 24 added a Pandemic Preparedness Portal with high-confidence predictions for 8,028 viral dimers—pairs of identical or different proteins—from 2,812 proteomes spanning 23 virus families relevant to human health. The release also includes more than 50,000 viral monomer and dimer models, with lower-confidence predictions available for bulk download. nature +2

The underlying study used AlphaFold2 and AlphaFold-Multimer to make about 1.7 million pairwise predictions from roughly 42,000 viral protein sequences. Only 2,749 homodimers and 5,279 heterodimers cleared the project’s confidence thresholds. The scale matters because many viral proteins work in assemblies rather than alone, but those interactions are difficult to determine experimentally, especially for viruses that are hard or unsafe to grow. research +1

Why the structures could matter

The models offer researchers a starting map for proteins involved in viral entry, assembly and replication. The study grouped the high-confidence results into 1,598 interface clusters; 471 had no detectable similarity to an experimentally determined interface in the Protein Data Bank. The authors highlight predicted complexes involving poxvirus entry proteins and viral proteases, which could help generate hypotheses for vaccine or antiviral research. research

The Swiss Institute of Bioinformatics says the work also adds first-time structural predictions for nearly 7,800 viral proteins derived from polyproteins—long precursors that viruses cut into mature proteins. That annotation step is important for viruses such as flaviviruses, because treating a whole polyprotein as one unit can obscure where the functional proteins begin and end. nature +1

A hypothesis engine, not a diagnosis

The release is not a catalogue of confirmed molecular behavior. AlphaFold’s models omit some biological details, including sugar molecules attached to many viral proteins, and the campaign focused largely on dimers even though important assemblies can contain three or more proteins. Nature reported that the dataset’s dimer predictions do not fully capture structures such as coronavirus spike trimers. nature

The paper’s benchmark found 31 high-confidence predictions among 215 experimentally supported viral pairs, with no false positives in its negative comparison set—but that was a stringent, limited test, not a guarantee for every model. The database itself says the predictions are intended to accelerate research, while experiments remain necessary to establish how a protein behaves. The next step is therefore practical: researchers can use the open models to choose experiments, then test which predicted interfaces survive contact with real viral proteins. research +2