Overview
- On Thursday, Sept. 24, 2026, a coalition led by Google DeepMind, EMBL-EBI and NVIDIA published AI predictions for roughly 2,800 viral proteomes after analysing about 41,774 viral protein sequences and generating millions of dimer models.
- The AlphaFold database now includes 2,749 high-confidence homodimers and 5,279 high-confidence heterodimers drawn from that larger set, while the full archive of lower-confidence predictions has been released for researchers to examine.
- NVIDIA’s BioNeMo inference runtime was used to scale AlphaFold2 predictions and the BioNeMo Structure Prediction Pipeline has been made public so labs can run GPU-accelerated structure inference themselves.
- The dataset is presented as a hypothesis-generating tool because predictions omit features such as glycosylation and many biologically important assemblies like trimers or larger multimers still lack reliable models and need experimental validation.
- About 30% of the newly reported interactions are not previously documented in the Protein Data Bank, offering fresh targets for study and lowering barriers for scientists in low-resource settings to explore viral structure and drug or vaccine leads.