This research develops SAKE-PP to enhance protein-protein interaction model ranking, suggesting improved prediction accuracy for structural biology applications.
Key Points
The aim is to improve the accuracy of ranking near-native protein-protein interaction models using SAKE-PP.
Developed a spatial-attention equivariant graph neural network (SAKE-PP)
Utilized a hierarchical iRMSD-guided sampling strategy
Trained on the PDBBind dataset to assess protein interactions
Evaluated using the 2024PDB benchmark with 176 heterodimers
Conducted zero-shot evaluations on 139 antibody-antigen complexes.
SAKE-PP improved AF3-decoy selection by 13.75% in iRMSD and 12.5% in DockQ
Exhibited better rankings than AF3 in overlap, hit-rate, and correlation metrics
Increased correlation by 0.4 in zero-shot evaluation on antibody-antigen complexes
Enhanced model selection by favoring geometrically near-native and energetically plausible interfaces