DPI-score: A deep learning-based metric for assessing protein-protein interfaces in cryo-EM derived assemblies

Abstract

Advances in cryoEM have led to a surge in high-resolution structures; however, model building and refinement at resolutions (>=3 [A]) and in regions with variable local resolution remain challenging, making validation essential. CryoEM derived assemblies often contain extensive protein-protein interfaces, however, most existing validation metrics focus on density fit or overall geometry without directly assessing interface quality. To address this, we present DPI-Score, a deep learning-based metric for assessing protein-protein interfaces in cryoEM derived complexes. The method uses only raw structural coordinates of interface atoms, without requiring engineered features, and achieves 87.53% validation accuracy. DPI-Score was applied to 29,120 interfaces from 6,011 fitted entries with resolutions worse than 3 [A] in the Electron Microscopy Data Bank. Here, we show that DPI-Score provides complementary information to existing validation metrics and can identify interface errors in modelled assemblies that are not detected by density-based scores alone.