Geometric deep learning-enabled metal-binding site identification and grafting

Metal-binding sites can participate in enzymes' catalytic reactions, as well as in protein folding, stability, and protein-protein interactions. In-depth research on metal-binding sites can aid in identifying new protein functions and designing new metalloproteins. We present MeSiteIG, a geometric deep learning-driven bifunctional tool with capabilities for both Metal-binding Site Identification and Grafting, established based on the geometric conservation of metal-binding sites. It comprises three main modules: residue triplet searcher (ResTriS), metal-binding site identifier (MeSI), and metal-binding site alignment (MeSA). As the core module, MeSI adopts E3-equivarient graph neural networks to predict potential metal-binding site residues, without the coordinating residue types and side-chain information. MeSI achieved superior performance on the independent test set, with an impressive speed of approximately 300 samples per second. Using MeSiteIG, we identified previously neglected protein metal-binding sites and potential protein bimetallic sites. Furthermore, it could graft metal sites onto antibody surfaces and protein pockets, creating new potentially valuable metalloproteins.