Abstract
ABSTRACT
Metal binding to proteins is crucial for many biological functions, yet the prediction of metal‐binding sites and metal types remains underexplored compared to other biomolecular interactions. Despite advances in deep learning, computational methods have achieved limited success in this area; even the best‐reported approaches reach only moderate localization accuracy on high‐resolution crystal structures, and even worse when applied to medium‐resolution data. We hypothesize that the main challenge lies in the nature of metal binding. While the putative sites span the entire protein surface, actual binding requires the atomic‐level coordination geometry. We present a structure‐based metal binding site and type predictor, PrOMet, in this work, which leverages a Prune‐and‐Optimization strategy built upon two hierarchically connected machine learning architectures. This strategy allows the method to effectively narrow the search space while maintaining robustness to AlphaFold models and medium‐resolution cryo‐EM structures. PrOMet outperforms existing methods, including structure‐based and language model‐based predictors, achieving a PR‐AUC of 0.739 for metal ion localization and an average F1‐score of 0.743 for metal type classification—Zn, Ca, Mg, Mn, Fe, Cu, respectively. PrOMet can be an effective tool for allocating uncertain metal positions and types within protein structures, as well as for metalloenzyme design studies.