A protein surface-aware multimodal framework for residue-level metal-binding site recognition

Accurate identification of metal-binding sites is essential for elucidating metalloprotein functions in cellular processes. Here, we introduce MetalBind, a protein surface-aware, multimodal deep learning framework tailored for residue-level recognition of metal-binding sites across diverse biologically relevant metals. By integrating advanced protein sequence embeddings, atom-specific physicochemical features, and a geometry-informed surface point learning module, MetalBind achieves higher predictive accuracy than existing methods. Our benchmarks across 13 metal-specific datasets reveal MetalBind's superior performance in several metrics. Notably, its efficacy persists even with sparse training data, highlighting its potential for large-scale metalloproteome annotation and guiding the rational design of engineered metalloproteins. These advancements significantly enhance metalloproteome annotation capabilities and expedite functional metalloprotein discovery. To promote accessibility, an online implementation of MetalBind is publicly available for the research community.