Abstract
Many proteins function through ligand binding. Yet, reliable experimental binding data remains limited. Recent advances predict binding residues from sequences using protein Language Model embeddings. The AlphaFold Protein Structure Database, which has reliable 3D structure predictions from AlphaFold2, opens the way for graph neural networks that predict binding residues. Here, we introduce bindNode24, a new method using Graph Neural Networks to predict whether a residue binds to any of three ligand classes: small molecules, metal ions, and nucleic macromolecules. Compared to state-of-the-art, this approach reduces the number of free parameters by almost 60 % at similar performance. Our findings also suggest that secondary and tertiary structure features from AlphaFold2 are easy to integrate into protein function prediction tasks that previously solely relied on protein Language Model embeddings.