LINKER: Learning Interactions Between Functional Groups and Residues With Chemical Knowledge-Enhanced Reasoning and Explainability

Accurate mapping of protein-ligand interactions is key to understanding molecular recognition and guiding drug design. Existing deep learning methods typically depend on 3D structures or distance-based labels, limiting their scope and interpretability. We present LINKER, the first sequence-based model that predicts residue-functional group (FG) interactions in biologically defined types directly from protein sequences and ligand SMILES. Trained with structure-supervised attention using FG-derived labels from experimental complexes, LINKER focuses on chemically meaningful substructures and predicts interaction types such as hydrogen bonds and $\pi$-stacking. Crucially, it requires only sequence input at inference, enabling scalable application without structural data. Experiments on the LP-PDBBind benchmark show that LINKER achieves accurate, biochemically interpretable interaction maps aligned with ground-truth annotations.