TUNA: A Target-aware Unified Network for Protein-Ligand Binding Affinity Prediction via Multi-Modal Feature Integration

The accurate prediction of protein-ligand binding affinity is crucial for early-stage drug discovery. Sequence-based deep learning models offer scalability and broader applicability than structure-based methods, but typically ignore the local binding site context, limiting predictive power. Recent advances in protein structure prediction and binding pocket detection have enabled for the integration of pocket-level information into sequence-based models. We present TUNA, a novel deep learning model that integrates multi-modal features to predict binding affinity. The TUNA integrates global protein sequences, localized pocket representations, ligand features derived from the SMILES, and molecular graph structures. We used three dimensional structure inference and pocket detection tools for proteins lacking experimentally determined binding sites. The pocket and protein sequences were encoded using embeddings from pre-trained models, including a model pretrained specifically on pocket-derived sequences. Ligands were represented through a fusion of Chemformer encoded SMILES and graph diffusion-based features, then unified via an alignment strategy to preserve both symbolic and structural information. TUNA achieved consistent improvements over sequence-based models across the PDB bind and BindingDB datasets while remaining competitive with structure-based methods for the PDBbind dataset. Its interpretable cross-modal attention mechanism enables the inference of potential binding sites, thus enhancing biological interpretation. These results demonstrate that TUNA is an effective affinity prediction method, especially for targets without known structures.