TransDTAP: A multimodal transformer architecture for drug-target affinity prediction using sequence and biochemical properties.

Accurate drug-target affinity (DTA) prediction is essential for virtual screening and early-stage drug discovery. We propose TransDTAP, a multimodal Transformer-based architecture that integrates ligand SMILES sequences, protein amino-acid sequences, molecular physicochemical descriptors, and protein biochemical features within a unified regression framework. A curated dataset comprising 4793 protein targets, 23,531 ligands, and 33,457 experimentally measured IC50 interactions was constructed from ChEMBL and UniProt sources. Under a standard random split, TransDTAP achieved an R2 of 0.7794, MSE of 0.2142, and MAE of 0.2827 on the pIC50 scale (0.8556nM after inverse transformation). Robustness was further evaluated using Bemis-Murcko scaffold-based splitting, where the model maintained competitive performance (R2=0.689), indicating effective generalization to chemically novel scaffolds. Fair benchmarking against re-implemented baseline models demonstrated consistent performance advantages under both random and scaffold settings. Ablation analysis confirmed that predictive improvements arise from complementary integration of sequence-based encoders and descriptor-level features, while SHAP-based interpretability analysis revealed that influential features align with established determinants of ligand-protein binding, including hydrophobicity, polarity, and electrostatic properties. These results demonstrate that multimodal fusion of learned sequence representations and domain-informed descriptors provides a robust, interpretable, and scalable framework for drug-target affinity prediction, supporting its application in large-scale virtual screening pipelines.