EGA-DTA: An Energetic-Geometric Augmented Graph Neural Network With Target-Conditional Gating for DTA Prediction

Drug-target affinity (DTA) prediction plays an important role in virtual screening. However, many graph-based DTA models represent molecular bonds solely as discrete bond types and combine drug and protein features through simple concatenation. We propose EGA-DTA, a graph neural network augmented with energetic and geometric edge features and target-conditional gating. EGA-DTA characterizes ligand bonds using bond dissociation energy (BDE) and conformer-derived bond length (BL), integrates graph-based and fingerprint-based drug representations, and employs a protein-derived gating mechanism to modulate the drug representations before fusion. On the Davis, KIBA, and Metz benchmarks, EGA-DTA achieves competitive mean squared error (MSE) and concordance index (CI), together with a favorable $r_{m}^{2}$ value on Davis. On the same fixed KIBA cold-start splits, EGA-DTA demonstrates competitive performance, yielding the lowest MSE point estimate among the reproduced baselines in the cold-target, cold-drug, and cold-pair settings. Controlled standard-split ablations further suggest that both the physicochemical edge encoding and the target-conditional gating mechanism may contribute to the observed performance.