Drug-conditioned Residue Gating with Bidirectional Cross-attention for Drug-target Binding Affinity Prediction

Abstract

Drug-target affinity (DTA) prediction remains challenging because drug and protein representations are frequently compressed before fine-grained atom-residue relationships and drug-specific residue relevance can be adequately modeled.In this study, a multimodal DTA framework is proposed in which molecular graphs are encoded using three residual Graph Convolutional Network layers, while contextual protein representations are obtained from frozen Evolutionary Scale Modeling 2 (ESM-2) embeddings.Drug-conditioned residue gating is applied before full cross-modal interaction, followed by local protein self-attention, two-layer bidirectional atom-residue cross-attention, sparse Entmax-1.5 pooling, gated multimodal fusion, and continuous affinity regression.The framework was evaluated on Davis, Kinase Inhibitor BioActivity (KIBA), and BindingDB-𝐾 𝑑 under held-out pair-wise protocols.The final validation-selected prediction procedures, which include dataset-specific post-hoc components external to the standalone neural architecture, achieved root mean squared error values of 0.400878, 0.383522, and 0.617746 for Davis, KIBA, and BindingDB-Kd, respectively, within their corresponding target scales.For Davis, the standalone neural architecture itself achieved RMSE = 0.421551 and MSE = 0.177705 before the external post-hoc prediction procedure was applied.In the controlled Davis comparison, replacing pooled bilinear fusion with bidirectional crossattention reduced test mean squared error from 0.256492 to 0.187028, corresponding to a 27.08% reduction, while both concordance index and Pearson correlation increased.Adaptive residue gating further reduced validation mean squared error by 8.37% and increased preferred 5-30% residue-selection coverage from 24.21% to 66.98%.Perturbation-based analyses additionally indicated substantially greater neural prediction sensitivity to attributionranked residues than to matched random positions.The results support drug-conditioned residue selection followed by explicit bidirectional atom-residue interaction as an effective interaction-modeling strategy with perturbationsupported interpretability under the evaluated pair-wise setting.