Attentive Multi-Channel Molecular Representation in Drug–Target Affinity Prediction

Efficient and accurate drug–target affinity (DTA) prediction is fundamental to computational drug discovery. Drug–target binding affinity is influenced by multiple factors, including structural conformations, functional groups, and molecular flexibility. However, existing graph neural network (GNN)-based approaches often fail to explicitly capture these fine-grained features, limiting both performance and interpretability. To address this gap, we propose a model that integrates structure-aware protein embeddings with GNN-based multi-channel molecular features. Specifically, three molecular channels capture complementary aspects: the molecule channel models global topology, the scaffold channel preserves the core backbone, and the context channel captures local patterns such as functional groups and substructures. These channels are fused via a cross-channel attention mechanism, which dynamically aligns protein features with molecular representations and adaptively assigns weights across channels, thereby leveraging complementarity while mitigating redundancy. Experiments on the Davis dataset demonstrate that our model outperforms strong baselines. Beyond performance gains, our approach advances interpretability in DTA prediction by providing biologically meaningful insights into structural and chemical determinants. Overall, these results highlight the effectiveness of attentive multi-channel molecular representation and establish a powerful and interpretable framework for advancing drug–target affinity prediction.