MAGNN-DTA: Cross-Modal Mutual Attention Graph Neural Networks for Drug-Target Affinity Prediction

Drug-Target Affinity (DTA) prediction is a crucial step in discovering and developing new pharmaceuticals. However, existing approaches remain limited in modeling multimodal features and capturing cross-modal interactions, often failing to fully characterize complementary information of drugs and proteins at the sequence, structural, and topological levels. To address these limitations, we propose MAGNN-DTA, a cross-modal mutual attention graph neural network framework for DTA prediction. First, MAGNN-DTA adopts a dual-path feature extraction module to extract sequence features and graph features of drug and protein. For drug representation, a Conformer block is used to capture long-range dependencies and local functional patterns, while a Graph Isomorphism Network (GIN) and multi-head attention are used to extract local and global features from molecular graphs. The drug features are further refined via a Kolmogorov-Arnold Network (KAN) to improve representational capacity and interpretability. For protein representation, sequence and structural information are integrated through a multi-scale deep convolutional network and a gated graph attention network (Gated-GAT) derived from residue distance maps. Subsequently, sequence feature and graph feature are concatenated by a gated pooling mechanism to form the representations of the drug and the protein. Finally, the model employs a cross-modal mutual attention module to explicitly capture bidirectional interactions between drug and protein modalities, enabling precise identification of key functional groups in drugs and critical binding residues in proteins. Extensive experiments on the Davis and KIBA benchmark datasets demonstrate that MAGNN-DTA achieves highly competitive performance compared with recent advanced methods across standard evaluation metrics. The source code and datasets are publicly available at https://github.com/HNUBIGroup/MAGNN-DTA.