CGA-DTA: interaction-conditioned graph topology and cross-graph attention for drug--target affinity prediction

Abstract

Drug–target affinity (DTA) prediction supports virtual screening in computer-aided drug discovery, but graph-based models still face two linked limitations. Adaptive-topology methods usually learn drug and protein graphs independently and merge them only at the prediction head, whereas cross-attention methods model drug–protein interactions after fixed or separately learned topology. Thus, topology refinement is rarely conditioned on the specific binding partner. In addition, many graph-based pipelines rely on multiple-sequence-alignment (MSA)-derived protein features, which limits scalability for large or newly annotated target sets. We propose CGA-DTA , a dual-graph framework built around interaction-conditioned topology . Adaptive Multi-head Graph Learning (AMGL) first learns multiple soft adjacency heads for each graph; an Interaction-Conditioned Topology (ICT) gate then uses a global summary of the binding partner to re-weight the head mixture. ICT therefore changes the effective row-stochastic message-passing kernel, rather than predicting physical contacts or reconstructing topology at the node or edge level. A Cross-Graph Attention Module (CGAM) subsequently applies bidirectional node-level attention to the conditioned representations. CGA-DTA uses ESM-2 residue embeddings as alignment-free protein node features, while the initial PconsC4 contact-map topology remains MSA-dependent. Across Davis, KIBA, Metz, and ToxCast, CGA-DTA obtains MSE/CI values of 0.168/0.915, 0.116/0.910, 0.231/0.854, and 0.149/0.920, respectively. Component ablations consistently assign the largest individual gain to CGAM and a smaller additional gain to ICT. CGA-DTA shows that coupling adaptive graph learning with cross-graph attention is effective for DTA prediction. The evidence supports partner-conditioned re-weighting of learned message-passing kernels, but does not establish that these kernels are physical protein–ligand contact maps. Replacing the remaining MSA-based contact-map stage and validating the learned kernels against experimentally resolved complexes are important next steps.