A dual-branch graph neural network architecture for drug-target binding affinity prediction

Abstract Graph Neural Networks have emerged as a powerful paradigm for artificial intelligence driven drug discovery, offering molecular representation learning that surpasses many conventional approaches. Traditional experimental pipelines are both time and resource-intensive, modern computational strategies—particularly those that integrate curated libraries of FDA-approved drugs—can accelerate target identification and candidate prioritization. In this work we introduce a dual-branch GNN architecture that synergistically combines Graph Convolutional Neural Networks, the GraphSage framework, and Jumping-Knowledge modules. This network jointly encodes structural topology and functional attributes, generating enriched embeddings for molecular graphs. We evaluated the proposed model against 45 state-of-the-art drug and target encoding baselines across well known Davis and KIBA datasets.The Proposed Model demonstrates a quantitative improvement over the GCN model, achieving a reduction in MSE (33.98 vs. 35.24), a slightly higher Pearson index (76.49 vs. 76.19), and a better Concordance index (85.41 vs. 84.41), indicating superior performance in terms of both prediction accuracy and ranking, demonstrating superior accuracy and robustness for candidate screening and establishing a new reference point for cheminformatics tasks. To illustrate practical impact, we performed a case study on COVID-19 drug repurposing: the top-ranked drugs have been also found potential drugs including Imunovir and Remdesivir from existing antiviral drugs.