Domain adversarial gated bilinear attention networks for cross domain drug target interaction prediction.

Accurate prediction of drug-target interactions (DTIs) is crucial for accelerating drug discovery. However, current computational methods often fail to learn discriminative interaction features and generalize effectively to novel, distribution-shifted data. To address these challenges, we propose GBAN-DA (Gated Bilinear Attention Network with Domain Adaptation), a novel deep learning framework that integrates three key components: (1) a hybrid molecular encoder combining a graph convolutional network (GCN) with feature attention for drugs and a synergistic CNN-Transformer architecture for proteins; (2) a gated bilinear attention mechanism that explicitly models substructure-level interactions; and (3) a conditional domain adversarial network (CDAN) that aligns feature distributions across domains to improve generalization. Comprehensive evaluations show that GBAN-DA achieves state-of-the-art performance. On in-domain tasks, it attains AUROC/AUPRC scores of 0.964/0.950 (BindingDB) and 0.909/0.905 (BioSNAP), outperforming eight baseline models. Notably, under rigorous cold pair splits for cross-domain prediction, GBAN-DA with CDAN achieves substantial improvements of 11.9% AUROC on BindingDB and 12.9% AUROC on BioSNAP over its non-adapted variant. Case studies further validate high-confidence predictions against experimental databases for targets such as ABL1 kinase (P00519) and Cathepsin K (P43235). GBAN-DA establishes a robust framework for accurate and generalizable DTI prediction in real-world drug screening. The source code is available at https://github.com/LWHao1999/GBAN-DA.