Motif-Attention Enhanced Graph Convolutional Networks for Drug-Target Interaction Prediction: Improved Generalization Under Cold-Start Conditions

Muhammad Yusuf ML (School of Computing Telkom University Bandung, Indonesia)
Isman Kurniawan (School of Computing Telkom University Bandung, Indonesia)

ARMADA : Jurnal Penelitian Multidisiplin
Vol. 4 No. 9 (2026): ARMADA : Jurnal Penelitian Multidisplin, September 2026

Accurate prediction of drug-target interactions (DTI) remains challenging, especially in cold-start scenarios involving unseen drugs or proteins. This study proposes a motif-attention augmented Graph Convolutional Network (GCN) framework that integrates structural motif information through an attention mechanism to improve molecular and protein representation learning. The proposed model was evaluated on the Davis kinase inhibitor dataset using four settings: random split, cold-drug, cold-protein, and cold-both. Results show that motif-attention significantly improves generalization in difficult scenarios, achieving an 11.19% reduction in Mean Squared Error (MSE) and a 15.70% increase in Concordance Index (CI) under cold-both evaluation, along with an 11.04% MSE reduction under cold-drug evaluation. Pearson correlation also improved from 0.0694 to 0.2234 in cold-both and from 0.3606 to 0.4686 in cold-drug settings. However, improvements in random and cold-protein splits were limited. These findings indicate that motif-level attention enhances robustness when prior structural information is limited, providing a promising approach for realistic DTI prediction in computational drug discovery.