MAGC-DTI: modality-shared space and adaptive gated interactive cross-attention for drug–target interaction prediction

Abstract

Abstract Accurate drug–target interaction (DTI) prediction is crucial for drug repurposing and accelerating drug development. Although deep learning has advanced DTI prediction, existing methods struggle with two key challenges: (i) capturing complex hierarchical patterns in protein sequences, and (ii) enabling effective bidirectional information exchange between drug and protein modalities. We propose MAGC-DTI, an end-to-end cross-modal framework that integrates bidirectional information exchange into both feature extraction and fusion stages through three key innovations: (i) multi-scale attention aggregation (MSAA) for hierarchical protein pattern capture, (ii) adaptive gated interactive cross attention (AGICA) for context-aware cross-modal interaction, and (iii) multi-path residual classifier (MPRC) for modality-preserving fusion. Comprehensive evaluations on six benchmark datasets show that MAGC-DTI generally achieves favorable performance relative to seven state-of-the-art baselines, with competitive results in cold-start and cross-domain scenarios. The model also provides interpretable insights through attention visualization and case studies confirm the biological relevance of learned representations.