fig1
From: Drug-target interaction prediction via hierarchical gated attention and information bottleneck
Figure 1. (A) Overall architecture of HGMAIB: (a) Node representation initialization - Node2Vec extracts drug and target features. (b) Adaptive meta-path search - A dynamic search strategy identifies informative semantic paths. (c) Multi-semantic structural modeling - Multi-step residual graph convolution and hierarchical gated multi-head attention (HGMA) capture deep dependencies. (d) Information bottleneck - Learned embeddings are refined to filter redundant noise. (B) HGMA stacking process: The hierarchical architecture performs weighted aggregation within each attention head, followed by a gating mechanism that integrates information across all heads.
