Fig. 1: Schematic diagram of the RareDDIE architecture.

a The calculation process of RareDDIE based on our five constructed modules. Feature Initialization (FIN) utilizes the knowledge graph DRKG⁵⁶ and the knowledge graph embedding method TransE⁵⁹ to initialize features; Chemical Substructure Information Extraction (CSE) employs message passing mechanism and Self-Attention Graph Pooling (SAGPooling) to learn drug chemical granular structure information; Neighborhood Adaptive Integration with Task Guidance (NAI); Pair Variational Representation (PVR) and the Comparator Module. b The details of the NAI module, showing how weak event representations, built from chemical structure information, guide the aggregation of neighborhood information. This results in the creation of dual-granular structure features by integrating drug chemical structure information with biological neighborhood structure information. c The illustration of the PVR module, which uses a Variational Autoencoder (VAE)-based approach to construct an effective general relation metric space and automatically form latent event semantic information. Std represents the standard deviation, and N(0,1) denotes the standard normal distribution, which is used for noise sampling.