KNDM: A Knowledge Graph Transformer and Node Category Sensitive Contrastive Learning Model for Drug and Microbe Association Prediction

It has been proven that the microbiome in human bodies can promote or inhibit the treatment effects of the drugs by affecting their toxicities and activities. Therefore, identifying drug-related microbes helps in understanding how drugs exert their functions under the influence of these microbes. Most recent methods for drug-related microbe prediction are developed based on graph learning. However, those methods fail to fully utilize the diverse characteristics of drug and microbe entities from the perspective of a knowledge graph, as well as the contextual relationships among multiple meta-paths from the meta-path perspective. Moreover, previous methods overlook the consistency between the entity features derived from the knowledge graph and the node semantic features extracted from the meta-paths. To address these limitations, we propose a <b>k</b>nowledge-graph transformer and <b>n</b>ode category-sensitive contrastive learning-based <b>d</b>rug and <b>m</b>icrobe association prediction model (KNDM). This model learns the diverse features of drug and microbe entities, encodes the contextual relationships across multiple meta-paths, and integrates the feature consistency. First, we construct a knowledge graph consisting of drug and microbe entities, which aids in revealing similarities and associations between any two entities. Second, considering the heterogeneity of entities in the knowledge graph, we propose an entity category-sensitive transformer to integrate the diversity of multiple entity types and the various relationships among them. Third, multiple meta-paths are constructed to capture and embed the semantic relationships based on similarities and associations among drug and microbe nodes. A meta-path semantic feature learning strategy with recursive gating is proposed to capture specific semantic features of individual meta-paths while fusing contextual relationships among multiple meta-paths. Finally, we develop a node-category-sensitive contrastive learning strategy to enhance the consistency between entity features and node semantic features. Extensive experiments demonstrate that KNDM outperforms eight state-of-the-art drug-microbe association prediction models, while ablation studies validate the effectiveness of its key innovations. Additionally, case studies on candidate microbes associated with three drugs-curcumin, epigallocatechin gallate, and ciprofloxacin-further showcase KNDM's capability to identify potential drug-microbe associations.