Comparing Community Structures in Knowledge Graphs Across Similarity Measures and Clustering Algorithms

Community detection in biomedical knowledge graphs (KGs) holds promise for uncovering functional groupings of drugs, yet its evaluation remains challenging due to the heterogeneity of node types and edge semantics. In this study, we investigate how different definitions of drug similarity—structural, topological, and semantic—influence the community structure of drug-centric graphs and their biological interpretability. We compare clustering outcomes obtained from both homogeneous drug–drug similarity networks and a heterogeneous KG integrating drugs, proteins, and side effects. Using methods such as Louvain, we analyze the extent to which detected communities align with known drug interactions, particularly those associated with adverse drug reactions. Evaluation is performed through structural metrics (e.g., modularity), pairwise agreement measures (ARI, NMI), and functional coherence, quantified by the intra- vs. inter-community distribution of side-effect-inducing drug pairs. Our results offer a systematic assessment of community structures in KGs and provide insights into the utility and limitations of unsupervised clustering in pharmacological network analysis.