PD-KG: A Multi-Omic Knowledge Graph for Parkinson’s Disease Discovery

Parkinson’s disease (PD) is a heterogeneous neurodegenerative disorder with substantial variation in clinical progression, pathology, and underlying molecular drivers. There remains a limited set of robust biomarkers and no unified resource that adequately integrates the disparate multi-omic data needed to study this complexity. To address this gap, we developed PD-KG, a comprehensive Parkinson’s disease knowledge graph that integrates 15 external public databases together with proteomic, transcriptomic, clinical, and GWAS-derived datasets into a large heterogeneous network comprising 222k nodes and 6.7 million edges. We applied network-based target prioritization using Random Walk with Restart and MultiXrank to validate the underlying knowledge represented in this resource. We show that PD-KG recovers known PD targets and further supports mechanistic contextualization through pathway-level analyses. We next apply graph representation learning to PD-KG using a Heterogeneous Graph Transformer to perform edge prediction across diverse biological entity types, showing that the graph captures neurodegenerative disease-relevant biological structure in learned embeddings. Together, these results suggest that PD-KG provides a useful foundation for genomics- and phenotype-driven discovery in Parkinson’s disease. Future directions include applying the graph and learned embeddings to patient subtyping, improved biomarker discovery, and identification of novel therapeutic targets.