KG2ML: Integrating Knowledge Graphs and Positive Unlabeled Learning for Identifying Disease-Associated Genes with Case Studies for 12 Diseases

Objective: To develop a machine learning (ML) pipeline for identifying potential gene-disease associations, even when explicit connections are absent in knowledge graphs. Background: Biomedical knowledge graphs (KGs), such as the Data Distillery Knowledge Graph (DDKG), capture known relationships among entities (e.g., genes, diseases, proteins), providing valuable insights for research. However, these relationships are typically derived from prior studies, leaving potential unknown associations unexplored. Traditional methods, like linkage analysis and genome-wide association studies (GWAS), can be time-consuming and resource-intensive. Recently, network-based methods and KGs, enhanced by advances in ML frameworks, have emerged as powerful tools for inferring these unexplored associations. Methods: We propose KG2ML (Knowledge Graph to Machine Learning), a pipeline utilizing the PULSNAR algorithm (Positive Unlabeled Learning Selected Not At Random) to infer disease-associated genes not explicitly recorded in the DDKG. CondensedKG, a subset of DDKG with 8 node types and 1,042 relationship types, was created to improve data interpretability. KG2ML transforms CondensedKG into a feature matrix using ProteinGraphML’s path-based method, and then PULSNAR generates probabilistic scores for gene-disease associations. KG2ML was applied to 12 diseases, including Bipolar Disorder, Coronary Artery Disease, and Parkinson's Disease. The top 15 high-probability genes per disease were validated through expert review and reference databases (e.g., TINX). XGBoost models were evaluated with 5-fold cross-validation to measure the impact of incorporating PULSNAR-imputed genes as positives. Results: For some of those 12 diseases, 14 of the 15 top-ranked genes lacked prior explicit associations but were supported by literature and TINX evidence. Incorporating PULSNAR-imputed genes as positives enhanced XGBoost classification, highlighting PU learning's potential in identifying hidden gene-disease relationships. Conclusions: KG2ML demonstrates that PU learning can effectively discover disease-gene associations missing from current KGs. By combining KG data with ML-based inference, it offers a scalable, interpretable strategy to advance biomedical research and address KG limitations.