KG2ML: Integrating Knowledge Graphs and Positive Unlabeled Learning for Identifying Disease-Associated Genes.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Biomedical knowledge graphs (KGs) like DDKG map known gene-disease relationships but miss unexplored associations.
• Identifying novel disease-associated genes is crucial but challenging with traditional methods.
• Machine learning (ML) and KGs offer promising avenues for inferring unknown associations.

Purpose of the Study:

• To develop an efficient computational approach for predicting novel disease-associated genes.
• To overcome limitations of existing ML pipelines for KG analysis.
• To identify previously unknown gene-disease links.

Main Methods:

• Developed KG2ML, a novel ML pipeline integrating Positive and Unlabeled (PU) learning (PULSNAR) with path-based feature extraction.
• Applied KG2ML to 12 diseases to infer missing gene associations from the Data Distillery Knowledge Graph (DDKG).
• Utilized ProteinGraphML for feature extraction and XGBoost for classification.

Main Results:

• KG2ML identified potential disease-associated genes for 12 diseases, with many lacking prior explicit links in DDKG.
• Top-ranked imputed genes were supported by literature and TINX evidence.
• Incorporating PULSNAR-imputed genes improved XGBoost classification performance, validating PU learning's potential.

Conclusions:

• PU learning, via KG2ML, effectively uncovers disease-gene associations absent in current KGs.
• The KG2ML pipeline offers a scalable and interpretable framework for biomedical research.
• This approach enhances KG utility and advances the discovery of novel gene-disease relationships.