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 store known relationships but miss unexplored associations.
• Identifying unknown disease-associated genes is critical for advancing biomedical research.
• Existing methods are time-consuming, necessitating efficient computational approaches.

Purpose of the Study:

• To develop an efficient computational approach for identifying novel disease-associated genes.
• To overcome limitations of existing machine learning pipelines for knowledge graph analysis.
• To infer previously unknown gene-disease relationships using advanced machine learning.

Main Methods:

• Developed KG2ML (Knowledge Graph to Machine Learning) pipeline, utilizing a novel Positive and Unlabeled (PU) learning algorithm, PULSCAR (Positive Unlabeled Learning Selected Completely At Random).
• Incorporated path-based feature extraction from ProteinGraphML within the KG2ML pipeline.
• Applied KG2ML to 12 diseases to infer novel disease-associated genes not present in the DDKG.

Main Results:

• KG2ML identified top-ranked novel disease-associated genes for 12 diseases, with 14 out of 15 lacking prior explicit associations in DDKG.
• Identified genes showed support from literature and TINX (Target Importance and Novelty Explorer) evidence.
• Incorporating PULSCAR-imputed genes as positives improved XGBoost classification performance.

Conclusions:

• Positive and Unlabeled (PU) learning effectively uncovers disease-gene associations missing from existing knowledge graphs (KGs).
• The KG2ML pipeline offers a scalable and interpretable framework for biomedical research.
• Integrating KG data with ML-based inference advances biomedical discovery by addressing KG limitations.