Knowledge-graph embeddings for osteoarthritis candidate prediction

Area of Science:

• Biomedical Informatics
• Computational Biology
• Genetics

Background:

• Osteoarthritis (OA) is a widespread and debilitating joint condition lacking disease-modifying treatments.
• Current therapeutic strategies primarily manage OA symptoms rather than addressing underlying disease mechanisms.
• Integrating large-scale biomedical data offers a promising avenue for discovering novel OA treatments.

Purpose of the Study:

• To develop and validate a knowledge-graph based computational approach for identifying candidate OA treatments.
• To leverage causal genetic associations from genome-wide association studies (GWAS) to enhance drug repurposing for OA.
• To create a specialized Osteoarthritis Knowledge-graph (OKG) integrating diverse biomedical information.

Main Methods:

• Construction of the Osteoarthritis Knowledge-graph (OKG) by combining the Drug Repurposing Knowledge-graph (DRKG) with causal genetic associations from a large-scale OA GWAS.
• Development of CausalPathKG, a novel knowledge-graph embedding model based on RotatE, incorporating weighted edges, path-based regularization, multi-hop graph attention, and adversarial negative sampling.
• Training and evaluation of CausalPathKG for link prediction, specifically predicting missing OA treatments, using held-out known OA-related edges for testing.

Main Results:

• CausalPathKG demonstrated superior performance in predicting held-out OA treatments compared to baseline models (TransE, RotatE), achieving higher link prediction accuracy.
• The model exhibited strong classification performance in identifying potential OA therapeutics.
• Case studies confirmed that top-ranked repurposed drugs identified by CausalPathKG target key OA-associated genes and pathways implicated by human genetics.

Conclusions:

• Integrating human genetic evidence into knowledge-graph models significantly enhances the discovery of therapeutics for complex diseases like OA.
• CausalPathKG provides a robust computational strategy for drug repurposing by effectively bridging genomic data with existing drug knowledge.
• This approach offers a promising pathway to accelerate the development of disease-modifying treatments for osteoarthritis.