Biomedical knowledge graph verification with multitask learning architectures.

Area of Science:

• Biomedical Informatics
• Knowledge Representation and Reasoning
• Machine Learning

Background:

• Large-scale biomedical knowledge graphs (KGs) are crucial for research but often contain errors due to automated construction.
• Erroneous triplets in biomedical KGs can compromise research validity and lead to inaccurate conclusions.
• Effective knowledge graph verification (KGV) methods are needed to ensure the quality of biomedical KGs.

Purpose of the Study:

• To design and evaluate an effective knowledge graph verification (KGV) method for biomedical KGs.
• To enable the identification and removal of erroneous biomedical triplets.
• To improve the overall quality and reliability of biomedical knowledge graphs for downstream applications.

Main Methods:

• A multitask-learning-based KGV (MTL-KGV) method was proposed, involving KG embedding (KGE) learning and triplet classification.
• Three multitask learning (MTL) architectures were explored: hard parameter sharing (HPS), multi-gate mixture-of-experts (MMoE), and customized gate control (CGC).
• The method was evaluated using SemMedDB for KG construction and a dataset of 6,427 expert-annotated triplets.

Main Results:

• All three versions of the MTL-KGV method consistently outperformed benchmark methods in empirical evaluations.
• The MTL-KGV method with the MMoE architecture demonstrated the highest effectiveness in detecting erroneous biomedical triplets.
• The proposed approach significantly improves the accuracy of biomedical knowledge graph verification.

Conclusions:

• This work introduces a novel multitask learning framework specifically tailored for biomedical KGV.
• The MTL-KGV method enhances the quality of biomedical KGs by effectively identifying and removing erroneous data.
• Improved biomedical KG quality supports downstream applications and advances biomedical research reliant on accurate knowledge graphs.