Alzheimer's disease knowledge graph enhances knowledge discovery and disease prediction

Area of Science:

• Biomedical Informatics
• Computational Biology
• Neuroscience

Background:

• Alzheimer's disease (AD) research involves complex, interconnected data from diverse sources.
• Integrating information on genes, drugs, and diseases is crucial for advancing AD understanding and treatment.
• Existing knowledge extraction methods often struggle to capture the full spectrum of AD-related relationships.

Purpose of the Study:

• To construct a comprehensive Alzheimer's Disease Knowledge Graph (ADKG) by extracting and integrating relationships from biomedical literature.
• To identify potential therapeutic targets, existing treatments, and diagnostic methods for Alzheimer's disease (AD).
• To leverage the ADKG for predictive modeling and advancing precision medicine in AD research.

Main Methods:

• Annotated 800 PubMed abstracts (ADERC corpus) with entities and relationships, augmented using GPT-4.
• Employed a SpERT model (SciBERT-based) for relation extraction, integrating biomedical databases and entity linking.
• Trained graph embedding models on the ADKG to predict novel relationships and validated utility with UK Biobank data.

Main Results:

• The ADKG comprises over 3.1 million entity mentions and 633,733 triplets, linking more than 5,000 unique entities.
• Graph embedding models generated evidence-supported predictions, facilitating the formulation of testable hypotheses.
• ADKG-enhanced predictive models for UK Biobank data achieved a higher AUROC (0.928) compared to non-enhanced models (0.903).

Conclusions:

• The ADKG provides a computable framework linking molecular mechanisms to clinical phenotypes in Alzheimer's disease.
• This structured knowledge synthesis advances precision medicine by accelerating therapeutic discovery and improving risk stratification for AD.
• The ADKG demonstrates significant potential for enhancing predictive modeling and guiding future Alzheimer's research efforts.