Graph embedding-based link prediction for literature-based discovery in Alzheimer's Disease

Area of Science:

• Computational biology
• Bioinformatics
• Artificial intelligence in medicine

Background:

• Literature-based discovery (LBD) aims to extract novel insights from existing scientific literature.
• Alzheimer's Disease (AD) research generates a vast and complex body of literature, making manual knowledge extraction challenging.
• Framing LBD as link prediction on knowledge graphs offers a computational approach to uncover hidden relationships.

Purpose of the Study:

• To explore literature-based discovery (LBD) for Alzheimer's Disease (AD) using link prediction and graph embedding techniques.
• To evaluate the impact of prediction window length within a time-sliced evaluation methodology for LBD.
• To assess the generalizability of the proposed approach to other disease contexts.

Main Methods:

• Constructed an AD-specific corpus and knowledge graph (∼11k nodes, ∼394k edges) from over 16k papers (1977-2021).
• Employed a four-stage approach including corpus collection, knowledge graph construction, time-sliced dataset creation (20 pairs), and link prediction.
• Utilized graph embedding-based link prediction methods, comparing performance across different temporal evaluation settings.

Main Results:

• The Structural Deep Network Embedding (SDNE) model demonstrated superior performance in link prediction accuracy over a 20-year progression.
• Model performance significantly improved when evaluating predictions over longer future time windows, reflecting knowledge accumulation.
• The study achieved a graph density of less than 1%, indicating the challenging nature of the link prediction task.

Conclusions:

• Neural network graph-embedding link prediction methods show promise for LBD, despite inherent challenges.
• The choice of prediction window length in time-sliced evaluations critically impacts LBD model interpretation.
• The developed methodology is adaptable for knowledge discovery in other disease domains.