Abstract
Alzheimer’s Disease (AD) poses a massive global health challenge. While pharmaceutical research is extensive, non-pharmacological interventions (NPIs) like dietary supplements and complementary therapies remain under-explored. This study introduces ADInt, a comprehensive biomedical knowledge graph, and utilizes Relational Graph Convolutional Networks (R-GCN) to predict novel NPIs for AD. By integrating specialized knowledge graphs with semantic relations from PubMed, the researchers identified promising candidates and used discovery patterns to provide biological explanations for these predictions.
Introduction
Alzheimer’s Disease (AD) is a progressive neurodegenerative disorder with limited treatment options. While drug repurposing often focuses on existing pharmaceuticals, NPIs—such as Vitamin E, aerobic exercise, and light therapy—offer potential for improving cognitive function with fewer side effects. However, the vast amount of biomedical literature makes it difficult to systematically identify and validate these interventions. This research addresses this gap by applying link prediction on a massive knowledge graph to discover and explain new NPI candidates.
Methodology: Building the ADInt Knowledge Graph
The researchers constructed ADInt, a heterogeneous knowledge graph containing 162,213 nodes and over 1 million edges. ADInt was built by merging SuppKG (dietary supplements) with semantic relations from SemMedDB, a database of triples extracted from millions of PubMed abstracts. To ensure data quality, the team used rule-based filtering and a fine-tuned PubMedBERT model to prune incorrect relations.
The study compared six modeling architectures, including embedding-based models (TransE, RotatE) and Graph Neural Networks (R-GCN, CompGCN). The models were evaluated using a "Time Slicing" test set (chronological split) and a "Clinical Trials" test set derived from real-world interventional studies.
Results: R-GCN and Top NPI Candidates
The R-GCN model emerged as the superior architecture, achieving a Hits@10 of 0.9848 on the Clinical Trials test set. This indicates that the model's top predictions highly correlate with interventions currently being tested in human trials. Top-ranked candidates included dietary supplements like Lutein and Artichoke, as well as complementary therapies like Photodynamic therapy and Mindfulness Relaxation.
Explainability: Uncovering Biological Mechanisms
A key contribution of this work is the use of Discovery Patterns to solve the "black box" nature of neural networks. By identifying multi-hop paths within the graph, the researchers could explain why an intervention might work. For instance, the model predicted that Choerospondias axillaris (Nepali hog plum) could prevent AD by inhibiting pro-inflammatory cytokines like Interleukin-6 and TNF protein.
Similarly, Photochemotherapy was linked to AD prevention through the modulation of immune responses. These pathways provide a biological rationale that bridges the gap between computational prediction and clinical application.
Discussion and Future Work
While the results are promising, some interventions face practical hurdles. For example, delivering Photodynamic therapy to the human brain is surgically complex. Furthermore, validating these AI-generated pathways still requires expert manual review. Future research will focus on expanding ADInt with molecular networks like DrugBank and KEGG and exploring reinforcement learning to further refine link prediction accuracy.
Reference: Xiao, Y., et al. (2023). Repurposing Non-pharmacological Interventions for Alzheimer's Diseases through Link Prediction on Biomedical Literature. medRxiv preprint.
Summary updated