Abstract
Alzheimer’s Disease (AD) presents a significant global health challenge with limited therapeutic options. This paper details a computational framework that leverages literature mining and knowledge graphs (KGs) to identify existing substances—including drugs, chemicals, and dietary supplements—that may be repurposed for AD. By integrating semantic triple extraction from PubMed with advanced graph embedding techniques and BERT-based denoising, the researchers identified several high-probability candidates, many of which have been validated by subsequent clinical literature.
Introduction
Alzheimer’s Disease (AD) remains a global health crisis with no effective cure. Developing new drugs from scratch is an incredibly slow, expensive, and high-failure-rate process. Drug repurposing—finding new uses for already approved substances—offers a faster, safer alternative. However, the "knowledge bottleneck" is a significant hurdle: millions of biomedical research papers exist, and no human can synthesize every connection between disparate chemicals and neurodegenerative pathways. This research addresses the challenge of automatically mining this vast "dark data" to find hidden therapeutic links that researchers might have missed.
Methodology
The authors developed a multi-stage pipeline to transform unstructured biomedical text into a searchable, predictive map of medical knowledge.
- Extraction and Normalization: The team extracted over 113 million "semantic triples" (e.g., Drug A — treats — Disease B) from SemMedDB, a database derived from PubMed abstracts. They utilized the Unified Medical Language System (UMLS) to normalize medical terms and the Integrated Dietary Supplement Knowledge Base (iDISK) to categorize supplements.
- Denoising with BERT: To ensure data quality, the researchers employed a two-step filtering process. After initial rule-based filtering to remove generic concepts, they fine-tuned a PubMedBERT classifier on ~6,000 manually labeled triples. This step verified if the extracted triples were accurately supported by their original source sentences.
- Knowledge Graph Completion: The resulting graph was analyzed using three embedding algorithms: TransE, DistMult, and ComplEx. These models project entities into a vector space, allowing the system to predict missing links (potential new drug-disease pairs) based on the geometric distance between entities.
- Evaluation: To validate the model's predictive power, the authors used a "time-slicing" approach, training the model on data published before 2019 and testing its ability to predict findings that only appeared in literature published after that date.
Results and Model Performance
The evaluation revealed that simpler models often outperformed more complex ones in this specific biomedical context. TransE, a translational model, achieved the best results, suggesting that the geometric relationship between entities is a powerful signal for biomedical knowledge graphs.
As shown in the performance table, TransE achieved a Mean Rank (MR) of 10.53 and a Hits@10 of 0.58, indicating that the correct therapeutic relationship appeared in the model's top 10 suggestions 58% of the time.
Beyond performance metrics, the model identified several high-probability candidates for AD repurposing. Notable drug candidates included Valacyclovir (an antiviral) and Prednisolone (a steroid). In the category of chemicals, Amifostine—typically used in cancer therapy—emerged as a top-ranked candidate for AD prevention. The study also highlighted the role of lifestyle factors, identifying dietary fiber, tea, and honey as having significant "affect" relationships with AD progression.
Validation against recent literature confirmed the framework's efficacy: supporting evidence was found for 7 out of 10 drug predictions and 8 out of 10 chemical predictions.
Discussion and Future Work
One of the key insights of this study is the essential role of the "BERT filter." Raw NLP-extracted data is inherently noisy; using a transformer-based model like PubMedBERT to calibrate the graph significantly improved the reliability of the resulting predictions. Furthermore, by explicitly including dietary supplements and chemicals, the framework acknowledges that AD management often involves a combination of clinical and lifestyle interventions.
Despite these successes, limitations remain. Computational constraints limited the final graph to 2.8 million triples out of the 113 million available. Future work will focus on scaling the model to incorporate more data and investigating embedding strategies that better handle "one-to-many" relationships, such as a single drug affecting multiple biological pathways. Additionally, the researchers aim to refine the detection of specific relationship types, moving from general "affect" links to more precise "treat" or "prevent" classifications for dietary supplements.
Summary updated