The Alzheimer’s Knowledge Base: A Knowledge Graph for Alzheimer Disease Research

Joseph D. Romano, Van Truong, Rachit Kumar, Mythreye Venkatesan, Britney E. Graham, Yun Hao, Nicholas Matsumoto, Xi Li, Zhiping Wang, Marylyn D. Ritchie, Li Shen, Jason H. MooreJournal of Medical Internet Research 2023 · DOI

TL;DR

Researchers have developed the Alzheimer's Knowledge Base (AlzKB), a comprehensive graph-based resource integrating diverse biomedical data to accelerate Alzheimer's disease (AD) drug discovery. By leveraging machine learning on this knowledge graph, they've successfully identified potential new therapeutic targets and proposed existing drugs for repurposing, offering novel hypotheses for experimental validation.

What Problem Does This Paper Address?

Alzheimer's disease (AD) is a devastating neurodegenerative illness with a significant global health and economic burden. Despite decades of research, effective treatments remain elusive, and existing data resources for drug discovery often fail to capture the complex relationships central to AD's etiology and drug responses. This gap hinders the identification of new therapeutic strategies and the repurposing of existing drugs. The authors aim to bridge this gap by creating a unified, semantically consistent knowledge base that can facilitate more robust and explainable AD research.

Methods

The core of this work is the Alzheimer's Knowledge Base (AlzKB), a large, heterogeneous graph knowledge base. It's built by integrating data from 22 diverse public biomedical databases. To ensure semantic consistency and enable ontological inference, AlzKB utilizes a Web Ontology Language (OWL) 2 ontology. This ontology defines the types of entities (e.g., genes, diseases, drugs) and their relationships within the AD domain.

The construction process involved:

  1. Ontology Design: Manually creating a class hierarchy and defining data and object properties based on expert domain knowledge.
  2. Data Source Integration: Identifying and collecting data from 22 public databases, mapping them to the ontology.
  3. Graph Database Population: Serializing the ontology into RDF/XML format and loading it into a Neo4j graph database using the neosemantics library.

To demonstrate AlzKB's utility, two machine learning-based use cases were implemented:

  1. Predicting Therapeutic Targets: A random forest (RF) classifier was trained using topological graph features (common neighbors, preferential attachment, etc.) to predict whether genes associated with Parkinson's disease (PD) are also linked to AD etiology.
  2. Drug Repurposing: Five knowledge graph completion models (TransE, RotatE, DistMult, ComplEx, and ConvE) were employed to predict new links between drugs and AD, suggesting potential repurposing candidates. The best-performing model, RotatE, was used for final predictions.

Datasets

AlzKB is built upon data from 22 diverse external public biomedical databases. These include sources like the Disease Ontology, DisGeNET, DrugBank, Gene Ontology, GWAS Catalog, PubChem, and Reactome, among others. The paper provides a detailed table (Table 1) listing each source, the data elements used, and their approximate size. AlzKB itself is a newly introduced benchmark resource, differing from existing AD-specific knowledge graphs by its larger scale and comprehensive integration of diverse data sources.

Key Insights and Ideas

  • Unified Knowledge Representation is Crucial: The paper highlights that integrating disparate data sources into a single, semantically consistent knowledge graph is essential for uncovering complex, non-obvious relationships in diseases like AD.
  • Ontology as a Semantic Backbone: Using an OWL 2 ontology is key to enforcing consistency and enabling inferential reasoning within the graph, making the knowledge more robust and interpretable.
  • Graph Topology Holds Predictive Power: Topological features of the graph (how nodes are connected) can be powerful predictors for biological relationships, as demonstrated by the RF model's success in identifying potential AD genes.
  • Drug Repurposing Potential is High: The success of knowledge graph completion models in identifying existing drugs with potential AD efficacy underscores the value of this approach for accelerating drug discovery pipelines.
  • Explainability is a Key Advantage: Unlike "black box" ML models, the graph-based approach allows for the visualization of pathways and intermediate nodes (genes), providing mechanistic explanations for predictions, which is crucial for experimental validation.

Key Results

  • AlzKB Size: The initial release of AlzKB contains 118,902 distinct nodes and 1,309,527 relationships.
  • Therapeutic Target Prediction: The RF model achieved an average balanced accuracy of 96.2% in predicting AD-associated genes. It identified 8 out of 73 PD genes as likely to be associated with AD, including genes like FYN and SNCA.
  • Drug Repurposing Predictions: The RotatE model identified 10 potential drugs for AD repurposing. Notably, 30% (3 out of 10) of these drugs have already been investigated in clinical trials for AD symptoms. Examples include Sumatriptan, Risperidone, and Sertraline.
  • Mechanistic Explanations: Visualizations of the shortest paths between predicted drugs and AD in the knowledge graph revealed AD-associated genes that mediate these relationships, offering plausible biological mechanisms.

Limitations and Future Work

  • Data Provenance and Noise: While efforts were made to clean the graph, the inherent noise and potential inconsistencies from diverse data sources remain a challenge.
  • Negative Sample Definition: Defining "negative samples" (pairs of entities with no known relationship) in knowledge graphs is inherently difficult, as absence of evidence is not evidence of absence. New methods are being explored to address this.
  • Gene vs. Protein Distinction: A key future direction is to better distinguish between genes and the proteins they encode, as AD pathogenesis involves complex protein interactions.
  • Heterogeneous Graph ML: Developing ML models that can effectively generalize to and learn from highly heterogeneous graphs like AlzKB is an ongoing area of research.
  • Explainability of LLMs: Integrating knowledge graphs with large language models (LLMs) to improve their understanding of complex domains like AD is a promising future avenue.
  • Community Effort: The authors emphasize that AlzKB is intended as a community resource, encouraging contributions and discussions to expand its scope and utility.

Summary updated

Abstract

AlzKB is a new, publicly available knowledge resource that enables researchers to discover complex translational associations for AD drug discovery. Through 2 use cases, we show that it is a valuable tool for proposing novel therapeutic hypotheses based on public biomedical knowledge.
BibTeX
@article{romano2023alzheimers,
  author = {Joseph D. Romano and Van Truong and Rachit Kumar and Mythreye Venkatesan and Britney E. Graham and Yun Hao and Nicholas Matsumoto and Xi Li and Zhiping Wang and Marylyn D. Ritchie and Li Shen and Jason H. Moore},
  title = {{The Alzheimer’s Knowledge Base: A Knowledge Graph for Alzheimer Disease Research}},
  journal = {Journal of Medical Internet Research},
  year = {2023},
  doi = {10.2196/46777},
}
Figure