Abstract
Effective and successful clinical trials are essential in developing new drugs and advancing new treatments. However, clinical trials are very expensive and easy to fail. The high cost and low success rate of clinical trials motivate research on inferring knowledge from existing clinical trials in innovative ways for designing future clinical trials. In this manuscript, we present our efforts on constructing the first publicly available Clinical Trials Knowledge Graph, denoted as CTKG. CTKG includes nodes representing medical entities in clinical trials (e.g., studies, drugs and conditions), and edges representing the relations among these entities (e.g., drugs used in studies). Our embedding analysis demonstrates the potential utilities of CTKG in various applications such as drug repurposing and similarity search, among others.
Introduction
Clinical trials are often described as the "valley of death" for drug development. They are staggeringly expensive—with costs reaching up to $52.9 million per phase—and carry a high risk of failure. In fields like oncology, success rates can dip as low as 3.4%. While the US government maintains the Access to Aggregate Content of ClinicalTrials.gov (AACT) database, its traditional relational structure is optimized for specific fact retrieval rather than discovering hidden patterns across the clinical ecosystem.
To address this, we developed the Clinical Trials Knowledge Graph (CTKG). By transforming flat relational data into a multi-dimensional network of medical entities and their relationships, CTKG enables machine learning models to perform complex reasoning and inference, potentially accelerating drug repurposing and improving future trial designs.
Methodology
Data Normalization and Cleaning
Raw clinical trial data is notoriously heterogeneous. To create a functional graph, we performed extensive normalization:
- Entity Standardization: We utilized the Medical Text Indexer (MTI) and MeSH codes to standardize drug names.
- Hierarchical Mapping: Adverse events were mapped to the MedDRA hierarchy, allowing the graph to group specific symptoms into broader clinical categories.
- NLP Processing: We employed NLTK and Stanza to clean statistical method names and other free-text fields.
Outcome Clustering
Trial outcomes are frequently recorded in non-standardized free-text (e.g., "Change from baseline in Platelet Count"). We used the CLUTO clustering toolkit to group these outcomes into 200 distinct clusters. This process successfully linked 56.96% of all outcome records to standardized nodes, significantly reducing noise and allowing the graph to connect studies that measured similar clinical endpoints using different terminology.
Graph Embedding
To enable mathematical analysis of the graph, we used the TransE algorithm (via the DGL-KE library) to project the CTKG into a 200-dimensional vector space. In this embedding space, the "closeness" of two nodes (measured by cosine similarity) represents their clinical or functional similarity.
The CTKG Framework
The resulting CTKG is a comprehensive map of the clinical trial landscape, featuring:
- 18 Node Types: Including Study, Condition, Drug Term, Adverse Event, and Organ System.
- 21 Edge Types: Representing relationships such as
StudiedDrug,Outcome-Measurement, andReportedAdverseEvent. - Dataset Scope: The graph is built from 8,210 studies registered between 1999 and 2020 that included both drug interventions and reported outcomes.
Results and Validation
Semantic Similarity
Our embedding analysis confirms that CTKG captures real-world clinical context without explicit instruction. For example, two independent studies on Chronic Obstructive Pulmonary Disease (COPD)—NCT01431274 and NCT01431287—exhibited a high cosine similarity of 0.826.
Link Prediction and Drug Repurposing
In similarity searches, the top-10 most similar drug pairs in the embedding space were almost all clinically related. A notable example is the relationship between the drug ABT-267 and Macrocyclic Compounds, which showed a 0.997 similarity; both are integral to the treatment of Hepatitis C. Such findings suggest that CTKG can identify potential new uses for existing drugs by analyzing their proximity to specific conditions and outcomes within the graph topology.
Discussion and Future Work
CTKG represents a significant step toward making clinical trial data "inference-ready." However, several limitations remain. The current version focuses exclusively on drug interventions, omitting medical device trials and incomplete studies. Furthermore, the graph currently lacks molecular-level data, such as protein-protein interactions or drug-gene pathways.
Future iterations of this work will focus on integrating CTKG with larger biomedical knowledge bases (such as DRKG). By linking clinical trial results directly to molecular biology, we aim to create a "master graph" that spans the entire spectrum from laboratory research to clinical application.
Summary updated