This research, published in BMC Bioinformatics, represents a significant advancement in medical data science by bridging the gap between Western biomedicine and Traditional Chinese Medicine (TCM). By leveraging advanced natural language processing (NLP) on a massive corpus of research, the authors developed StrokeKG, a comprehensive knowledge graph designed to unify these disparate medical traditions.
The Challenge: Siloed Medical Knowledge
Stroke is a leading cause of mortality and long-term disability globally. While both Western biomedicine and TCM offer extensive research and treatments, they typically exist in silos. Western databases focus on synthetic drugs and molecular pathways, whereas TCM databases emphasize complex herbal formulas and natural ingredients. With over 327,000 papers on PubMed related to stroke, manual curation is no longer feasible. There is a critical need for automated tools that can extract, normalize, and link entities across these two traditions to uncover new therapeutic directions.
Methodology
The researchers designed a multi-stage computational workflow to transform unstructured literature into a structured knowledge graph. The process began with the acquisition of over 124,000 abstracts from PubMed using stroke and TCM-related keywords.
Following data acquisition, the team employed an ensemble of Named Entity Recognition (NER) tools to identify specific concepts:
- DNorm and tmChem for diseases and chemicals.
- GNormPlus and PWTEES for genes and biological pathways.
- BiLSTM-CRF and PKDE4J (a rule-based system) were specifically adapted to identify TCM entities like Chinese Patent Medicines (CPMs) and herbs, which standard biomedical tools often overlook.
Relation Extraction
To determine how these entities are linked (e.g., identifying which drug treats a specific symptom), the authors combined three distinct approaches: co-occurrence analysis, rule-based parsing using dependency trees, and deep learning. A pre-trained BioBERT model was fine-tuned to classify 11 distinct relationship types, such as Treatment, Cause, and Positive Association.
Key Insights and Results
The final StrokeKG was integrated into a Neo4j graph database, consisting of 46,983 nodes and 157,302 relationships.
Bridging the Silos
The most innovative aspect of StrokeKG is its "common ID space," which allows a TCM herb to be linked directly to a Western molecular pathway. For instance, the graph can demonstrate how a specific ingredient in Danggui (Angelica) interacts with genes like VEGF, which are known to be involved in stroke recovery. This multi-level connectivity allows for "indirect" discovery—finding traditional herbs that target specific genetic risk factors identified in Western medicine.
Performance Metrics
- NER Accuracy: The rule-based tool for identifying Chinese Patent Medicines achieved an impressive F1-score of 90.06%.
- Relation Extraction: The hybrid PKDE4J + BioBERT approach achieved an F1-score of 84.26% for classifying CPM-Disease relationships.
- Entity Discovery: The system successfully mined 4,210 distinct diseases and 728 types of symptoms, revealing a complex "comorbidity web" linking stroke to conditions like diabetes and atrial fibrillation.
Limitations and Future Outlook
Despite its success, the authors noted challenges regarding linguistic complexity, particularly the lack of standardized abbreviations and varied transliterations for Chinese herbs. Additionally, the model occasionally struggled to distinguish between "Treatment" and "Cause" in highly complex sentences.
Future work will focus on optimizing the mining model to cover other disease domains, with the ultimate goal of building a comprehensive, cross-tradition medical knowledge map.
Source: Yang et al. "Mining a stroke knowledge graph from literature." BMC Bioinformatics (2021) 22:387. https://doi.org/10.1186/s12859-021-04292-4
Summary updated