tbkg

is a Knowledge Graph.

A weighted heterogeneous knowledge graph containing four types of entities (Tumor, Biomarker, Drug, and ADR) extracted from MEDLINE corpus for adverse drug reaction discovery in antitumor drugs. TBKG uses a naive Bayesian model to explore correlations and provides explainable predictions through tumor-biomarker-drug pathways. The knowledge graph contains 1,179 tumors, 2,550 biomarkers, 1,806 drugs, and 756 ADRs with six types of relationships totaling 139,254 edges.

Compare

Infores ID

Unknown

FAIRsharing ID

Unknown

Product Summary

Publications

Adverse Drug Reaction Discovery Using a Tumor-Biomarker Knowledge Graph Preferred

Products

From this Resource

ID	Name	URL	Category	Format	Description
tbkg.data	TBKG Knowledge Graph Data	full#supplementary-material	Graph	mixed	Weighted heterogeneous knowledge grap...
tbkg.osimertinib_case_study	TBKG Osimertinib ADR Case Study Data	full#supplementary-material	Product	mixed	Clinical validation dataset with calc...

Relevant Taxa

Human (NCBITaxon:9606)

Details

Overview

TBKG (Tumor-Biomarker Knowledge Graph) is an explainable knowledge graph-based approach for discovering potential adverse drug reactions (ADRs) of antitumor drugs. The system extracts entities from biomedical literature (MEDLINE database with 22+ million citations) using the UMLS Metathesaurus 2020AA and Apache cTAKES natural language processing tool.

Key Features

Entity Types: Four node types (Tumor, Biomarker, Drug, ADR) with minimum frequency threshold of 50
Knowledge Graph Structure: Weighted heterogeneous graph with undirected edges representing correlations
Relationship Discovery: Naive Bayesian model combining prior and posterior probabilities to avoid bias
ADR Discovery: Depth-first search algorithm to find all paths between drug-ADR pairs
Explainability: Provides tumor-biomarker-drug pathways explaining predicted ADRs
Performance: 0.81 accuracy in three-fold cross-validation, outperforms co-occurrence analysis

Data Sources

MEDLINE database (1928-2020, filtered with “cancer therapy” keyword)
UMLS Metathesaurus 2020AA for entity dictionaries
WHO source dictionary for ADR coding
Clinical validation from 3rd Xiangya Hospital (8 patients, 2017-2020)

Technical Implementation

Entity Extraction: cTAKES NLP system with UMLS concept mapping
Importance Measure: log(p(biomarker tumor present)) - log(p(biomarker tumor absent))
Graph Construction: Binary matrix representation (0/1 for entity presence in abstracts)
Cross-validation: Three-fold to prevent overfitting

Clinical Validation

Osimertinib case study demonstrated:

Moderate consistency with official manual (Kappa=0.68)
Better specificity than co-occurrence methods (Kappa=0.4)
Discovery of rare/unreported ADRs (e.g., renal failure requiring dialysis)
Identification of mediating biomarkers (e.g., Macrophage-Activating Factors for nephrosclerosis)

Applications

Early ADR detection before drug development
Mechanism research through biomarker pathways
Clinical decision support for oncologists
Literature-based drug safety surveillance
Rare ADR discovery and explanation