Context-Aware Multilevel Classification of Semantic Relations in Drug-Adverse Drug Reaction (ADR) Networks-Predicting...

Area of Science:

• * Computational toxicology
• * Bioinformatics
• * Drug safety science

Background:

• * Adverse drug reactions (ADRs), particularly drug-induced liver injury (DILI), pose significant risks to public health and patient safety.
• * Computational models, including deep learning and knowledge graphs, are crucial for understanding the complex biological mechanisms underlying drug toxicity.
• * Existing knowledge graphs offer valuable data but require sophisticated methods to effectively model biological systems and predict toxic effects.

Purpose of the Study:

• * To develop and evaluate a hierarchical, context-aware knowledge graph model for detecting ADRs, using DILI as a specific case study.
• * To integrate diverse life-science data sources into a comprehensive knowledge graph for enhanced predictive capabilities.
• * To assess the model's performance using established classification metrics and compare it with existing computational approaches.

Main Methods:

• * Construction of a knowledge graph integrating data from over 200 life-science databases, publications, and patents.
• * Application of combined translation- and path-based methods to identify two-hop paths linking drugs to ADRs via intermediate biological concepts (genes, proteins, pathways).
• * Utilization of multilevel statistical modeling on path bundle probabilities to predict DILI risk, accounting for contextual dependencies.

Main Results:

• * The four-level knowledge graph model achieved high performance in DILI prediction, with an AUC of 0.850 and an MCC of 0.639.
• * The model demonstrated good separation in prediction probabilities between DILI-positive (0.641) and DILI-negative (0.402) drugs.
• * A cascaded Quantitative Structure-Activity Relationship (QSAR) and knowledge graph approach successfully reclassified 29 out of 37 misclassified drugs from a previous QSAR study and rescued additional DILI-positive cases from a deep learning model.

Conclusions:

• * The developed hierarchical, context-aware knowledge graph model provides a robust framework for predicting drug-induced liver injury (DILI).
• * Integrating knowledge graphs with QSAR or deep learning methods enhances the accuracy and reliability of adverse drug reaction (ADR) prediction.
• * This approach offers significant potential for improving drug safety assessments and mitigating risks in drug discovery.