Investigating ADR mechanisms with Explainable AI: a feasibility study with knowledge graph mining.

Area of Science:

• Pharmacogenomics
• Bioinformatics
• Computational Toxicology

Background:

• Adverse drug reactions (ADRs) are common, but their molecular mechanisms are often unknown, even for well-monitored toxicities like liver and skin reactions.
• Existing knowledge graphs contain rich information on drug properties, interactions, and pathways, alongside classifications of drugs causing specific ADRs.

Purpose of the Study:

• To mine knowledge graphs for biomolecular features that can automatically classify drugs as causative or not for specific ADRs.
• To utilize Explainable AI (XAI) techniques, like decision trees and classification rules, to generate human-readable models for ADR classification and mechanistic insights.
• To identify and evaluate features that are both accurate in classification and interpretable by domain experts.

Main Methods:

• Mining knowledge graphs to extract relevant biomolecular features.
• Training classification models (Decision Trees, Classification Rules) to distinguish drugs associated with drug-induced liver injuries (DILI) and severe cutaneous adverse reactions (SCAR).
• Isolating and evaluating interpretable features such as Gene Ontology terms, drug targets, and pathway names for their explanatory power.

Main Results:

• Classification models achieved good accuracy in identifying drugs associated with DILI (0.74) and SCAR (0.81).
• Domain experts found a significant portion of the most discriminative features to be potentially explanatory for DILI (73%) and SCAR (38%), with high agreement for partial explanation (90% and 77%).

Conclusions:

• Knowledge graphs offer diverse features suitable for developing simple, explainable models to predict ADRs.
• The identified discriminative features serve as valuable candidates for further investigation into the molecular mechanisms of ADRs.