Mobile Menu

Identifying Adverse Drug Reaction Mechanisms

A recent article proposed the mining of knowledge graphs to identify biomolecular adverse drug reaction mechanisms, which at present largely remain unknown.

Adverse drug reaction mechanisms

Researchers are able to statistically characterise adverse drug reactions within randomised clinical trials and post marketing pharmacovigilance. However, the molecular mechanisms behind the harmful or beneficial effects of drugs are largely unknown. Understanding the mechanism of a drug can help to guide drug development, improve drug safety and enable precision medicine through better dosing or combination of drugs.

Knowledge graphs hold important information about drugs characteristics, such as their chemical and physical properties, their interactions with biomolecules (such as their targets), or their involvement in biological pathways or molecular functions. In this study, the researchers considered knowledge graphs represented using Semantic Web technologies, including RDF (Resource Description Framework) and URI (Uniform Resource Identifier). In these knowledge graphs, nodes represent entities and named individuals of a domain, classes of individuals, or literals. Nodes are connected by directed edges that are labelled with predicates.

This study proposes that knowledge graphs can be mined to identify biomolecular features that may enable automatically reproducing expert classification that distinguish between a drugs beneficial effects and adverse drug reactions.

Mining knowledge graphs to identify adverse drug reaction mechanisms

Using an Explainable AI perspective, the researchers explored simple classification techniques such as decision trees and classification rules because they provide human-readable models, which explain the classification itself, as well as also providing elements of explanation for molecular mechanisms behind adverse drug reactions. Explainable AI usually refers to research on methods that provide explanatory elements to results of sub-symbolic approaches, such as deep neural networks.

The researchers focused on the knowledge graph named PGxLOD, which encompasses and connects drug, pathway and biomolecule data; and two types of adverse drug reactions: drug induced liver injury (DILI) and severe cutaneous adverse reactions (SCAR). The adverse drug reactions were chosen since hepatic or skin toxicities are commonly monitored during drug development due to their importance in pharmacovigilance.

Firstly, the researchers identified biomolecular features from their knowledge graph, which enabled automatic reproduction of expert classifications. The team mined the graphs for neighbours of drugs, paths and path patterns rooted by drugs and passing by at least one entity of the following types: pathway, gene/protein, Gene Ontology (GO) term or MeSH term.

Following that, the researchers isolated both predictive and interpretative features as they hypothesised that, as well as the features being explanatory for the classification, they may also explain adverse drug reaction mechanisms. To this aim, the researchers considered explanatory classification techniques, such as decision tree and propositional rule learner over extracted features, since they provide human-readable models.

Finally, the researchers used independent experts to manually evaluate the isolated features to identify whether they could explain adverse drug reactions.

Using knowledge graphs to identify adverse drug reaction mechanisms

The features that the researchers isolated reproduced classifications of drugs that causative or not for DILI and SCAR, with an accuracy of 0.74 and 0.81, respectively. The team calculated accuracy measurements as AUC, with results higher than 0.70 demonstrating high accuracy. The experts who evaluated the features agreed that 73% and 38% of the most discriminative features are possibly explanatory for DILI and SCAR, respectively. Moreover, they partially agreed for 90% and 77% of the features. These findings suggest that knowledge graphs provide diverse features, which allow identification of explainable models to distinguish between drugs that are causative or not for ADRs. Moreover, these discriminative features appear to be good candidates for investigating adverse drug reaction mechanisms further.


The study developed a method to mine knowledge graphs to identify features that could help to explain the largely unknown adverse drug reaction mechanisms. Their results show that they were able to identify a number of good candidates, which researchers can explore further to identify the mechanisms behind adverse drug reactions. Moreover, this work illustrates that simple models, fed with diverse and explicit knowledge, can be used as an alternative to complex models, which are efficient but hard to interpret.

Image credit: pch.vector – FreePik

Share this article