A knowledge graph embedding based approach to predict the adverse drug reactions using a deep neural network

https://doi.org/10.1016/j.jbi.2022.104122Get rights and content
Under an Elsevier user license
open archive

Highlights

  • A knowledge graph is constructed with drugs, ADRs, indications, genes, and pathways as entities and edges between them.

  • Node2vec algorithm is used to generate the node sequences from the knowledge graph.

  • A Deep Learning model called Continuous Bag of Words model is used to get the embeddings(descriptions) of drugs and ADRs from those node sequences.

  • A new Deep Neural Network( KGDNN) has been proposed to predict the ADR for given drugs.

  • The proposed method is validated with different case studies which include covid19 recommended drugs.

Abstract

Recently Artificial Intelligence(AI) has not only been used to diagnose the disease but also to cure the disease. Researchers started using AI for drug discovery. Predicting the Adverse Drug Reactions(ADRs) caused by the drug in the manufacturing stage or in the clinical trial stage is a very important problem in drug discovery. ADRs have become a major concern resulting in injuries and also becoming fatal sometimes. Drug safety has gained much importance over the years propelling to the forefront investigation of predicting the ADRs. Although prior studies have queried diverse approaches to predict ADRs, very few were found to be effective. Also, the problem of having fewer reports makes the prediction of ADRs more difficult. To tackle this problem effectively, a novel method has been proposed in this paper. The proposed method is based on Knowledge Graph(KG) embedding. Using the KG embedding, we designed and trained a custom-made Deep Neural Network(DNN) called KGDNN(Knowledge Graph DNN) for predicting the ADRs. A KG has been constructed with 6 types of entities: drugs, ADRs, target proteins, indications, pathways, and genes. Using the Node2Vec algorithm, each node has been embedded into a feature space. Using those embeddings, the ADRs are classified by the KGDNN model. The proposed method has obtained an AUROC score of 0.917 and significantly outperformed the existing methods. Two case studies on drugs causing liver injury and COVID-19 recommended drugs have been performed to illustrate the model efficacy.

Keywords

Medical Informatics
Bioinformatics
COVID-19
Adverse Drug Reaction
Machine Learning
Deep Learning
Node2Vec
Knowledge Graph

Cited by (0)