Supervised machine learning-based prediction for dry mouth oral adverse drug reactions

Abstract

Adverse drug reactions (ADRs) are defined as an unintended and harmful response that occurs with the ingestion of a certain drug. ADRs result in an appreciably harmful or unpleasant reaction that determines the success or failure of a drug. In this way, the generation of effective models for the prediction of ADR during the drug development process is of high relevance for human health. In this work, we present a complete proposal based on supervised machine learning to study dry mouth oral ADRs for the first time. Our approach integrates different drug properties, such as, chemical (fingerprint), biological (target protein, transporters and enzymes) and phenotypic (therapeutic indications and other known adverse reactions), all of them obtained from public databases that are combined on two and three levels. We employ different tree- based classification algorithms (AdaBoost and Random Forest), with the aim of obtaining the best predictors of the oral RAM studied. 14 models were generated, which gave an average AUC of 0.82 and an accuracy of 78%, where the best model with AdaBoost gave and accuracy and AUC of 87% and 0.89, respectively for the prediction of dry mouth oral ADR