Drug-Drug Interaction, Interaction Type and Resulting Severity Forecasting by Machine Learning-Based Approaches

Abstract

Drug-drug interactions can be discovered by wet-lab experiments. Considering huge amount of possible drug combinations, knowing the drugs that are likely to interact before starting these scientific investigations will provide great savings in terms of time, labor and cost. In this study, it is aimed to identify the best machine learning algorithm to detect not only unknown drug-drug interactions but also to assign type and severity of the predicted interactions. We have applied the molecular structure of drugs, their targets, classification codes, and enzyme effects as inputs and tested logistic regression, naive bayes, K-nearest neighbor, decision trees and deep learning methods. The best performance was attained by deep neural networks (DNN). 88% average success rate has been reached by cross training of the DNN model. The lowest success rates were received for test sets including many severe interactions. Our study demonstrates that DNN can predict not only potential drug-drug interactions but also whether the interaction type is pharmacokinetics or pharmacodynamics. Our model differentiates minor and moderate interaction severities fairly but weak in differentiating severe interactions because of small number of known severe interactions in the training set.