ARCHIVES

The ability to predict side effects of drugs is an important problem with direct implications on patient health and the pharmaceutical industry's ability to develop new drugs. Unpredictable adverse drug reactions (ADR) have caused many drugs to be pulled from the market and have caused more than 100 million injuries or deaths worldwide. In this article, I will be using data from the Mizutani dataset, which has 658 drugs and 1368 target protein structures, to predict drug side effects using the targeted protein interaction profiles as the primary features. A Naive Bayes (NB) model with all 1339 side effects as output variables was created; and the classification accuracy of the NB classifier was 91.14% in the Mizutani dataset as a whole. In addition, 15 of the more commonly reported side effects will be identified; and the 15 selected side effects will be subjected to 5 machine learning classifiers (Logistic Regression (LR), Random Forest (RF), Naive Bayes (NB), Support Vector Machine (SVM), and Decision Trees (DT)) to compare their respective accuracies, precisions, and F-scores. The classifier with the highest average accuracy of 70.29% across the selected 15 side effects was the LR classifier. In addition to this work, the proposed method will also be used to predict side effect profiles of 92 previously uncharacterized drugs using the LR classifier; and the LR classifier achieved a 76.45% match with publicly available reference sources for those drugs. These results indicate that targeted interaction protein interaction profiles provide sufficient and useful predictive features for predicting drug side effects.