Application of Supervised Machine Learning Models for Drug-Action Prediction Towards Nuclear Type I Receptors

Abstract Interactions between drugs can lead to adverse side effects for patients taking combination therapies to treat complex diseases such as cancer. Knowledge of drug-action towards a receptor would allow these drug-drug interactions to be predicted, and in this study, we trained a total of 5 different machine learning models to classify whether a given drug was an agonist (activator), antagonist (blocker), or a decoy (non-binder) to each of the androgen, estrogen, glucocorticoid, and progesterone receptors. The classification performance and efficiency, measured in training time, of the decision tree, naïve Bayes, neural network, random forest, and support vector machine models for each receptor were then compared. The results showed that the decision tree and naïve Bayes models were best suited for drug-action prediction across all receptors while only requiring minutes of training time at most. Future work will focus on increasing the prediction accuracy of antagonist drugs, integrating experimental data during training, and using other targets outside of nuclear type I receptors.