Prediction of Drug-Target Interactions with CNNs and Random Forest

Abstract

Drug molecules interact with target proteins to influence the pharmacological action of the target to achieve the phenotypic effect, which can facilitate the identification of novel targets for current drug. Traditional biological experiments for discovering new drug-target interactions are expensive and time-consuming. Therefore, it is crucial to develop new prediction methods for identifying potential drug-target interactions. Computing methods have been increasing developed which can quickly and effectively predict drug-target interactions. In particular, machine learning methods have been widely used due to high predictive performance and computational efficiency. This paper first uses MACCS substructure fingerings to encode the drug molecules, then uses CNNs to extract the biological evolutionary information of target protein sequences, and finally uses random forest algorithm to predict drug-target interactions. Four datasets of drug-target interactions including Enzymes, Ion Channels, GPCRs and Nuclear Receptors, are independently used for building models with random forest. The results demonstrate our proposed method has a general compatibility, which is effective and feasible to predict drug-target interactions.