Machine-learning-approaches-to-select-candidate-molecules

Drug discovery can become a time-consuming and expensive process, with average development time exceeding a decade. Many methods have been developed to optimize drug candidate selection, such as rule-based methods for selecting compounds with values within set bounds on select properties. Machine learning techniques have also been employed in drug discovery, from predicting the values of drug properties, to classifying a compound as a drug or as not a drug. In this study, multiple machine learning methods are evaluated on their ability to classify approved drugs from non-approved drugs using physiochemical properties available on the public ChEMBL database. Further, these approaches are evaluated on their ability to classify approved drugs still on the market from drugs that have been withdrawn. The approaches evaluated and compared include logistic regression, random forest, support vector machine, and neural network. Drug-like properties, from those available in the ChEMBL database, are also evaluated to determine which are most important in explaining drug variance.