Learning Molecule Drug Function from Structure Representations with Deep Neural Networks or Random Forests

Abstract

Empirical testing of chemicals for desirable properties, such as drug efficacy, toxicity, solubility, and thermal conductivity, costs many billions of dollars every year. Further, the choice of molecules for testing relies on expert knowledge and intuition in the relevant domain of application. The ability to predict the action of a molecule in silico would greatly increase the speed and decrease the cost of prioritizing molecules with desirable function for experimental testing. We explore the use of molecule structures to predict their drug classification without any explicit biological model (e.g. protein structure or cell system). Two approaches were compared: (1) a two-dimensional image representation of molecule structure with transfer learning from a pre-trained convolutional neural network (CNN), and (2) Morgan molecular fingerprints (MFP) with a random forest (RF). Both methods using only molecule information as input achieved significantly better accuracy than previous work that used transcriptomic measurements of molecular effects. Because these data suggest that much of a molecules function is encoded by its chemical structure, we explore which general molecular properties are drug class-specific and how misclassification of structures might provide drug repurposing opportunities.