Deep Learning Methods to Help Predict Properties of Molecules from SMILES.

Abstract

Machine learning methods have been proposed in lieu of simulations to predict chemical properties of molecules. The trade-off here is paying for the training time once, in exchange for instant predictions on the input data. However, many of these methods rely heavily on feature engineering to prepare the data for these models. Moreover, the use of molecular structural information has been limited, despite having such information encoded in the Simplified Molecular Input Line Entry System (SMILES) format. In this paper we present a framework that relies on SMILES data to predict molecular properties. Our methods are based on 1-D Convolutional Networks and do not require complex feature engineering. Our methods can be applied to learn molecular properties from base data, thus making them accessible to a wider audience. Our experiments show that this method can predict the molecular weight and XLogP properties without any encoding of complex chemical rules.