A Deep Learning Approach for Molecular Crystallinity Prediction

Abstract

With the success of Convolutional Neural Networks (CNN) in computer vision domain, cheminformatics is slowly moving away from feature Engineering towards Network Engineering. New deep networks and approaches are being proposed to explore the chemical behavior and their properties. In this paper, we propose a deep learning approach using Convolutional Neural Network for predicting the crystallization propensity of an organic molecule. The work is inspired from Chemception and architecture is based on the Inception-Resnet v2 model. The proposed approach only requires a 2D molecular drawing to predict if the molecule has a good probability of forming crystals, without the need of any molecular descriptor, any advanced chemistry knowledge or any study of crystal growth mechanisms. We have evaluated our approach on the Cambridge Structural Database (CSD) and the ZINC datasets. Compared with the machine learning approach of generating molecular descriptors plus SVM classification, our proposed approach gives a better classification accuracy.