Molecular image-based convolutional neural network for the prediction of ADMET properties

Abstract

Convolutional neural network (CNN), is one of the most representative architectures in deep learning and is widely adopted in many fields especially in image classification and object detection. In the last few years, CNN has been aroused more and more attentions in drug discovery domain. In this work, molecular 2-D image-based CNN method was used to establish prediction models of the ADMET properties, including CYP1A2 inhibitory potency, P-glycoprotein (P-gp) inhibitory activity, Blood-Brain Barrier (BBB) penetrating activity, and Ames mutagenicity. The results showed that the predictive power of the established CNN models is comparable to that of the available machine learning models based on manual structural description and feature selection. It can be inferred that CNN can extract efficiently the key image features related to the molecular ADMET properties and offer a useful tool for virtual screening and drug design researches.