A convolutional neural network for virtual screening of molecular fingerprints

Abstract

In the last few years, Deep Learning (DL) gained more and more impact on drug design because it allows a huge increase of the prediction accuracy in many stages of such a complex process. In this paper a Virtual Screening (VS) procedure based on Convolutional Neural Networks (CNN) is presented, that is aimed at classifying a set of candidate compounds as regards their biological activity on a particular target protein. The model has been trained on a dataset of active/inactive compounds with respect to the Cyclin-Dependent Kinase 1 (CDK1) a very important protein family, which is heavily involved in regulating the cell cycle. One qualifying point of the proposed approach is the use of molecular fingerprints as a suitable embedding for describing molecules; up to our knowledge there is no Deep Learning approach for VS that makes use of such descriptor. Several kinds of fingerprints are reported in the scientific literature to address different aspects of both the structure and the local properties of a molecule. Both 1D and 2D CNNs have been trained to test the performance of each single descriptor separately, along with suitable ensembles of multiple descriptors for the same compound; the best performing architecture has been used for prediction. The CNN architectures are described in detail, and the results are compared with some recent approaches for Virtual Screening with respect to Cyclin-Dependent Kinase proteins that do not use molecular fingerprints as their descriptor.