Graph Convolutional Neural Network-Based Virtual Screening of Phytochemicals and In-Silico Docking Studies of Drug Compounds for Hemochromatosis

Abstract

Machine learning based Virtual Screening has proved as an important intermediate process that helps in the field of drug discovery in reducing the cost and manpower of classical drug discovery process. This work proposes a deep learning based virtual screening model for the early discovery of drug compounds for the disease named Hemochromatosis, which is the excess absorption of iron in the human body. Our study focuses on finding possible drug compounds from medicinal plants to cure Hemochromatosis. The proposed method uses Graph Convolutional Neural Networks (GCN) for Ligand Based Virtual Screening (LBVS). Deep Learning algorithm, GCN outperformed all other experimented models in LBVS with an accuracy of 98.26% and F-score of 98% respectively. A small set of biologically active compounds was identified from the phytochemical dataset after performing the LBVS. The selected ligands after LBVS are taken for In-Silico Structure Based Screening (SBVS) called molecular docking and the best compounds that have high binding affinity towards the disease protein for Hemochromatosis is selected and recommended for in-vitro studies. Ablation studies are done with 12 different machine learning models including ensemble models. The proposed model exhibited a related percentage improvement of around 0.5% in accuracy and F-score, when compared to the tree based ensemble model, XGBoost. This study aims to suggest in-silico studies for ligand based and structure based screening to identify potential drug molecules from medicinal plants which can be tested in in-vitro analysis and studies.

Keywords

• Computational drug discovery
• virtual screening
• molecular descriptors
• molecular fingerprints
• graph convolutional neural networks