AI Driven Computational Drug Discovery using Graph Neural Networks for Dengue virus DENV-2

Abstract

Dengue Virus Serotype 2 (DENV-2) poses a critical global health threat due to its severe complications such as dengue hemorrhagic fever and dengue shock syndrome. The absence of effective antiviral therapies and limitations in mosquito control strategies used currently is another concern. This study introduces an Artificial Intelligence (AI)-driven computational drug discovery pipeline by using Graph Neural Networks (GNNs) to identify and design potent antiviral compounds targeting this disease. The main aim of this study is to accelerate drug discovery process by identifying drug candidates. The methodology involved collecting bioactivity datasets from ChEMBL database, drug molecules from DrugBank databases and a target protein on DENV-2 from Protein Data Bank (PDB) database. Then preprocessing the bioactivity data with Simplified Molecular Input Entry Line System (SMILES)-to-graph conversion, Absorption Distribution Metabolism Excretion (ADME) filtering, and Lipinski’s Rule of Five (Ro5) compliance and developing a Hybrid Graph Convolution Network-Graph Attention Network (GCN-GAT) architecture to predict negative logarithm (base 10) of the Half-Maximal Inhibitory Concentration (pIC50) values for Non-Structural Protein 3 (NS3) inhibitors. The GNN model trained on augmented datasets achieved a state-of-the-art R² score of 0.9938, demonstrating exceptional predictive accuracy for bioactivity. Additionally, a Variational Graph Autoencoder (VGAE) enabled de novo molecule generation. Through Virtual screening of Food and Drug Administration (FDA)-approved drugs from DrugBank database, we identified potential drugs for repurposing. Then VGAE model is employed to generate novel inhibitors for drug designing purpose. Also, we used molecular docking to validate binding affinities (≤ -5 kcal/mol) of top candidates. This work establishes a robust framework for AI-driven computational drug discovery for combating DENV-2 and other challenging diseases. Future directions include expanding datasets with diverse chemical libraries, refining the generative model to identify better drug candidates to bridge the gap between computational predictions and clinical translation.