Machine learning and molecular docking prediction of potential inhibitors against dengue virus

Abstract

Introduction: Dengue Fever continues to pose a global threat due to the widespread distribution of its vector mosquitoes, Aedes aegypti and Aedes albopictus. While the WHO-approved vaccine, Dengvaxia, and antiviral treatments like Balapiravir and Celgosivir are available, challenges such as drug resistance, reduced efficacy, and high treatment costs persist. This study aims to identify novel potential inhibitors of the Dengue virus (DENV) using an integrative drug discovery approach encompassing machine learning and molecular docking techniques.

Method: Utilizing a dataset of 21,250 bioactive compounds from PubChem (AID: 651640), alongside a total of 1,444 descriptors generated using PaDEL, we trained various models such as Support Vector Machine, Random Forest, k-nearest neighbors, Logistic Regression, and Gaussian Naïve Bayes. The top-performing model was used to predict active compounds, followed by molecular docking performed using AutoDock Vina. The detailed interactions, toxicity, stability, and conformational changes of selected compounds were assessed through protein-ligand interaction studies, molecular dynamics (MD) simulations, and binding free energy calculations.

Results: We implemented a robust three-dataset splitting strategy, employing the Logistic Regression algorithm, which achieved an accuracy of 94%. The model successfully predicted 18 known DENV inhibitors, with 11 identified as active, paving the way for further exploration of 2683 new compounds from the ZINC and EANPDB databases. Subsequent molecular docking studies were performed on the NS2B/NS3 protease, an enzyme essential in viral replication. ZINC95485940, ZINC38628344, 2',4'-dihydroxychalcone and ZINC14441502 demonstrated a high binding affinity of -8.1, -8.5, -8.6, and -8.0 kcal/mol, respectively, exhibiting stable interactions with His51, Ser135, Leu128, Pro132, Ser131, Tyr161, and Asp75 within the active site, which are critical residues involved in inhibition. Molecular dynamics simulations coupled with MMPBSA further elucidated the stability, making it a promising candidate for drug development.

Conclusion: Overall, this integrative approach, combining machine learning, molecular docking, and dynamics simulations, highlights the strength and utility of computational tools in drug discovery. It suggests a promising pathway for the rapid identification and development of novel antiviral drugs against DENV. These in silico findings provide a strong foundation for future experimental validations and in-vitro studies aimed at fighting DENV.

Keywords: dengue virus; drug discovery; machine learning; molecular docking; molecular dynamics simulation.

FIGURE 1

Graphical illustration of study workflow methods and instruments. The study developed five models on data from PubChem and used it for predicting new compounds. The predicted hits were screened through molecular docking, in silico pharmacological and toxicity tests, structural assessment using MD simulations, and estimation of binding free energies.

FIGURE 2

3D plot showing the correlation between the active and inactive compounds based on ALogP, XLogP, and Zagreb. [Labels: Blue = Actives, Red = Inactives, Pink = Outliers].

FIGURE 3

PyMOL visualization of NS2B/NS3 protease structure (A) Light-green cartoon structure representation; (B). Light-green surface representation of the protein with ZINC000095486052 (blue) docked in the active site.

FIGURE 4

Ligand ZINC38628344 docked in NS2B/NS3 binding pocket; 3D pose and 2D protein-ligand interaction diagram generated using PyMOL and LigPlot, respectively.

FIGURE 5

Inhibitor Prednisolone docked in the NS2B/NS3 binding pocket, showing the protein-ligand interactions visualized in LigPlot and the 3D pose in PyMOL.

FIGURE 6

RMSD versus time graph of unbound protein and NS2B/NS3pro-ligand complexes generated over a 100 ns MD run.

FIGURE 7

Rg graph of the NS2B/NS3pro-ligand complexes and unbound protein.

FIGURE 8

Examination of the RMSF trajectories of the NS2B/NS3pro-ligand complexes and the unbound protein residues.

FIGURE 9

MMPBSA plot of binding free energy contributions per residue for NS2B/NS3-ZINC14441502 complex.