i-DENV: development of QSAR based regression models for predicting inhibitors targeting non-structural (NS) proteins of dengue virus

Abstract

Introduction: Dengue virus (DENV) is a significant global arboviral threat with fatal potential, currently lacking effective antiviral treatments or a universally applicable vaccine. In response to this unmet need, we developed the "i-DENV" web server to facilitate structure-based drug prediction targeting key viral proteins.

Methods: The i-DENV platform focuses on the NS3 protease and NS5 polymerase of DENV using machine learning techniques (MLTs) and quantitative structure-activity relationship (QSAR) modeling. A total of 1213 and 157 unique compounds, along with their IC50 values targeting NS3 and NS5 respectively, were retrieved from the ChEMBL and DenvInD databases. Molecular descriptors and fingerprints were computed and used to train multiple regression-based MLTs, including SVM, RF, kNN, ANN, XGBoost, and DNN, with ten-fold cross-validation.

Results: The best-performing SVM and ANN models achieved Pearson correlation coefficients (PCCs) of 0.857/0.862 (NS3) and 0.982/0.964 (NS5) on training/testing sets, and 0.870/0.894 (NS3) and 0.970/0.977 (NS5) on independent validation sets. Model robustness was supported through scatter plots, chemical clustering, statistical analyses, decoy set etc. Virtual screening identified Micafungin, Oritavancin, and Iodixanol as top hits for NS2B/NS3 protease, and Cangrelor, Eravacycline, and Baloxavir marboxil for NS5 polymerase. Molecular docking further confirmed strong binding affinities of these compounds.

Discussion: Our in-silico findings suggest these repurposed drugs as promising antiviral candidates against DENV. However, further in vitro and in vivo studies are essential to validate their therapeutic potential. The i-DENV web server is freely accessible at http://bioinfo.imtech.res.in/manojk/idenv/, offering a structure-specific drug prediction platform for DENV research and antiviral drug discovery.

Keywords: QSAR; algorithm; antivirals; artificial intelligence; machine learning; web server.

FIGURE 1

Curated data from ChEMBL and DenvInD, resulting in 1213 unique entries targeting the NS3 protein and 157 unique entries targeting the NS5 protein. Using PaDEL software, descriptors in one-dimensional, two-dimensional, and three-dimensional formats were computed. Recursive feature elimination method from the sklearn module was employed for feature selection. The data was then divided into training and testing datasets, and various MLTs were applied. Model performance was assessed using MAE, MSE, RMSE, R2, and PCC, and validated with applicability domain, scatter plot, chemical clustering, box plot, statistical test such as paired t-tests or Wilcoxon signed-rank tests and decoy set analysis. Potential repurposed drugs were identified through an analysis of the DrugBank database using the best-developed models for both NS3 and NS5 inhibitors. For further validation, the top predicted drugs were docked against NS3 and NS5 proteins using AutoDock Vina. The best-performing models were integrated into the web server “i-DENV”.

FIGURE 2

Performance comparison of five (A) NS3 and (B) NS5 machine learning models (SVM, RF, ANN, kNN, and XGBoost) using five evaluation metrics: PCC, MAE, MSE, RMSE, and R ². Each boxplot illustrates the distribution of metric values across multiple runs, with the central line denoting the median, the box representing the interquartile range (IQR), and whiskers extending to 1.5 times the IQR. Outliers are depicted as individual points beyond this range.

FIGURE 3

SHAP-Based Analysis of Key Features in the Best Predictive SVM Model for pIC50 Prediction: (A) Beeswarm Plot for NS3 Protein, (B) Beeswarm Plot for NS5 Protein. Each point represents a single sample (compound) and illustrates the impact of a feature on the model’s output. The x-axis indicates the SHAP value, representing the magnitude and direction of a feature’s contribution to the prediction. Features are ranked vertically by their overall importance, measured by the mean absolute SHAP value. The color gradient (blue to red) denotes the feature value for each sample, with blue indicating low values and red indicating high values. Points spread further from zero along the x-axis indicate greater influence on the prediction. These plots help visualize both the importance of each feature and the direction of its effect across the dataset.

FIGURE 4

The applicability domain analysis using william plot for both (A) NS3 and (B) NS5 protein between the leverage and standardized residuals of the molecules. The x-axis represents the leverage values, indicating the influence of each compound on the model, while the y-axis shows the standardized residuals, reflecting the prediction error for each compound. Data points are color-coded: blue circles represent compounds from the training/testing dataset, and orange triangles represent those from the independent validation dataset. The horizontal dashed red lines at ±3 delineate the threshold for standardized residuals; points outside this range are considered outliers. The vertical dashed red line represents the critical leverage value (h*), beyond which compounds may be deemed influential and outside the model’s applicability domain. These plots help identify both outliers and structurally influential compounds, supporting the reliability and robustness of the developed models.

FIGURE 5

Reliability of the best predictive models for NS3 and NS5 was evaluated by generating scatter plots comparing the actual pIC50 values of molecules with their predicted values. The models assessed include: (A) SVM for NS3, (B) ANN for NS3, (C) SVM for NS5, and (D) ANN for NS5.

FIGURE 6

To validate the accuracy of the predicted models using the top-performing SVM model for NS3 protein, we created scatter plots. These plots were used to compare the actual pIC₅₀ values with the decoy values from four distinct decoy sets: (A) Set 1, (B) Set 2, (C) Set 3, and (D) Set 4.

FIGURE 7

To validate the accuracy of the predicted models using the top-performing SVM model for NS5 protein, we created scatter plots. These plots were used to compare the actual pIC₅₀ values with the decoy values from four distinct decoy sets: (A) Set 1, (B) Set 2, (C) Set 3, and (D) Set 4.

FIGURE 8

Chemical Diversity Analysis using (A) 2D MDS Plot for NS3, (B) 3D MDS Plot for NS3-Targeting Compounds, (C) 2D MDS Plot for NS5, and (D) 3D MDS Plot for NS5-Targeting Compounds.

FIGURE 9

Docking poses of selected drugs (A) Delavirdine; (B) Raltegravir; (C) Rimexolone and (D) Fluoxymesterone with DENV NS2BNS3 protease represented as 2-D line models.

FIGURE 10

Docking poses of selected drugs (A) Baloxavir marboxil; (B) Latamoxef; (C) Cyclothiazide and (D) Diphemanil with DENV NS5 polymerase represented as 2-D line models.

FIGURE 11

Docking poses of selected drugs (A) Delavirdine; (B) Raltegravir; (C) Rimexolone and (D) Fluoxymesterone with DENV NS2BNS3 protease in the form of ribbon structures.

FIGURE 12

Docking poses of selected drugs (A) Baloxavir marboxil; (B) Latamoxef; (C) Cyclothiazide and (D) Diphemanil with DENV NS5 polymerase in the form of ribbon structures.