Dengue virus (DENV), an arbovirus from the Flaviviridae family, causes tropical diseases and is a major global health concern. The World Health Organization (WHO) lists Dengue among the top ten global threats, with nearly half the world’s population at risk and about 390 million infections annually (Waggoner et al., 2016; Bhatt et al., 2013; Brady et al., 2012). Globalization, urbanization, and climate change are expanding the range of Aedes aegypti and Aedes albopictus, potentially placing 60% of the global population at risk by 2080 (Ebi and Nealon, 2016; Messina et al., 2019; de Almeida et al., 2017). DENV encompasses four serotypes (DENV-1 to DENV-4), each capable of causing the full range of disease. Infection with a different serotype can lead to severe conditions like dengue hemorrhagic fever (DHF) and potentially fatal dengue shock syndrome (DSS) due to antibody-dependent enhancement (ADE), which complicates vaccine development (Goethals et al., 2023; Dash et al., 2006). Currently, no approved antivirals exist and only symptomatic treatment is available (Kaptein et al., 2021).

DENV possesses an 11 kb single-strand RNA genome, yielding seven non-structural proteins (NS1, NS2A, NS2B, NS3, NS4A, NS4B, NS5) and three structural proteins (CP, EP, MP) (Khan et al., 2023). Its life cycle reveals key steps such as endocytosis, viral fusion, transcription, and release of new viral particles (Behnam et al., 2016). The envelope (E) protein mediates host cell attachment and membrane fusion. The membrane (M) protein stabilizes the mature virion and aids assembly after prM cleavage. The capsid (C) protein packages viral RNA and initiates particle formation. NS1 supports replication, immune evasion, and viral assembly. NS2A is involved in replication, polyprotein processing, and cytopathogenesis. NS2B, as a cofactor for NS3, forms a protease complex crucial for polyprotein processing and immune suppression. NS4A modulates host membranes and promotes viral protein oligomerization for replication. NS4B dimerizes, interacts with NS5, and helps form the replication complex (Nasar et al., 2020; Nath et al., 2024). NS3 and NS5 proteins have diverse enzymatic activities. NS5 contains a methyltransferase domain (N-terminal) for mRNA capping and an RNA-dependent RNA polymerase (RdRp) domain (C-terminal) for genome replication. NS5 is highly conserved among the four DENV serotypes, making it a promising target for anti-dengue drug development due to the absence of similar RdRp activity in human enzymes (Shimizu et al., 2019; Coulerie et al., 2013). The NS3 protease, consisting of NS2B and NS3, is crucial for DENV replication, functioning as a trypsin-like serine protease, with catalytic residues His51, Asp75, and Ser135. As a result, disrupting NS3 protein proves fatal to the virus, underscoring its potential as a key target for antiviral drug development (Tomlinson et al., 2009).

Several experimental studies have targeted NS3 and NS5 proteins to combat DENV infection. For example, - Abdullah et al. used computational methods to identify Zileuton, trimethadione, and linalool as novel NS3 inhibitors, with IC₅₀ values of 3.3 mM, 25.97 mM, and 1.12 mM. They further proposed Ziltri and zilool based on docking results (Abdullah et al., 2023). Likewise, Balasubramanian et al. identified curcumin as a DENV2 NS2B/NS3 protease inhibitor, synthesizing analogs (CC1-CC5) with IC₅₀ values between 36.23 and 66.01 μM, EC₅₀ values between 8.07 and 29.25 μM, and CC50 values between 25.50 and 87.40 μM (Balasubramanian et al., 2019). Salleh et al. investigated 21 Malaysian medicinal plants and found that Dryobalanops aromatica methanol extract showed 99.7% inhibition at 200 μg/mL with an IC50 of 0.30 μg/mL targeting NS2B-NS3 protease (Salleh et al., 2019). Shimizu et al. identified RK-0404678 as a potent DENV2 NS5 RNA polymerase inhibitor with an EC₅₀ of 6.0 μM, screened from 16,240 compounds (Shimizu et al., 2019). Jarerattanachat et al. identified Isoquercitrin as a dual-binding NS5 Methyltransferase inhibitor, effectively suppressing DENV with minimal toxicity (CC50 > 20 μM) (Jarerattanachat et al., 2023). However, only a few of these antiviral candidates have advanced to clinical trials, highlighting the need for novel DENV-targeted treatments.

In this concern, combining computational approaches with experimental studies offers a more effective strategy for developing antivirals against viral structural and non-structural proteins, expediting drug discovery. For example, - Indu et al. screened 7,000 phytocompounds against DENV proteins, identifying astragaloside II, III, and IV as potential inhibitors based on strong binding energies, which were further tested in Vero cell line (Indu et al., 2021). Similarly, Khan et al. assessed diterpenoids against dengue viral proteins (Envelope, NS1, NS3, NS5) using molecular docking, dynamics simulation, and network pharmacology (Khan et al., 2021). Cabarcas-Montalvo et al. docked 210,903 PubChem molecules against NS2B/NS3 and tested the top 5 candidates in antiviral assays (Cabarcas-Montalvo et al., 2016). In another study, Mirza et al. screened 18 million compounds from the ZINC database against NS3 protease using various in silico methods and tested 4 potent compounds through in vitro studies (Mirza et al., 2018). Furthermore, machine learning techniques (MLTs) have been widely applied in drug development. For example, - Gupta et al. developed DDPM, an early diagnostic model using MLTs to aid in dengue diagnosis and prognosis (Gupta et al., 2023). Similarly, Natali et al. used MLTs to identify rare antibody sequences capable of neutralizing pathogens (Natali et al., 2024). In this context, our group has developed several machine learning-based antiviral prediction tools utilizing quantitative structure–activity relationship (QSAR) information of molecules and peptides. Quantitative Structure–Activity Relationship (QSAR) is a computational modeling technique employed to establish relationships between the structural or physicochemical properties of chemical compounds and their biological activities. The fundamental assumption of QSAR is that variations in molecular structure lead to differences in biological behavior, including key pharmacokinetic parameters such as absorption, distribution, metabolism, excretion, and toxicity (ADMET) (Kwon et al., 2019). These include AVCpred for prediction and design of antiviral compounds (Qureshi et al., 2017), AVPpred for prediction of highly effective antiviral peptides (Thakur et al., 2012), AVP-IC50Pred for prediction of peptide antiviral activity in terms of half maximal inhibitory concentration (IC50) (Qureshi et al., 2015), and HIVprotI for prediction and design of HIV proteins inhibitors (Rajput and Kumar, 2022), among others. Furthermore, we have created MLT-based platforms such as anti-Flavi (Rajput and Kumar, 2018), anti-Nipah (Rajput et al., 2019), and anti-corona (Rajput et al., 2021a) to predict antiviral compounds against various pathogens. Additionally, we created a database of repurposed drugs effective against 23 epidemic and pandemic-associated viruses (Rajput et al., 2021b).

Recently, we developed the Anti-Dengue platform to predict repurposed drugs targeting DENV based on their IC₅₀/pIC₅₀ values (in μM) using various MLTs. However, this platform focuses on the entire virus without specifying the targeted site of the drug (Gautam et al., 2024). To address this gap, we introduce a new machine-learning-driven pipeline named “i-DENV”, which enables the identification of inhibitors targeting NS3 and NS5 proteins of DENV. The platform employs diverse MLTs (Support vector machine (SVM), Random Forest (RF), k-nearest neighbor (kNN), Artificial neural network (ANN), Extreme Gradient Boosting (XGBoost) and Deep Neural Network (DNN)). Using our top-performing models, we screened the DrugBank database to predict promising repurposed drug candidates for NS3 and NS5 proteins. Selected drug candidates underwent additional support through molecular docking, confirming strong binding affinities. Overall, this study contributes to the discovery of antiviral drugs specifically targeting NS3 and NS5 proteins, offering potential therapeutic benefits in combating DENV infection.

For developing the “i-DENV” predictive algorithm, the overall workflow is given in Figure 1.

For the predictor, experimentally validated compounds were taken from the CHEMBL and DenvInD databases. CHEMBL provides data on bioactive molecules with drug-like properties, integrating chemical, bioactivity, and genomic data to aid in developing new pharmaceutical treatments (Mendez et al., 2019; Davies et al., 2015). DenvInD is a comprehensive database encompassing known inhibitors targeting potential therapeutic targets of the DENV (Dwivedi et al., 2021).

• ChEMBL provided information on 1,372 compounds for NS3 and 67 for NS5 protein, while DenvInD offered data on 95 compounds specifically targeting NS5 protein (Mendez et al., 2019) (Davies et al., 2015) (Dwivedi et al., 2021).

Supplementary Table S1 provides the dataset of drugs and inhibitors used to develop the model for NS3 and NS5 protein.

The compound structures from the NS3 and NS5 dataset were converted from SMILES to structure-data file (3D-SDF) format using Open Babel version 3.1.1 (O’Boyle et al., 2011). These converted files were subsequently employed as input to extract chemical descriptors and fingerprints.

To develop QSAR based predictive models for the NS3 and NS5 proteins of DENV, we used PaDEL software to compute molecular descriptors and fingerprints (Yap, 2011). We computed 17,968 chemical descriptors for each molecule in the NS3 and NS5 protein dataset. Descriptors are numeric representations of molecular features, while fingerprints capture structural fragments, connectivity, bonds, and functional groups using binary sequences, where 1 denotes presence and 0 denotes absence (Grisoni et al., 2018). Descriptors and fingerprints are essential for analyzing drugs and chemicals, as they help assess their QSAR (Perkins et al., 2003).

Feature selection is the process of selecting the most important features from a larger set to enhance model performance and interpretability (Taghizadeh et al., 2022). Feature selection used the Recursive Feature Elimination (RFE) module from scikit-learn with Perceptron, Support Vector Regression (SVR), and Decision Tree (DT) methods. The goal was to identify the top 50, 100, 150, and 200 features for each dataset. All selected features were then used as input for MLTs on the NS3 and NS5 datasets (Lin et al., 2012) (Gholami et al., 2012).

Random selection processes were used to create a training/testing (TT) and independent validation (IV) dataset. To achieve this, 10% of the total data were randomly allocated as an IV dataset, with the remaining 90% utilized for the TT dataset. This process was repeated 5 times to get different 5 sets of TT and IV dataset. The NS3 dataset comprised 1213 molecules - T₁₀₉₂ and V₁₂₁, while the NS5 dataset consisted of 155 molecules - T₁₄₀ and V₁₅.

Ten-fold cross-validation is employed to assess the performance of a machine learning model. This method entails dividing the dataset randomly into ten equal subsets, or “folds”. The model undergoes ten rounds of training and evaluation, where each round uses a different fold as the validation set and the remaining nine folds as the training set. This systematic approach ensures that each data point is validated exactly once. Afterward, the performance metrics obtained from each round are averaged to provide a more accurate evaluation of the model’s performance.

We evaluated the model’s performance using metrics such as Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Coefficient of Determination (R2), and Pearson’s Correlation Coefficient (PCC or R). These metrics were calculated using the formulas provided below- P C C = n ∑ n = 1 n   E i a c t E i p r e d − ∑ n = 1 n   E i a c t ∑ n = 1 n   E i p r e d n ∑ n = 1 n   E i a c t 2 − ∑ n = 1 n   E i a c t 2 − n ∑ n = 1 n   E i p r e d 2 − ∑ n = 1 n   E i p r e d 2 M A E = 1 n ∑ n = 1 n   E i p r e d − E i a c t R M S E = 1 n ∑ n = 1 n   E i p r e d − E i a c t 2

Where, ‘n' represents the size of the dataset, ‘Eact’ denotes the actual values, and ‘Epred’ corresponds to the predicted values.

To ensure the robustness of predictive models, 95% confidence intervals (CIs) were computed for key performance metrics (PCC, MAE, MSE, RMSE, and R ²) using 10-fold cross-validation across all machine learning techniques (MLTs). Statistical significance of performance differences among models was assessed through pairwise comparisons of these metrics. The Shapiro-Wilk test was used to evaluate normality, determining the choice of statistical tests: a paired t-test for normally distributed data or the Wilcoxon signed-rank test for non-normally distributed data. Additionally, box plots were generated using Seaborn and Matplotlib libraries to visualize the distribution and variability of performance metrics across different models (Hazra, 2017; Shapiro and Wilk, 1965; Rosner et al., 2006; Ross and Willson, 2017; Williamson et al., 1989).

SHAP (SHapley Additive exPlanations) was employed to analyze feature importance in the best predictive Support Vector Machine (SVM) model for pIC₅₀ prediction targeting the NS3 and NS5 protein. A Support Vector Regression (SVR) model with optimized hyperparameters was trained using 10-fold cross-validation. Features were standardized using StandardScaler, and SHAP values were computed using a KernelExplainer with a k-means-selected background dataset of 50 representative training samples. The SHAP values were aggregated across all folds to generate a comprehensive beeswarm and summary plot, highlighting the most influential features (Lundberg and Lee, 2017).

In addition to model performance, accuracy for new predictions is crucial. Applicability domain (AD) analysis establishes the model’s boundaries to ensure reliable predictions. A query molecule’s chemical properties must fall within the AD of the trained model to ensure precise predictions (Kar et al., 2018).

To assess this, William’s plot was employed using a leverage approach for NS3 and NS5 proteins, illustrating the relationship between leverage and standardized residuals. The leverage threshold (h*) is: Leverage threshold  h * = 3 p + 1 n where p - the number of descriptors used in developing the model. n is the number of compounds used in the training dataset.

The leverage threshold (h∗) is a key component of AD analysis, used to identify influential or outlier compounds in a QSAR model.

Leverage (h) quantifies a compound’s influence on the model. If h > h∗, the compound is flagged as an outlier or highly influential, indicating it may fall outside the AD and require further evaluation. Compounds within h∗ are considered reliably predicted by the model. The factor 3 in the formula is an empirically chosen value commonly used in QSAR modeling to set a practical threshold (Khosrokhavar et al., 2010).

To affirm the models’ effectiveness, scatter plots were created to compare predicted pIC₅₀ values with actual pIC₅₀ values for the top-performing models in both NS3 and NS5 datasets.

Using the DecoyFinder 2.0 tool, Decoys were created for these drug candidates (Cereto-Massagué et al., 2012). We used a molecular weight-based method to create these decoys, utilizing the ZINC20 database, which includes 4.78 million drug-like molecules (Irwin et al., 2020).

Four distinct decoy datasets were created, containing 1213 decoys for the NS3 dataset and 155 for the NS5 dataset. These decoys, randomly generated to correspond to active drug candidates, were processed further through format conversion and molecular descriptor calculations to obtain their pIC₅₀ values. A correlation analysis was then performed to measure the Pearson’s Correlation Coefficient (PCC), evaluating the relationship between the pIC₅₀ values of the decoys and the actual pIC₅₀ values of the corresponding active drug candidates.

We performed chemical clustering analysis using two methods: the ChemMine tool and t-distributed Stochastic Neighbor Embedding (t-SNE) visualization. In the ChemMine tool, molecular SMILES were used as input to assess drug heterogeneity, applying both multidimensional scaling (MDS) and binning clustering with a similarity threshold of 0.6 (Backman et al., 2011). For t-SNE, the dataset containing molecular descriptors and pIC₅₀ values was loaded using Pandas and standardized with StandardScaler to ensure uniform feature scaling. t-SNE from sklearn. manifold was applied to reduce dimensionality to two components for visualization. A scatter plot was generated using matplotlib. pyplot, with pIC₅₀ values as the color gradient to illustrate activity distribution (Van Der Maaten and Hinton, 2008).

We used top-performing SVM and ANN models for both NS3 and NS5 proteins to identify potent repurposed drug candidates. We then screened 2,150 approved drugs from the Drugbank database to identify promising candidates (Knox et al., 2024). As a preliminary step, transformed the file format and calculated molecular descriptors of these drugs with PaDEL software. These descriptors were then used to identify potential repurposed drug candidates for DENV.

After identifying highly promising drug candidates targeting dengue NS proteins, the top 4 drugs in each category, based on 3D structure availability and predicted pIC₅₀ values, were selected for molecular docking studies. For ligand preparation, the selected drugs 3D structures were acquired from PubChem in SDF format, converted to PDB format using PyMOL software, and then import the file in auto dock tools to convert it into PDBQT format (Eberhardt et al., 2021) (Trott and Olson, 2010).

To prepare the receptors, the 3D crystal structures of the DENV NS2B/NS3 protease (PDB ID: 2FOM) and NS5 polymerase (PDB ID: 4V0Q) were retrieved in PDB format from the RCSB Protein Data Bank. These structures were then loaded into Discovery Studio to remove the original ligands—glycerol (GOL) from the NS2B/NS3 protease and S-adenosyl-l-homocysteine (SAH) from the NS5 polymerase.

Prior to docking analyses, the 3D protein structures were refined by removing extraneous ions, ligands, and non-essential water molecules. Polar hydrogen atoms and Kollman charges were added to the receptor proteins. The refined structures were then saved in PDBQT format for the subsequent docking experiments.

To perform the docking simulations, AutoDock Vina (version 1.1.2) with default settings was used to create grid boxes for the dengue NS3 and NS5 proteins. Nine optimal docking conformations were then generated for both proteins and the inhibitor molecules (Eberhardt et al., 2021) (Trott and Olson, 2010).

The exhaustiveness parameter was set to 8 to compute the minimum binding affinity between proteins and ligands. Subsequently, the interactions were then analyzed and visualized using PyMOL and Discovery Studio Visualizer (DSV) (Rigsby and Parker, 2016) (Biovia et al., 2016).

The “i-DENV” web server, designed to predict a chemical’s pIC₅₀ and IC₅₀ (μM) for inhibiting DENV, integrates the most effective SVM models for NS3 and NS5 proteins. It is built using the LAMP software stack, which includes Linux (OS), Apache (web server), MySQL (database), and PHP (scripting language). The system runs on Ubuntu 20.04.6 LTS (Focal Fossa), a Debian-based Linux distribution, with Apache 2.4.41 as the web server and MySQL 8.0.40-0ubuntu0.20.04.1 on a 64-bit Linux system. For server-side scripting, it utilizes PHP 7.4.3-4ubuntu2.28 (CLI), incorporating Zend Engine v3.4.0 and Zend OPcache v7.4.3-4ubuntu2.28 for optimized performance.

The “i-DENV” web server’s user interface is built with HTML, CSS, and PHP, while the backend is powered by Python, Perl, and JavaScript to handle computations and data processing. To improve accessibility, it includes dedicated “Help” and “FAQs” pages, ensuring users have guidance and support while using the platform.

Dengue remains a major global health challenge due to the lack of effective antivirals and vaccine development difficulties (Norshidah et al., 2021). The genetic diversity among DENV serotypes (DENV-1–4) ranges from 60% to 75% amino acid homology, while viruses within the same serotype share over 97% (Guzman and Harris, 2015). Effective antivirals must target all serotypes to prevent antibody-dependent enhancement (ADE). Vaccine development is hindered by limited knowledge of protective factors, inadequate animal models, and the need for a tetravalent vaccine (Wilder-Smith, 2020). Further, natural products and phytochemicals have also been explored as an alternative to synthetic drugs for managing dengue fever (Sagar et al., 2024; Saleh and Kamisah, 2020). However, developing new small molecule antivirals is crucial, and employing a computational approach in this process would significantly accelerate drug discovery research. To address this, we developed “i-DENV”, a QSAR-based machine-learning pipeline predicting repurposed drugs against NS3 and NS5 proteins.

In this study, we employed SVM, RF, ANN, kNN, XGBoost, and DNN algorithms to develop predictive models, which have been widely utilized in previous studies. For instance, Auto-Kla identifies lysine lactylation sites (Lai and Gao, 2023), and IML-TYLCVs assesses TYLCV symptom severity (Bupi et al., 2023). Similarly, MLTs have been applied in various antiviral prediction platforms, including AVPpred (Thakur et al., 2012), AVP-IC50Pred (Qureshi et al., 2015), SMEpred (Dar et al., 2016), HIVprotI (Qureshi et al., 2018), and Anti-nipah (Rajput et al., 2019), anti-Ebola (Rajput and Kumar, 2022), anti-corona (Rajput et al., 2021a), and anti-dengue (Gautam et al., 2024). But these servers primarily focus on predicting antiviral candidates against entire viruses, lacking target protein specificity. However, few research groups have also employed in silico methods like molecular docking, ligand interaction analysis, and QSAR-based models to identify repurposed drugs against DENV (Panchal et al., 2021) (Pathak et al., 2017) (Tahir ul Qamar et al., 2019). For instance, Chongjun et al. developed a classification model using 591 NS3-targeting compounds, generating nine molecular fingerprints. They implemented SVM, RF, and XGBoost models for evaluating accuracy and ROC_AUC (Valencia et al., 2021). Similarly, Kurniawan et al. created a classification model using 845 compounds, 2,603 molecular descriptors and applied ensemble RF, AdaBoost, and ERT models. Further, they evaluated performance using accuracy, AUC, MCC, etc. (Kurniawan et al., 2020). However, our study advances previous work by using a larger dataset and extending the analysis by including NS5 protein. Unlike classification approaches, we employed regression models evaluated by MAE, MSE, RMSE, R ², and PCC. Our comprehensive validation includes applicability domain analysis, scatter plots, decoy validation, chemical clustering, and further support by molecular docking approach. The best SVM-SVR models were integrated on the web server to predict inhibitors specifically targeting DENV NS3 and NS5 proteins. Thus, “i-DENV” represents the first regression-based web server to identify potential drug candidates against the NS3 and NS5 proteins of DENV.

We have also predicted potential repurposed drug candidates from the approved DrugBank database using our best models for both NS3 and NS5 proteins. The top predicted drugs have also shown antiviral effects against other viruses. For example, in case of NS3, Micafungin and its derivatives are effective against SARS-CoV-2 (Nakajima et al., 2023), chikungunya (Ho et al., 2018), PRV (Hondo et al., 2024), and enterovirus 71 (Kim et al., 2016). Trametinib inhibits influenza A (Schräder et al., 2018), while Ergotamine blocks TMPRSS2 activity relevant to COVID-19 (Wang et al., 2022). Candesartan cilexetil, and Montelukast targets Zika (Loe et al., 2019) (Chen et al., 2020), SARS-CoV-2 (Copertino et al., 2021), and MERS-CoV (Gan et al., 2021). For NS5, Doxycycline inhibits COVID-19 (Gendrot et al., 2020), vesicular stomatitis virus (Wu et al., 2015), and JEV (Topno and Khan, 2018). Ribavirin is effective against HCV (Dixit and Perelson, 2006), PPRV (Zhang et al., 2023), and IBDV (Akram et al., 2023). Minocycline, combined with favipiravir, may treat COVID-19 (Itoh et al., 2022) and inhibits RSV (Bawage et al., 2019) and West Nile virus (Michaelis et al., 2007). Heparan sulfate mimetics are active against enterovirus 71 (Pourianfar et al., 2012), African swine fever virus (Garcfa-Villal and Gil-Fern, 1991), and HIV (Baba et al., 1988). Doxorubicin shows anti-chikungunya activity (Kasabe et al., 2023), while Posaconazole inhibits the SARS-CoV-2 helicase (Abidi et al., 2021).

Further, several top candidates for both targets have been previously reported in the literature through experimental or in silico analyses for DENV. For NS3, Micafungin (IC₅₀–10.23 μM) and its analogs Caspofungin (IC₅₀–20.78 μM) and Anidulafungin (IC₅₀–3.24 μM) exhibit antiviral effects (Chen et al., 2021). Desmopressin was identified as a potential NS5 inhibitor via pharmacophore modeling (Kumar et al., 2022). AR-12, a celecoxib derivative, inhibits all DENV serotypes (Hassandarvish et al., 2017). Trametinib suppresses flaviviral replication (Valencia et al., 2021). Ergotamine and conivaptan exhibit high in silico binding affinity (Montes-Grajales et al., 2020). Candesartan cilexetil inhibits DENV-2 (IC₅₀–1.602 μM) (Loe et al., 2019) and Cyclosporine disrupts NS5-Cyclophilin interaction, hindering replication (Qing et al., 2009). Montelukast inhibits NS2B-NS3 protease (IC₅₀–25.65 mM) (Jiang et al., 2022). For NS5, doxycycline inhibits all DENV serotypes (IC₅₀–52.3 μM at 37°C, 26.7 μM at 40°C), with greater efficacy against DENV2 and DENV4 (Rothan et al., 2014). Mycophenolic acid (IC₅₀–0.4 mM) and ribavirin (IC₅₀–50.9 mM) inhibit DENV2 replication (Takhampunya et al., 2006). Minocycline targets multiple DENV life cycle stages (Leela et al., 2016), while sulfated polysaccharides show varying antiviral activity (PPS < suramin < PI-88) (Lee et al., 2006). Doxorubicin (CC₅₀–116.9 μM, EC₅₀–6.573 μM) and posaconazole block DENV RNA replication (Punekar et al., 2022) (Meutiawati et al., 2018). Thus, the repurposed drug candidates predicted by our algorithm hold promise as antiviral agents, potentially accelerating DENV drug discovery efforts. Further, molecular docking validated the top predicted drugs’ interactions with NS3 and NS5 proteins. Delavirdine showed the lowest NS3 binding energy (−8.3 kcal/mol), interacting with catalytic residues (His51, Ser135), while Baloxavir marboxil had the lowest NS5 energy (−8.4 kcal/mol), forming stable hydrogen bonds and hydrophobic interactions. These results confirm the model’s accuracy in predicting DENV inhibitors.

The key feature of the current study is the development of the “i-DENV” web server, specifically designed to predict the antiviral activity of molecules against DENV. For algorithm development, we utilized an experimentally tested dataset of molecules targeting NS3 and NS5 proteins of DENV. Additionally, we also predict potential repurposed drug candidates that were also previously reported in the literature. “i-DENV” is the first regression-based web server designed to assess the efficacy of user-defined molecules targeting both NS3 and NS5 proteins.

The limitation of the current study is the relatively small dataset used for the NS5 protein; expanding the dataset could further enhance the model’s predictive performance. Another limitation is that the “i-DENV” algorithm is designed solely to predict the antiviral efficacy of a molecule in terms of its pIC₅₀ value. In the future, this approach could be extended to multi-task learning (MTL) by incorporating additional properties beyond antiviral activity prediction. To ensure the reliable prediction of our model, experimental validation of a molecule is essential to confirm their efficacy against DENV.

“i-DENV” algorithm was developed using QSAR properties of compounds and employing various MLTs (SVM, RF, kNN, ANN, XGBoost and DNN). We created 360 unique configurations (5 random states × 4 feature sets × 3 feature selection methods × 6 models per set) for each protein dataset (NS3 and NS5). For NS3, the SVM-SVR and ANN-SVR models demonstrated PCC values of 0.857 and 0.862 on the TT dataset, and 0.870 and 0.894 on the IV dataset, respectively. For NS5, SVM-SVR and ANN-SVR showed PCC values of 0.982 and 0.964 on the TT dataset, and 0.970 and 0.977 on the IV dataset, respectively. These model’s robustness was confirmed through various analyses like applicability domain, chemical clustering, decoy set, statistical tests, etc. We used the approved category of the DrugBank database to identify potential repurposed drugs. Further, the top candidates were validated through molecular docking that confirmed the reliability of the predictive models. While the findings provide promising insights, it is important to note that the study is entirely based on in silico methodologies. Therefore, further in vitro and in vivo experimental validation is essential to confirm the therapeutic efficacy of the predicted compounds. Overall, the freely accessible “i-DENV” web server serves as a valuable resource for identifying novel antiviral candidates targeting the non-structural proteins of the DENV.