Conformational dynamics and binding free energy analyses unveil a stable flavonoid inhibitor of dengue virus NS5 polymerase

Abstract

The persistent global spread of dengue virus (DENV) necessitates the identification of novel antiviral therapeutics. The RNA-dependent RNA polymerase (RdRp) domain of the non-structural protein 5 (NS5) is a validated target for therapy due to its critical roles in viral genome replication. In this study, we used a structure-guided virtual screening approach to identify potent phytochemical inhibitors of DENV RdRp from the Plant Secondary Compound Database (PSC-db) by focusing on 326 flavonoids. Following drug-likeness filtering and hierarchical docking with MM-GBSA analysis, five top candidates were identified, with PSCdb01560 emerging as the most promising. Subsequent molecular dynamics simulations over 500 ns revealed PSCdb01560’s exceptional binding stability, characterised by low protein–ligand RMSD and sustained binding stability. Free energy landscape analysis demonstrated its occupation of deep, well-defined conformational basins. At the same time, post-MD MM-GBSA yielded a binding free energy of –91.65 kcal/mol—outperforming the reference compound. Structural superimposition confirmed strong conformational fidelity between docked and minimum-energy poses (RMSD 1.68 Å). Together, our approach points toward PSCdb01560 as the putative inhibitor molecule that outperforms existing flavonoid leads in thermodynamic and kinetic descriptors. The results thus provide the avenue for experimental validation, such as enzymatic and antiviral screens, in hopes of turning these in silico observations into viable therapeutic development for dengue.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-38864-2.

Introduction

The dengue virus continues to pose a formidable global health threat, with recent years witnessing unprecedented surges in cases and fatalities. In 2023, over 6.5 million cases and more than 6,800 deaths were reported worldwide—the highest figures on record—with South America and Southeast Asia bearing the heaviest burden¹. The situation escalated further in 2024: WHO data recorded over 7.6 million cases by April, including 3.4 million confirmed infections and more than 3,000 deaths, with the Americas the most affected region². By mid-2024, the CDC confirmed over 13 million cases across the Americas, with outbreaks persisting into 2025 in areas such as Puerto Rico and the US Virgin Islands³. As of June 2025, Europe and Southeast Asia collectively recorded over 3 million fresh infections and 1,400 deaths⁴.

Several interrelated factors are fuelling the intensification of dengue transmission, including the expansion of Aedes aegypti and Aedes albopictus habitats, climate change–induced weather anomalies—particularly linked to El Niño—urban overcrowding, and global travel⁵. Disturbingly, outbreaks have emerged in regions previously considered non-endemic. For example, the United States has observed a resurgence of local transmission in Florida, California, and Texas between 2024 and 2025⁵. Bangladesh faced its most severe outbreak in 2023 with nearly 750,000 suspected cases and 1,705 deaths, while Argentina experienced its deadliest epidemic in 2024, reporting over 333,000 cases and 238 fatalities⁶.

Dengue is caused by four serologically distinct serotypes of the dengue virus (DENV-1 to DENV-4), and clinical presentations can range from asymptomatic to severe forms involving plasma leakage and haemorrhagic symptoms, particularly during secondary infections due to antibody-dependent enhancement⁷. The viral non-structural protein 5 (NS5) is essential for viral replication and mRNA capping, functioning through its C-terminal RNA-dependent RNA polymerase (RdRp) domain and its N-terminal methyltransferase domain. These properties make NS5, and in particular the RdRp domain, a prime target for antiviral drug development⁷.

Pharmacological interventions against the dengue virus have seen some progress, particularly with the identification of NS5 inhibitors. Synthetic inhibitors, such as pyridobenzothiazolones, have demonstrated potent inhibition of NS5 polymerase activity in vitro⁷. Other compounds, like those disrupting NS3–NS5 interactions, have shown efficacy across flaviviruses⁸. Plant-derived secondary metabolites have garnered attention, with species such as Myrtopsis corymbosa yielding bioactive compounds that inhibit RdRp⁹. Additionally, recent reviews emphasize the potential of phytochemicals in flavivirus drug discovery, although many still face challenges related to bioavailability and toxicity¹⁰.

Several small molecules and natural compounds have been investigated as inhibitors of DENV NS5 RdRp. Early high throughput screening identified N sulfonylanthranilic acid derivatives as allosteric inhibitors of the polymerase. The prototype compound NITD 1 showed an IC50 of approximately 7.2 µM in primer extension assays¹¹. More recently, a study described two synthetic molecules named SW b and SW d, which inhibited DENV 2 RdRp with IC50 values of 11.54 ± 1.30 µM and 13.54 ± 0.32 µM. These compounds demonstrated better potency than the nucleoside analogue control, 3’-prime dATP, which had an IC50 of 30.09 ± 8.26 µM¹².

Natural flavonoids and flavonoid-rich plant extracts have also shown activity against DENV NS5 RdRp. Leaf extract of Carpolepis laurifolia produced several flavonoids including quercetin, isoquercitrin and hyperoside which inhibited RdRp in enzyme-based assays. The reported IC50 values for these isolated flavonoids were in the low micromolar range between 1.7 and 2.1 µM¹³. Another study using extract of Scutellaria baicalensis observed dose dependent suppression of dengue virus replication in Vero cells. The IC50 values were reported as extract concentrations ranging from 56 to 95 µg µg/mL which limits direct comparison with purified compound assays but still demonstrates antiviral potential¹⁴.

In addition to experimental investigations, computational studies have also explored bioflavonoids as potential inhibitors of dengue virus NS5 RNA-dependent RNA polymerase. A docking-based study by Sivaraman and Pradeep evaluated plant-derived flavonoids including apigenin, kaempferol, myricetin, naringenin, and hesperidin against the DENV RdRp and identified apigenin and kaempferol as promising candidates based on binding affinity. These compounds were reported to interact with conserved polymerase residues such as Ser710, Arg729 and Arg737, which are critical for de novo RNA initiation. While this work highlighted the relevance of flavonoids in targeting functionally important regions of NS5, the analysis was primarily limited to static docking and did not assess long-timescale conformational stability or allosteric pocket specificity¹⁵. In contrast, the present study explicitly targets the conserved allosteric N-pocket and integrates extended 500 ns molecular dynamics simulations with principal component and free energy landscape analyses to provide a more rigorous evaluation of ligand binding persistence and thermodynamic stability.

Based on these findings, phytochemicals were selected for the present study because they offer extensive structural diversity and contain multiple functional groups such as phenolic hydroxyl groups and aromatic rings that can establish stable interactions with RdRp binding pockets. Flavonoids in particular have been repeatedly noted for antiviral activity, favourable safety characteristics and drug friendly chemical features which support their use as natural scaffolds for antiviral discovery. Although no approved drug currently exists for the treatment of dengue virus infection, several natural product derived molecules have become successful therapeutics in other diseases. Artemisinin derived from Artemisia annua is a widely used antimalarial agent, and Paclitaxel from the Pacific yew tree is an established anticancer drug. These examples show that natural product scaffolds can be developed into safe and effective therapeutic agents and justify continued exploration of plant derived compounds as potential inhibitors of DENV NS5 RdRp.

The novelty of this work lies in the integration of a structured natural compound library with a comprehensive multi-level computational pipeline specifically tailored to the dengue NS5 RdRp N pocket. To our knowledge, no earlier study has systematically screened flavonoids from the Plant Secondary Compound Database against this conserved allosteric region of NS5, nor evaluated their stability through extended 500 ns molecular dynamics simulations. Most previous reports focused either on docking-based screening or on shorter simulation windows and lacked deeper conformational analyses. In contrast, the present study combines hierarchical Glide screening, MM-GBSA based free energy filtering, long scale molecular dynamics, principal component analysis, free energy landscape mapping and post MD superimposition to characterise ligand behaviour at both structural and energetic levels. The combined use of extended molecular dynamics, PCA, and free energy landscape analyses provides a level of conformational and thermodynamic precision not previously reported for PSC-db flavonoids in the context of dengue NS5 inhibition. In light of these findings, the current study was conceived to explore novel inhibitors of DENV NS5 RdRp, explicitly focusing on phytochemicals derived from the PSC-db. The rationale for this approach rests on the underexploited chemical diversity of natural compounds and their structural resemblance to known antiviral scaffolds. The central objective was to identify plant-derived small molecules with high theoretical binding affinity and favourable drug-likeness, offering viable leads for future preclinical development.

Methodology

Protein and ligand library preparation

The 3D structure of the dengue virus NS5 RdRp was obtained from the Protein Data Bank using the PDB identifier 5K5M^16,17. This structure corresponds to the crystallised RdRp domain of DENV NS5 and was chosen because of its high resolution (2.01 Å) and suitability for structure-based drug discovery. Protein preparation was carried out using the Protein Preparation Wizard in Schrödinger¹⁸. Missing hydrogen atoms were added, bond orders were corrected, and appropriate protonation states of ionisable residues were assigned at neutral pH. Water molecules beyond the binding site were removed. The protein structure was then subjected to restrained energy minimisation using the OPLS 2005 force field to relax strained geometries while maintaining the integrity of the crystallographic coordinates¹⁹.

A curated library of plant secondary metabolites was retrieved from the Plant Secondary Compound Database (PSC db), which contains structurally diverse natural compounds in three dimensional format²⁰. This is a free, searchable, web-based resource providing structural, physicochemical and pharmaceutical information for over 2,800 plant-derived secondary metabolites. Compared with larger mixed natural product repositories such as ZINC or NPASS, PSC-db is specifically focused on plant secondary metabolites and offers curated three-dimensional structures with reduced redundancy, making it particularly suitable for structure-based screening^21,22. From this large library, a focused subset of 326 flavonoids was selected for the present screening. The selection was based on the documented antiviral properties of flavonoids, their structural diversity, favourable drug like features and their suitability for stable interactions within protein binding pockets. Only compounds belonging strictly to the flavonoid class with complete 3D structural information and appropriate physicochemical parameters for molecular docking were retained for the virtual screening workflow. All ligand structures were prepared using the LigPrep module of the Schrödinger suite²³. During ligand preparation, proper bond orders were assigned, hydrogen atoms were added, and all possible ionisation states at physiological pH were generated. Stereochemical ambiguities were resolved, and each ligand was subjected to energy minimisation using the OPLS 2005 force field to obtain a low energy and geometrically optimised conformation suitable for docking studies. This standardised preparation ensured that all ligands entered the screening pipeline in a chemically consistent and comparable form.

Virtual screening of plant-derived ligands

For structure-based screening, a receptor grid was generated around the conserved allosteric N pocket located at the interface of the thumb and palm subdomains of the NS5 polymerase. This pocket is lined with highly conserved residues (including Ser710, Arg729, and Arg737) across all dengue virus serotypes. This pocket was selected for grid generation due to its conserved structure, ligand accessibility, and lower mutation rates compared to the active site, enhancing druggability. Previous studies have also demonstrated its potential to modulate polymerase activity allosterically, supporting its relevance as a viable antiviral target^17,24. A hierarchical virtual screening cascade was employed using the Glide module²⁵. In the first stage, High Throughput Virtual Screening (HTVS) was used to rapidly eliminate poorly fitting ligands, and the top 50 percent of compounds based on Glide score were selected for the next stage. These selected ligands were then subjected to Standard Precision (SP) docking, from which the top 25 percent of compounds were retained. Finally, Extra Precision (XP) docking was carried out on the top 10 percent of ligands to obtain highly refined binding poses and reliable docking scores. At each stage, compounds were ranked based on Glide score, visual inspection of binding orientation and interaction with key N pocket residues. The final shortlisted compounds were selected for further analysis.

To validate the docking protocol, the native inhibitor 68 T was extracted from the crystal structure and re docked into the NS5 RdRp binding pocket using Glide XP with the same receptor grid generated during the virtual screening procedure. The redocked pose reproduced the original crystallographic orientation with a ligand RMSD of 0.8 Å using UCSF Chimera, confirming the reliability of the docking protocol²⁶. Accordingly, 68 T was used as the reference compound for subsequent docking and molecular dynamics analyses (Supplementary Fig. 1).

MM-GBSA free energy filtering

Following molecular docking, binding free energy estimation was performed using the Molecular Mechanics Generalised Born Surface Area (MM-GBSA) method to obtain a more physically meaningful assessment of protein ligand interaction strength. The Prime MM-GBSA module of the Schrödinger suite was used for this purpose^27,28. This method combines molecular mechanics energy terms with an implicit solvent model to account for electrostatic, van der Waals, lipophilic and solvation contributions to binding. Unlike docking scores, which primarily reflect geometric complementarity and scoring function approximations, MM-GBSA provides a better approximation of the thermodynamic stability of the complex under solvated conditions. For each shortlisted docked pose obtained from XP docking, the protein–ligand complex was subjected to local energy minimisation to remove minor steric strain. The free energies of the complex, the isolated protein and the isolated ligand were then calculated. The binding free energy was determined as the difference between these energy terms. This procedure allowed us to quantitatively rank the ligands based on their predicted interaction strength with the dengue NS5 RdRp. MM-GBSA calculations were used as a final energetic filter in the screening pipeline to eliminate false positive docking hits and to prioritise only those ligands that displayed both favourable geometric fit and strong thermodynamic stability. Only compounds that satisfied dual selection criteria of high docking score and strongly negative MM GBSA binding free energy were considered suitable for subsequent molecular dynamics simulations.

Molecular dynamics simulations

To evaluate the dynamic behavior and structural integrity of the dengue virus NS5 RdRp in complex with selected ligands, molecular dynamics (MD) simulations were performed using the Desmond simulation suite^29,30. The top-scoring ligand–protein complexes from docking and MM-GBSA filtering were subjected to all-atom MD simulations in explicit solvent conditions. Each protein–ligand system was embedded in an orthorhombic simulation box filled with TIP3P water molecules, maintaining a minimum distance of 10 Å between the solute and box boundaries to avoid artefactual interactions³¹. The system was electrically neutralised by the addition of counter-ions (Na⁺ or Cl⁻) as needed, followed by the incorporation of 0.15 M NaCl to emulate physiological ionic strength. Parameterisation was carried out using the OPLS_2005 force field for both protein and ligand components³². Ligands were assigned force field parameters through the LigPrep and Force Field Builder tools to ensure consistency in topology and charges. Initial energy minimisation was conducted using a steepest descent algorithm until convergence, ensuring the removal of steric clashes and high-energy contacts³³. System equilibration was performed in multiple restrained stages under NVT (constant Number of particles, Volume, and Temperature) and NPT (constant Number of particles, Pressure, and Temperature) ensembles. This stepwise relaxation facilitated the gradual adaptation of the system to the simulation environment while maintaining structural constraints. The production MD run spanned 500 ns for each complex, maintaining a temperature of 300 K via the Nose–Hoover thermostat and a pressure of 1.013 bar using the Martyna–Tobias–Klein barostat^34,35. A 500 ns simulation timescale was selected to ensure adequate conformational sampling and full equilibration of the protein–ligand complexes, as shorter simulations may not capture slow collective motions and long-timescale stability of ligand binding. This extended timescale also enables post-simulation analysis, providing a more detailed assessment of binding stability. A time step of 2 femtoseconds was used, and coordinates were saved every 10 picoseconds for post-simulation analysis. All simulations were initiated using independent random seeds, resulting in distinct initial velocity distributions for each system to ensure unbiased conformational sampling.

Trajectory analysis included evaluation of Root Mean Square Deviation (RMSD), Root Mean Square Fluctuation (RMSF), hydrogen bonding persistence, radius of gyration (Rg), and centre of mass stability. These metrics allowed for comprehensive insight into the conformational stability, protein flexibility, and integrity of ligand binding throughout the simulation duration.

Principal component analysis (PCA)

Trajectory data from the MD simulations were subjected to Principal Component Analysis (PCA) to discern dominant motions and conformational transitions of the protein–ligand systems using the Geo-measures plugin in PyMOL^36,37. Covariance matrices were constructed from the atomic fluctuations of the protein backbone, and eigenvectors corresponding to major principal components were extracted. The first two principal components (PC1 and PC2) were used to project the conformational space sampled during simulations. Trajectory projections revealed distinct clustering patterns indicative of energy minima and conformational stability.

Free energy landscape mapping

To explore the thermodynamic stability and conformational distribution of the protein ligand complexes, Free Energy Landscapes were constructed based on Principal Component Analysis. The first two principal components PC1 and PC2 were used as reaction coordinates. The Gibbs free energy of each conformational state was calculated using the Boltzmann distribution derived from the probability density in PC1 and PC2 space. The Geo measures plugin in PyMOL was used to generate and visualise the energy basins. These basins represent the most thermodynamically favourable conformational states sampled during the simulation^38,39. A schematic overview of the PCA and free energy landscape workflow is provided in Supplementary Figure S2.

Structural superimposition analysis

The lowest energy conformations corresponding to the deepest free energy basins were extracted as representative minimum structures. These minimum energy poses were then used for further structural comparison with the original docked complexes to evaluate conformational stability after molecular dynamics simulation. Structural superimposition was carried out using UCSF Chimera to align the extracted minimum energy structures with their respective initial docked poses. Alignment was based on the protein backbone, and the RMSD of the ligand was calculated to quantify positional deviation within the binding pocket. This analysis allowed direct assessment of how much each ligand retained or altered its docking orientation after dynamic relaxation. The superimposition also enabled visual comparison of conserved and rearranged ligand residue interactions following the simulation²⁶.

Post-MD binding free energy estimation

Post molecular dynamics binding free energy estimation was performed using the MM GBSA algorithm with the Prime module on representative frames extracted from the last 10 ns of each MD trajectory, where the systems had attained full equilibrium. This ensured that binding energies were evaluated using only structurally stable and dynamically relaxed conformations of the protein ligand complexes. Frames were selected from low energy and highly populated conformational states to obtain reliable ΔGbind values under realistic dynamic and solvated conditions. This post MD MM-GBSA analysis served as the final thermodynamic validation step to confirm the stability and strength of binding of the selected protein ligand complexes after long time scale simulation. The MM-GBSA calculations were performed without explicit inclusion of entropic contributions. Consequently, the resulting binding free energy values represent approximate estimates and were used primarily for comparative ranking of ligand binding stability rather than as absolute free energy measurements.

Results

Hierarchical virtual screening and binding free energy profiling

An initial cohort of 326 flavonoid compounds, extracted from the PSC-db repository, was refined using LigPrep to generate ionisation states and stereoisomers suitable for receptor-based screening. To enhance pharmacokinetic viability, an early-stage physicochemical filter employing Lipinski’s Rule of Five was implemented, excluding 155 molecules with unfavourable drug-likeness. This pruning step ensured the selection of candidates with an optimal balance of molecular weight, hydrophobicity, hydrogen bond capacity, and permeability.

The remaining subset (171 molecules) underwent a tiered structure-based screening pipeline using the Glide docking algorithm, progressing from High-Throughput Virtual Screening (HTVS) to Standard Precision (SP), and culminating with Extra Precision (XP) docking. Subsequent HTVS and SP docking reduced the pool to 85 and 21 compounds, respectively. Finally, XP docking retained 6 top candidates. This sequential narrowing enabled the prioritisation of ligands based on spatial fit and scoring performance within the N-pocket of the DENV NS5 RdRp.

Out of the filtered candidates, six ligands exhibited docking scores ranging from –8.058 to –6.717 kcal/mol. These were further evaluated via MM-GBSA binding free energy estimation to capture solvent and entropic contributions to binding affinity. Three flavonoids—PSCdb01560, PSCdb00428, and PSCdb01521—were shortlisted, each demonstrating negative ΔG_bind values between –57.34 and –45.3 kcal/mol. Although PSCdb01560 shows a moderate docking score of − 6.717 kcal per mol, its highly negative MM GBSA value of − 57.34 kcal per mol indicates strong stabilising non bonded interactions after energy minimisation. Hence, PSCdb01560 remains energetically more favourable despite not having the lowest docking score. The docking score and MM GBSA binding free energy describe ligand binding at two different levels. Docking scores mainly reflect geometric complementarity and empirical scoring during pose generation, whereas MM GBSA captures the thermodynamic stability of the complex by incorporating van der Waals, electrostatic and solvation contributions. Hence, PSCdb01560 remains energetically more favourable despite not having the lowest docking score (Table 1).

Table 1.

Docking scores and MM-GBSA binding free energy values (ΔG_bind) of the top five screened flavonoid compounds from PSC-db.

Sr. No	Compound ID	Docking score	MMGBSA ΔG_bind
1	PSCdb01560	− 6.717	− 57.34
2	PSCdb00428	− 8.058	− 48.85
3	PSCdb01521	− 6.762	− 45.3
4	PSCdb01706	− 7.471	− 44.97
5	PSCdb00061	− 7.125	− 39.46

Molecular interaction analysis of selected flavonoid–NS5 RdRp complexes

To delineate the binding characteristics of the top-ranking compounds with DENV NS5 RNA-dependent RNA polymerase, detailed interaction profiling was performed on the docked complexes. The analysis revealed distinct hydrogen bonding and hydrophobic contact patterns that support the binding affinities inferred from docking and MM-GBSA results.

Compound PSCdb01560 established hydrogen bonds with key residues Arg729, Arg737 and Tyr766, in addition to two interactions with Thr794. These polar contacts were supported by extensive hydrophobic interactions involving Met341, Leu512, Cys709, Leu734, Tyr758, Met761, Met765, Tyr766, Ala799 and Trp803. Notably, π-cation interactions with Arg729 further stabilised the ligand within the allosteric site (Fig. 1a and b).

Fig. 1.

Molecular interaction diagrams of docked complexes for (a-b) PSCdb01560, (c-d) PSCdb00428, (e–f) PSCdb01521, and (g-h) reference compound, showing hydrogen bonding, hydrophobic, and π-interactions within the DENV NS5 RdRp N pocket.

Compound PSCdb00428 exhibited a diverse hydrogen bonding network comprising Lys460, Cys709, Ser710, Arg729, Tyr758 and Ser796. The hydrophobic landscape included Leu512, Ile740, Leu754, Tyr758, Trp795 and Ile797, indicating a snug fit within the non-polar pocket. However, π-based interactions were not detected for this ligand (Fig. 1c and d).

Compound PSCdb01521 demonstrated hydrogen bonding with Arg729, Thr794 and His798. Hydrophobic contacts were formed with Leu512, Leu734, Tyr758, Met761, Met765, Tyr766, Trp795, Ala799 and Trp803. No aromatic stacking or π-cation contacts were recorded (Fig. 1e and f).

The reference molecule (68 T) interacted via hydrogen bonds with Arg729, Arg737, Thr794, Lys800 and Glu802. Hydrophobic residues included Leu512, Leu515, Cys709, Tyr758, Met761, Leu764, Met765, Tyr766, Trp795, Ala799 and Trp803. A notable π–π stacking interaction with Arg737 was also observed, enhancing aromatic anchoring within the site (Fig. 1g and h).

These interaction profiles suggest that all selected flavonoids are capable of anchoring within the RdRp N-pocket through a combination of electrostatic and hydrophobic interactions, with varying degrees of π-interaction contributions (Table 2).

Table 2.

Pre-MD intermolecular interaction profiles of selected compounds with DENV NS5 RdRp, detailing hydrogen bonds, hydrophobic contacts, and π interactions.

S no	Complex	H-Bond	Hydrophobic	π-π stacking/ π-π cation
1	PSCdb01560	Arg729, Arg737, Tyr766, (Thr794)2	Met341, Leu512, Cys709, Leu734, Tyr758, Met761, Met765, Tyr766, Ala799, Trp803	Arg729
2	PSCdb00428	Lys460, Cys709, Ser710, Arg729, Tyr758, Ser796	Leu512, Cys709, Ile740, Leu754, Tyr758, Trp795, Ile797	–
3	PSCdb01521	Arg729, Thr794, His798	Leu512, Leu734, Tyr758, Met761, Met765, Tyr766, Trp795, Ala799, Trp803	–
4	Control	Arg729, Arg737, Thr794, Lys800, Glu802	Leu512, Leu515, Cys709, Tyr758, Met761, Leu764, Met765, Tyr766, Trp795, Ala799, Trp803	Arg737

Molecular dynamics simulation analysis

To elucidate the conformational dynamics and stability of the DENV NS5 RdRp–ligand complexes, a comprehensive molecular dynamics (MD) simulation analysis was undertaken over a 500 ns trajectory. Protein–ligand stability, local fluctuations, and atomic flexibility were evaluated using Root Mean Square Deviation (RMSD) and Root Mean Square Fluctuation (RMSF) metrics, separately for protein backbone and ligand atoms. The simulations provide a dynamic view of how each flavonoid interacts with the target site, in comparison with a reference inhibitor.

Root mean square deviation (RMSD)

The protein backbone and ligand RMSD was analysed over the 500 ns simulation period to assess the conformational stability of all complexes (Fig. 2a–d). The PSCdb01560 complex (Fig. 2a) exhibited a highly stable dynamic profile throughout the simulation. The protein backbone RMSD rapidly equilibrated and fluctuated within a narrow range of approximately 2.0–2.8 Å, while the ligand RMSD remained consistently below 2.0 Å, indicating tight binding and negligible positional drift within the NS5 binding pocket. In the PSCdb00428 complex (Fig. 2b), the protein backbone RMSD remained stable around ~ 2.0 Å; however, the ligand exhibited substantial instability after approximately 200 ns, with RMSD values exceeding 5.5 Å. This behaviour suggests significant ligand displacement and reduced dynamic retention within the binding cavity. A similar but more pronounced ligand mobility was observed for the PSCdb01521 complex (Fig. 2c). Although the protein backbone maintained structural stability with RMSD values below 3.0 Å, the ligand RMSD increased sharply beyond 7.0 Å after the equilibration phase, indicating extensive rearrangement and weaker positional placemnet during the simulation. The reference complex (Fig. 2d) displayed a balanced dynamic trend, where both protein and ligand RMSD values initially increased and subsequently stabilised around ~ 3.0 Å and ~ 2.5 Å, respectively. The stable plateau confirms overall equilibration; however, the RMSD magnitude remained higher than that of PSCdb01560, indicating comparatively reduced structural restraint.

Fig. 2.

Protein–ligand Root Mean Square Deviation (RMSD) plots over 500 ns MD simulation for (a) PSCdb01560, (b) PSCdb00428, (c) PSCdb01521, and (d) reference compound 68 T, showing the relative stability of protein and ligand across time.

To provide a quantitative comparison, statistical RMSD analysis was performed over the equilibrated region from 50 to 500 ns. The PSCdb01560 complex showed a mean protein RMSD of 3.30 ± 0.13 Å and a very low mean ligand RMSD of 0.37 ± 0.07 Å, confirming exceptional complex stability. In comparison, PSCdb00428 exhibited a mean protein RMSD of 3.12 ± 0.16 Å with a higher ligand RMSD of 2.13 ± 0.35 Å, indicating increased ligand mobility. The PSCdb01521 complex displayed a mean protein RMSD of 3.45 ± 0.11 Å and ligand RMSD of 0.73 ± 0.36 Å, reflecting stable protein dynamics with moderate ligand fluctuation. The reference complex showed the highest mean protein RMSD of 4.01 ± 0.14 Å and ligand RMSD of 1.41 ± 0.19 Å, indicating a larger conformational shift relative to the phytochemical complexes. Although the reference exhibited a stable RMSD plateau, its higher mean RMSD values reflect comparatively lower dynamic stability than PSCdb01560. Overall, the RMSD analysis clearly establishes PSCdb01560 as the most dynamically stable complex among all tested systems, exhibiting minimal protein deviation and the strongest ligand positional retention. To establish a baseline dynamic reference, the intrinsic stability of the apo NS5 protein was first examined using backbone RMSD analysis and directly compared with the PSCdb01560-bound complex. The apo NS5 RMSD profile showed a gradual increase from approximately 1.4 Å at the beginning of the simulation to around 3.5 Å toward the end of the 500 ns trajectory, reflecting continuous conformational adjustment in the absence of ligand binding (Supplementary Figure S3). In contrast, the PSCdb01560-bound complex exhibited a more restrained and stable RMSD profile, with fluctuations largely confined between 2.0 and 2.8 Å and a mean backbone RMSD of 3.30 ± 0.13 Å, indicating ligand-induced stabilization of the polymerase structure. Given the extended simulation timescale and the inherent flexibility of NS5 RdRp, stability was assessed based on RMSD convergence and plateau behaviour rather than a fixed numerical threshold. Accordingly, absolute RMSD values were interpreted in a relative manner across systems. Given the long simulation length of 500 ns and the inherent flexibility of the NS5 RdRp, stability was evaluated based on the overall convergence of RMSD profiles rather than a fixed numerical cut-off. Complexes that reached a clear plateau and maintained consistent fluctuations without progressive drift over time were considered stable. Absolute RMSD values were therefore interpreted in a relative manner across systems.

Protein RMSF

Protein RMSF analysis was used to examine residue-level flexibility, where reduced fluctuations across key functional and binding-site regions, compared with the apo protein, were taken as an indication of ligand-induced stabilization rather than complete rigidity. The residue-wise flexibility of NS5 in all complexes was evaluated using backbone RMSF analysis over the 500 ns trajectories (Fig. 3a–d). The PSCdb01560 complex (Fig. 3a) exhibited well-regulated fluctuations across the protein, with most residues fluctuating within ~ 0.6–1.5 Å. Higher deviations approaching ~ 3.0 Å were confined to surface-exposed loop and terminal regions, while residues forming the N pocket and catalytic core remained conformationally stable. The PSCdb00428 complex (Fig. 3b) showed a similar fluctuation pattern, with moderate RMSF peaks at flexible loop regions reaching ~ 3.0 Å. However, structured secondary elements and active-site residues retained low fluctuation values, indicating preservation of the NS5 structural framework. For the PSCdb01521 complex (Fig. 3c), controlled backbone flexibility was observed, with several surface loops showing transient peaks close to ~ 3.5 Å. The central catalytic region remained comparatively rigid, reflecting stable protein dynamics despite ligand mobility observed in RMSD. The control complex (Fig. 3d) displayed predominantly restrained fluctuations below ~ 2.5 Å for most residues, with a few flexible loop regions showing higher motion. This indicates an equilibrated but dynamically distinct conformational behaviour relative to the phytochemical-bound complexes. Statistical analysis further confirmed comparable global flexibility across all systems. The mean protein RMSF values were 1.06 ± 0.45 Å for PSCdb01560, 1.00 ± 0.42 Å for PSCdb00428, 1.04 ± 0.53 Å for PSCdb01521, and 1.02 ± 0.54 Å for the control. These closely matched values demonstrate that none of the ligands induced abnormal local flexibility or destabilisation of NS5. Overall, the RMSF analysis confirms that PSCdb01560 maintains optimal backbone rigidity while preserving native dynamic features of the enzyme. Residue-wise flexibility analysis further highlighted clear differences between the apo and ligand-bound states. The apo NS5 RMSF profile displayed pronounced fluctuations at several regions, with peaks exceeding 3.0 Å, particularly within loop and terminal segments, highlighting the higher intrinsic flexibility of the unbound enzyme (Supplementary Figure S4). In contrast, the PSCdb01560-bound system exhibited more regulated residue-level motion, with most residues fluctuating below 2.0 Å and a mean protein RMSF of 1.06 ± 0.45 Å. Notably, residues forming the N-pocket and catalytic core showed reduced flexibility in the ligand-bound state relative to apo NS5, indicating localized stabilization upon ligand binding.

Fig. 3.

Protein backbone Root Mean Square Fluctuation (RMSF) profiles for (a) PSCdb01560, (b) PSCdb00428, (c) PSCdb01521, and (d) reference complex, highlighting residue-level flexibility throughout the trajectory.

Ligand RMSF

Ligand RMSF was examined to assess the internal flexibility of bound compounds within the pocket. Ligands displaying limited atomic fluctuations throughout the trajectory were interpreted as being stably accommodated in the binding site, whereas higher fluctuations suggested weaker or more transient binding.

Ligand RMSF data revealed atom-specific flexibility throughout the MD run. PSCdb01560 (Fig. 4a) exhibited uniformly low RMSF values, typically below 1.0 Å, indicating a well-anchored ligand conformation with minimal intramolecular flexibility. Conversely, PSCdb00428 (Fig. 4b) showed widespread fluctuations, with RMSF peaking between 3.0 and 4.0 Å across multiple atoms, suggesting considerable torsional rotation or loop detachment. PSCdb01521 (Fig. 4c) displayed the highest fluctuation, with atom-specific RMSF rising to 5.5 Å, reinforcing the previously observed RMSD-derived instability. The reference ligand (Fig. 4d) maintained modest fluctuations, mostly between 1.0 and 2.5 Å, reflecting moderate but acceptable conformational adaptability without detachment or disruption. Among the studied ligands, PSCdb01560 demonstrated the most favourable dynamic profile, combining low ligand RMSD, stable protein–ligand interaction, and subdued flexibility, indicative of sustained engagement with the RdRp catalytic pocket. The reference compound performed comparably well, albeit with slightly higher protein RMSD. In contrast, PSCdb00428 and PSCdb01521 exhibited significant ligand displacement and elevated atomic fluctuations, casting doubt on their long-term stability within the binding site. Ligand RMSF analysis revealed distinct stability trends among the investigated compounds. PSCdb01560 exhibited the lowest atomic fluctuation with a mean ligand RMSF of 0.19 ± 0.06 Å, indicating exceptionally strong conformational stability within the NS5 binding pocket. In contrast, PSCdb00428 displayed substantially higher ligand fluctuation with a mean RMSF of 0.84 ± 0.58 Å, indicating reduced conformational restriction inside the pocket. PSCdb01521 also showed limited ligand flexibility with a mean RMSF of 0.48 ± 0.40 Å, supporting stable binding behaviour. The control ligand showed a mean RMSF of 0.51 ± 0.26 Å, reflecting moderate flexibility. Collectively, these results confirm that PSCdb01560 exhibits the most stable dynamic behaviour among all tested ligands. Moreover, to provide chemical context to the RMSF analysis, atomic fluctuations were interpreted with respect to the ligand’s structural features. Atoms forming the rigid aromatic scaffold and hydroxyl-bearing moieties exhibited consistently lower RMSF values, indicating restricted internal motion and stable anchoring within the binding pocket. In contrast, comparatively higher fluctuations were associated with peripheral and solvent-exposed substituents, reflecting localized flexibility of non-core regions without compromising overall ligand retention.

Fig. 4.

Ligand RMSF plots indicating atomic-level fluctuations in (a) PSCdb01560, (b) PSCdb00428, (c) PSCdb01521, and (d) reference ligand during simulation, reflecting the conformational adaptability within the binding site.

Solvent accessible surface area (SASA) analysis

Solvent Accessible Surface Area (SASA) analysis was performed to quantify the degree of solvent exposure for each protein–ligand complex throughout the 500 ns molecular dynamics simulations. Variations in SASA offer insights into the dynamic rearrangements of protein surfaces and the potential conformational shielding or exposure of the ligand–protein interface⁴⁰. The PSCdb01560 complex (Fig. 5a) demonstrated moderate SASA values ranging from approximately 60 to 120 Ų. The profile exhibited subtle undulations without drastic deviations, indicating consistent surface compactness over the simulation. The histogram confirmed this trend, showing a dominant distribution in the upper-mid SASA range. The lack of abrupt surface exposure events reflects the conformational resilience of the complex, reinforcing its previously noted structural stability. The PSCdb00428 system (Fig. 5b), on the other hand, exhibited highly fluctuating SASA values, peaking near 200 Ų. These wide oscillations suggest substantial surface rearrangements and transient unfolding events, possibly correlating with the previously observed ligand displacement. The broader histogram spread further indicates that the protein underwent variable conformational states, likely impacting ligand retention. Similarly, the PSCdb01521 complex (Fig. 5c) presented extensive SASA variability, with values frequently exceeding 150 Ų. This trend aligns with the elevated ligand RMSD and RMSF values, suggesting solvent exposure of internal regions due to ligand disengagement or protein-pocket expansion. The histogram corroborated persistent structural breathing throughout the trajectory. In contrast, the reference complex (Fig. 5d) depicted a gradual decline in SASA, stabilising below 100 Ų towards the latter half of the simulation. This contraction of solvent-exposed area is indicative of increasing protein compactness and potentially tighter ligand encapsulation over time. The histogram supports this interpretation, showing a pronounced peak in the lower SASA spectrum. Overall, the SASA results reinforce the superior stability of PSCdb01560 and the reference compound, while suggesting weaker or unstable protein–ligand encapsulation for PSCdb00428 and PSCdb01521. These findings, combined with RMSD and RMSF analyses, strengthen the case for PSCdb01560 as the most promising candidate for further optimisation. From a binding perspective, stable and lower SASA values reflect effective burial of the ligand within the NS5 binding pocket and reduced solvent exposure at the protein–ligand complex, which are characteristic of stronger and more persistent interactions. Conversely, large SASA fluctuations and elevated values indicate partial ligand exposure or pocket opening events, consistent with weaker binding and reduced encapsulation. In this context, the restrained SASA behaviour observed for PSCdb01560 supports its enhanced binding strength and sustained engagement with NS5 RdRp, whereas the pronounced SASA variability of PSCdb00428 and PSCdb01521 suggests compromised binding stability.

Fig. 5.

Solvent Accessible Surface Area (SASA) plots and histograms for (a) PSCdb01560, (b) PSCdb00428, (c) PSCdb01521, and (d) the reference complex, representing structural compactness and solvent exposure.

Radius of gyration (Rg) analysis

The radius of gyration (Rg) provides insight into the overall compactness and structural folding behaviour of protein–ligand complexes throughout molecular dynamics simulations⁴¹. A stable Rg profile indicates conformational consistency, while large deviations suggest global unfolding or significant structural rearrangements. The PSCdb01560 complex (Fig. 6a) demonstrated a tightly maintained Rg value between 4.2 and 4.4 Å over the entire 500 ns simulation period. The minor fluctuations and the relatively narrow histogram distribution confirm that the protein retained a compact, well-folded state while accommodating the ligand. This observation aligns with the compound’s stable RMSD and low solvent exposure, reinforcing its suitability for further development.

Fig. 6.

Radius of gyration (Rg) plots over 500 ns simulation for (a) PSCdb01560, (b) PSCdb00428, (c) PSCdb01521, and (d) reference complex, demonstrating the global folding and compactness of the protein–ligand systems.

The PSCdb00428 complex (Fig. 6b) exhibited a higher Rg profile, fluctuating between 5.2 and 5.6 Å. While the trajectory remained stable without dramatic collapse or expansion, the elevated Rg values may reflect a more extended protein conformation or partial structural flexibility required to house the ligand. This interpretation is supported by the broader SASA and RMSF values seen in earlier analyses. In the PSCdb01521 system (Fig. 6c), the Rg hovered around 3.6 to 3.8 Å, maintaining a relatively stable and compact state. Despite ligand instability noted in RMSD and SASA plots, the protein backbone retained its core folding, suggesting that the ligand’s mobility did not induce global protein unfolding. The reference complex (Fig. 6d) presented a gradual reduction in Rg from approximately 5.5 Å at the start to about 5.2 Å toward the end of the simulation, with consistent fluctuations thereafter. This trend may indicate increasing structural compactness over time, possibly due to induced fit or ligand-driven tightening of the binding site. The histogram further supports the predominance of lower Rg values in the latter part of the simulation. Collectively, Rg analysis supports the conclusion that PSCdb01560 and the reference compound maintained optimal protein compactness, reinforcing their dynamic stability. The other complexes remained structurally coherent but exhibited greater flexibility, which may influence their binding efficacy and persistence under physiological conditions. From the analsyis, sustained Rg stability reflects preservation of target protein compactness upon ligand binding, which is often associated with effective pocket accommodation and stable protein–ligand interactions. In contrast, elevated or variable Rg values indicate increased structural flexibility or partial expansion, which may reduce binding efficiency or persistence. Accordingly, the restrained Rg behaviour observed for PSCdb01560 supports its strong and stable association with NS5 RdRp, whereas the comparatively flexible profiles of the other complexes may contribute to reduced binding robustness.

Post-MD interaction analysis of protein–ligand complexes

To assess the persistence and adaptability of key protein–ligand interactions under physiological conditions, post-simulation contact profiling was conducted for each complex. Changes in hydrogen bonding, hydrophobic interactions, and π-related contacts following 500 ns molecular dynamics (MD) simulations were evaluated and compared with pre-MD interaction data. To ensure that only dynamically stable and biologically relevant interactions were considered, protein–ligand contacts persisting for more than 30 percent of the total 500 ns simulation time were selected for detailed analysis. This threshold allowed the identification of sustained hydrogen bonds and non-bonded interactions contributing significantly to ligand retention within the NS5 binding pocket.

PSCdb01560 retained stable binding characteristics after MD. Notably, hydrogen bonds with Leu512, Arg729 (dual contacts), Arg737, and Tyr766 were preserved, suggesting a robust polar interaction framework. Hydrophobic contacts with Leu512, Tyr766, and Ala799 were also observed, with Arg729 interactions still detectable. No π–π or π–cation contacts were detected, implying the compound’s stability arises primarily from classical non-covalent interactions rather than aromatic stacking. The persistence of key interactions from the initial docking indicates that this complex remained well-seated within the N pocket (Fig. 7a). PSCdb00428 exhibited a broadened interaction profile, engaging in hydrogen bonds with Lys457, Glu464, Arg737, Ile740, Gln742, and Tyr758. This suggests that the ligand adjusted its binding orientation during simulation to form additional polar contacts. Hydrophobic interactions with Ile740, Leu754, Tyr758, and Trp746 further reinforced this new configuration. Importantly, π–π stacking and π–cation interactions emerged with Arg737 and Trp746, which were absent in the pre-MD phase, indicating dynamic optimisation of binding geometry during simulation (Fig. 7b). PSCdb01521 showed retention of several critical interactions, particularly hydrogen bonds with Cys709, Leu734, Trp762, and Tyr766. Additional hydrophobic contacts with Glu464, Arg729 (two instances), Thr793, and previously noted residues supported the binding conformation. π–stacking with Arg737 remained intact, implying some degree of aromatic stabilisation. The binding appeared stabilised but required structural reorganisation during simulation (Fig. 7c). The reference compound demonstrated a highly networked interaction ensemble post-MD, preserving multiple hydrogen bonds with Glu464, Arg472, Arg729 (two contacts), Arg737, Tyr758, Tyr766, Thr794, Lys800, and Glu802. Its hydrophobic footprint included Leu512, Tyr758, Met761, Tyr766, and Trp795, while aromatic interactions with His711, Arg737, and Trp795 remained consistent. This confirms the high affinity and entrenchment of the ligand in the catalytic pocket throughout the trajectory (Fig. 7d). Overall, the interaction landscape post-MD reaffirms the high stability of PSCdb01560 and the reference ligand. Meanwhile, PSCdb00428 displayed an adaptive binding mode with new stabilising interactions, and PSCdb01521 maintained essential contacts despite moderate fluctuations (Table 3). In addition to the residue-wise schematic interaction mapping presented in the main manuscript, a quantitative residue-level interaction profile represented as stacked bar charts was provided (Supplementary Figure S5). These bar plots depict the temporal occupancy of hydrogen bonds, hydrophobic contacts, ionic interactions, and water bridges, normalized over the entire simulation trajectory. Together, these analyses offer both qualitative and quantitative insight into the persistence and stability of key interactions governing ligand binding.

Fig. 7.

Protein–ligand interaction profiles derived from MD simulations for (a) PSCdb01560, (b) PSCdb00428, (c) PSCdb01521, and (d) reference compound. Only interactions persisting for more than 30% of the total simulation time are shown.

Table 3.

Post-MD interaction analysis of selected ligands with DENV NS5 RdRp, identifying dynamic changes in hydrogen bonding and hydrophobic/π interactions.

S no	Complex	H-Bond	Hydrophobic	π-π stacking/π-π cation
1	PSCdb01560	Leu512, (Arg729)2, Arg737, Tyr766,	Leu512, Tyr766, Ala799,	–
2	PSCdb00428	Lys457, Glu464, Arg737, Ile740, Gln742, Tyr758,	Ile740, Leu754, Tyr758, Trp746	Arg737, Trp746
3	PSCdb01521	Cys709, Leu734, Trp762, Tyr766	Glu464, Cys709, (Arg729)2, Trp762, Tyr766, Leu734, Thr793	Arg737,
4	Control	Glu464, Arg472, (Arg729)2, Arg737, Tyr758, Tyr766, Thr794, Lys800, Glu802,	Leu512, Tyr758, Met761, Tyr766, Trp795,	His711, Arg737, Trp795

Principal component and free energy landscape (FEL) analysis

The conformational diversity and thermodynamic stability of each protein–ligand complex were further assessed using Principal Component Analysis (PCA) coupled with Free Energy Landscape (FEL) mapping. Principal Component Analysis was performed to capture the dominant collective motions of the protein–ligand complexes. For the PSCdb01560 complex (Fig. 8a), PC1 and PC2 accounted for 39.12% and 22.96% of the total variance, respectively, with a cumulative contribution of 62.09%, indicating that the majority of essential dynamics is well represented within the first two components. The compact clustering observed in the PCA projection suggests restricted conformational sampling and a stable dynamic behaviour of the complex. For PSCdb00428 (Fig. 8b), PC1 and PC2 explained 38.37% and 19.95% of the variance, respectively, with a combined contribution of 58.32%, and the broader dispersion of conformations indicates increased structural flexibility relative to PSCdb01560. In the case of PSCdb01521 (Fig. 8c), PC1 and PC2 contributed 32.71% and 21.85% of the total variance, capturing 54.55% of the overall motion, and the spread of conformers reflects moderate conformational transitions during the simulation. The control complex (Fig. 8d) showed 37.40% and 27.28% variance contribution from PC1 and PC2, respectively, with the highest cumulative value of 64.67%, indicating extensive conformational sampling during the trajectory. Overall, the PCA results demonstrate that PSCdb01560 exhibits the most restricted and stable essential dynamics, whereas PSCdb00428 and the control display comparatively broader conformational flexibility.

Fig. 8.

PCA cluster analysis of (a) PSCdb01560, (b) PSCdb00428, (c) PSCdb01521, and (d) reference compound. The colour gradient represents the progression of simulation frames over the 500 ns trajectory, illustrating the conformational space sampled during molecular dynamics simulation.

The 2D FEL plots were generated using the first two principal components (PC1 and PC2), with colour gradients representing Gibbs free energy levels, with the global minimum normalized to ΔG = 0 kcal/mol and higher-energy conformational states extending up to approximately 14 kcal/mol across all complexes. Differences in basin depth and confinement within this range distinguished the stability profiles of individual systems. These plots provide a visual depiction of the accessible conformational states and their relative stability during the simulation^36,39. The PSCdb01560 complex (Fig. 9a) exhibited a well-defined, low-energy basin surrounded by minor energy barriers of ΔG ≈ 0 kcal/mo, with two primary minima occupying distinct regions along PC1 and PC2. The sharp and deep contours observed in the darker red to black zones indicate that the complex largely resided in stable conformational states with minimal thermodynamic perturbation. This supports the earlier findings from RMSD and Rg analyses, suggesting that PSCdb01560 consistently maintained a folded and energetically favourable conformation.

Fig. 9.

2D Free Energy Landscapes (FEL) generated from PCA for (a) PSCdb01560, (b) PSCdb00428, (c) PSCdb01521, and (d) reference compound, showing conformational energy basins and thermodynamic accessibility.

In contrast, the PSCdb00428 complex (Fig. 9b) displayed a multi-basin landscape with three distinct energy wells. Although these basins were separated by moderate energy barriers, the transitions between them suggest that the complex frequently sampled alternative conformations. This flexibility may be attributed to the ligand’s dynamic nature and its adaptive repositioning during simulation, as seen in the RMSF and SASA profiles. The PSCdb01521 complex (Fig. 9c) showed the broadest conformational space among the three candidates, with multiple shallow minima and wide transitions. The extended landscape and gradient shifts point towards a system with substantial conformational entropy and limited entrenchment in any single low-energy state. This aligns with the elevated ligand RMSF and fluctuating Rg, reflecting structural instability. The reference compound (Fig. 9d) presented a funnel-like energy landscape, dominated by a major energy basin and a few peripheral sub-states. The compactness and depth of the global minimum suggest a stable conformation, with only occasional excursions into higher energy states. This thermodynamic behaviour is consistent with the persistent post-MD interactions and reduced SASA. Altogether, the FEL results reinforce the structural and energetic stability of PSCdb01560 and the reference complex, while highlighting the dynamic and less stable nature of PSCdb00428 and PSCdb01521 within the DENV NS5 binding pocket (Supplementary Figure S6).

Superimposition analysis of minima structures with initial docked poses

To assess the conformational fidelity of ligands after long-term dynamics, the lowest energy structures from the Free Energy Landscape (FEL) were extracted and superimposed onto their respective pre-simulation docked poses. This comparison reveals the extent of positional deviation and ligand adaptability within the DENV NS5 RdRp N pocket.

PSCdb01560 showed high structural retention, with an RMSD of 1.68 Å between the extracted minima and its initial pose. The triplet images (Supplementary Figure S7a1–a3) highlight minimal realignment of the core scaffold, with the ligand remaining deeply buried within the pocket. The multi-pose overlay (Fig. 10a) further confirms excellent spatial overlap, indicating that the binding conformation was stable and energetically favourable throughout the simulation.

Fig. 10.

Superimposition of extracted minima structures and their initial docked poses for (a) PSCdb01560, (b) PSCdb00428, (c) PSCdb01521, and (d) reference compound. RMSD values indicate the extent of conformational deviation post-MD.

PSCdb00428 exhibited slightly better structural conservation, with a post-MD RMSD of 1.46 Å. As shown in images S7b1–b3, the ligand undergoes modest repositioning, primarily through peripheral rotations. Despite earlier observed ligand instability from RMSD and SASA plots, the overall shape and orientation remain largely consistent. This suggests transient fluctuations without permanent displacement. Superimposition (Fig. 10b) shows the ligand realigning back into a low-energy binding state.

PSCdb01521 presented the highest deviation among the three candidates, with an RMSD of 1.89 Å. The conformational change is visible in panels S7c1–c3, where parts of the ligand exhibit greater spatial spread. Although it remains in the binding cleft, the variation suggests flexibility in peripheral substituents or ring torsion. The overlay (Fig. 10c) reflects these structural shifts, consistent with the wider FEL landscape and elevated ligand RMSF noted earlier.

The reference compound yielded a deviation of 1.86 Å, slightly higher than that of PSCdb01560 and PSCdb00428. Nonetheless, its binding core remained largely aligned (S7d1–d3), and the superimposition in Fig. 10d illustrates its ability to retain a conserved pose while tolerating minor adaptive shifts. Taken together, these superimposition analyses reinforce the dynamic resilience of PSCdb01560, which not only maintained strong interactions and favourable energy profiles but also preserved its initial binding geometry to a significant extent.

MM-GBSA binding free energy analysis

To quantitatively evaluate binding affinity and molecular contributions to complex stability, Molecular Mechanics Generalised Born Surface Area (MM-GBSA) calculations were performed on the selected ligand–DENV NS5 RdRp complexes. Binding free energy (ΔG_bind) and associated energy terms were extracted post-MD simulation to reflect interaction energetics under dynamic conditions (Table 4).

Table 4.

MM-GBSA binding free energy breakdown post-MD simulation, including contributions from Coulombic, hydrogen bond, van der Waals, lipophilic, and solvation terms, along with ligand strain energy.

MM/GBSA components	PSCdb01560	PSCdb00428	PSCdb01521	Control
ΔGBind	− 91.65 ± 4.02	− 45.39 ± 3.16	− 41.25 ± 5.46	− 87.77 ± 4.98
ΔGBind Coulomb	− 25.42 ± 9.34	− 19.84 ± 2.45	− 30.34 ± 5.46	− 28.16 ± 22.00
ΔGBind Covalent	4.14 ± 2.11	3.47 ± 2.88	1.62 ± 1.15	4.04 ± 1.24
ΔGBind Hbond	− 3.50 ± 1.02	− 0.96 ± 0.18	− 2.62 ± 0.47	− 2.60 ± 0.47
ΔGBind Lipo	− 19.22 ± 1.22	− 14.85 ± 0.91	− 8.68 ± 0.92	− 29.22 ± 1.04
ΔGBind Packing	− 7.55 ± 0.60	− 1.80 ± 0.53	− 2.36 ± 0.57	− 3.90 ± 0.39
ΔGBind Solv GB	38.17 ± 4.88	26.37 ± 1.79	28.16 ± 3.49	44.38 ± 20.31
ΔGBind vdW	− 57.75 ± 2.83	− 37.77 ± 1.53	− 27.01 ± 2.30	− 72.29 ± 2.94
Ligand Strain Energy	4.72 ± 1.54	10.39 ± 2.46	2.81 ± 1.00	4.00 ± 1.07

PSCdb01560 displayed the most favourable binding profile, with a ΔG_bind of –91.65 ± 4.02 kcal/mol, surpassing the reference compound (–87.77 ± 4.98 kcal/mol). Its strong van der Waals (–57.75 kcal/mol) and lipophilic contributions (–19.22 kcal/mol), along with stabilising electrostatic interactions (Coulombic energy –25.42 kcal/mol), account for its potent affinity. Ligand strain was modest (4.72 kcal/mol), indicating minimal conformational distortion upon binding. These values underline a thermodynamically stable and tightly bound complex.

PSCdb00428 presented a significantly reduced ΔG_bind of –45.39 ± 3.16 kcal/mol, suggesting a comparatively weaker affinity. While van der Waals (–37.77 kcal/mol) and lipophilic terms (–14.85 kcal/mol) contributed to stabilisation, the lower electrostatic (–19.84 kcal/mol) and hydrogen bond (–0.96 kcal/mol) energies, along with elevated ligand strain (10.39 kcal/mol), suggest partial destabilisation or poor geometric compatibility.

PSCdb01521 exhibited the weakest binding among the three, with a ΔG_bind of –41.25 ± 5.46 kcal/mol. Its modest van der Waals (–27.01 kcal/mol), lipophilic (–8.68 kcal/mol), and hydrogen bond (–2.62 kcal/mol) contributions were insufficient to overcome desolvation penalties (+ 28.16 kcal/mol). Notably, Coulombic stabilisation was higher (–30.34 kcal/mol), yet the overall affinity remained inferior, corroborating its observed conformational instability.

The reference compound demonstrated strong binding, driven by extensive van der Waals (–72.29 kcal/mol), lipophilic (–29.22 kcal/mol), and Coulombic (–28.16 kcal/mol) interactions. Its total ΔG_Bind of –87.77 ± 4.98 kcal/mol validates its efficacy and reliability as a benchmark for comparison. In conclusion, MM-GBSA results highlight PSCdb01560 as a lead candidate, outperforming both comparator compounds and aligning with dynamic stability metrics across MD analyses.

Discussion

Our current findings provide strong support for PSCdb01560 as a promising inhibitor of DENV NS5 RdRp, exhibiting superior dynamic stability, compactness, and energetics compared with both PSCdb00428 and PSCdb01521. When benchmarked against recent in silico studies targeting dengue RdRp, several key consistencies and distinctions emerge⁴².

A PLOS ONE study by Huq et al. (2024) explored 29 phenolic compounds from Theobroma cacao against DENV3 RdRp using docking, DFT, and MD simulation, ultimately identifying catechin as a lead thanks to its stability in RMSD, RMSF, and SASA profiles alongside favourable MM-GBSA energies⁴³. Much like PSCdb01560, catechin maintained low backbone RMSD and ligand RMSF, indicating a stable binding mode. However, PSCdb01560 surpassed catechin in our longer-duration MD simulations (500 ns vs 100 ns), as reflected in its consistent Rg, minimal SASA fluctuations, and a deeper FEL basin, pointing to an even stronger persistence of binding conformation.

Phunyal et al. (2024) conducted an extensive screen of 2,000 flavonoids, where five hits demonstrated negative MM-PBSA energies (–38 to –17 kcal/mol), correlated with favourable structural metrics over 200 ns simulations⁴⁴. In comparison, PSCdb01560 delivered a significantly more negative ΔG_bind (≈ –92 kcal/mol), placing it well ahead of those earlier flavonoid-derived candidates. It’s pronounced van der Waals, lipophilic, and electrostatic contributions underline its more potent interaction profile.

Another recent comparative screen of 270 phytochemicals from local vegetables (Hasan et al., 2025) identified apigenin-7-glucoside and rutin as strong RdRp inhibitors with docking scores below –8.7 kcal/mol. Subsequent 100 ns MD validated their binding, with apigenin-7-glucoside and rutin showing lower RMSD, stable Rg and SASA, and more favourable MM-PBSA than the favipiravir control⁴⁵. Our work extends these observations: PSCdb01560 not only meets but exceeds their stability benchmarks across extended simulation time-and exhibits lower ligand displacement, sharper FEL wells, and better Rg consistency.

The interaction analysis is biologically relevant, as recent dengue-specific computational studies have highlighted residues Arg729 and Arg737 as key contributors to stable ligand binding within the NS5 RNA-dependent RNA polymerase. Docking and MD Simulation investigations have reported persistent engagement of these conserved residues by flavonoid and phenolic inhibitors, linking their involvement to effective disruption of polymerase function^15,43. In the present study, PSCdb01560 maintained consistent interactions with Arg729 and Arg737 both before and after molecular dynamics simulation, indicating engagement of functionally important regions of NS5. This persistence supports the biological significance of the observed binding mode and reinforces the potential of PSCdb01560 to interfere with dengue NS5 RdRp activity.

Comparative MD metrics across these studies emphasize a clear trend: compounds with sustained low RMSD and RMSF, stable Rg, and minimal solvent exposure tend to yield more favourable binding free energies. PSCdb01560 clearly aligns with this pattern, outperforming PSCdb00428 and PSCdb01521 in all regards. PSCdb00428 and PSCdb01521, whilst initially promising in docking, showed ligand drift, broad FEL landscapes, and higher solvent accessibility—indicating less robust entrenchment—features also observed among less potent flavonoids in previous reports⁴⁴.

The FEL analysis for PSCdb01560 revealed a deep, well-defined energy basin with minimal transitions. This is comparable to the narrow energy wells reported for rutin and apigenin-7-glucoside in the vegetable-based study, though our FEL minima were better resolved due to the longer simulation time⁴⁴. Furthermore, the RMSD of the extracted minima to docked poses (1.68 Å) indicates excellent conformational retention, on par or better than previous flavonoid studies which generally reported deviations above 2 Å.

Taken together, PSCdb01560 not only replicates but improves upon key success criteria established in recent literature, including comparatively stronger MM-GBSA binding trends, enhanced structural stability across extended MD simulations, compact and energetically confined conformational space, and robust pose conservation. In contrast, PSCdb00428 and PSCdb01521 exhibited instability trends reminiscent of lower-performing hits reported in other flavonoid and phenolic compound screening studies.

This comprehensive comparative overview underscores PSCdb01560’s strong candidature as a lead flavonoid-derived inhibitor of DENV NS5 RdRp. It exhibits both improved thermodynamic characteristics and dynamic resilience relative to other naturally derived candidates reported to date. Future work to validate its antiviral efficacy in vitro and in vivo would thus be well justified.

Conclusion

This study identified PSCdb01560 as a compelling candidate for inhibiting the RNA-dependent RNA polymerase (RdRp) domain of the dengue virus NS5 protein. Through a multi-layered computational approach—encompassing molecular docking, MM-GBSA binding energy estimation, extensive molecular dynamics simulations, free energy landscape mapping, and structural superimposition—the compound demonstrated superior binding stability, minimal conformational drift, and the most favourable thermodynamic profile among all tested ligands, including a known reference inhibitor.

The dynamic resilience, compact interaction profile, and low-energy conformational occupancy of PSCdb01560 suggest its strong potential for further development. However, in silico findings, while highly indicative, must be experimentally validated. Future work should focus on:

• In vitro enzymatic assays to confirm direct inhibition of NS5 polymerase activity, ideally comparing kinetic parameters (e.g., IC₅₀, Ki, etc.) with standard antivirals.

• Cell-based antiviral assays using dengue virus-infected human cell lines to evaluate cytopathic effects and viral load reduction.

• Structure–activity relationship (SAR) studies to optimise PSCdb01560’s pharmacophoric features for enhanced potency and specificity.

• Pharmacokinetic profiling including ADMET screening to assess absorption, distribution, metabolism, excretion, and toxicity.

• Crystallographic or cryo-EM studies to determine the binding mode of PSCdb01560 within the NS5 RdRp N-pocket in atomic detail.

Such experimental validation will not only confirm the computational predictions but also pave the way for rational optimisation and preclinical advancement of this phytochemical as a novel antiviral against dengue.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

I.M.A., H.S.G., S.M.A., M.H.A., V.D.D., and E.I.A. contributed towards conceptualization. I.M.A., H.S.G., and V.D.D. contributed to the methodology and data analysis. I.M.A. H.S.G., and S.M.A., contributed to the writing—original draft preparation. V.D.D. and E.I.A. supervised the work. I.M.A., H.S.G., S.M.A., M.H.A., V.D.D., and E.I.A. contributed to data visualization and validation. I.M.A., H.S.G., S.M.A., M.H.A., V.D.D., and E.I.A. contributed towards Writing, review & editing.

Funding

This work was funded by Community Jameel—Saudi Arabia under Jameel Fund for Infectious Disease Research and Innovation under grant no JML: 002–141-2024. The authors would also like to thank the Deanship of Scientific Research (DSR) at King Abdulaziz University, Jeddah, Saudi Arabia, for their support.

Data availability

The original contributions presented in the study are included in the article/Supplementary Materials, further inquiries can be directed to the corresponding authors.

Declarations

Competing interests

The authors declare no competing interests.

Informed consent

Not applicable.

Institutional review board statement

Not applicable.                                                                                                                                                                                                                                                                                        Declaration of generative AI in scientific writing: During the preparation of this work the authors used ChatGPT in order to improve the readability and language of the manuscript. Subsequent to using this tool, the authors reviewed and edited the content as needed, assuming full responsibility for the content of the publication.